Short-term Power Load Forecasting Based On Similar Day And Support Vector Machine

Short-term power load forecasting has a strong influence on the operation, controlling and planning of power systems. The accurate forecasting provides maintaining security and stability to the operation of power systems. This paper started with analyzing the characteristics of short-term load and took daily peak load and its effect factors as the research objects. Two short-term load forecasting models were put forward to solve the forecasting of special holidays and normal days.First, this paper discussed a method to select similar days (similar-day method, SD in short) in order to improve the accuracy of forecasting. With meteorological factor, distance to the predicting day and weekday type considered, a quantitative model was given to calculate the similarity degree. In particular, The similarity in meteorology was obtained with Grey Correlation Analysis Method while the similarity in time took into consideration of"near greater far smaller"and"cyclical"principal. And weekday type similarity function was proposed according to the week cycle pattern effect on power load. Moreover, the product of the similarity degree on meteorology, time as well as weekday type was to evaluate the overall similarity.And then, the approach based on similar days and support vector machine (SD-SVM in short) was presented to predict peak load of normal days. Several similar days were selected from history data according to their similarity to forecasting day, which were trained by SVM. With the advantages of both SD and SVM, SD-SVM solved the small sample problem effectively. It not only overcame noise interference on SVM but also avoided subjectivity of SD. Applying it into Huizhou short-term power load forecast, this method offers a more accurate result, with comparison to SD, SVM and the method based on similar-day and neural network. Its average relative error of SD-SVM is 1.47%Finally, considering that the the load characteristic of special holidays is different from that of normal days and the holiday of the same type is far away from the predicted holiday, an algorithm was proposed based on the similar days'load modification.In order to overcome the difficulty due to few samples, recent weekends were added to the samples from which similar days were selected.Moreover, the product of the similarity degree on meteorology and time was to access the overall similarity. In addition, considering that the day type and annual cycle have an impact on the load, the load of similar weekends was modified on the basis of the proportion between the previous holiday and corresponding weekend while that of similar holidays was amended on grounds of the rate at which power load grows. When this algorithm is put into practical operation, in Huizhou power load forecast, it offers a desirable result, with an average relative error of 2.29%.