Wind Power Combination Forecasting Based On Wavelet Packet Transform

Wind power is the most widely developed and the fastest growing clean energy. With the rise of wind power capacity in power grid, the intermittent and randomness of wind power increasingly affects the quality and stability of power grid, which limit wind power installed capacity in power grid. Consequently, it is of significance to improve the wind power prediction accuracy in order to enhance grid scheduling and wind farm power generation planning.For the wind power output forecasting problem, this paper puts forward a novel combination forecasting model by using wavelet packet transform(WPT), in which, the original wind power sequence is decomposed into a series of subsequences with WPT, then each sub-sequence is forecasted by combination forecasting model, all the subsequence forecasting outputs are superposed to obtain the final forecasted results. For improving the accuracy of combination forecasting, three kinds of forecasting methods including time series method, neural network and SVM are adopted in order to fully exploit their respective advantages in data mining for wind power forecasting. The accuracy of combination forecasting method depends on the determination of weight of each forecasting method. A new optimization method called Crisscross Optimization Algorithm (CSO) is proposed and applied to the identification of the weights in the combination model.The crisscross optimization algorithm is a new search algorithm inspired by Confucian doctrine of gold mean. CSO search algorithm consists of horizontal crossover and vertical crossover, which mimics social learning and self-learning of human beings, respectively. By adopting the horizontal crossover—competition—vertical crossover—competition evolutionary mechanism, CSO elegantly addresses the ubiquitous premature convergence and local optimum problem existing in other swarm-based optimization algorithms. Because of its excellent global search ability, CSO provides a new solutions to solve large-scale complex optimization problems, the results on12benchmark functions show that the CSO has overwhelming advantages in terms of solution accuracy and convergence rate, compared to other heuristic search algorithm. In the experimental simulations, The CSO-ANN algorithm proposed in this paper is used to optimize the large-scale wind power forecasting neural network, which has318coefficients including weights and biases, the results show that the proposed method has better generalization ability and better forecasting accuracy compared to other learning algorithms. In the WPT based combination model, the reverse variance method and CSO algorithm are used to determine the weights of combination forecasting, respectively. The experimental results show that combination forecasting model can obtain better prediction results than single prediction model, and the CSO algorithm has obvious advantage over the compared inverse variance method.