Research On Parameters Optimization Of SVM Model For Short-term Load Forecasting

Support vector machine (SVM) is a new machine learning algorithm based on statistical learning theory. It has been successfully applied to the short-term load forecasting area by some remarkable characteristics such as good generalization performance, the absence of local minima and fast computing speed. The regression accuracy and generalization ability of forecasting model depend on a proper setting of its parameters, but it has no general theory and method yet that constrains the application a certain extent.In this paper, the load forecasting model of SVM is constructed based on the analysis of characteristic of electrical load, features selection of input samples and preprocessing of historical datas. Traditional cross validation method is applied to determine the parameters of the model. Compared with BP neural network, the results show that SVM is much more excelent in forecasting accuracy and speed. With respect to the forecasting course and results, the paper analyses the performance of SVM with different parameters which also have different influences and the disadvantages like blindness of man-made choice, low efficiency in the parameters selection.To achieve greater forecasting accuracy, aiming at the deficiency in the parameters selection of SVM, the selection of parameters is considered as a compound optimization problem in view of the integrative influence by each parameter and the objective function is set. Two intelligence optimization techniques have been employed into the optimization of parameters for a short-term load forecasting model based on SVM that makes automatical selection of parameters come true.Modified particle swarm optimization (PSO), the approach guidances initial population extracting and estimates premature convergence according to the information on population diversity. It makes particles distribute in the search space evenly and updates positions of stagnating particles to avoid areas of local optima when the particles are easily being trapped in local optima during late evolutionary process. Thus, it maintains the particle activation, overcomes the problem of premature convergence and enhances global search capability.Improved mutative scale chaos optimization algorithm, the approach makes the best use of ergodicity property of chaotic motion and considers its local searching ability. A couple of cycles are set during the searching course, chaos searches in the inner cycle and the interval is dynamically reduced on basis of searching state in the outer cycle. Thus, it avoids blind and repeated searching in searching space and improves searching efficiency.The simulation results prove that compared with conventional method, the proposed methods can reduce modeling error and forecasting error of SVM model effectively and have superior learning and generalization performance. In addition, they are convenient and fast for calculating and programming. Therefore, the approaches have higher practicability and adjust to the need of precision and rapidness for short-term load forecasting.In conclusion, a Chaos-Particle Swarm Optimization algorithm is proposed which organically combines each advantage of the two algorithms.