Power load forecasting based on the improved TimeMixer

Electricity plays a crucial role in economic and social development. Under the dual-carbon target, the integration of distributed energy makes power load forecasting more challenging and essential. Traditional prediction models, like those based on statistics, machine learning, and deep learning, have limitations despite their contributions. This research presents an improved framework based on the traditional TimeMixer model. The key is introducing the Frequency Enhanced Channel Attention Mechanism (FECAM) and the Fourier Analysis Network (FAN). FECAM applies discrete cosine transform to power load data and generates channel attention weights according to spectral features, accurately highlighting key seasonal and trending details and enhancing feature representation. FAN, replacing the original predictor, utilizes Fourier transform to explore spectral information, optimize feature representation, capture long and short-term interaction precisely, and integrate the complementary prediction capabilities of multi-scale data, avoiding information loss. The two collaborate. Through data preprocessing, core feature extraction and processing, and prediction output modules, they overcome traditional model weaknesses, enhance power load forecasting accuracy and efficiency, support the stable and efficient operation of the power system, and drive the intelligent development of the power industry.