Research On Wind Power Prediction And Optimal Day-ahead Generation Schedule

The wind power is clean, renewable and easy to achieve so that many countriesin the world strive to develop this technology and improve wind power capacity intheir own power grid. While, due to the improvement of wind power penetration inthe power system and the uncertainty and volatility of the wind, creating an optimalday-ahead generation schedule is a great challenge. However, accurate wind powerprediction can help us wrestle with it.Based on former research, this paper builds the BP network wind powerprediction model and proposes two improvements. Firstly, in order to filter therandom fluctuation that is included in the measured wind speed, the paper uses thewavelet packet to decompose the wind speed into different frequency section andremove the high frequency section. Because the random fluctuation always hashigher frequency than others. Secondly, to the problem that to determine the modelinputs is difficult, the paper puts forward the principal component analysis (PCA)which uses a linear combination of the meteorological parameters to replaces oforiginal data and select the main component on the basis of relevance. The methodcan reduce the input parameters as far as possible, while assures the integrality ofinformation.Then, the BP network is improved by accepting a novel training algorithm anda network ensemble. In terms of training network, a new particle swarmoptimization is raised by adding the one dimensional minimum disturbance andfuzzy parameters adaptive. The experimental result shows the new PSO can speedup training. For avoiding net over fitting, a network ensemble is proposed whichchanges training samples by modifying weight according to the former net trainingresult.In the last part of the paper, the one-day ahead generation schedule becomesmore robust and economic by describing the uncertainty of wind power forecast andusing it in the unit commitment. Three categories of wind forecast information aredescribed. And the simulation shows the more information is counted, the better thegeneration schedule is.This research work is supported by the National High Technology Research andDevelopment of China (863Program)(NO.2011AA05A105).