Wavelet-machine learning in wind prediction: A hybrid approach

Abstract

Wind power relies heavily upon the availability of wind at a specific speed; ultimately, wind power forecasting relies on accurate wind speed forecasts. Although theoretically, wind power increases eightfold as its speed doubles, wind speed is often characterised by linear and nonlinear patterns, nonstationary behaviour as well as intermittency on both location and time scales such that it requires immediate and continuous adjustment to maintain power grid stability. Consequently, predicting wind power from the multidimensional wind speed resource can be a cumbersome and skillful task that cannot be performed effectively with a single forecasting model. However, literature has shown that hybrid approaches that simultaneously combine the strengths of data pre-processing, data optimisation, and data post-processing methods to efficiently, accurately, and reliably quantify wind data are very scant. Using high-resolution wind speed data from the Southern African Universities Radiometric Network (SAURAN) and Wind Atlas for South Africa (WASA), this work proposes improved wavelet machine learning hybrids to fill this gap. In the first module, an efficient and robust combination model leveraging wavelet transform (WT), autoregressive integrated moving average (ARIMA), extreme gradient boosting decision tree (XGBoost), and support vector machine (SVR) denoted by WT-ARIMA-XGBoost-SVR is developed and validated at three different SAURAN locations and at a short-to-long-term forecasting horizon. The proposed hybrid strategy efficiently reduces wind speed forecasting error accumulation caused by the use of linear models by reconciling nonlinear wavelet subseries forecasts. The second module proposed a highly advanced, robust hybrid approach blending WT, neural network autoregression (NNAR), stateless long short-term memory (LSTM) learning, gradient boosting machine (GBM), and sample entropy (SampEn) denoted by WT-NNAR-LSTM-GBM for short-term wind speed forecasting at four different SAURAN stations. With this approach, gradients were effectively mitigated from vanishing and exploding, while improving wind speed forecasting accuracy. Furthermore, this approach emphasised the classification and modelling of wavelet subsignals based on similar complex and deterministic features. The third module developed a hybrid approach by combining maximal-overlap discrete WT (MODWT) filters with gated recurrent units (GRUs) and differential evolution (DE) algorithm denoted by wavelet-MODWT-GRU to forecast wind speed at three different WASA locations in the medium to long-term forecast vi horizon. The module provides a more efficient and reproducible method for selecting the most appropriate filters and wavelet decomposition levels to improve wind speed forecasts. As a secondary aim, the work also provided a comprehensive overview of the status of South Africa's electricity supply, focusing on load shedding and unplanned power outages. The review emphasised the importance of scaling up investment in wind power in the country. Moreover, the module also evaluated the generalisability and adaptability of the typical wavelet-machine learning (ML) hybrids in short-term power outage forecasting in order to enhance power grid reliability. Overall, the proposed wavelet-ML strategies demonstrated higher generalisability, reliability, and robustness, and were highly accurate than their counterparts based on various point and probabilistic error indicators, as well as various statistical tests. As a result, these approaches can help operate power grids in real-time, optimise wind power output, minimise energy losses, and distribute wind power effectively. As a result of the four modules, key contributions were summarised and potential directions for future research were highlighted to improve forecasting.