Power load forecasting based on canonical-correlation analysis and LSTM networks

Under the premise of limited data, to improve the accuracy of short-term power load forecasting, a data-enhanced load forecasting method using a long-short-term memory network combined with canonical-correlation analysis is proposed. Firstly, canonical-correlation analysis is performed on the data to determine the correlation of each influencing factor in the data set to load prediction. The data is then augmented based on the correlations to reconstruct the dataset. Finally, the prediction model of the long short-term memory network is created, and the reconstructed data is added for power load prediction. The model load prediction methods such as RNN, LSTM, and GRU are compared. The results show that LSTM has a better stability. After adjusting and training with the canonical-correlation analysis method, the prediction accuracy of the model is further improved.