Short-Term Power Load Forecasting Based on MSCNN-BiGRU-Attention

To address the problem of difficult extraction of key features in power load, a multi-scale convolutional neural network-bi-directional gated recurrent unit-Attention (MSCNN-BiGRU-Attention) hybrid model is proposed for short-term power load forecasting. Firstly, the Spearman correlation coefficient was used to analyze the correlation of power load data set, and the features with high correlation were screened out to build the power load data set. Secondly, the data was input into the multi-scale convolutional neural network (MSCNN) to extract the multi-scale time sequence of power load data. Then, the extracted temporal features were input into the bidirectional gated recurrent unit (BiGRU) neural network for temporal prediction, and the temporal features were filtered and screened by attention mechanism. Finally, the outputs are integrated through a fully connected layer to generate the predicted values. With the 3 years of multidimensional power load data from a region in Australia as a data set, five control models were established. Meanwhile, we selected two years of multidimensional power load data from a region in southern China as the validation dataset for the models. The results show that MSCNN-BiGRU-Attention hybrid model can achieve better prediction effects than other models, thus effectively solving the problem of difficult extraction of the key features of regional power load.