Analog Ensemble Probabilistic Forecasting using Deep Generative Models

Alessandro Fanfarillo

Preprint · 16 January 2020 · NSF NCAR / UCAR

In one sentence. A conditional VAE replaces the historical archive in analog ensemble wind speed correction for NAM, matching AnEn skill with up to 25x less memory.

Abstract

Deterministic Numerical Weather Prediction (NWP) models are the state-of-the-science models to provide reliable weather forecasts that provide indispensable actionable data to society. However, each NWP model run carry uncertainty caused by errors in initial conditions and assumptions in the model. As a result, probabilistic forecasts can be used to quantify and even correct the model bias. The Analog Ensemble (AnEn) technique uses a historical dataset of past deterministic forecasts and their associated observations to generate an ensemble of future outcomes. One of the main advantages of the AnEn technique, along with other related statistical ensemble techniques, is that it is not necessary to run multiple NWP runs by varying initial conditions or model settings. However, all these techniques require access to the entire historical dataset to generate analogs. Moreover, these techniques require reading the dataset for every forecast which is computationally expensive. In this work, the whole historical dataset is replaced by a model that has the capability to learn the Probability Density Function (PDF) of that dataset. Specifically, we utilize a Conditional Variational Autoencoder (CVAE) deep generative machine learning model in order to correct the wind speed forecasts of North American Mesoscale (NAM) forecasting system. As a result, we feed the values forecasted by the NWP model as a condition to our CVAE and generate an ensemble used to correct the forecasted value in constant time and with small memory usage. Initial results show that CVAE probabilistic performance is comparable to AnEn while CVAE can be up to 25 and 2 times smaller in memory and runtime, respectively, for 5 years of historical data.

Labels

Labels and summary are generated by Claude and reviewed weekly. Wrong? Edit this entry.