Two-Stage Short-Term Wind Speed Prediction Based on LSTM-LSSVM-CFA

Accurately predicting wind speed is vital to the large scale connection of wind power grids. This paper proposed a two-stage prediction way called LSTM-LSSVM-CFA (LSTMs-long short term memory networks, LSSVM-least squares support vector machine and CFA-chaos firefly algorithm) which may be an effective method to improve the accuracy of short-term wind speed prediction. Integrated learning of wind speed sequence based on LSTMs and LSSVM-CFA is the core technology in this new method. The prediction section include two main procedures. In the first stage, a couple of LSTMs with different hidden layers and different neurons processing guarantee the implied information of wind speed time series can be fully collected and avoid the weak generalization capability of single deep neural network learning. In the second stage, the prediction result of LSTMs-layer is transmitted to the LSSVM-based regression layer and its parameters are optimized by the CFA which has global search capabilities. Finally, the final integrated learning result of prediction is given by the LSSVM-layer. Based on actual historical wind speed data, we made use of single-LSTM, LSTM-LSSVM-PSO and LSTM-LSSVM-CFA to perform short-term wind speed prediction one hour ahead. Compared with single learning and other popular integrated learning models, experimental results show that the proposed LSTM-LSSVM-CFA can perform better.