Offshore wind power forecasting based on WPD and optimised deep learning methods

Accurate wind power forecasting is vital for (i) wind power management, (ii) penetration increment of the power generated into the power grid, and (iii) making maintenance more efficient. Motivated by the recent application of wavelet transforms and advancements in deep learning methods, a hybrid forecasting method is developed based on the wavelet packet decomposition [WPD], Long Short-Term Memory Network [LSTM], and Convolutional Neural Network [CNN] to improve the accuracy of wind power forecasting. WPD is employed to decompose pre-processed wind power data into sublayers with different frequencies. Sequential Model-Based optimisation (SMBO) with the Tree Parzen Estimator (TPE) is then used to tune the hyper-parameters of LSTM and CNN, efficiently. The optimised LSTM is employed to predict the low-frequency sub-layer that has both long-term and short-term dependencies, and CNN is used to forecast the high-frequency sub-layers with short-term dependencies. To evaluate the prediction performance of the developed method, seven forecasting models, including random forest (RF), feed-forward neural network (FFNN), CNN, LSTM, WPD-FFNN, WPD-CNN, and WPD-LSTM models, are considered as comparison models. Comparing the prediction results of all involved models proves that the developed model improves the prediction accuracy by at least 77.4% compared to methods that do not use WPD. In addition, the proposed combination of optimised CNN and LSTM improves the forecasting accuracy by 26.25% compared to methods that use only one deep learning model to forecast all sub-series.