Improved Stacked Ensemble based Model For Very Short-Term Wind Power Forecasting

The high intermittency and uncertainty associated with wind energy have brought serious challenges to the power system operators, such as increasing the operating costs by increasing the requirements on primary reserves. Hence, the need to apply more robust and accurate forecasting model has become a persistent need to meet these challenges. This paper proposes a generic two-layer stacked ensemble based model for a very-shortterm wind power forecasting. The proposed approach combines a heterogeneous machine learning algorithms composed of three well-established models SVR, RBFNN, and RF heuristically with a parameter tuned SVR on the upper layer. First, the base learners are trained and tested on the original data, and then their outputs are used to train the upper layer learner to predict the final output. In this paper, a criterion for selecting the most efficient base predictors is used based on a reasonable decision for the number and type of the combined learners that takes into consideration: (i) the reliability of predictor(s) for forecasting the wind power data set; and (ii) the number of predictors that produce high performance. Many experiments had been conducted and the results prove that the proposed model is consistently better than the state-of-the-art models. The proposed model can assist network operators to make the right decision to regulate and schedule their resources optimally.