Balancing Cost and Risk in High-Load Power Systems: An Integrated Prediction–Optimization Strategy

Accurate medium-horizon load forecasting and risk-aware unit commitment are critical for high-load power systems. This study develops an integrated prediction–optimization framework that couples 744 h recursive load forecasting with uncertainty-aware scheduling. In the forecasting stage, a CNN-LSTM model is tuned by the Dung Beetle Optimizer (DBO), while Monte Carlo Dropout is retained during inference to generate probabilistic trajectories and time-varying prediction intervals. In the scheduling stage, these forecast-derived intervals are embedded into a mixed-integer linear robust unit commitment model through a dynamic uncertainty budget. Using real-world load data from Southern China, the proposed method achieves average RMSE, MAE, MAPE, and R2 values of 2941 kW, 2137 kW, 4.33%, and 0.97, respectively. Relative to SARIMA and Informer, the average RMSE is reduced by 48.1% and 26.0%, respectively, while point-forecasting performance remains competitive with XGBoost. The proposed model also provides the best overall interval quality, with average PINAW and Winkler Score values of 0.19 and 17,049, outperforming XGBoost, CNN-LSTM, and Informer. In the scheduling study, the proposed robust strategy reduces average EENS and LOLH to 68.6 kWh and 0.0454 h, respectively, and yields the lowest average generalized total cost of CNY 97.30 million, compared with 124.69 million CNY for the deterministic benchmark and CNY 99.66 million for the chance-constrained benchmark. These results show that forecast uncertainty can be effectively translated into more reliable and economical scheduling decisions.