Research Of Short-term Wind Power Prediction Algorithms

With the exhaustion of traditional energy, the deterioration of human living environment and increasingly improved environmental awareness, more and more countries focus on the utilization of renewable energy. Among kinds of renewable energy, wind energy has the advantages of large reserves, cleanliness, renewable and extensive distribution, which makes the reserves of wind power generation rapidly increase. Especially in our country, China has become the country of largest wind power installed capacity of the world. However, wind power generation has the features of randomness and fluctuation, which is a severe challenge to safe operation of electrical power system and electrical energy quality when large scale wind power is combined to the grid. As a result, research of wind power prediction is necessary to develop wind power in large scale, insure safe and economic electrical power system operation and arrange scheduling operation of the system reasonably.To predict wind power of wind farm effectively, this paper makes some in-depth researches on the technology of wind power prediction. The main works are as follows;Model WT-LSSVM based on wavelet analysis theory and least squares support vector machine is proposed to predict wind power of wind farm in future 4 hours. Technology of wavelet analysis is introduced. During modeling, the series of history wind power used to SVM training is decomposed firstly, then SVM predicted model of each component is built. During prediction, the input of model is decomposed firstly, then get each result via different predicted models of each component, at last, get final result via adding each component predicted result.Short-term wind power prediction model is established to predict wind power of the wind farm in future 24 hours. The results of BP neural network and radial basis function neural network used in short-term wind power prediction are compared and analyzed. Traditional neural network has defects of low convergence rate and easily falling into local minimum, at the same time, when training the network, different initial weight values and thresholds may cause different networks. As a result, it is very difficult to realize basic algorithm. To solve above problems, this paper utilizes particle swarm optimization to optimize the initial weight value and threshold of neural network. Experiments prove that optimized network has better predicted result.Online sequential extreme learning machine is introduced to solve model online updating problem of wind power prediction model of wind farm. OS-ELM overcomes the defects of traditional single-hidden layer feedforward neural network, realizes adding batch training data and solves online updating problem of wind power prediction model.Make statistics and correction of the predicted result of model. Firstly, correct wind data predicted by numerical weather prediction (NWP) through OS-ELM to reduce error of front wind data of wind power prediction model. Secondly, make second correction of predicted result of short-term wind power prediction model, and substitute those predicted values which are beyond confidence interval of wind power with values calculated by wind power curve, which improves the prediction accuracy.