Matrix-based wavelet transformation embedded in recurrent neural networks for wind speed prediction

Accurate wind speed forecasts are essential for wind energy applications. It is challenging to accurately anticipate wind speed because of its high volatility and non-stationarity. Although wavelet transformation-based hybrid models, the most representative hybrid models, are extensively used, the endpoint effect and fixed decomposition parameters limit their prediction accuracy. To enhance the performance of such models, a special end-to-end model architecture combined with wavelet decomposition in matrix form and recurrent neural networks (RNNs) is proposed. In addition, a two-stage training optimization approach is presented to prevent the characteristics of the wavelet decomposition coefficients from vanishing during model training. In this framework, three conventional RNNs, consisting of the standard RNN, the long short-term memory network (LSTM), and the gated recurrent units network (GRU), are utilized to construct three innovative models, referred to as MW-RNN, MW-RNN, and MW-GRU, respectively. Then, using two years of data from a site, the recommended models are contrasted with the standard ones. The results show that the recommended models beat traditional ones in a variety of error indices with just a slight increase in training costs. Additionally, the impacts of wavelet decomposition parameters on forecast accuracy are contrasted, and suggested settings for actual wind speed predictions are offered.