Statistical post-processing of NWP or ML forecasts: bias correction, ensemble calibration, distributional regression, station forecasts.
Search and filter this topic Plain list on GitHub
Review of statistical postprocessing methods for NWP forecasts and their challenges in big-data settings, with little explicit machine learning focus.
Deutscher Wetterdienst (DWD)ECMWFKarlsruhe Institute of Technology (KIT)KNMI (Royal Netherlands Meteorological Institute)MET Norway (Norwegian Meteorological Institute)Met OfficeMétéo-France (incl. CNRM)MeteoSwiss
Neural networks for ensemble post-processing outperform benchmark distributional regression for 2-m temperature at German stations, using embeddings and auxiliary predictors.
Self-supervised encoder pretraining on atmospheric variables, then transfer to a labelled precipitation task, improves regional NWP precipitation correction over other approaches.
Hybrid ANN plus quantile-mapping post-processing of GEFSv12 reforecasts reduces warm bias and improves summer Tmax and extreme-day forecasts over Taiwan at extended ranges.
Operational multi-task neural network post-processes HRRR forecasts with a CRPS loss, improving RMSE by 3–25% and CRPS by 31–45% over HRRR.
A residual U-Net postprocesses RAP forecasts with satellite data, improving dewpoint profiles and CAPE by up to 40% with well-calibrated uncertainty for severe-weather forecasting.
Wavelet-based bias correction outperforms quantile mapping for climate model precipitation in Iran, with wavelet-ANN predictor screening for downscaling.
Proposes an explainable two-stage classification-then-regression ML post-processor for daily station precipitation, improving on single-stage black-box baselines.
CNN-1D and CNN-LSTM bias correction reduces CMIP6 surface air temperature warm bias over India's monsoon core to 0.35°C RMSE in the historical period.
Conditional GANs generate synthetic ensemble members from HRRR convection-allowing forecasts, and a CNN turns them into severe weather probabilities with up to 20% Brier skill score gain.
GNN-based local bias correction of NWP forecasts exploits meteorological and spatial dependencies, improving RMSE by 4.75% over the best baseline.
An XGBoost calibration of the VMF3-FC zenith wet delay forecast cuts bias by 93% and RMSE by 6% versus the operational product, validated on 492 radiosondes.
PostRainBench and CAMT post-process NWP for precipitation, with the first deep model beating NWP on heavy-rain CSI by 15.6–31.8%.
A generalised regression neural network corrects GRAPES and ERA5 weighted mean temperature, nearly removing bias and cutting RMSE by 19–21% for GNSS water vapour.
Uses neural networks to correct GFDL-ESM2G CMIP5 surface air temperature over Andhra Pradesh against IMD observations, with RMSE of 0.97–1.33 °C.
Trains a conditional GAN on ERA5 to generate downscaled 1 km precipitation ensembles over UK and US regions, competitive with IFS ensembles at 8–16 h lead.
CNN post-processing of ECMWF-HRES forecasts at 390 North China stations raises threat scores by up to 296% at high precipitation thresholds.
Institute of Atmospheric Physics, Chinese Academy of SciencesChina Meteorological Administration (CMA)
Vine copulas model dependence among weather variables at three African sites, proposed as a dependency layer for subseasonal forecast calibration.
Permutation-invariant neural networks postprocess raw ensemble forecasts as unordered member sets, achieving state-of-the-art calibrated distributions for surface temperature and wind gusts.
Combines machine learning and wavelet simulations with reweighting to bias-correct CMIP6 NEX-GDDP precipitation for Cork, projecting a decline to about 116–119 mm/month by 2076–2100.
Evaluates Pangu-Weather against ECMWF and MEPS at 183 Norwegian stations; probabilistic post-processing helps global models most, though MEPS remains best.
Neural network corrects COSMO 2 m temperature forecast errors at eight Kyiv-region stations, improving forecasts in over 50% of cases.
Fuses random forest, KNN, extra trees and GBDT into a CMIP6 multimodel ensemble for Hanjiang Basin precipitation, reaching correlation 0.88 and Taylor skill 0.764.
Machine learning corrects WRF urban air-temperature biases per grid cell using 402 crowdsourced personal weather stations over London.
U-Net calibrates GEFS 10-m wind forecasts over Southwest China, raising the share of errors under 0.5 m/s by 46-50% at day 1 and beating CNN and linear baselines.
Nanjing University of Information Science & Technology (NUIST)
LightGBM interpolates plains forecasts and observations to improve temperature and precipitation forecasts in Japanese mountain areas, with partial RMSE gains.
CycleGAN bias correction of NUIST-CFS summer precipitation forecasts in southeastern China outperforms quantile mapping and better captures extreme precipitation frequency and intensity.
Shanghai Artificial Intelligence LaboratoryNanjing University of Information Science & Technology (NUIST)
Machine learning corrects WRF hub-height wind forecasts for black-start operations, cutting day-2 wind MSE by over 90% with a multioutput network and nowcasting with LSTM.
DABU-Net, a data-augmented deep learning correction of GFS wintertime precipitation over southeastern China, cuts RMSE by 19–25% at 24–72 h.
Tsinghua University (NowcastNet)Nanjing University of Information Science & Technology (NUIST)
Ensemble learning with focal loss improves 60-hour sea fog forecasts over WRF and NOAA FSL baselines off the Yangtze estuary, raising detection and cutting false alarms.
D-vine copula quantile regression postprocesses 10 m wind ensembles at 60 German stations, outperforming gradient-boosted tEMOS globally.
Sliding-window DNN bias correction of 100 m LDAPS forecasts at Pyeongchang reduces RMSE by 33% for temperature and 45% for wind speed.
CNN post-processing of ECMWF ensemble means yields calibrated global tercile forecasts for weeks 3-6, outperforming recalibrated ECMWF sub-seasonal forecasts.
CNN combines multi-GCM ensembles to improve monthly precipitation predictions at local scales over China compared with other methods.
Analog ensemble variant that applies past analog errors reduces dry bias and improves heavy 3-hourly precipitation reliability over British Columbia, and can forecast record values.
Compares quantile mapping with an LSTM bias correction of WRF summer rainfall over South Korea; LSTM best tracks interannual variability, but underestimates rain above 50 mm/day.
UNet CNN fuses GOES-16 cloud observations with COAMPS forecasts to correct high-cloud forecasts at 3-12 h lead times, improving accuracy and sensitivity.
An SVR and MLP ensemble calibrates GFS hourly solar irradiation forecasts in Romania, cutting normalized RMSE by about 16% using sky imagery.
UNet corrects TSMP European simulation mismatches against COSMO-REA6 for precipitation and pressure, improving RMSE by about 22%.
Compares proportional-odds logistic regression and MLP classifiers for calibrating ECMWF visibility ensembles; POLR generally beats MLP and beats raw ensembles.
SOM-CNN-LSTM post-processing of 15-day precipitation forecasts in the Huaihe basin improves correlation by up to 32.3% and reduces RMSE by 26.6%.
Ensemble neural networks correct nonstationary sea surface temperature bias in NCEP CFSv2 seasonal forecasts, reducing RMSE by 0.35 °C, with a deeper network giving further gains.
Institute of Atmospheric Physics, Chinese Academy of Sciences
Machine-learning post-processing of hourly HARMONIE and ERA5 forecasts predicts 1-minute irradiance variability for Dutch stations, with NWP radiation errors causing about half the inaccuracy.
Multimodel ensemble machine learning for regional climate assessment and prediction; no abstract was supplied, so this classification is tentative.
An explainable ANN corrects ECMWF extended-range European summer temperature exceedance probabilities and traces errors to western Pacific SST teleconnections, improving skill.
PoET, a hierarchical transformer, calibrates ensemble members directly and improves global 2m temperature skill by up to 20% over the raw ensemble.
Normalizing-flow neural post-processing jointly calibrates ensemble temperature forecasts across locations and lead times, reaching state-of-the-art on the EUPPBench station benchmark.
Neural-network post-processing of global 1 to 7 day forecasts across 30 levels improves RMSE by up to 12% for 850 hPa temperature at day 7, and discusses metric limits.
Stacking ensemble of neural networks corrects ECMWF-IFS maximum temperature forecasts at about 2400 stations, cutting RMSE by 52% versus IFS.
ANN bias correction of monthly and daily precipitation at one US station works well monthly but poorly daily, compared with conventional methods.
CNN and FNN multi-model ensembles of ECMWF and CMA forecasts reduce 12-72 h surface air temperature errors in Henan, with the CNN best at short range.
Nanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
A U-Net postprocesses 4 km West-WRF precipitation reforecasts into censored shifted gamma distributions, beating analog, regression and meta-Gaussian benchmarks at all lead times.
Compares CNN1D and LSTM encoder-decoder deep learning for bias correcting CMIP6 monsoon precipitation over India, with LSTM-ED performing best against IMD observations.
Applies unpaired UNIT image-to-image translation, combined with quantile mapping, to bias-correct HadGEM3 monsoon simulations toward ERA5, preserving cross-variable correlations and spatial structure.
XGBoost corrects short-term wind speed forecasts from a local NWP system at Hangzhou stations, cutting RMSE 26.1% versus NWP and outperforming MOS.
Institute of Atmospheric Physics, Chinese Academy of Sciences
Convolutional autoencoder corrects 3–6 month seasonal forecast errors for Greece using NCEP FNL analysis, improving 6-month-ahead accuracy.
XGBoost and random forest models trained on 25 years of Southern Ocean surface observations cut ERA5 shortwave mean bias to near zero and reduce RMSE by about 25%.
Weighted U-Net corrects ECMWF-IFS precipitation over China, improving heavy-rainfall forecast skill over the ordinary U-Net, with gains growing at longer lead times.
Institute of Atmospheric Physics, Chinese Academy of SciencesChina Meteorological Administration (CMA)
U-Net CNN post-processes West-WRF precipitation over the US West Coast, cutting RMSE 12.9–15.9% versus raw model at 1–4 day leads.
Probabilistic neural networks post-process ensemble precipitation at four stations, outperforming logistic regression and GLM for hourly precipitation and rain probability.
Proposes a probabilistic clear-sky benchmark and neural post-processing methods for NWP-based solar irradiance forecasts, with proper scoring and calibration checks.
Convolutional autoencoder bias correction of TRMM and PERSIANN-CDR satellite rainfall over the Mekong basin, using APHRODITE gauge data as benchmark.
Linear regression, random forests and neural networks correct ECMWF IFS 2 m temperature and 10 m wind errors, improving RMSE by 10–15% with uncertainty estimates.
Deep-learning latent space replaces hand-tuned similarity weights in analog ensembles for near-surface wind and solar forecasts, improving RMSE in Pennsylvania.
Grid-to-point deep learning corrects 3 km PRUFS forecasts at 311 stations, cutting RMSE 19% for temperature, 24% for humidity and 42% for wind.
Nanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
Machine-learning correction of ECMWF and grid forecasts for Shenyang station maximum and minimum temperature improves accuracy by up to 186% in some seasons.
U-Net trained on WRF rainfall outputs cuts district-scale heavy rainfall MAE by 64.78% over Assam and lifts categorical accuracy to 77.9%.
NFC-Net fuses FY-4A satellite, DEM and ERA5 data with a UNet to correct ECMWF forecasts, cutting 2m temperature and 10m wind speed RMSE by about 50%.
Nanjing University of Information Science & Technology (NUIST)
Attention U-Net with multi-scale residual blocks bias-corrects CFSv2 sub-seasonal rainfall over Thailand, improving RMSE by 8.65% over linear downscaling.
Neural networks ensembling eight North American NWP precipitation models cut MAE 17% and false alarm ratio 41% versus the mean baseline.
Compares tree-based probabilistic post-processing (NGB, QRF, DRF) with EMOS for 1-10 day ensemble 2 m temperature over the Czech Republic; QRF and NGB lead at short lead times.
A multi-task double-encoder TrajGRU corrects ECMWF 10-day wind forecasts over the western North Pacific in real time, cutting speed and direction biases by 8-11% and 9-14%.
Physically constrained cGANs correct CMIP6-class Earth system model precipitation biases in both local distributions and spatial patterns, outperforming existing methods on extremes' spatial intermittency.
Potsdam Institute for Climate Impact Research (PIK)Technical University of Munich (TUM)
EasyUQ converts single-valued model output into calibrated predictive distributions via isotonic regression, with a smooth variant and weather forecast examples; the method is general across models.
Deep multi-model ensemble method combining NWP outputs with window attention and a rolling recurrent network for the two-dimensional initial-lead time structure, applied to sea level pressure.
Neural network post-processing of Swiss surface forecasts enforces thermodynamic equations, yielding physically consistent temperature and humidity without losing accuracy, especially with scarce data.
Two deep-learning bias-correction pathways improve 4-month-lead summer precipitation forecasts from a Chinese GCM, with the deterministic pathway better for extremes.
Nanjing University of Information Science & Technology (NUIST)Alibaba (DAMO Academy)
ML post-processing of lagged ensemble forecasts improves two-week-ahead monthly precipitation and temperature forecasts over the continental US, beating climatology and ensemble means.
Swin-transformer U-Net corrects NWP 2 m temperature and 10 m wind biases and produces hourly temporal downscaling, cutting RMSE by 0.78 versus the original forecasts.
CatBoost bias-corrects WRF wind-speed forecasts at a Gobi wind farm, reducing a four-day mean deviation from 3.21 to about 2.50 m/s.
Institute of Atmospheric Physics, Chinese Academy of SciencesChina Meteorological Administration (CMA)
Deep-learning error correction of day-ahead NWP wind speed forecasts improves accuracy without per-site modelling, compared with ARIMA and smoothing baselines.
Convolutional autoencoder bias-corrects seven SubX S2S rainfall forecasts over Java, raising weekly correlations (e.g. 0.76 to 0.81) and CRPSS by 20–30%.
Machine learning classifiers and committee models correct ECMWF S2S precipitation category forecasts over three regions, improving on the raw ECMWF baseline.
Study of loss functions to counter class imbalance when deep learning corrects precipitation forecasts; abstract unavailable, so based on title only.
Crowd-sourced personal weather station data reveal a spatially varying WRF cool bias over London; a machine-learning per-cell bias correction improves daily min, mean and max temperatures.
HFIT statistically fits NWP track forecasts per lead hour, cutting 2014–2021 Atlantic hurricane track errors 8% at 24 h versus the NHC official forecast.
ConvMOS, a convolutional bias-correction model for REMO precipitation with per-location and global parameters, reduces errors and better captures extreme precipitation.
DLCR corrects and remaps FourCastNet gridded forecasts onto sparse, non-uniform observation locations, bias-correcting them against ground truth without gridded targets.
Compares MLR and ANN downscaling of ERA5 and MERRA2 winds against South China tall-tower observations; ERA5 with ANN gives r=0.69 and RMSE 2.11 m/s.
Tree-based XGBoost trained on LDAPS NWP output beats raw LDAPS for Seoul visibility (MAE 2.04 km vs 4.66 km, R 0.88 vs 0.40).
STCNet combines 2D and 1D convolutions to post-process NWP for station weather, outperforming a Seq2Seq RNN and yielding hourly forecasts from 3-hourly input.
Generative neural network post-processing of ensemble forecasts captures multivariate dependencies better than copula methods for German station temperature and wind.
Machine-learning adaptive bias correction of ECMWF subseasonal forecasts raises contiguous-US temperature skill 60-90% and precipitation skill 40-69%.
Microsoft Research (AI for Science)Massachusetts Institute of Technology (MIT)
CNN post-processing of GEFS re-forecasts beats CSG EMOS and raw ensembles for 1–7 day China precipitation, especially for extremes above 50 mm/day.
Nanjing University of Information Science & Technology (NUIST)China Meteorological Administration (CMA)
Neural network and GLM correct COSMO-REA6 reanalysis wind speeds at 100 m after diagnostic wind-model downscaling, reducing RMSE by almost 30% in Central Europe.
Deep-learning latent similarity metric for analog ensembles outperforms the original AnEn in RMSE, bias and CRPS for wind and solar forecasts in Pennsylvania.
CNNs on GEFS precipitation fields for 1–14 day Sacramento precipitation occurrence; skill matches bias-corrected ensemble mean but cost-tunable thresholds help water management.
Transformer post-processing of NWP 2 m temperature, trained on crowd-sourced stations, cuts STD to 1.13 °C versus 1.96 °C for raw NWP.
Neural networks trained on MOS, HRRR and NWS guidance cut 1-day temperature error 20–30% versus MOS at 11 US cities, though gains over linear regression were not significant.
Random forest and ANN post-processing of four NMME models gives better monthly and seasonal precipitation forecasts over Southwest Asia than individual models.
ML exospheric temperature model calibrates NRLMSIS 2.0 against satellite density, cutting model-satellite differences by 25% with calibrated uncertainty intervals.
Machine learning post-processing of ECMWF forecasts improves MJO prediction, outperforming multiple linear regression in location and intensity skill.
Machine learning forecasts residuals of NWP wind speed and direction at an Arctic wind farm, enabling bias correction of NWP inputs for wind power forecasting.
Machine learning ensemble of 27 CMIP6 models improves China extreme precipitation simulation and projects increases of up to 13.7% by late century.
CNN bias correction of ERA5 and PWRF downward longwave radiation in the Arctic reduces RMSE by 13–18% on independent 2020 data.
Bayesian neural network weights CMIP6 models per location and season, improving monthly US precipitation predictions over simple ensemble baselines with interpretable weights.
Compares nine machine-learning site-adaptation methods to calibrate the STRÅNG gridded solar irradiance product in Sweden using three stations; no single method wins everywhere.
Neural network blends two precipitation ensembles into calibrated hourly exceedance probabilities for +1 to +6 h, outperforming both inputs at all lead times.
Uses tree-ring chronologies and machine learning to reconstruct precipitation and temperature back to 1886, then ranks CORDEX models and bias-corrects them in the Tien Shan.
GAN-based post-processing corrects rainfall forecasts from a numerical weather prediction model; abstract unavailable, so scope is inferred from the title.
ANN, random forest and SVM multimodel ensembles improve CMIP6 precipitation and temperature projections for an Indian river basin over simple averaging.
Neural networks post-process Norwegian 2.5 km air temperature forecasts with station data, beating GRIDPP in 74.8% of cases where its error exceeds 3 K.
Two-step neural post-processing of ensemble forecasts for 100m wind and irradiance, improving on raw ensembles up to 48 h at the Hungarian Met Service.
Uses regression, decision trees and deep networks to merge 16 CMIP6 models into a new ensemble; it outperforms single models (R=0.81 precipitation, 0.99 temperature) and projects 1.5 °C by 2024 under SSP5-8.5.
ANN-based postprocessing of GEFS forecasts with short 60-day training windows improves probabilistic 24-h precipitation forecasts over CONUS, especially for heavy rainfall thresholds.
UNet post-processing of CFSv2 cuts global precipitation bias from +5–7 mm/day to within ±1 mm/day over land at 1–3 day leads.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of California, Los Angeles (UCLA)
Facebook Prophet and ML regressions using CMIP5 decadal hindcasts improve monthly rainfall prediction for the Brisbane River catchment.
Hybrid BA-KNEA post-processing of GEFSv12 forecasts improves 1–3 day solar radiation forecasts in Xinjiang, with RMSE 3.2–3.7 MJ m-2 d-1.
Random forest post-processing of CFSv2 precipitation forecasts over Iran (1982-2017) reaches correlation above 0.87, packaged as a GUI decision support tool.
Bayesian neural network learns spatiotemporally varying weights to combine climate-model precipitation ensembles, outperforming three ensemble methods and quantifying epistemic uncertainty.
A dendritic neural network corrects CWRF summer precipitation forecasts against station data, raising anomaly correlation by 0.1 and cutting MSE by about 26%.
Nanjing University of Information Science & Technology (NUIST)
Machine learning bias correction of simulated hourly precipitation, combined with quantile mapping, improves spatial frequency distribution but not temporal distribution of local precipitation.
A neural network post-processes ECMWF forecasts of 2 m temperature and 24 h precipitation using station observations in 10 European cities, improving on raw ECMWF.
A CNN-LSTM predicts cloud-cluster evolution from geostationary IR imagery to rolling-correct 1-6 h GRAPES_Meso precipitation forecasts over China, improving rain/no-rain scores.
Neural network post-processing of ECMWF ensemble mean corrects Brazilian wind-speed bias to within ±0.1 m/s and cuts RMSE by 10–30% for 1–46 day lead times.
Random forest post-processing of CFSv2 precipitation forecasts over Iran, with correlation above 0.87, packaged as a decision support system.
A benchmark of statistical learning methods for post-processing NWP solar irradiance forecasts; abstract not available, so scope is inferred from the title.
Hybrid deep and machine learning models use global climate model and observational predictors to boost solar radiation predictions; abstract not available, so scope is inferred from the title.
Deep learning model forecasts quantitative precipitation from basic NWP variables; no abstract available, so the method and gains are unverified.
Convolutional autoencoders compress spatial NWP fields into features that augment station-based neural post-processing of 2 m temperature ensemble forecasts over Germany.
Hybrid analog-ensemble and CNN post-processing of GEFS precipitation raises CRPSS about 10% over quantile-mapped baseline out to 7 days in British Columbia.
Compares ML ensembling (RF, SVR, ETR, LSTM) of 34 GCMs for daily precipitation and temperature in an Indian basin; LSTM reaches R² of 0.9 for precipitation.
Neural networks post-process 2 km WRF output over the NW Himalaya and interpolate it to 90 m, with NSE above 0.75 at five stations.
Hybrid deep learning corrects numerical weather forecasts of short-term wind speed; the abstract is unavailable, so the gain is unquantified.
Support vector quantile regression post-processes mesoscale ensemble solar irradiance forecasts, reducing serious overestimation at five Kanto sites.
Skill-score-weighted ensemble of MLR, SVR, GRU and CNN post-processors corrects LDAPS next-day maximum temperature, with CNN reaching 1.41 °C RMSE in South Korea.
A generalized regression neural network bias-corrects CNRM-CM5 near-surface temperature over India, and the trained model is applied to RCP4.5 and RCP8.5 projections to 2100.
Thesis develops probabilistic post-processing of ensemble forecasts for water level, solar irradiance and total cloud cover using statistical and machine learning methods.
A neural network corrects empirical typhoon model wind fields toward JMA mesoscale model output, enabling more realistic storm surge simulation for Typhoon 10 (2020).
Applies deep learning to post-process ECMWF forecasts of heavy rainfall, correcting NWP output for extreme precipitation events.
A Bayesian neural network ensemble weights CMIP6 models by location and season, calibrating projections to observations with interpretable weights and uncertainty growing away from constrained periods.
Video CNNs using GloSEA5 pressure and temperature fields forecast monthly regional rainfall in Great Britain; an ensemble with GloSEA5 reduces RMSE to 31-32 mm.
Machine learning refines hydrological quantitative precipitation forecasts for rainfall impact prediction; the abstract is missing, so the scope is inferred from the title.
Deep MLP corrects RACMO2 Antarctic surface melt using automatic weather station and MODIS albedo data, outperforming random forest and XGBoost on cross-validation.
CNN-based statistical post-processing corrects ensemble precipitation forecasts, improving on raw ensemble output for rainfall prediction.
Institute of Atmospheric Physics, Chinese Academy of Sciences
ML error correction of WRF solar irradiance forecasts in the tropics cuts NRMSE by 40-50% and makes results across four radiation schemes comparable.
DeepSTF combines CNN and attention encoder-decoder to post-process multi-site forecasts at 226 Beijing stations, improving on MOS benchmarks.
Institute of Atmospheric Physics, Chinese Academy of Sciences
An XGBoost model post-processes ECMWF EPS outputs to correct systematic negative biases in tropical cyclone maximum wind forecasts over the western North Pacific, 2016–2019.
An adaptive neural-network postprocessor evolved in a predator–prey ensemble matches or beats linear regression and standard ANNs on a demonstration circulation model.
Deep-learning postprocessing turns deterministic West-WRF atmospheric river IVT forecasts into reliable probabilistic forecasts that rival calibrated GEFS out to 120 h.
Five tree-based ML models correct NMME seasonal precipitation over the Himalayas, cutting the 20 mm/day multi-model wet bias with RMSE near 2 mm/day.
An ANN post-processing of 3-km WRF forecasts raises 10 m wind direction accuracy in a small southeast French valley from 56% to 79%.
Conditional GAN and dense networks postprocess COSMO-E and IFS cloud cover ensembles over Switzerland, with cGAN producing realistic scenarios without copula or shuffle templates.
A comparison of 21 machine-learning post-processors for GEFS precipitation reforecasts up to 8 days finds extra-trees regression performs best in the Yalong basin.
Bayesian neural network post-processes NWP inputs to forecast medium-term wind speed, beating LASSO on MAPE and NMAE across several horizons.
Online sequential-learning combinations of NWP and statistical extended-range forecasts improve probabilistic skill by 5-15% over standard multimodel methods up to lead week 5.
Columbia University (LEAP, Lamont-Doherty)University of Reading
Two machine-learning and statistical up-scaling methods for convection-permitting ensemble rainfall forecasts reduce the double-penalty problem and improve severe-weather forecast skill.
Spatio-temporal deep learning corrects day-ahead wind speed forecasts built on WRF predictions for energy use; abstract missing, judged from title.
Neural network combination of 15 GCMs for Tehran temperature and precipitation outperforms the best single model and multi-model mean in present-day simulations and future projections.
U-Net-based bias correction of Yin-He global model 3-hour precipitation re-forecasts over China improves threat scores, especially for heavy rain and rainstorms.
AutoML post-correction of WRF forecasts at 13 Xinjiang stations cuts temperature RMSE from 3.24 to 2.34 °C and humidity RMSE from 19.54% to 11.54%.
Deep networks trained on an NWP ensemble with a frequency-weighted loss correct rainfall, improving extreme-event forecast skill by factors of two to above six.
Potsdam Institute for Climate Impact Research (PIK)Technical University of Munich (TUM)
Gaussian process model fuses station observations with NWP forecasts to improve local short- and long-term weather forecasts for dispersion modelling.
Random forest combines CMIP6 multi-model ensembles over China, projecting less warming and wetter conditions with better spatial skill than the arithmetic mean.
Nanjing University of Information Science & Technology (NUIST)
A simple machine learning bias correction conditioned on weather regimes improves precipitation frequency and local distributions in forecasts beyond 24 hours.
CU-net, a CNN, corrects gridded ECMWF-IFS 24–240 h forecasts of four surface variables over North China, beating the ANO correction at all lead times.
Colorado State University (incl. CIRA)China Meteorological Administration (CMA)
XGBoost downscales NWP outputs to medium-range daily reference evapotranspiration across China, a weather-derived quantity for water and agriculture planning.
ML-enhanced Model Output Statistics post-processes NWP output to predict 24-hour temperature and wind speed at four Chinese cities, beating neural network and SVM baselines.
A self-attentive ensemble transformer post-processes ECMWF ensemble members member-by-member, calibrating spread and producing spatially coherent 2m temperature ensembles.
Systematic comparison of eight statistical and ML wind gust ensemble postprocessing methods on German convection-permitting ensembles; locally adaptive neural networks perform best.
An end-to-end spatiotemporal scale-selection model improves ECMWF precipitation bias correction, especially for heavy rain, beating 11 published methods on two benchmarks.
An ordinal distribution autoencoder corrects ECMWF and SMS-WARMS precipitation forecast bias and outperforms classical ML and deep baselines.
A multivariable denoising-autoencoder network corrects hub-height wind speed NWP forecasts at 15 Japanese wind farm sites, cutting RMSE using extra atmospheric variables.
Fully convolutional networks post-process multi-model precipitation forecasts; they did not beat regularized logistic regression, but combining global and regional NWP inputs improved skill.
FOCUSED corrects short-term NWP wind speed forecasts for dangerous bora gusts on the Croatian Krk bridge, outperforming ANN, random forest and SVM baselines.
Bayesian joint probability calibration of ML and GCM monthly precipitation forecasts over China yields reliable ensembles, though ML outputs lack skill beyond climatology.
Nonlinear neural network bias adjustment of CORDEX regional projections for summer precipitation in southeast Iran. No abstract was provided.
FogNet, a 3D CNN with attention, post-processes NAM forecasts to predict fog visibility at a US airport, outperforming HREF on eight metrics.
Deep learning improves wind forecasts from reanalysis, which then force a wave model, giving about 10% gains in wind and wave height over the Mediterranean and over 35% in the Aegean.
CNNs postprocess HARMONIE-AROME wind speed forecasts over the Netherlands using full spatial patterns, improving CRPS and Brier skill over fully connected networks and quantile regression forests at 48 h.
Hybrid neural network and parametric scheme for post-processing medium-range precipitation forecasts; abstract unavailable, so classification rests on the title.
A neural-network similarity metric replaces hand-tuned distances in analog ensembles, improving wind speed and solar irradiance forecasts over Euclidean-distance analogs.
An encoder-decoder CNN post-processes NWP output to produce gridded surface temperatures over Kanto, correcting front position errors and extreme temperatures better than the operational Kalman-filter guidance.
Oxford workshop position on AI post-processing for weather and climate, releasing five archived ML-ready datasets with baseline notebooks to encourage trustworthy, FAIR and physically informed methods.
The Alan Turing InstituteKarlsruhe Institute of Technology (KIT)Met OfficeNSF NCAR / UCAR
Random forest bias-corrects global ensemble precipitation forecasts; the abstract is missing, so this rests on the title alone.
BLSTM and CNN fuse station observations with RDAPS model images to predict Korean temperatures up to 14 days ahead, beating single-source models.
A neural network corrects 3 km WRF forecasts of valley winds in complex terrain, significantly improving low-level wind speed and direction.
Neural networks post-process NCEP CFS week-3–4 precipitation and 2-m temperature forecasts, improving skill over multiple linear regression correction.
Applies machine learning to correct surface wind speed forecasts from a numerical weather prediction model using observations.
Neural networks combine RAP forecasts, surface analyses and GOES-16 satellite data to correct temperature and moisture profiles for severe weather nowcasting.
A flexible machine learning method calibrates 100 m wind speed ensembles into truncated normal distributions, outperforming three EMOS variants and raw ensembles in a Hungarian case study.
XGBoost with probability matching post-processes LENS and Dong-Nae forecasts, improving Korean heavy-rainfall QPF skill (Korean-language version).
XGBoost with probability matching post-processes LENS and Dong-Nae forecasts, improving Korean heavy-rainfall QPF skill over 14 events.
A machine-learning bias correction for a 25-km regional climate model raises simulated violent typhoon frequency from zero to match observations and projects increases under warming.
Tree-based machine learning corrects climate model biases against Indonesian station observations, with ENSO included; savanna regions show the largest corrections, reaching about 36.67% improvement.
ConvMOS, a CNN for model output statistics, reduces precipitation errors in the REMO regional climate model and outperforms three common MOS approaches.
A hybrid model selects the best climate model forecast with a probabilistic MLP, then uses a deep LSTM to produce next-day rainfall in Australian climate zones.
Random forests, gradient-boosted trees and logistic regression calibrate NOAA Warn-on-Forecast ensemble guidance for tornado, hail and wind, beating updraft-helicity baselines.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
Applies GANs and deep belief networks to improve ECMWF short-term intense rainfall forecasts; the abstract is missing, so reported gains are unverified.
Neural network predicts and corrects COSMO regional model 2m temperature errors, improving forecasts in about 50% of cases.
End-to-end CNN post-processing (MODL) of TIGGE ECMWF forecasts improves Tianjin 2 m temperature RMSE over ECMWF, MOS and other ML baselines, especially in the first three days.
A CNN post-processes one-day WRF forecasts over South Korea, raising annual index of agreement for wind from 0.67 to 0.85 at 93 stations.
Neural networks and NGBoost post-process SubX GEFS 2-m maximum temperature forecasts, beating EMOS by about 0.2 in CRPS over lead times of 1 to 35 days.
Neural networks postprocess ECMWF reforecasts into probabilistic 7-day California precipitation outlooks out to week 4, beating a state-of-the-art method, though week 3-4 skill stays low.
Stacked LSTM and DRNN models using WRF outputs predict typhoon wind speeds in northeastern Taiwan for 1 to 6 hours, outperforming memoryless MLPs.
Distilling an analog ensemble into a deep network postprocesses ECMWF stratospheric wind forecasts, cutting latency to near zero and beating ENS wind-speed error by 20% against balloon data.
K-means station clustering and decision-tree transfer learning post-process operational temperature forecasts at 301 Chinese stations, cutting 7-day RMSE by up to 39.4%.
Validates GFS day-ahead solar irradiance at 18 China stations; a gradient-boosting correction of GFS gives the best accuracy, beating persistence.
K-means station clustering and decision-tree transfer learning post-process operational temperature forecasts at 301 Chinese stations, cutting 7-day RMSE by up to 39.4%.
A U-Net with a learned predictor-weighting layer post-processes ARPEGE cloud cover forecasts over Europe, outperforming random forests and logistic quantile regression.
Compares calibration of heat index ensembles, finding direct GEV-based EMOS gives the best skill up to day 6; machine learning performs comparably only for warning categories.
A general regression neural network corrects zenith tropospheric delay from ECMWF and NCEP reanalysis over Japan, improving accuracy for GNSS positioning.
Random forests convert convection-allowing ensemble forecasts into day-1 severe weather probabilities, beating calibrated updraft helicity and SPC outlooks in Brier skill, especially for wind and hail.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
Operational quantile regression forests and ensemble copula coupling improve station temperature and gridded hourly rainfall forecasts in France, better capturing cold and heat waves than raw ensembles.
A deep neural network post-processes a subset of ensemble trajectories, improving CRPS by over 14% and allowing fewer members to match the full ensemble.
Quantile regression forests post-process and combine multiple NWP models across lead times into calibrated probabilistic road surface temperature forecasts.
Neural networks post-process the North American multi-model ensemble to improve monthly precipitation forecasts; abstract unavailable.
Differentiable versions of probability of detection and false alarm rate let neural precipitation forecasts be trained directly on categorical skill scores.
Review of statistical postprocessing methods for NWP forecasts and their challenges in big-data settings, with little explicit machine learning focus.
Deutscher Wetterdienst (DWD)ECMWFKarlsruhe Institute of Technology (KIT)KNMI (Royal Netherlands Meteorological Institute)MET Norway (Norwegian Meteorological Institute)Met OfficeMétéo-France (incl. CNRM)MeteoSwiss
STAS, an end-to-end deep model with adaptive spatio-temporal feature selection, corrects ECMWF precipitation forecast bias and beats eight published BCoP methods on threat scores.
Road weather station network with ML that flags faulty sensor data and corrects physical road-ice forecasts, aiming to improve winter road safety.
ZiANN combines feature selection, ensemble learning and neural networks to post-process NWP output for six-hour nowcasts and 40-hour point wind forecasts in Austria, outperforming operational INCA and AROME.
Machine-learning regression diagnoses horizontal visibility from mesoscale NWP forecasts; abstract not available, so the scope is inferred from the title.
Neural-network correction of COSMO-Ru13 surface forecasts up to 117 hours, tested on 23 million 2018 forecasts, reduces errors even at unobserved points.
Random forest on HRRR surface fields produces gridded CONUS probabilistic subfreezing road temperature forecasts, with a 2-h Brier skill score of 0.66.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
Distills the analog ensemble post-processing into a small neural network, enabling fast on-the-fly ECMWF lower-stratosphere wind corrections with a model of only hundreds of kilobytes.
Introduces Label Distribution Learning with Climate Probability to post-process ensemble precipitation forecasts in East China, improving RMSE and CRPS over baselines.
Multistage ensemble of linear regression, ANN and SVM refines NWP wind speed for wind-farm forecasts, reaching 0.13% MAPE at 72 hours.
Deep learning applied to correct NWP model cloud forecasts; abstract unavailable, so classification rests on the title alone.
XGBoost corrects KMA quantitative precipitation forecasts for heavy rain in Seoul, cutting MAE by up to about 30% at some sites.
Compares MLP, gradient boosting and random forest calibration of ECMWF total cloud cover ensembles as oktas classification; MLP and GBM give the best skill gains.
Multimodal machine learning corrects 1 km gridded winter air temperature biases at 88 Hokkaido sites, reducing RMSE by about 0.1°C with a nested model.
U-Net with shortcut connections corrects NWP precipitation outputs for Kansai heavy rain, improving strong-rain representation over plain CNNs.
ML-tuned blending of multiple ensembles gives probabilistic thunderstorm forecasts from 9 hours to 4 days, clearly better than raw ensemble output.
Neural-network quantile function regression with Bernstein polynomials postprocesses 60 h ensemble wind speed at 125 Norwegian stations, improving quantile skill score by nearly 1%.
Random forest postprocesses HREFv2 ensembles for hail, with isotonic calibration matching SPC outlook reliability for operational day-1 forecasts.
Neural network uses automatic weather station sensor data to correct ECMWF forecasts of temperature and precipitation for regional accuracy gains.
Random forest calibrates 12–36 h SREF and HREFv2 precipitation ensembles over CONUS, improving reliability and Brier skill over raw and smoothed probabilities.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)University of Oklahoma (AI2ES)
Expert-free denoising autoencoder with ordinal regression corrects ECMWF precipitation forecasts, outperforming classical ML and deep baselines on threat scores.
Deep generative models replace stored archives in the analog ensemble method, cutting memory needs by orders of magnitude for probabilistic NWP post-processing at stations.
A CNN postprocesses GFS integrated vapor transport forecasts, cutting RMSE by 9-17% from 3 hours to 7 days and improving atmospheric river forecasts over the western US.
A genetic algorithm learns optimal weights for 24 SINTEX-F2 ensemble members, significantly improving boreal summer 2m air temperature anomaly forecasts over several regions versus the unweighted mean.
Coevolutionary evolutionary-programming postprocessor improves 72-h surface temperature forecasts and 60-min convection nowcasts over bias-corrected NWP and logistic regression.
NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
Machine learning post-processing of model output to produce grid-point temperature forecasts for the Beijing area.
Institute of Atmospheric Physics, Chinese Academy of Sciences
GRU-based correction of wind speed forecast errors to improve short-term wind power forecasting.
Compares seven statistical post-processing methods for probabilistic HARMONIE-AROME solar radiation forecasts; quantile regression and generalized random forests perform best.
Japanese conference paper proposing machine learning to predict the accuracy of weather forecasts; abstract unavailable, so details are unverified.
Regression-based regionalization trains an ANN on a few pixels per region to bias-correct CCSM temperature and precipitation over northern South America.
Combines wavelet decomposition and machine learning to improve local precipitation forecasts from the NMME multi-model seasonal ensemble.
Uses machine learning to correct wind speed predictions from a numerical weather prediction model, improving site-level forecast accuracy.
Neural networks (MLP, TLFN, NARX) correct daily near-surface temperature from NA-CORDEX output at Big Trout Lake, reaching R2 above 0.84.
A CNN corrects cloud-resolving model precipitation output for very short-term rainfall forecasts, trained against JMA analysed rainfall over urban cells.
Surveys artificial neural network methods for rainfall prediction and argues they improve accuracy over traditional NWP and statistical models.
Nanjing University of Information Science & Technology (NUIST)
Ensemble of ANN, SVR and Gaussian process models blends NWP wind forecasts with weather-station data via Bayesian model averaging for wind power forecasting at ERCOT.
Thesis applying neural networks to statistically post-process numerical weather prediction output; abstract unavailable, so significance is judged from title only.
Random forest and SVM multi-model ensembles outperform simple averaging and Bayesian averaging in reproducing observed Australian monthly rainfall and temperature from CMIP5 GCMs.
Compares ELR, ZAGA and quantile regression forests for postprocessing HARMONIE-AROME hourly precipitation; QRF is most skillful for high thresholds.
Artificial neural network statistically downscales MM5 temperature forecasts at three Western Himalaya stations, reducing RMSE to about 2 °C for day-2 forecasts.
Extra-Trees and AdaBoost correct seasonal forecasts and WRF downscaling to predict 6-month drought (SPI) in data-sparse Fiji, outperforming bias-corrected forecasts.
Neural networks for ensemble post-processing outperform benchmark distributional regression for 2-m temperature at German stations, using embeddings and auxiliary predictors.
Neural network post-processing of three years of WRF output improves short-term cloud ceiling and cover forecasts for aviation use.
Random forests trained on 11 years of GEFS reforecasts give better CONUS extreme precipitation exceedance forecasts than raw ensembles and ECMWF.
A differential polynomial network corrects 24-hour numerical relative humidity forecasts using local observations, improving forecasts in most cases.
Quantile regression forests calibrate Météo-France ensemble temperature and wind forecasts, outperforming EMOS and giving better-calibrated, sharper probabilistic forecasts from 3 to 54 h.
Early work using artificial neural networks to produce localized precipitation forecasts from NWP model output. Abstract not available; classification based on title.
WeatherGFM is a generalist weather foundation model using in-context visual prompting to handle up to ten weather tasks, including forecasting, super-resolution and post-processing.
Foundation modelsShanghai Artificial Intelligence Laboratory
Combining three NWP sources with ML wind-downscaling reduces Nordic wind power forecast errors by 8-30%, with turbine-level aggregation improving accuracy by 10-15%.
CNN-BiLSTM-attention model forecasts convective and heavy precipitation from GFS data, reaching threat score 0.336 versus 0.252 for forecasters.
Extremes & hazardsNanjing University of Information Science & Technology (NUIST)Institute of Atmospheric Physics, Chinese Academy of Sciences
A denoising diffusion model restores fine-scale detail in FuXi 5-day forecasts, improving extreme precipitation and wind over HRES, but not TC intensity.
A neural network predicts parametric wind speed distributions, with the M-Rice distribution outperforming alternatives and ERA5 neighbour data improving 6-hour station forecasts.
SwinIR and Uformer transformers downscale and bias-correct ECMWF HRES 2 m temperature and 10 m wind to 0.01°, outperforming UNet and RCAN baselines.
Downscaling & super-resolutionFudan University (FuXi)Alibaba (DAMO Academy)
Proposes deep ensembles to improve uncertainty in deep-learning statistical downscaling of climate projections, tested with pseudo-reality experiments for non-stationary future climates.
MODWT wavelet decomposition combined with machine learning improves daily rainfall forecasts three days ahead at three Turkish stations, judged by MSE and Nash-Sutcliffe efficiency.
Proposes a stacking ensemble of machine learning algorithms to statistically downscale three CMIP6 GCMs for rainfall at two Iranian basins under SSP scenarios.
Random Forest and BP networks with causally screened circulation predictors forecast western China precipitation 1, 3 and 6 months ahead, improving 1- and 3-month skill over an institute's forecasts.
Compares four ML models trained on public weather forecasts versus observations for one-day-ahead reference evapotranspiration; forecast-trained models perform better.
On-site adaptive probabilistic wind power forecasting combined with NWP forecasts improves short-term reliability for grid operation.
Proposes eleven stacking and ensemble learners that blend PERSIANN and IMERG satellite precipitation with gauge data over CONUS, improving monthly accuracy.
Releases EUPPBench, a time-aligned forecast-observation dataset with evaluation methods and a first benchmark of 2 m temperature postprocessing.
Benchmarks, datasets & evaluationDeutscher Wetterdienst (DWD)ECMWFKarlsruhe Institute of Technology (KIT)Met OfficeMétéo-France (incl. CNRM)MeteoSwisscode
Hybrid model corrects NWP wind speed with multi-scale deep learning for seven-day wind power forecasts, cutting MSE by 65% versus a state-of-the-art baseline.
Bayesian deep learning downscales daily wind speed for climate monitoring; the abstract is missing, so the scope is inferred from the title.
GMDH model using lagged sea surface temperature predicts monthly rainfall in Pará, Brazil, with R 0.96 versus 0.82 for a multilayer perceptron.
Intercomparison of deep-learning super-resolution bias-correction methods for Indian summer monsoon rainfall in CORDEX South Asia simulations; the abstract is unavailable, so the classification relies on the title.
S2S models overestimate heatwave frequency over northern Nigeria at 30-day lead; deep-learning bias correction improves forecast reliability.
Extremes & hazardsECMWFNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
XGBoost corrects precipitation gauge undercatch (51.4% at a 4147 m Tibetan site), beating traditional methods and transferring to Norway and the US.
Data-driven ML models forecast stratospheric sudden warmings better than ECMWF S2S beyond 25 days and boost ensemble skill by up to 20%.
Sub-seasonal to seasonalETH ZurichSwiss Data Science Center (SDSC)
A two-step U-Net (TU-Net) predicts 4-month-lead summer precipitation over the Yangtze and beats most NMME dynamical models on correlation and RMSE.
Sub-seasonal to seasonalNanjing University of Information Science & Technology (NUIST)
Compares statistical downscaling methods, including XGBoost, for temperature and precipitation in Spain and Central America, finding model output statistics best for temperature.
Probabilistic Output Extreme Learning Machine builds multi-model ensembles for Indian summer monsoon rainfall, aiming to avoid overconfidence and Gaussian assumptions. No quantitative skill score given.
Sub-seasonal to seasonalNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
SALAMA, a feedforward network post-processing convection-resolving NWP fields, gives calibrated thunderstorm probabilities with skill beyond NWP reflectivity up to eleven hours lead.
LSTM post-processing uses MJO and jet-stream quasi-periodic signals to weight operational S2S ensembles for Asian minimum 2 m temperature.
Sub-seasonal to seasonalNanjing University of Information Science & Technology (NUIST)
Conditional GAN and U-Net generate high-resolution short-range precipitation maps from full NWP variables, reducing over-forecast bias; CGAN gives the most realistic output.
Downscaling & super-resolutionDeutscher Wetterdienst (DWD)Forschungszentrum Jülich / Jülich Supercomputing Centre
Neural network and Gaussian process models fill gaps and bias-correct ERA5-Land meteorological series against station data, reducing NRMSE to 2–11%.
Studies predictor selection for CNN-based statistical downscaling of monthly precipitation, aiming to improve regional precipitation fields from large-scale predictors.
Downscaling & super-resolutionInstitute of Atmospheric Physics, Chinese Academy of SciencesChina Meteorological Administration (CMA)
Evaluates near-surface air temperature reanalyses over Turkey and downscales them with a random forest for complex terrain; no abstract available.
Applies LightGBM to probabilistically merge contiguous-US gauge data with PERSIANN and IMERG satellite precipitation, outperforming quantile regression forests at extreme quantiles.
Customized deep learning with weighted losses, multitask learning and covariates improves hourly precipitation bias correction and downscaling, better capturing extremes than quantile mapping.
Recurrent neural networks learn model bias from analysis increments in an idealized Lorenz-96 setup, outperforming linear regression and reducing extended-range forecast error growth.
A deep CNN estimates site-specific surface wind speed from gridded ECMWF ERA5 across eastern China, reducing RMSE 23–26% versus linear interpolation.
Downscaling & super-resolutionChina Meteorological Administration (CMA)
A CNN-RNN model combining multi-NWP grid inputs with farm histories improves day-ahead wind power forecasts over LightGBM when trained globally across farms.
Explores replacing precipitation ensemble prediction systems with machine learning; no abstract is available, so the approach and results are unclear.
Hybrid GWO-WNN, VMD-LSTM and Q-learning model combines NWP forecasts with wind power series to improve short-term wind power prediction accuracy.
CNN downscales 25 km regional climate precipitation to 1 km in mountain terrain, cutting RMSE 34% versus raw output and 12% versus quantile mapping.
A U-Net corrects CFSv2 one-month-lead summer precipitation forecasts over China, cutting RMSE by about 43-50% on test data while leaving correlation skill largely unchanged.
Applies machine learning to improve tropical cyclone intensity forecasts from the NCMRWF ensemble prediction system; abstract not provided, so results are unknown.
Comparison of eight machine learning algorithms correcting PERSIANN satellite precipitation against GHCN gauges over CONUS finds XGBoost and random forests most accurate.
WMO prize challenge winners used AI to post-process ECMWF S2S forecasts, beating bias-corrected references for weeks 3-6 precipitation and temperature.
Sub-seasonal to seasonalColumbia University (LEAP, Lamont-Doherty)Environment and Climate Change Canada (ECCC)ECMWFMax Planck Institute for MeteorologySwiss Data Science Center (SDSC)
Spatial Mixture-of-Experts layer routes experts by location and sets new state of the art for medium-range weather prediction and ensemble post-processing.
Compares GEV models with GLM or neural network spatial interpolation of extreme Mediterranean precipitation parameters, with the ANN hidden-unit complexity key to performance.
CNN using ECMWF SEAS5 pressure and temperature forecasts beats the raw ECMWF output for monthly regional rainfall in Great Britain at 1 to 6 month leads.
Generates future weather files for India from CanESM2 using spatial and temporal downscaling plus ML bias correction across three RCP scenarios and five future years.
Early wildfire alert system combining a 3D CNN on satellite data and a CNN bias correction of WRF, feeding a Fire Weather Index for Colombian regions.
Archived ECMWF ensemble dataset for 2017 to 2020 over Europe and North America supports probabilistic solar forecasting, with calibration and ML-versus-physics case studies.
Contrastive-learning GAN performs super-resolution and bias correction of CMIP6 daily precipitation and temperature, reaching double NEX-GDDP resolution with comparable or better bias reduction.
Genetic-programming AutoML wind power forecasts combining three NWP sources cut errors by 8–22% at a Norwegian wind farm using direct and indirect turbine-level schemes.
Random forest trained on GFS forecasts and radiometer data predicts zenith atmospheric noise temperature at NASA Deep Space Network sites, enabling up to 2.43 dB downlink gain.
Applications & impactsNASACalifornia Institute of Technology (Caltech)University of California, Los Angeles (UCLA)
Combines CWB WRF-based NWP wind speeds with PCA/EMD preprocessing and XGBoost for wind power forecasts, reaching 5.43% NRMSE at one hour ahead.
ANN classifies monthly and daily rainfall in Java with up to 89% accuracy, then Ordinary Kriging maps six-month rainfall estimates.
Proposes supervised SLMVP dimensionality reduction for NWP-based solar irradiance forecasts, achieving about 3-5% lower MAE than PCA and LPP at Seville and Lisbon.
Stochastic weight averaging improves a neural radiation emulator in a Korean NWP model, giving a 60-fold radiation speedup and 12.6% week-long flux gains.
Hybrid physics-ML & parameterizationKorea Meteorological Administration (KMA)
Hybrid CNN-RNN-MLP network post-processes GEFS western North Pacific tropical cyclone tracks at 2–6 day leads, reducing track error and correcting speed and direction bias.
Tree-based ML (decision tree, random forest, AdaBoost) combines four seasonal forecast systems for China summer rainfall, reaching ACC 0.30 versus 0.21 for simple averaging in 2019–2021.
PSO-tuned extended RBF network estimates NWM zenith tropospheric delay accuracy against GNSS references, with average error better than 4 mm.
Machine and deep learning models predict intensity of newly formed blocking events at onset with error comparable to GEFS, using less computation.
Machine learning pipeline to improve NWP-based wind power forecasts in Portugal, targeting under 10% error; results are modest and the method is not novel.
VMD decomposition with an attention-LSTM encoder-decoder improves wind power forecasts from NWP inputs over several benchmark models.
Deep recurrent networks trained on Croatian wind farm output and NWP forecasts are compared with operational tools, with a focus on configuration sensitivity.
Partially-interpretable neural networks for extreme quantile regression of U.S. wildfire risk, plus a point-process extreme value model avoiding finite lower-endpoint issues.
Differential polynomial neural network with PDE-modular structure predicts intra-day and day-ahead PV output from weather and plant records, compared against statistical and probabilistic baselines.
Backpropagation neural network fuses GNSS precipitable water vapour with ERA5, cutting ERA5 RMSE by 18.5% in China and raising correlation to 0.99.
Random forests trained on GEFS reforecasts and severe weather reports give probabilistic US severe-weather forecasts for days 4-8, comparable to SPC outlooks at days 4-5.
Extremes & hazardsColorado State University (incl. CIRA)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Evaluates Dark Sky daily forecasts against 124 Florida and Georgia stations; finds good yes/no rain detection (POD 0.91–1.0) but poor rainfall amounts.
Systematic PRISMA review of 110 studies on statistical, NWP and ML short-term rainfall forecasting, highlighting hybrid models for nowcasting.
Presents XCast, an open Python toolkit that applies scikit-learn-style methods gridpoint-by-gridpoint to gridded climate data, simplifying statistical and ML climate forecasting workflows.
Software, scaling & compressionNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Columbia University (LEAP, Lamont-Doherty)
Linear regression with lagged ENSO and IOD indices matches or beats an MLP for seasonal spring rainfall in Western Australia.
KoMet benchmark pairs GDAPS-KIM NWP outputs with Korean station observations to test deep post-processing of precipitation, with open-source code.
Compares analog, GLM and neural network statistical downscaling of daily Tx and Tn at 101 Argentine stations across three reanalyses, including a warm-period extrapolation test.
GA-RBF neural network downscales CMIP5 GCM output to monthly rainfall at Taichung and Hualien, Taiwan, with separate typhoon and non-typhoon rainfall modelling.
Conditional GANs predict ensemble spread of 500 hPa geopotential from a single deterministic forecast, avoiding costly operational ensemble runs.
Convolutional networks learn IFS model bias from radio occultation temperatures, pretrained on ERA5 and transfer-learned on one year of IFS data, as an alternative to weak-constraint 4D-Var.
Machine learning combines NWP outputs with local observations to forecast wind speed and power 10 minutes to 4 hours ahead, comparing variable selection and linear versus nonlinear models.
Deep learning bias correction and downscaling of multiple CMIP climate model simulations; abstract not available, so scope is inferred from the title.
Ensemble-learning neural networks interpolate neighbouring station data, terrain and NDVI to predict next-hour minimum temperature and frost at sites without local sensors, detecting 92.55% of frost events.
A deep learning model pre-trained on GCM hindcasts and transfer-learned on ERA5 raises seasonal China precipitation correlation by 0.10-0.13 over raw GCM output; used operationally for guidance.
Sub-seasonal to seasonalTsinghua University (NowcastNet)China Meteorological Administration (CMA)
Knowledge distillation and transfer learning of forecast errors improve turbine-level wind power predictions at an Arctic wind park by 3.3% to 23.9%.
Chains dimensional reduction with normalizing flows to produce conditional probabilistic forecasts of surface wind from 22 years of WRF simulation data.
CorrectorGAN jointly bias-corrects and super-resolves global ensemble precipitation forecasts over the US against radar, approaching an operational regional model in seconds.
Downscaling & super-resolutionThe Alan Turing InstituteUniversity of Oxford
Spatially enhanced loss functions using neighbourhood and spectral filters improve neural network convection forecasts, with spectral-filter models giving better-calibrated probabilities than pixelwise training.
Benchmarks, datasets & evaluationColorado State University (incl. CIRA)NOAA (incl. GFDL, PSL, GSL, NCEP/EMC)
Deep learning probabilistic day-ahead wind power forecasts for an Arctic site improve when measured weather and NWP are used as covariates to correct NWP biases.
A deep neural network corrects bias in daily satellite precipitation estimates over tropical regions, using ground observations as the reference.
Generative deep learning turns weather forecasts into multi-step solar irradiation predictions for energy use; the abstract is unavailable, so details are limited.
Wavelet coherence predictor screening improves ANN statistical downscaling of CMIP5/CMIP6 precipitation and temperature, projecting precipitation declines of 7–29% for the study region.
Couples ERA5 and ECMWF HRES inputs with a CatBoost model and grid selection to forecast Turkey's day-ahead national wind power, reaching 7.6% NRMSE.
Compares ML and ensemble methods for day-ahead wind farm output using lagged NWP point forecasts, with a new averaging ensemble giving the lowest error at two farms.
Hybrid neural-network and regression models using GloSea5 and SEAS5 pressure indices forecast Irish winter and summer rainfall up to 4 months ahead.
NN-GPR combines multiple climate models into a 50 km projection without regional models, rivaling RCMs for temperature and precipitation tails.
Model residual machine learning uses station forecasts to fit ECMWF residuals, producing high-resolution grids of temperature, humidity and wind that beat interpolation.
Downscaling & super-resolutionInstitute of Atmospheric Physics, Chinese Academy of Sciences
An M5' decision tree with model output statistics downscales BNU-ESM CMIP5 output, projecting North Atlantic wind speed declines and a 13–19% drop in wind energy by 2030–2040.
A convolutional LSTM with backward stepwise feature selection forecasts the Oceanic Niño Index 1 to 12 months ahead, with SST the best predictor.
Introduces ENS-10, a 20-year, 10-member perturbed NWP ensemble dataset at 0.5° for ML post-processing, with baselines for 48-hour correction and extremes.
Super-resolution deep learning merges 31 CMIP6 models into a 0.5° monthly near-surface temperature dataset (1850–2100) with bias correction and downscaling.
Downscaling & super-resolutionNanjing University of Information Science & Technology (NUIST)
Late fusion of multiple neural networks with a denormalization layer improves long-range 2m temperature forecasts, beating 30-year climate normals; no headline number given.
Deep ensemble of self-normalizing networks with Gaussian outputs wins top-3 in Yandex Shifts weather temperature regression challenge under distribution shift.
Wavelet packet denoising with CatBoost combines station observations and NWP priors to beat LSTM and random forest for short-term Beijing weather forecasts.
Hybrid machine learning with rough-set feature reduction and WRF outputs improves solar radiation prediction in Saudi Arabia, with classification rates up to 8.5% higher.
Random forest predicts Yangtze summer precipitation with 0.473 skill from December, outperforming eight climate models in several respects.
Sub-seasonal to seasonalNanjing University of Information Science & Technology (NUIST)
Artificial neural network downscales five GCMs' rainfall over Pune, India, giving a higher correlation than distribution-based scaling for the best models.
Gradient boosting and analog ensembles extrapolate NWP wind-speed uncertainty from one year of ensemble runs to 20 years for offshore wind siting, with R2 above 0.75.
A machine learning approach to precipitation level prediction and mapping is proposed on a cloud big-data environment. The abstract is truncated and gives no quantitative results.
Autoencoder correction of regional NWP feeds a hybrid neural network that forecasts regional wind power, needing only about three months of training data.
A 3D CNN from one meteorology frame predicts daily CONUS precipitation, beating operational weather models up to 5-day leads and improving them when combined.
Machine learning bias correction of billions of smartphone pressure observations cuts errors 43-57% and yields analyses 21-31% better than MADIS networks.
ML trained on a large climate model ensemble yields seasonal precipitation forecasts over the western US competitive with NMME dynamical models.
Sub-seasonal to seasonalCalifornia Institute of Technology (Caltech)NASA
Ridge and lasso models using Niño3.4 give modest week 3-4 skill for US winter temperature, only slightly beating CFSv2 and climatology-based baselines.
A feedforward neural network grid model of GNSS weighted mean temperature over China improves on GPT2w-Tm by about 8.6–10.9% in validation.
Random forests trained on convection-allowing WRF output probabilistically forecast excessive rainfall over CONUS, but they consistently underperform WPC's day-1 Excessive Rainfall Outlooks.
Compares analogue, neural network and GLM statistical downscaling of daily precipitation extremes over southeastern South America; no single method performs best across all aspects.
Compares analog, regression, ANN, SVM and kernel ridge downscaling over Spain with perfect predictors, finding nonlinear methods add value for temperature.
A residual U-Net trained on reanalysis, with cubed-sphere inputs, learns global precipitation that doubles grid-point skill versus the operational IMD NWP model, as a proof of concept.
Support vector regression using JMA mesoscale ensemble forecasts improves day-ahead Kanto solar irradiance prediction and reduces the maximum error.
Deep learning bias correction of multi-model dynamical forecasts cuts MJO amplitude error by about 90% and phase error by 77% over four-week forecasts.
CSU-MLP random-forest excessive rainfall forecasts, built from ensemble reforecasts, were transitioned into operational use at NOAA's Weather Prediction Center.
Extremes & hazardsNOAA (incl. GFDL, PSL, GSL, NCEP/EMC)Colorado State University (incl. CIRA)
Compares linear regression, KNN, trees and neural networks to reconstruct ERA5 wind fields from 1% of points, with k-means sensor placement improving accuracy.
Neural network combines RadVOR radar nowcasts with postprocessed COSMO-DE-EPS ensembles, giving calibrated precipitation threshold probabilities that beat both inputs for 1 to 6 h.
ML-optimized ensemble selection and weighting improves preseason Atlantic hurricane forecasts over simple averaging for major hurricane counts.
ML-enhanced analogue method using LLE reduces NWP precipitation error for tropical cyclones at Shenzhen, improving on raw NWP and traditional analogue approaches.
Proposes XGBoost-based statistical downscaling of ACCESS1.0 GCM output for Indian river basin, with bias near zero and NSE 0.75-0.85 against 35 stations.
Machine-learning bias correction of 2016 smartphone pressure data reduces bias, and dense analysis resolves convective structure in a Beijing hailstorm better than station data.
Observations & remote sensingNSF NCAR / UCARChina Meteorological Administration (CMA)
Reviews how artificial intelligence is applied across numerical weather prediction, including data assimilation, error correction, and end-to-end neural forecasting models.
XGBoost wind power forecaster uses autoencoder-compressed 20-member NWP ensembles from Taiwan CWB, improving short-term deterministic and probabilistic accuracy.
Regression, ANN and GWR corrections of IMERG late-run rainstorm precipitation in southern China slightly improve accuracy, with the best method varying by region.
Hybrid NWP and ANN models forecast wind farm power over complex terrain beyond 6 hours, with the direct power model reaching 8.76% normalized MAE.
Combines numerical weather prediction and machine learning to characterise and forecast wind power ramp events; no abstract is available.
ANN ensemble combines five regional climate models to correct Korean precipitation against observations, improving on individual RCMs, most strongly inland.
EALSTM-QR produces interval wind-power forecasts from numerical weather prediction inputs using deep learning; the title suggests an energy-sector application (no abstract).
Object-based verification of storm-scale mesocyclone probabilities, plus evaluation of ML calibrations of Warn-on-Forecast severe weather guidance.
Neural network maps GDAS numerical weather data to photovoltaic power output at an Italian plant with under 10% error, even without local weather measurements.
Feedforward neural networks statistically downscale GEFS precipitation forecasts, assessing which predictors matter at each lead time up to 16 days.
Compares SRCNN variants and DeepSD for 4x downscaling of Indian summer monsoon rainfall, finding DeepSD gives the best spatial distribution and lowest RMSE.
XGBoost combines NWP turbulence indices into a multi-index consensus, raising aviation turbulence forecast skill by 3–17% over Asia-Pacific against pilot reports.
Random forest post-processing of radar nowcasts and NWP for hourly precipitation occurrence at 41 Norwegian stations, improving skill and reducing false alarms.
Applies a temporal convolutional network with transfer learning to forecast offshore visibility at Haikou Station, improving 24 h forecast scores and beating the CUACE model on TS score by 0.16.
Local & station forecastingNanjing University of Information Science & Technology (NUIST)
Modular fuzzy neural network with information-gain predictor selection improves 24 h heavy-rainfall forecasts from ECMWF output over Guangxi stations.
Deep learning predicts short-term photovoltaic power using weather forecast inputs; abstract not available, so classification is low confidence.
Neural network models with ENSO indices forecast monthly rainfall anomalies in Southwestern Colombia five months ahead, with the data-driven approach showing suitable skill.
XGBoost corrects MAIAC satellite column water vapor errors, cutting RMSE by 26.9% for Terra and 16.5% for Aqua against AERONET sites.
Deep neural networks predict wind power generation from weather forecast data; abstract not available, so classification is low confidence.
Machine learning models predict precipitation from large-scale climate indices in an arid-land region; abstract not available, so classification is low confidence.
Random forests and gradient-boosted regression trees using HRRR forecast features improve upper-level aviation turbulence (EDR) forecasts and reduce false alarms versus regression baselines.
Generalized regression neural network with dual-factor feature extraction for rainstorm prediction; only the title is available, so the abstract-based assessment is limited.
Convolutional autoencoder corrects pixel-wise bias of PERSIANN-CDR daily satellite precipitation against APHRODITE over the Mekong, outperforming a standard-deviation method.
Machine-learning multi-model ensemble and bias correction of GCMs project Nigerian meteorological drought changes under RCP scenarios through 2099.
Random forests classify continuous versus intermittent precipitation textures in AROME high-resolution forecasts, outperforming other classifiers and transferring to gridded observations for verification.
Machine learning with PCA, ANN and quantile mapping downscales GCM output to 30 m sub-daily temperature, humidity and wind in complex terrain, improving on GCM accuracy.
Random forests trained on GEFS reforecasts give calibrated Day 1-3 severe weather outlooks that beat SPC at Days 2-3, and blending with SPC improves Day 1 outlooks.
Neural-network-augmented quantile mapping bias corrects GCM-downscaled precipitation at 18 Florida gauges, reproducing extreme precipitation indices better than existing quantile methods.
NCAR's updated wind power system combines NWP with ML power conversion, analog ensembles, radar ramp forecasts and fuzzy-logic icing alerts for Xcel Energy.
Random forest, SVR, ANN and a multi-model ensemble correct next-day LDAPS extreme temperature forecasts in Seoul, cutting maximum-temperature RMSE from 2.08 to about 1.6 °C.
Data-driven multi-model ensemble combining seasonal forecasts for deterministic and probabilistic precipitation prediction; abstract missing, so classification is based on the title only.
Extends concentrated solar dispatch planning with probabilistic weather forecasts and an ML calibration step, reporting up to 15% higher revenue than a persistence approach.
Compares statistical and machine learning ensembles for daily temperature downscaling; abstract unavailable, so judgment rests on the title.
A conditional VAE replaces the historical archive in analog ensemble wind speed correction for NAM, matching AnEn skill with up to 25x less memory.
Ranking and combining spatially distributed NWPs with gradient boosting and quantile regression improves day-ahead probabilistic wind farm power forecasts by up to 1.8%.
Kernel-density analogue ensemble forecasts PV output from weather forecasts in Portugal, beating a neural network in most cases and running 32 times faster.
Machine learning corrects a coupled climate model's seasonal forecasts of daily air temperature for field-scale agricultural management.
Sub-seasonal to seasonalKorea Meteorological Administration (KMA)
Deep 3D U-Net variants estimate weather forecast ensemble spread from a small set of numerical simulations, reducing the compute needed for uncertainty quantification.
Statistical downscaling of monthly precipitation to 1 km for the Tarim River Basin, combining polynomial regression with batch gradient descent; it outperforms ordinary polynomial regression.
Random forests post-processing HREFv2 ensembles with isotonic calibration give severe hail probabilities comparable to SPC outlooks over CONUS.
Quadratic discriminant analysis and neural networks classify Vaigai catchment rainfall classes from ENSO indices over 1959–2016 at seven rain gauges.
Instance-based transfer learning borrows records from neighboring cities to extend scarce station data for temperature and precipitation forecasts in Zhaoqing, China.
GRU combined with weather forecasts predicts 24-hour-ahead solar irradiance, cutting RMSE 11.9% versus RNN and 23.3% versus backpropagation.
Geographically weighted ensemble of ML learners plus residual-network downscaling of reanalysis yields high-resolution wind speed over China, improving R2 by 12–16%.
Feed-forward network on HWRF output beats operational intensity guidance at 24–57 h in cross-validation and quadruples rapid-intensification hits.
Neural networks trained on past BRAMS forecasts and lightning density fields predict cloud-to-ground discharge occurrence as an ancillary convective-alert tool.
CNN-LSTM statistical downscaling of ECMWF precipitation over a monsoon basin beats quantile mapping and SVM, with gains up to two-week lead times.
Downscaling & super-resolutionTsinghua University (NowcastNet)
A wavelet-denoised multicellular gene expression programming model (WTGEPRP) forecasts precipitation at Zhengzhou, Nanning and Melbourne, outperforming several baseline hybrid models.
Ordinal neural network refines a persistence baseline to predict airport low visibility, outperforming persistence and autoregressive models at longer horizons on station data.
Local & station forecastingUniversity of Cambridge (Aardvark)
A CNN trained on past forecasts assigns a confidence score to medium-range forecasts without running an ensemble; it is cheaper but less skilful than ensembles.
Neural networks post-process PNU CGCM seasonal hindcasts with sea-ice and SST indices to downscale winter 2-m temperature over Mongolia to 20 km, with improved skill.
Ensemble of regression models with the SubseasonalRodeo dataset beats the top Rodeo entry and improves debiased CFSv2 precipitation skill by 129-169%.
Sub-seasonal to seasonalMicrosoft Research (AI for Science)Stanford University
Compares ANN, K-nearest-neighbour and ANFIS for downscaling GCM precipitation at a Latyan basin station; KNN performed best monthly.
Review of gust forecasting methods, comparing physical boundary-layer parametrizations on NWP fields with statistical and machine learning approaches trained on observations, and outlining challenges from higher resolution.
An MLP downscales arid-region monthly precipitation, and four bias correction methods are compared, with power transformation found most reliable.
Gaussian process regression with a multi-scale anisotropic kernel interpolates NWP 10 m wind fields, cutting RMSE by up to 95% versus neural networks in cyclone cases.
A metaheuristic bat-optimised neural network downscales climate model predictors to Kuching station precipitation, reporting average r of 0.96 across three GCMs.
Wavelet-entropy and SOM predictor selection with ANN downscales multi-GCM station rainfall, projecting a 21–29% rainfall decrease for 2017–2050.
Generalized regression neural network consensus of dynamical models improves Atlantic tropical storm intensity forecasts over 2012–16, outperforming individual models and the ensemble mean.
Machine learning learns WRF precipitation errors from past observations to correct systematic bias and identify which physics schemes drive forecast uncertainty.
Neural-network decision support system calibrates weather forecasts for wind farm orography and predicts power output, improving prediction at ten real wind farms.
Machine learning corrects smartphone barometer biases using ancillary sensor data, cutting average pressure bias by 82% and supporting use in NWP.
Machine-learning post-processing gives calibrated probabilistic forecasts of flash floods, extreme rain, tornadoes, hail and convective winds across the US.
Machine learning error correction makes low-cost Mini-AWS pressure readings meet WMO ±0.1 hPa accuracy, offering a cheap complement to automatic weather stations.
Storm-based machine learning produces calibrated probabilistic severe hail forecasts from convection-allowing ensembles, achieving higher critical success index than existing hail forecasting methods.
Extremes & hazardsNSF NCAR / UCARUniversity of Oklahoma (AI2ES)
A stacked denoising autoencoder bias-corrects PERSIANN-CCS satellite rainfall, cutting precipitation-rate bias by 78–98% and false alarms by 33–43%.
An expanded Bernoulli-gamma neural density network downscales multisite daily precipitation from synoptic predictors, reproducing spatial covariances and generating realistic synthetic series in coastal British Columbia.
Downscaling & super-resolutionEnvironment and Climate Change Canada (ECCC)
