Integration of Deep Learning and Time Series Decomposition for Enhanced Short-Term Load Forecasting

This study aims to improve the accuracy of short-term power load forecasting through deep learning to ensure the dynamic balance and stable operation of the power grid system. In response to the poor prediction accuracy of traditional serial combination models and the issue of data leakage in modal decomposition prediction models, this paper proposes a parallel combination model based on deep learning. The model integrates a self-developed time series decomposition (TD) module, Bidirectional Gated Recurrent Unit (BiGRU), Temporal Convolutional Network (TCN), and Multi-Head Self-Attention (MHSA). First, the time series decomposition module decomposes power load data into trend and periodic components to extract long-term and short-term features. These features are then fed into parallel BiGRU and TCN deep networks for processing. The multi-head self-attention mechanism is incorporated to further enhance the model’s expressive capabilities. Additionally, the model leverages the Beluga Whale Optimization (BWO) algorithm to optimize hyperparameters, fully exploiting the complex correlations within the data. Experimental results show that the proposed BWO-TD-BiGRU-TCN-MHSA deep learning model significantly outperforms baseline models in terms of MAPE, RMSE, MAE, and R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup>. This verifies the model's high accuracy and practicality in power load forecasting, providing reliable support for the planning and stable operation of power systems.