DOI: 10.5176/ 978-981-07-7531-5_PEEE.14
Authors: Kanna Bhaskar and Sri Niwas Singh
An accurate wind power forecasting plays an important role in a competitive electricity market. It enables wind power producers to bid in a day ahead electricity market and to minimize their imbalance costs. It also helps system operators to include wind power in economic schedule and unit commitment problems. This paper deals with the problem of long-term wind power forecasting using a combination of two statistical models. In first model, the Numerical Weather Prediction (NWP) models’ hourly wind speed forecasts are mapped to wind power outputs using a Recurrent Neural Network (RNN) and a Node Decoupled Extended Kalman Filter (NDEKF) technique is used for training the RNN. However, due systematic errors in NWP wind speed forecasts, the forecast errors for first look-ahead hours could not be comparable to the persistence model results. To mitigate this problem, another statistical model, which regress up on the available historical values is suggested. In this second model, a wavelet decomposed of wind power series is carried and an Adaptive Wavelet Neural Network (AWNN) is used to forecast each decomposed signal up to 24 look ahead hours. The final wind power forecasts are the combination of these two model outputs. The results show that the significant improvement over persistence method can be achieved.
Keywords: Wind power forecast, extended Kalman filter, recurrent neural network, multi-resolution analysis, adaptive wavelet neural network