Enhancing short-term net load forecasting with additive neural decomposition and Weibull Attention

Accurately forecasting net load is important, yet it is becoming increasingly challenging due to the inherent variability of renewable energy sources. Effectively characterizing the uncertainty in net load forecasting is critical. To address this, we propose the AND-Weibull model featuring a novel additive neural decomposition (AND) that explicitly separates the load into stable (Shape) and high-frequency (Residual) components. The ’Shape’ component is built by an adaptive Fourier network, while the ’Residual’ is modeled via a Weibull-based attention LSTM, capturing non-uniform renewable-driven fluctuations. The final forecast is obtained through an adaptive, matrix-based fusion of the ’Shape’ and ’Residual’. To optimize the entire process—from decomposition to forecasting—we introduce a new Spectral Temporal Error (STE) loss function that operates in both the frequency and time domains, thereby enhancing prediction accuracy. Experiments on the OPSD dataset (Austria, Belgium, Germany) demonstrate AND-Weibull’s robust performance against existing deep learning methods (e.g. Autoformer, NHITS). In Austria, AND-Weibull reduces MAE by over 20% compared to Informer, and in Germany’s more volatile setting, it achieves approximately 41% lower MSE. Beyond deterministic metrics, the narrower interval predictions—measured by pinball-based WS, RWS, and PINAW—underscore AND-Weibull’s reliability in probabilistic forecasting, confirming its overall effectiveness.