Short Electric Load Prediction Of The Chaotic Character Using Support Vector Machines

Electric load forecasting is the foundation of power system operation and control and also it's the foundation of power market.Accuracy,real time,reliability,intelligence are the new request of electric load forecasting In power market.Its precision will influence the economic and secure operation of power systems and quality of power supply. Load forecast is one of the important works for power system plan and operation.Because load forecast results affect directly generation, transmission and distribution of electricity energy,load and electricity demand In a planning period decide the development scale and the development speed power of system.So load and electricity energy forecast of power system is the most important task for the development planning of power system. The key work for electricity demand forecast is the choice of correct forecast methods.Although there are many methods to choose,to decide which method is more reliable is real a difficult task.The reliability of forecast result is not only decided by the forecast methods, but also decided by the reliability of the basic data. The second one is not soluble by theory studies,but some instructive knowledge for the first problem can be get through the analysis and compare for different forecast methods. Power load system is a multi-dimensional nonlinear system.It is easy to get the chaotic time series of practical load data in the real situation,So in this paper we study the forecast of the chaotic time series at first and present a new forecast method based on support vector machines regress theory and apply it to the chaotic time series.Support vector machine is a new generation machine learn algorithm and based on statistical study theory,the essence of the training support vector machines is equal to solving the quadratic formulation problem.The advantages of this method include:high forecasting accuracy,global optima property and small time complexity in order to proof the performance in these sides,we have had done two works in this paper;the first is that we apply to chaotic time series by using the reconstruct theory and analyst result by comparing with artificial neural network in other literature. The second is that we add different level noise to chaotic time series and analyst result by comparing with back-perception and radial basis function network.The simulation shows the forecasting behavior of predicting chaotic time series by using support vector machines is better than others and the method has good generalization and solid character. Through the foundation work in chaotic time series,we are in combination with the character of short-time power load and apply the prediction conception to short-time power system in the first time;we also build the model of short-time power load short time prediction based on support vector machines and do the simulation experiment by using the practical short-time power load data. Utilizing the support vector machine model of chaos character, we carry on short-term electric load to predict, at first, we should first judge the system whether have chaotic character, then we confirm the embedding dimension m and delay time t of the restructuring phase space; use m and delay time t restructure the phase space , figure out the largest Lyapunov exponent , form study samples and predicted value, and then utilize the samples training the SVM , at last trained network predict to some time in future , according to drawn predicted value, we compare with real load value, judge whether it has the advantage . There proposed one predict method basing on the support vector machine regression theory, and apply it to predict real short electric load. For prove performance of this algorithm, I make two job, one is what using restructure theory of phase space predict the single step to short electric load , and make comparative analysis with the prediction result of the neural network in documents, which can be found from the form, as to the different fetching value m, an evaluation index while adopting SVM (RMSE, RMSPEE ) small corresponding value while adopting BP network and fuzzy neural network, it indicate that it is better than BP network and fuzzy neural network on model and prediction. to support the result in view of modeling of short-term electric load and predicting of the vector quantity machine, by seeing the running time , compared with BP network , support vector machine have the faster speed of convergence , its running time can be shortened greatly compared with the time which is adopted by BP network; another work is what execute many step predict on short electric load , comparing with the result of BP network and RBF network, artificial result indicate, support vector regression algorithm can make the better result than other methods, and there is the better on the stability and melting ability.　In real electric load artificial test, SVM method can make best use of the training samples'distributed character, according to partly training samples to structure the differentiating function , it do not need too much priori information and using skill , finally turn into the problem of twice seek excellent , in theory , it can get optimum through solving the overall situation , it can efficient prevent the neural network fall into the some extreme value issue , at the same time through non-linear exchange and kernel function it varies to be ingenious to solve high dimension problem , make its algorithm complexity have nothing to do with the dimension of the sample , and quicken the training speed of studying; In addition, it can seek to trade off best between complicated type of the model and learning ability according to limited sample information , guarantee it having the better meltingperformance.