Short-Term Multi-Load Forecasting Based on a Hybrid Model Coupling Internal and External Differentiated Characteristics

Abstract

Fixed feature sets and static coupling assumptions in existing short-term multi-load forecasting fail to capture seasonal variations in load interactions and meteorological sensitivity within integrated energy systems. This paper proposes an adaptive hybrid framework to improve point-forecasting accuracy for electric, cooling, and heating loads. Autocorrelation and Kendall rank analyses quantify seasonal periodicity, inter-load coupling, and load-weather relationships; XGBoost then selects load- and season-specific features. These features feed into a BiLSTM network, whose hyperparameters are optimized by an enhanced Newton-Raphson-Based Optimizer (NRBO) with a quasi-opposition-based initialization, following a multi-input single-output scheme. On hourly data from Arizona State University, the model achieves MAPE values of 0.87%-1.70% (average 1.14%) across twelve season-load tasks, reducing MAPE by 54.5%- 68.1% (average 64.8%) compared with BiLSTM-XGBoost. Diebold-Mariano and Wilcoxon tests yield p < 0.05 against ARIMA, BiLSTM, SVR, and BiLSTM-XGBoost. The statistically significant error reductions indicate that the proposed framework is robust and effective under seasonal fluctuations. The improvement is mainly attributed to the coordinated use of season- and load-specific feature construction, XGBoost-based feature screening, BiLSTM temporal modeling, and improved NRBO initialization, which together capture differentiated internal and external load characteristics. These findings support more refined energy management in integrated energy systems. The proposed framework, integrating XGBoost-based selection, BiLSTM modeling, and NRBO optimization, substantially enhances multi-load forecasting performance, offering a practical solution for refined energy management in integrated energy systems.