Data-driven method for the improving forecasts of local weather dynamics
In one sentence. Gaussian process model fuses station observations with NWP forecasts to improve local short- and long-term weather forecasts for dispersion modelling.
Abstract
This paper describes the modeling approach for lower atmosphere dynamics in a selected location. The purpose of this model is to provide short-term and long-term forecasts of the weather variables which are used as the input data for the model of the dispersion of radioactive air pollution. The information from this integrated system is important for the implementation of the population safety measures in the case of a nuclear accident with an atmospheric release. We developed a dynamical, probabilistic, and non-parametric model based on Gaussian processes (GPs). GP nonlinear autoregressive model with exogenous inputs and variational training principle was implemented for multi-output training. A Monte Carlo approach to multi-output simulation of the model for long-term forecasts is presented which allows arbitrary prior distributions over function values. The model encompasses all available measurements from the weather stations near the location of interest and combines them with the forecasts from the numerical weather prediction model. The contribution of the developed model is the harvesting of all available information and simultaneously providing interconnected forecasts. The key result of this investigation is the improvement of short-term and long-term weather variable forecasts over those of the numerical weather prediction model. Consequently, we significantly enhance the dispersion forecast of radioactive air pollution for the case study considered. The computationally demanding modeling is accelerated using general-purpose computing on graphics processing units. The proposed method represents a step forward in the assurance of safety in the case of a nuclear accident.
Labels
- Topic
- Post-processing & calibration, Local & station forecasting
- Type
- New model or method
- Method
- Classical ML (trees, GP, linear, kernels)
- Time scale
- 6 h–3 days
- Domain
- Local / station
- Data
- Station / in-situ, NWP analysis or forecasts (IFS, GFS, ...)
- Status
- Research
- Identifiers
- DOI:10.1016/j.engappai.2021.104423 OpenAlex:W3194951001
- Added
- 2026-10-22
Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.