Application and Performance Comparison of a CNN–LSTM Hybrid Neural Network for Power Load Forecasting

Abstract

Accurate and stable power load forecasting is essential for power system operation and planning. However, traditional single prediction models often exhibit limited performance when modeling complex load data characterized by strong nonlinearity and non-stationarity. To address this issue, this paper proposes a hybrid neural network model that combines a Convolutional Neural Network (CNN) with a Long Short-Term Memory (LSTM) network for short-term power load forecasting. First, considering the temporal dependency and nonlinear characteristics of power load time-series data, both LSTM and CNN–LSTM forecasting models are constructed. Then, the hybrid model is systematically optimized by adjusting the number of convolutional layers, the structure of LSTM hidden layers, and different activation function configurations. Finally, the forecasting performance of the proposed model is evaluated and compared with that of the single LSTM model using root mean square error (RMSE), mean absolute error (MAE), and mean absolute percentage error (MAPE). The experimental results indicate that the CNN–LSTM model achieves higher forecasting accuracy and better generalization performance, and its predicted load curves show greater consistency with actual load variations. The proposed approach provides an effective solution for power load forecasting and offers valuable support for power system dispatching and planning.