Correlation aware multi-step ahead wind speed forecasting with heteroscedastic multi-kernel learning

For the management of wind energy, wind speed forecasting is often required. Accurate multi-step ahead wind speed forecasts make the power system be adjusted timely and properly to ensure the stable and efficient operation of power system. Currently, various techniques have been developed for multi-step ahead wind speed forecasting. However, the correlation among different forecasting steps is often neglected in current multi-step ahead wind speed forecasting approaches, and the characteristic of heteroscedasticity in wind speed forecasting errors is usually not taken into consideration. In this work, a novel multi-step ahead forecasting method based on multi-kernel learning is developed. This method considers the task correlation, which represented by the covariance of multi-step ahead forecasting tasks, as well as the heteroscedasticity of forecasting errors. The optimization is solved within the framework of variational Bayesian. Thus, a correlation aware multi-step ahead wind speed forecasting technique with heteroscedastic multi-kernel learning is designed. In this paper, the experimental results in different wind farms and different seasons prove that the regression model considering the characteristics of multi-step ahead wind speed forecasting, task correlation and heteroscedasticity, will produce more accurate forecasts than the other models as for two to six-ahead wind speed forecasting. However, it is difficult to tell which characteristic is more important from the forecasting results. So, the regression model considering both of them will be more reasonable. Moreover, the training time of the proposed model is more than 10 min but less than 20 min. Thus, two to six-step ahead wind speed forecasts can be used in some practical applications, such as the load dispatch planning and the load increment/decrement decisions.