A medium- and long-term model for forecasting renewable energy demand: PSO-TPA-LSTM-Stacked

Accurate renewable energy demand forecasting is important for medium- and long-term power-system planning. This study proposes PSO-TPA-LSTM-Stacked, a compact forecasting framework that integrates particle swarm optimization (PSO), temporal pattern attention (TPA), and stacked long short-term memory (LSTM) networks. PSO optimizes a bounded set of structural hyperparameters, TPA refines temporal representations from LSTM hidden states, and a narrower second LSTM layer provides an additional nonlinear transformation while limiting model capacity. On the common 2015-2022 held-out Sichuan test period, the proposed model achieves RMSE = 60.651 GWh, MAE = 52.781 GWh, and MAPE = 1.82%. Relative to the best-performing Hybrid benchmark, RMSE decreases by 10.00%, MAE by 1.52%, and MAPE by 5.70% (0.11 percentage points in absolute terms). Rolling-origin validation and a cross-region experiment using Yunnan Province provide preliminary evidence of temporal stability and regional applicability, but the small annual sample limits strong generalization claims. Recursive forecasts for 2023–2029 are therefore reported together with 90% residual-bootstrap prediction intervals. The framework is intended as a quantitative scenario-analysis tool for medium- and long-term renewable-energy planning, including wind-power integration and wind-hydro coordination in hydropower-dominated provincial systems.