Short-Term Power Load Forecasting Based on CNN-LSTM Network

Abstract:Traditional neural networks have low accuracy in load forecasting with strong temporal correlation. To effectively improve the accuracy of short-term power load forecasting, a load forecasting method based on the combination of Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network is proposed. Five-dimensional load feature data are collected, using CNN convolutional and pooling layers as feature extraction units to extract spatial coupling interaction features; the reconstructed data is input into the LSTM network to mine load time series features, employing Dropout technique to enhance the model’s generalization ability; the model is trained using the Adaptive Moment Estimation (Adam) optimizer; the test data is input into the trained neural network model to predict future 1 hour and 12 hours of electric load. Experimental results show that this load forecasting model converges faster and achieves higher prediction accuracy compared to improved BP neural networks and LSTM forecasting models, with 1 hour load forecasting accuracy reaching 98.66% and 12 hours load forecasting accuracy reaching 96.81%, thus improving the accuracy of short-term power load forecasting.

Keywords:Long Short-Term Memory Network; Short-Term Load Forecasting; Dropout Technique; Convolutional Neural Network; Adaptive Moment Estimation