Wind Speed Prediction using Classical Time Series and Machine Learning Models: A Comparative Analysis



Journal Title

Journal ISSN

Volume Title



The need of delivering future accurate predictions of renewable energy generation has been recognized by stakeholders working in the field of renewable energy. It is the reason for recent improvements in the methods to provide more precise energy generation prediction. Wind power production is dependent on weather pattern variations, particularly wind speed, which are irregular in locations with unpredictable weather. Wind speed prediction in a given location is crucial for the evaluation of the wind power project; the accurate prediction improves the planning, reduces the cost, and improves the use of resources for wind power generation. Models such as Weibull probability density function based wind prediction (WBM), autoregressive integrated moving average (ARIMA), Kalman filter and support vector machines (SVR), artificial neural network (ANN), and hybrid methods classical time series and deep learning models have been used for accurate prediction of wind speed with different forecast horizons. For short and ultra-short terms that are two to three hours in the future, the ARIMA ensemble with ANN has demonstrated improved performance. For medium-term wind speed predictions, however, SVR, Kalman filters, and ensembles of both have demonstrated good performance. Recurrent neural networks (RNN) in particular have recently reported enormous success in time series forecasts, especially for medium- and long-term predictions. There has been growing interest in the field of deep learning and neural networks for the prediction of wind speed as it can overcome the issue of accurately forecasting the nonlinear patterns of wind speed data using classical time series methods. The main contribution of this dissertation research is the comparative analysis of the performance of each method for accurately predicting wind speed for different time horizons and proposing a Weibull distribution based featured engineered hybrid model for wind speed prediction. In this research, the wind speed generated from the Weibull probability density function is used as a feature in the wind speed prediction model and the prediction model is developed by optimizing the weight function for each model contributed to the hybrid prediction model. The demonstration of the accuracy of the 7 proposed model and comparative analysis of the different model is performed on the five different data set obtained from the National Oceanic and Atmospheric Administration and System Advisory Module (SAM) database.