Short-term Load Forecasting In Power Systems Based On Time Series

Short-term load forecasting is used to predict the loads in the coming hours, in the nextday and even the next several days, and has important impacts on maintaining the security andeconomics of the power system concerned.The time series forecasting model represents a classical prediction method, and has beenwidely used for short-term load forecasting in actual power systems around the world.However, the time series currently used widely in this field are mainly the single variablemodels. The model is only employed to handle past load data so as to find the rule of loadchanges and on this basis, to forecast the load. Obviously, one major disadvantage of the timeseries model lies in the fact that the weather factor cannot be taken into account, while itusually plays an important role in short-term load forecasting. Given this background, a lotwork has been done in this thesis. The main research outcomes are as follows:(1) A time series model for load forecasting with temperature taken into account isproposed. The model combines the regression model and Auto Regressive Integrated MovingAverage (ARIMA) Model. First, the regression variables are chosen by using predictedresidual sums of the square criterion, and then the regression model between load andregression variables was developed. Finally, the ARIMA model of the regression model’sresidual are built for modifying the regression model.(2) A transfer function model for load forecasting based on times series decomposition isdeveloped for explaining the influence of temperature on the load better. First, the periodicalcomponent is decomposed from the overall load data by using the time series decompositionmodel based on historical load data. On this basis, some major factors are selected from thosefactors having impacts on the non-periodic component of loads by using the stepwiseregression model. And then a transfer function model is developed for forecasting thenon-periodic component of loads.(3) A transfer function model for load forecasting based on spectral decomposition isdeveloped for improving the model prediction accuracy further. The modeling process is,firstly, the daily periodic component and weekly periodic component are decomposed fromthe overall load data by using the spectral decomposition method. Then, a transfer functionmodel is developed for forecasting the surplus component of loads by using the same modeling scheme as the non-periodic component of loads.Simulation results demonstrate that the three methods proposed in the thesis caneffectively improve the accuracy and weather adaptability of the time series model.