Short-term load forecasting based on different characteristics of sub-sequences and multi-model fusion

Rapid and accurate short-term load forecasting for distribution network is beneficial to ensure the safe and stable operation of power grid, reduce operating costs and improve the utilization rate of energy. Initially, through the data preprocessing minimizes the impact of outlier data on predictions. Subsequently, using variational mode decomposition and sample entropy methods separate modal components into high-frequency and low-frequency periodic sequences. Pearson correlation coefficient and principal component analysis are then employed to analyze feature parameter correlations, constructing distinct feature matrices for each Sub-sequence. High-frequency sequences are inputted into a prediction model combining time convolutional and bidirectional long short-term memory networks, while low-frequency periodic sequences are fed into a model combining auto regressive integral moving average and support vector regression. An illustrative analysis using January data from a Chinese province. Results indicate that compared with the 13-dimensional eigenmatrix, the proposed method saves 63 s in prediction time and improves the efficiency by 23.6 %. Mean absolute percentage error only decreased by 0.143 %, indicating that the method can ensure the prediction accuracy without losing robustness. Additionally, case analyses for different prediction durations (1 day and 1 week) exhibit promising results with mean absolute percentage error indices of 1.982 % and 2.022 %, indicating strong predictive performance.