Research on short-term power load forecasting method of microgrid based on machine learning

• Main Authors: Cheng, Tao, Xiao, Jiarong
• Format: Conference Proceeding
• Language: English
• Subjects: Accuracy; Adaptation models; Analytical models; Data models; Load forecasting; Load modeling; Long short term memory; microgrid; Microgrids; Optimal scheduling; Predictive models; sequential convolutional networks; short-term load forecasting
• ISSN: 2770-663X
• DOI: 10.1109/ICISCAE62304.2024.10761164

Reasonable optimal scheduling can effectively guarantee the economy, environmental protection and stability of microgrid operation, and reliable load prediction data is the most powerful basis for formulating optimal scheduling. Therefore, the accuracy of load prediction of microgrid directly determines the feasibility of scheduling scheme. To solve the problem that the traditional load forecasting model is not effective in time series modeling, and the TCN-Attention short-term load forecasting model is proposed. This paper takes the time convolutional network as the basis model, so the origin and structure of LSTM network and TCN network are analyzed in detail, and the TCN prediction model is built and compared with the LSTM prediction model. The simulation results show that the TCN model has higher prediction accuracy. In order to further improve the accuracy of the TCN model, considering the advantage of Atention mechanism to assign weights adaptively according to the importance of features, a short-term load prediction model of TCN-ATTENTION was constructed and experimental simulation was carried out. The results show that compared with the single TCN model, the accuracy of the TCN-attention model is increased by 0.49%, reaching 96.42%. Secondly, aiming at the problem of information hybridisation of load data and not obvious change rule, a data decomposition method based on WOA-VMD is proposed. The minimum envelope entropy of the variational mode decomposition algorithm is taken as the fitness function of the whale optimization algorithm, and the optimal parameter combination of the mode number k of the VD algorithm and the quadratic penalty factor a is obtained through iterative optimization. Then, the optimal parameter combination is substituted into the VMD algorithm, and the original one-dimensional load data is decomposed into multiple eigenmode functions with different frequencies but with regular changes. Each IMF component is taken as the input load feature of the TCN-Attention model. After that, the above two models were combined. At the same time, in order to further improve the prediction accuracy of the model, Pearson correlation coefficient was used to screen out the daily maximum temperature, minimum temperature and average temperature as the meteorological feature input, and the date type factor was reconstructed as the date type feature input. Finally, based on all the above methods, a short-term load forecasting method based on WOA-VMD