Multi-time-scale system load forecasting based on deep learning

Abstract

Many current prediction methods used in power system load forecasting are adversely affected by an imbalance of feature distributions at different time scales, yielding constrained prediction accuracy and elevated error rates. This paper introduces a multiscale prediction method integrating a multiscale attention convolutional network with a long short-term memory (LSTM) network. First, historical charging-load data from charging stations are obtained and structured into distinct temporal segments for multiscale convolutional feature learning. The multiscale attention convolutional network employs parallel convolutional branches with varying dilation depths to process the input sequence, integrating the outputs via feature concatenation and channel weighting to form unified conceptual representations. An LSTM network models the multilevel time-series data derived from the various convolutional filters, enabling predictions across multiple time scales. The attention mechanism assigns weighted importance scores to the fused features based on temporal relevance prior to sequence modeling, improving prediction stability across various load periods for electric vehicle charging loads. The multi-time-scale prediction experiment demonstrates consistent accuracy at the medium and long time scales, where the mean absolute percentage error across all periods does not exceed 5%, and the coefficient of determination remains stable above 0.9. Through ablation experiments, this paper validates that the integrated multiscale attention convolutional network and LSTM framework yield a mean absolute error of 8.8 kW and a root mean square error of 15.5 kW during peak periods, outperforming single-scale architectures. Accordingly, the model provides an accurate, quantifiable foundation for power system load forecasting.