Short-term wind generation forecasting and confidence interval estimation based on neural networks trained by extended Kalman particle filter

Short-term wind generation forecasting predicts wind power 24-hours into the future in hourly steps. Effective forecasting is important for reliability, electricity markets and transmission grids. It is however difficult in view of the intermittent nature of the wind generation. We previously presented a method of neural networks trained by extended Kalman filter. In this paper, the method of the neural networks trained by extended Kalman particle filter is developed to capture the stochastic feature of the wind generation. The key idea is to take differencing transformation on the wind generation since differenced data has the student t-distribution based on the data analysis. The method of neural network trained by extended Kalman particle filter is then developed. It runs several neural networks trained by extended Kalman filter in parallel, and dynamically weights them based on the particle filter algorithm. Because the forecasting accuracies can be volatile, the method of generalized autoregressive conditional heteroscedastic is used to model the changes in residuals to improve predictions. Individual predictions from each neural network are then statistically combined to form the final forecasting. To accurately estimate the confidence interval, the dynamic covariance matrices are derived based on the differencing transformation. The overall dynamic covariance matrix is calculated by statistically combining all the covariance matrices. Numerical testing based on EIRGRID and ISO New England data demonstrates the significant values of our methods.