Short-Term Wind Speed Forecasting With Principle-Subordinate Predictor Based on Conv-LSTM and Improved BPNN

Abstract

To improve the accuracy of short term wind speed forecasting, a novel principle-subordinate prediction model based on Conv-LSTM network and BPNN is designed in this paper. The proposed model combines deep learning algorithms and improved neural network to deal with the problem of wind speed forecasting. In this model, the prediction sequence of each subseries is obtained by Conv-LSTM to form more smoother and characteristic series; the BPNN, optimized by MWOA (Modified Whale Optimization Algorithm), is trained with the reconstruction sequence, which processed for the prediction sequence by invert-EMD (inverse empirical model decomposition). Before prediction, singular spectrum analysis (SSA) and complete ensemble empirical model decomposition adaptive noise (CEEMDAN) are adopted to de-noise and decompose the original wind speed data into several subseries. This process is beneficial to improve the ratio of signal to noise and simply the features of wind speed data. In addition, the Conv-LSTM is tested on three datasets, the results proved that data process is advantaged to obtain higher quality wind speed datasets, and convolutional layer can deep extract the characteristics of each subseries, it can facilitate LSTM to make accurate prediction. The final prediction results, compared with other five different models, demonstrate that the proposed model can achieve higher precision. Such as the performance evaluation matrix (MAPE = 2.62%, RMSE = 0.151) are smallest obtained from experiments on three wind different wind speed datasets.