Correcting Subseasonal Forecast Errors with an Explainable ANN to Understand Misrepresented Sources of Predictability of European Summer Temperatures

Abstract

Abstract Subseasonal forecasts are challenging for numerical weather prediction (NWP) and machine learning models alike. Forecasting 2-m temperature (t2m) with a lead time of 2 or more weeks requires a forward model to integrate multiple complex interactions, like oceanic and land surface conditions leading to predictable weather patterns. NWP models represent these interactions imperfectly, meaning that in certain conditions, errors accumulate and model predictability deviates from real predictability, often for poorly understood reasons. To advance that understanding, this paper corrects conditional errors in NWP forecasts with an artificial neural network (ANN). The ANN postprocesses ECMWF extended-range summer temperature forecasts by learning to correct the ECMWF-predicted probability that monthly t2m in western and central Europe exceeds the climatological median. Predictors are objectively selected from ECMWF forecasts themselves, and from states at initialization, i.e., the ERA5 reanalysis. The latter allows the ANN to account for sources of predictability that are biased in the NWP model itself. We attribute ANN corrections with two explainable artificial intelligence (AI) tools. This reveals that certain erroneous forecasts relate to tropical western Pacific Ocean sea surface temperatures at initialization. We conjecture that the atmospheric teleconnection following this source of predictability is imperfectly represented by the ECMWF model. Correcting the associated conditional errors with the ANN improves forecast skill. Significance Statement We want to understand occasions in which a numerical weather prediction (NWP) model fails to forecast a predictable event existing in the real world. For forecasts of European summer weather more than 2 weeks in advance, real predictable events are rare. When misrepresented by the model, predicted future states become needlessly biased. We diagnose these missed opportunities with an explainable neural network. The neural network is aware of the initial state and learns to correct the NWP forecast on occasions when it misrepresents a teleconnection from the western tropical Pacific Ocean to Europe. The explainable architecture can be useful for other applications in which conditional model errors need to be understood and corrected.