Short-term power load forecasting using bidirectional gated recurrent units-based adaptive stacked autoencoder
Jizhe Dong, Yiwen Jiang, Peiguang Chen, Jiulong Li, Ziheng Wang, Shunjie Han
Abstract
• A novel adaptive deep learning model improves power load forecasting accuracy. • The model combines bidirectional gated recurrent units and stacked autoencoders. • Bayesian optimization enhances feature extraction and hyperparameter selection. • Multi-head self-attention mechanism improves focus on critical temporal features. • Experiments demonstrate superior accuracy and robustness across four datasets. Short-term load forecasting plays a crucial role in ensuring the stability of power systems and enabling optimized allocation of resources. However, previous studies have struggled to propose an adaptive framework specifically tailored for stacked autoencoder models, which has led to relatively low prediction accuracy. Therefore, this paper introduces a novel deep learning approach that integrates a bidirectional gated recurrent units-based adaptive stacked autoencoder. The model is structured in three stages: (1) data preprocessing stage for structured input; (2) pretraining stage of bidirectional gated recurrent units-based adaptive stacked autoencoder with automatic hyperparameter optimization; (3) load forecasting stage that fine-tunes the bidirectional gated recurrent units-based adaptive stacked autoencoder, and integrates multi-head self-attention mechanism and deep residual network for enhanced load prediction. The proposed method has been validated on four datasets from the high renewable penetration 38-bus test system, China, Australia, and Malaysia, demonstrating strong adaptability, lower prediction errors, and enhanced robustness, making it widely applicable for load forecasting.