Multi-step Short-term Load Forecasting Based on Attention Mechanism, TCN-BiLSTM Network and Decomposition-based Error Correction

Aiming at the problems of regional-level power load forecasting accuracy and efficiency, this paper proposes a multi-step short-term load forecasting model based on attention mech-anism, TCN-BiLSTM network and decomposition-based error correction. Firstly, improve the combined prediction model TCN-BiLSTM with the attention mechanism to improve the model's ability to focus on important information in the input, and a new model structure, Attention-TCN-BiLSTM, is established. Secondly, obtain the preliminary load prediction results based on the Attention-TCN-BiLSTM model. Thirdly, calculate the difference between predicted and actual load sequence to obtain the error sequence. Then, Variational Mode Decomposition (VMD) is applied to decompose the error sequence into multiple subsequences, where the initial number of subsequences is selected through the central frequency method. Fourthly, predict the future values of each subsequence, and superimpose them to obtain the summed error prediction sequence. Finally, sum up the preliminary load prediction results and error prediction results to obtain the final power load prediction results. Example analysis show that attention mechanism combined with error correction significantly improved the prediction ability of TCN-BiLSTM Network, and the proposed model outperformed all related models, with the highest accuracy, good robustness and high applicability.