A hybrid approach to multi-step, short-term wind speed forecasting using correlated features

Abstract

Wind power is becoming a main alternative energy source to meet the growing electricity needs. Forecasting wind speed is important to mitigate generation uncertainty and optimize asset utilization. This paper proposes a hybrid wind speed prediction model with multivariate input and multi-step output capability. The model synthesizes linear time series regression with nonlinear machine learning algorithm. The input neurons of the hybrid model are determined by the number of lag observations in autoregressive integrated moving average (ARIMA), and also by correlated meteorological features, such as wind direction, air pressure, humidity, dew point, and temperature. The output neurons are further derived based on the forecasting horizon. The hybrid model is trained, validated, and tested by using 1.73 million hourly meteorological records from three cities with diverse wind profiles. The performance of the model is compared with several existing methods based on root mean square error and mean absolute error. Though the hybrid model does not show obvious advantage in 1-h ahead prediction, it outperforms persistence model, ARIMA, and univariate neural network models in 3-to-24 h ahead prediction. The hybrid model is able to reduce the prediction error by 20% in comparison with univariate neural networks.