Intrinsic Explainable Artificial Intelligence Using Trainable Spatial Weights on Numerical Weather Predictions

Addressing the volatility of renewable energies like solar and wind is crucial for the energy system’s stability and optimal utilization of renewable energies. Accurate energy forecasts are important to improve scheduling. Electrical demand and renewable energies are weather-dependent and Numerical Weather Predictions have proven to be beneficial for energy forecasts due to their fine-grained spatial resolution. State-of-the-art Deep Learning approaches for energy forecasting are black-box models. However, decisions in energy systems depend on energy forecasts, and, thus, it is important that models are explainable and trustworthy. Explainable Artificial Intelligence techniques exist that add explainability to energy forecasting models, but all existing methods are only post-hoc or do not use weather data on large spatial areas. This paper introduces a novel approach to forecast energy that scales and adds intrinsic explainability by design. Therefore, we use trainable spatial weights to make accurate forecasts on large spatial areas. The trained weights can be interpreted spatially to enhance explainability and increase trust. Furthermore, the spatial weights enable a wide range of future work, including postprocessing, subregion forecasting, hierarchical learning, and spatial-temporal weights.