A spatiotemporal deep learning framework for electricity price linkage forecasting using graph attention and reinforcement learning

Abstract

Medium- and long-term electricity spot-price forecasting is affected by temporal fluctuation, cross-market transmission, and unstable regional coupling, making single-market sequence models insufficient for linkage-aware prediction. This study proposes DL-GNN-LP, a spatiotemporal forecasting framework that combines LSTM temporal encoding, three-layer GAT market propagation, PPO-based binary edge control, and multi-task learning. The objective is to improve continuous price regression and auxiliary price-trend classification while preserving delayed cross-regional transmission consistency. A 3-year dataset from 10 electricity market regions during 2020–2022 was constructed. After cross-region alignment, cleaning, normalization, linkage-label construction, and 24-h sliding-window conversion, 26,000 hourly timestamps were transformed into 25,977 graph snapshots, each of which is defined as one model sample. Under the same chronological split, preprocessing procedure, prediction horizon, tuning budget, and evaluation protocol, DL-GNN-LP outperforms ARIMA, Random Forest, LSTM, GCN-Base, TCN, N-BEATSx, Temporal Fusion Transformer, TimesNet, DLinear, NLinear, PatchTST, Informer, and ST-GCN. The model achieves R 2 = 0.880 ± 0.006 for continuous price regression, 92.4 ± 0.5% Accuracy, and 88.4 ± 0.6% F1-score for auxiliary trend-direction prediction. Ablation and edge-control comparisons show that graph attention, dynamic edge weighting, PPO-guided propagation, and multi-task constraints jointly improve forecasting stability. These results indicate that DL-GNN-LP provides a reproducible linkage-aware framework for medium- and long-term electricity price forecasting and dispatching support.