A CNN-LSTM model for predicting wind speed in non-stationary wind fields in mountainous areas based on wavelet transform and adaptive programming

Abstract

Improving the accuracy of wind speed prediction is crucial for engineering applications and disaster warning due to the highly unstable and unpredictable nature of wind as an energy source. A wind speed prediction model (WT-CNN-LSTM) was constructed based on wavelet decomposition, long short-term memory network (LSTM), and convolutional neural network (CNN) to address the non-stationary characteristics of wind speed in mountainous areas. The wind speed sequence is decomposed into subsequence columns and tested for stationarity using an adaptive program. The data are then decomposed and reconstructed. A prediction model is established using CNN and LSTM. The final wind speed prediction value is obtained by overlaying the predicted values. The results indicated that compared with the WT-CNN and WT-LSTM models, the WT-CNN-LSTM combination model proposed in this paper reduced the MAE, MSE, and RMSE indicators by 0.10%–0.11%, 0.57%–0.63%, and 0.11%–0.13%, respectively. In addition, the adaptive program eliminates the need to rely on traditional manual empirical values to determine parameters, ensuring that the prediction results are not affected by changes in the number of hidden layer nodes. This information can serve as a reference for future mountainous engineering construction.