Response load prediction of demand response users based on parallel CNN

Abstract

YAs China advances its transition towards green and low-carbon energy, the proportion of new energy generation in the power grid is gradually increasing, leading to a significant rise in the demand for power resource scheduling. However, due to the scarcity of historical load response data from users, it is challenging to effectively predict user-responsive loads. To address this issue, this study proposes a method of augmenting historical load response data in a weakly supervised manner. Taking into account the unique circumstances of high-voltage users, a sparse CNN for anomaly detection is introduced, along with a multi-branch parallel CNN model capable of weighted output of prediction results from both global and local perspectives. Subsequently, effective iterative training of the model is performed using the EM algorithm. Ultimately, accurate prediction of user-responsive loads is achieved. Based on historical 96-point load data and load response data from high-voltage users in a specific city in China, the predicted results are compared with actual load response data, validating the rationality and accuracy of this method in predicting user-responsive loads.