The Research And Application Of Short-Term Forecasting Model For Time Series

Prediction based on time series is one of the universal problems in human being’s daily life. There have been a great deal of research at home and abroad, aiming at all kinds of time series prediction problems such as market trends, traffic flow, weather forecast, etc. On the other hand, due to the substantial growth of global population and the rapid development of economy, the non-renewable energy resources in the world gradually become exhausted. Global energy crisis has forced human being to explore and research all sorts of renewable energy sources like wind energy, solar energy, tidal energy, and so on. Wind energy has become one of the significant development directions of renewable energy source researches because of its prominent advantages like abundant resources, vast distribution and pollution-free, etc. It is a kind of energy equipped with intermittency and randomness features, so merging the wind energy with strong volatility into power grid will bring serious impact to the quality and security of electric power system operation. Therefore, accurately predicting the wind speed at wind power plant and the wind power is of vital practical significance to set proper operation plans and reduce operational risk and cost.At present, the applied researches on wind speed at wind power plant and short-term prediction model of wind power mainly consist of two aspects. One is constantly improving the accuracy of short-term prediction model. The accuracy of wind speed short-term prediction now can be up to about10%, but that of wind power can only be15%～20%. The other is making fully use of the advantages of each single prediction algorithm in combined prediction models in order to further promote its estimated performance.Aiming at the short-term prediction problems of wind speed and wind power time series, this paper takes the wind speed and measured wind power data of a wind power plant in Hubei Province as research object, and detailedly analyzes and compares the modeling approaches of thirty-minute-ahead and one-hour-ahead predicted wind speed and wind power from three aspects of data pre-processing, single prediction algorithm, and combined prediction algorithm. Specific to non-linear and non-stationary wind speed and wind power time series, this paper firstly further researches the pre-processing methods of time series data, highlights the principle and procedure of wavelet decomposition algorithm and empirical mode decomposition algorithm, contrasts their results from the perspective of practical application, and provides more detailed data for building wind speed and wind power short-term prediction models on basis of statistical theory.According to the practical application problems of time series single prediction algorithm of measured wind power plant data, this paper also separately researches the theoretical basis of support vector regression, wavelet neural network and grey prediction system, and detailedly discusses the modeling approaches of three single prediction algorithms when applied to time series short-term prediction. In the aspect of parameter selection of support vector regression, this paper brings in grid optimization algorithm and particle swarm optimization algorithm, and builds two kinds of short-term time series prediction models on basis of support vector regression; while in the application aspect of wavelet neural network prediction models, this paper deeply researches models’ learning algorithm, and takes genetic algorithm as learning algorithm to build wavelet neural network prediction models. As models’application results of measured data, the learning results of genetic algorithm are more optimized than traditional gradient descent. Basing on grey system theory, this paper also focuses on the improving problems of initial data and background value computing method of grey GM(1,1) prediction models. It separately applies traditional and developed GM(1,1) prediction models to measured data, and compares and analyzes the functions and application scope of GM(1,1) prediction algorithm.In order to improve the accuracy and stability of prediction models, this paper separately brings in two different model composition thoughts. In accordance with above three single prediction algorithms, it puts forward a weight combined prediction model of entropy theory, error sum of squares and minimum theory; in accordance with wavelet decomposition and empirical mode decomposition algorithm, it puts forward two kinds of combined prediction models aiming at decomposition vector characteristics. Through examples analysis, it compares and summarizes the prediction functions and features of four combined prediction models. The result shows that the prediction results of combined prediction models will not be confronted with retardation phenomenon of single prediction model, and the accuracy of combined prediction models which aims at decomposition vector’s detailed characteristics is higher than that of weight combined prediction models.