| With the rapid development of wind power in the world, the proportion of wind power in Power grid is increasingly larger. However, the accurate forecasting of wind speed is still a challenge for the present large-scale wind power generation. Therefore, accurate forecasting of the output of the wind power is necessary so as to achieve the optimal operation and dispatching of the Power system, as well as reducing the power system spinning reserve and operating costs.The Autoregressive Integrated Moving Average Model, ARIMA model, can obtain a stationary series by conducting difference on the original data. Moreover, on the basis of the ARIMA model, Generalized Autoregressive Conditional Heteroskedasticity Model, GARCH model, takes the conditional heteroskedasticity into consideration and adds a conditional heteroskedasticity equation. Based on grey model, Rolling Grey Model, RGM, continually adopts new data while eliminate old data to fulfill data update. All these three models are applied in wind speed forecasting wildly.Artificial Neural Networks, ANN, are adopted to replace the traditional time series methods for its good nonlinear fitting capacity and generalization ability. Back Propagation Neural Network, BPNN, and Wavelet Neural Network, WNN, have more frequent use. With the back propagation of its output error, BPNN can adjust its weights and thresholds between each layers to get a better performance. While WNN combines the network structure of BPNN and wavelet analysis, which draws great attention from researchers.In this paper, compared to the traditional time series model, ARIMA, GARCH and RGM included, and the single BPNN and WNN, the hybrid models based on the ensemble operators of neural networks is proposed on the basis of single BP network and single wavelet network and ensemble forecasting method, which avoids the flaws because of the uncertainty weights and local minimum of neural networks by adopting the ensemble operators. Under the conduction of three ensemble operators, Mean, Median and Mode, the performance of the ensemble models have been improved to some extent compared to single neural networks. And to increase the forecasting accuracy of ensemble model, the optimization algorithms:Particle Swarm Optimization Algorithm and Cuckoo Search Algorithm, are adopted to optimize the window width of the kernel function which is used to calculate the output of the Mode ensemble model. The experimental study of the wind speed in Inner Mongolia of China reveals that the results of the proposed hybrid model is much more better than the other five traditional single forecasting models. |
