Prototype-routed exponential-gated residual learning for multi-horizon regional electricity-load forecasting

Short-term electricity-load forecasting (STLF) supports reliable dispatch, reserve scheduling, renewable integration, demand-response planning, and low-carbon power-system operation. Existing forecasting models often struggle to jointly capture heterogeneous load-weather regimes, horizon-dependent residual errors, and causal reliability control for multi-horizon prediction. To address this gap, this study proposes a Prototype-Routed eXponential-Gated Residual Network (PRX-Net) for six-hour-ahead regional electricity-load forecasting. PRX-Net combines a prototype-routed randomized backbone, recurrent residual correction, and causal previous-error gating. The randomized backbone uses global and local enhanced edRVFL experts to model broad and regime-specific load-weather behavior, while the recurrent module estimates the remaining horizon-dependent residual. The exponential gate controls the residual contribution using matured error evidence from previous forecast origins. Experiments on three regional load-weather datasets show that PRX-Net achieves root-mean-square error (RMSE) = 7 4 0 . 8 2 , mean absolute error (MAE) = 5 2 8 . 8 9 , mean absolute percentage error (MAPE) = 2 . 0 2 % , 𝑅 2 = 0 . 9 8 0 8 , normalized RMSE (NRMSE) = 2 . 9 0 % , and normalized MAE (NMAE) = 2 . 0 9 % . Compared with edRVFL, iTransformer, and PatchTST, PRX-Net reduces average RMSE by 12.91%, 2.49%, and 11.69%, respectively. These results show that combining regime-aware randomized forecasting with gated residual correction improves multi-horizon regional load prediction.