A combined prediction model with multi-module integration for short-term power load forecasting

Abstract

The existing power load forecasting algorithms are constrained by preprocessing limitations and insufficient prediction accuracy. Temporal Convolutional Network (TCN), Bi-directional LSTM (BiLSTM), and Multi-Head Attention (MHA) for high-precision, real-time power load prediction were used in this paper to propose a hybrid model, which combined Improved Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (ICEEMDAN) to address these issues. The ICEEMDAN algorithm is initially employed for multi-layer decomposition to enhance data smoothness, thereby enabling the subsequent model to more effectively capture essential features. To overcome the BiLSTM's inability to capture long-term dependencies in extended sequences, this paper incorporates the TCN module and MH-Attention mechanism. The TCN improves local feature extraction through convolution, while the MH-Attention mechanism enables the model to focus on the most critical features for load prediction, thereby enhancing learning efficiency. The model is validated using an actual power plant in Quanzhou City, China, with a dataset of power load data obtained from real measurements. Experimental results demonstrate an R2 of 0.99802, RMSE of 0.00996, and MAE of 0.00694, showcasing exceptional forecasting accuracy. Compared to alternative models, the R2 improves by 1.5%-3.4%, RMSE decreases by 56.2%-80.6%, and MAE is reduced by 58.3%-84.1%. These results validate the model's superiority. The proposed combinatorial forecasting model framework effectively integrates the advantages of data decomposition, convolution, and attention mechanisms, allowing for an in-depth exploration of temporal features and critical patterns in power load data.