MMEMformer: A multi-scale memory-enhanced transformer framework for short-term load forecasting in integrated energy systems

As integrated energy systems (IES) increase in scale and complexity, achieving high-precision load forecasting has become essential for ensuring energy supply-demand balance, optimizing dispatch strategies, and enhancing energy utilization efficiency. Traditional deep learning methods often struggle with the multi-seasonal patterns, nonlinear couplings, and external disturbances inherent in IES load data, resulting in reduced accuracy and generalization. To address these challenges, this paper presents a multi-scale, memory-enhanced deep learning prediction framework. Drawing inspiration from the TimesNet approach, the proposed framework decomposes the input time series into multiple temporal scales, enabling the extraction and fusion of complex multi-seasonal dynamics. In parallel, a historical pattern retrieval memory module leverages a feature repository to incorporate long-term memory, thereby improving the model's ability to track persistent trends and irregular fluctuations. Additionally, cross-attention mechanisms are employed to integrate external features, increasing the model's sensitivity to environmental and behavioral factors. Finally, a multi-gating fusion strategy adaptively combines seasonal, trend, historical, and external information sources to produce more stable and flexible forecasts. Experimental results on a multi-energy load dataset demonstrate that the proposed model significantly outperforms other deep learning and state-of-the-art predictive approaches in terms of accuracy, stability, and generalization performance.