Load Forecasting using The Multiple Regression Model （Part l）* -Fundamental Considerations for Load Variation and Load Forecast?

摘要

Since it is difficult to store and keep electric energy, the supplier of electric power must try to generate just the right amount required by the consumers. A power station cannot change its output easily, because very big equipment must be controlled, for example a boiler, a turbine, a generator etc. Consequently, forecasting electric load becomes important. According to our survey, the forecasting error of existing or developed methods is almost 3％ of the load. This error is sometimes bigger than a copacity of a power unit in many power systems. We wanted to improve the precision by developing a new load-forecasting method. This report describes our fundamental investigations concerning our new method. The shapes of the load curve not only for Saturday and Sunday but also for Monday were different from the shapes of other weekdays. The load for Tuesday to Friday deviated about 100 MW from the average taken over other such weekdays with the same temperature and at the same time of day. Considering days with a temperature change from the previous day of a　given value, the deviation of the load increment was about 50 MW from the average of such days. This means that the forecast error should be 5 or 10％ if the electric load is forecasted only by temperature or temperature increase. We forecast the load using the multiple regression method. We attempted to forecast the load with various combinations of data. These data include past and present temperature and past load values. 0ne example of a combination of data used was load levels of 48 and 24 hours ago and the previous day’s temperature and the forecasted temperature for the following day. The load forecast errors of the result were about 1％. Because we forecasted the load only for the previously described weekdays (Tuesday to Friday) in only June and July of 1986, this error cannot be claimed to be small but we plan to apply artificial intelligence techniques to improve this proposed method.