(554f) An Unscented Kalman Filter Driven State-Space Support Vector Regression Method for Short-Term Wind Speed Forecasting Toward Optimized Wind Power Generation

The promising future of wind power as a renewable energy source to mitigate carbon emission and improve environmental sustainability has led to considerable research on forecasting wind speed and optimizing wind power generation. However, one of the most difficult challenges in increasing the penetration of wind power in electricity grid is caused by the stochastic nature and intermittency of wind speed, which may lead to the unstable wind power generation. Thus accurate short-term wind speed prediction techniques are highly desirable, since they are critical towards reliable wind turbine operation and control as well as wind farm optimization.

In literature study, wind speed prediction techniques can be divided into two main categories: physical model based approaches and statistical modeling methods. Nevertheless, the prediction ability of physical model based approaches degrades significantly when the random uncertainty of weather conditions is strong. Time-series models and artificial neural network (ANN) based models are two main subtypes of statistical modeling methods, which dynamically predict future wind speed by utilizing historical wind speed data to train the models. On the one hand, the essentially linear time-series models may not be well-suited to characterize the stochastic nature and uncertain dynamics of wind speed. On the other hand, the generalization ability of ANN models cannot be guaranteed so that a well trained ANN model may lead to poor prediction performance for new observations.

In this study, a novel hybrid predictive model is proposed by integrating unscented Kalman filter (UKF) with support vector regression (SVR) based nonlinear state-space framework in order to enhance the capacity of dealing with random uncertainty and minimize the prediction errors of multi-step-ahead wind speed forecasting. Since SVR employs the structural risk minimization principle to effectively overcome the drawback of over-fitting with strong generalization capability, it is employed to formulate a kernel function based nonlinear state-space model structure in the first stage. To tackle the nonlinearity in the state-space model built upon SVR, unscented transformation technique is utilized to approximate the nonlinear function in the state-space formulation through a subset of points termed as sigma points, which are propagated through the random system to update the statistics of the states and observations.  Therefore, UKF instead of regular Kalman filter is employed to conduct recursive state estimation on the SVR based nonlinear state-space model with strong stochastic uncertainty. As such, the presented SVR-UKF approach can mitigate prediction errors by accounting for stochastic and dynamic nature in wind speed.

The novel SVR-UKF method is compared with artificial neural networks (ANN), SVR, autoregressive (AR) and autoregressive integrated with Kalman filter (AR-Kalman) approaches for predicting short-term wind speed sequences collected in three sites in Massachusetts, USA. The forecasting results indicate that the proposed method has much better performance in both one-step-ahead and multi-step-ahead wind speed predictions than the other approaches across all the locations. The better performance of the SVR-UKF method in both prediction horizons for different sites proves its considerably improved reliability and robustness.