Spatiotemporal Joint Power Forecasting for Hydro-Wind-Solar Complementary Systems via a Hierarchical Complementary-Prior Regularized MSTD-HGFormer

Accurate joint forecasting of hydro-wind-photovoltaic bases is difficult because hydropower regulation, wind randomness and photovoltaic intermittency interact through both temporal complementarity and spatial-electrical coupling. This study aims to improve source-level and bundled-power forecasting by embedding the complementary mechanism of heterogeneous renewable sources into both representation learning and model training. A hierarchical complementary-prior regularized multi-scale seasonal-trend decoupled heterogeneous graph Transformer, denoted MSTD-HGFormer-HCPR, is proposed. The model first separates the historical multi-source sequence into low-frequency trend and high-frequency fluctuation components through learnable multi-scale moving-average kernels. The two components are then encoded by parallel heterogeneous graph Transformer branches using temporal patch attention and a static-dynamic fused adjacency that combines physical topology, historical correlation and learnable latent relations. A gated fusion module produces the final forecast, while a composite loss introduces capacity-weighted system- level and correlation-aware topology-level complementary-prior constraints. Experiments on a real 17-node hydro-wind-photovoltaic base show that the proposed method reduces rRMSE by 12.9–50.6% against seven baselines, improves the correlation coefficient by 0.011–0.013 over the strongest baseline, and keeps the forecasting accuracy at or above 96.2% for hydro, wind, photovoltaic and aggregate outputs. A second, 23-node cascade-radial base with five-year records is further examined under a rolling-origin and season- stratified protocol; the aggregate rRMSE is reduced by 13.2–44.0%, and the pooled accuracies remain 96.0% for hydropower, 96.0% for wind, 96.4% for photovoltaic and 96.2% for aggregate outputs. The proposed regularized joint forecaster provides stable and accurate spatiotemporal prediction for complementary renewable bases and reduces the accumulation of source-level errors in bundled output forecasting.