Wind Speed Forecasting at Different Time Scales Using Time Series and Machine Learning Models

Abstract

Wind energy is considered to be one of the fastest growing green energy resources. The time horizon of wind energy forecasting plays a crucial role in several end user applications. This study focuses on the short term (day ahead) and long term (multiple days to months ahead) forecasting of wind speed using time series and machine learning methods. For this, we first analyse time series plots of daily, weekly and monthly sampled wind speed data and perform stationarity test. Then, we implement time series SARIMA and window-sliding ARIMA models due to the presence of yearly seasonal patterns in the dataset. In addition, we implement two most popular machine learning models, namely MLP and LSTM, and compare their performance with the time series methods at different time scales. The experimental results based on 15 yr (2000–2014) of daily, weekly and monthly wind speed data at four different locations in India reveal that the window-sliding ARIMA has the best performance in terms of its lowest RMSE and MAPE values for daily data. For weekly forecasting, the performance of LSTM, MLP and the window-sliding ARIMA are very similar, whereas for monthly forecasting, the SARIMA model produces the least error values. In summary, the present study enables a generic guideline for the choice of wind speed forecasting models at daily, weekly and monthly time scales.