Abstract

Short-term power load forecasting is fundamental to the secure operation and optimal dispatch of modern power systems. This study proposes an Improved Pied Kingfisher Optimization–Temporal Convolutional Network–Bidirectional Gated Recurrent Unit (IPKO-TCN-BiGRU) model to address the challenges of strong non-stationarity, high randomness, and multi-factor coupling in load time series. The model employs a multi-scale TCN for simultaneous extraction of local and global temporal features, a BiGRU enhanced with an Improved Self-Attention (ISA) mechanism for bidirectional dependency modeling, and an Autoregressive (AR) module combined with an election mechanism to jointly capture linear and nonlinear load components. The Improved Pied Kingfisher Optimization (IPKO) algorithm—incorporating SPM chaotic initialization, a planetary optimization strategy, and adaptive t-distribution perturbation—is applied to globally optimize key hyperparameters, demonstrating superior convergence accuracy and global search capability over the original PKO and other benchmark optimizers. To ensure evaluation integrity, dataset splitting precedes all normalization operations, with StandardScaler fitted exclusively on the training set and applied to the test set without leakage. Validation is conducted on two benchmark datasets: a U.S. residential electricity load dataset (hourly, 2012, 13-dimensional features including HVAC and lighting systems) and a China Electrical Engineering Mathematical Modeling Competition dataset (15 min intervals, three years, enriched with five meteorological variables). The U.S. dataset exhibits a clear annual double-peak seasonal pattern, while the Chinese dataset shows strong intraday fluctuations significantly coupled with temperature and humidity, both posing substantial forecasting challenges. On the U.S. dataset, the proposed model achieves MAE = 0.0190 kW, RMSE = 0.0301 kW, MAPE = 1.7673%, and R2 = 0.9947; on the China dataset, MAE = 79.8125 MW, RMSE = 109.4154 MW, MAPE = 1.1124%, and R2 = 0.9955. The proposed model consistently outperforms six mainstream baseline models—including Transformer, Autoformer, and FEDformer—reducing RMSE by up to 34.4% and 18.9% on the two datasets, respectively, while maintaining a compact architecture of 15.2 MB and 74.6–78.9 MFLOPs. Ablation experiments confirm the significant and synergistic contribution of each module, and the direct comparison between PKO-TCN-BiGRU and IPKO-TCN-BiGRU validates that the algorithmic improvements translate into measurable forecasting gains beyond benchmark function optimization. The proposed model is most suitable for ultra-short-term to short-term single-step-ahead forecasting within a horizon of 15 min to 24 h, with an inference latency of 2.3–2.7 ms per sample, fully meeting the real-time requirements of practical power dispatching systems.

IPC Classification

Keywords