Research On Wind Power Forecasting Based On Extreme Learning Machine

Due to the randomness and intermittent of wind power, the volatility of wind farm’s output is very strong, which will cause great impact on the power grid and bring difficulties and challenges to the planning and scheduling of the power grid. Therefore, accurat e and reliable wind power prediction is very important to optimize the cost and the reliability of the power grid operation. At present, the prediction method with modern statistical approaches like neural network and support vector machine has been widely applied to the short-term and ultra short term wind power prediction. Extreme learning machine(ELM), a new single hidden layer feedforward neural networks(SLFNs) can provide better generalization performance at a much faster learning speed, and it is successfully applied in ultra short-term wind power prediction and short-term wind power prediction. In kernel extreme learning machine(KELM), based on ELM, the unknown hidden layer feature map is replaced by kernel function, the output weights is solved by regularized least square algorithm, and the SLFNs hidden layer node is no need to determined subjectively. Therefore, the KELM method is the main line of this paper, and the combined and single predictions are researched for further improve the precise forecasting of short-term and ultra short-term wind power.The main research contents of this thesis includes the following aspects:(1) Based on the ELM theory, basic KELM, KELM with gaussian kernel and KELM with wavelet kernel are deeply researched. Meanwhile, the regularized least square algorithm is considered applied to weights solving of the KELM network for further generalization performance improvement of the KELM network.(2) From the perspective of combination forecast model,firstly,the original wind power series were analyzed. Secondly,on the basis of empirical mode decomposition(EMD), complete ensemble empirical mode decomposition with adaptive noise(CEEMDAN) was studied. Furthermore, integrated with fuzzy entropy(FE), a new combination forecasting model based on CEEMDAN-FE and KELM is presented. It is applied to the real wind power prediction instance in one wind farm, with comparison to KELM, EMD-KELM and EEMD-KELM methods. The result of these experiments show that the combination forecast method of CEEMDAN-FE-KELM obtained highest prediction accuracy.(3) On the basis of KELM, integrated with optimization algorithm, a optimized algorithm based on optimization kernel extreme learning machine(O-KLEM) is proposed. Namely, the three opimization methods including genetic algorithm(GA), differential evolution(DE) and simulated annerling(SA) are respectively used to tune the set of input variables, regularized coefficient as well as hyperparameter of kernel function. Thus, three different algorithm of O-KELM, namely GA-KELM, DE-KELM and SA-KELM were obtained. The O-KELM method is firstly applied to directly six-step prediction for Mackey-Glass chaotic time series, under the same conditon, compared with existing optimized ELM method, From the analysis of the results it can be verified that the prediction accuracy of the proposed O-KELM method about one oeder of magnitude increases over the optimized ELM method,furthermore DE-KELM algorithm can obtained the lowest root mean square error(RMSE).The O-KELM method is the applied to real-world wind power prediction instance in one wind farm. The 10-minute ahead single-step prediction as well as 20-minute ahead,30-minute ahead, 40-minute ahead multi-step prediction for wind power are respectively implemented to evaluate the O-KELM method. Experimental results of very short-term wind power time series prediction at different time horizon confirm that the proposed O-KELM method,simultaneously,GA-KELM method for wind power prediction outperforms other method at future 10-,20-,40-min ahead prediction at the root mean absolute error(RMSE), DE-KELM method outperforms other method at future 30-min ahead prediction the normalized mean square error(NMSE)at the normalized mean square error(NMSE).The results from these applications demonstrate the effectiveness and feasibiity of the proposed O-KELM method.