Short-Term Wind Power Prediction Based on Wavelet Transform and Convolutional Neural Networks

Wind power prediction (WPP) is critical to the safe operation and economic dispatch of power systems, but it faces two challenges: 1) Information on different frequency bands is included in the numerical weather prediction (NWP) data, making it difficult to mine vital information if the original NWP data is directly applied as input data; 2) There is a strong nonlinear relationship between the input and output data in the WPP model, meaning that traditional models are difficult to fit accurately. This paper proposes a new short-term WPP model based on wavelet transform (WT) and convolutional neural networks (CNN). First, WT is applied to split the original NWP data and historical power data into multiple sets of different frequency components. Then, based on the NWP and historical power data sets of relevant frequencies, CNN models are established for prediction. Finally, based on the prediction results of different frequency components, inverse WT is applied to reconstruct the WPP results. This case study shows that 1) the WT method allows for effective mining of the information contained in the different frequency components of the NWP data; and 2) when the WT-CNN model is applied to predict the wind power in the next 12 hours, the NRMSE of the model is reduced by 1.55 percentage points compared with Back Propagation Neural Network.