An empirical study of combinational load forecasting in a city power company of China

Abstract

It is known that load forecasting plays an important role for the smart grids and mostly, the users have the combinational load forecast demands with different time-scales. However, due to various internal or external factors, such as poor management, the available load data set is limited. Therefore, the problem becomes as the high-precision combinational load forecast at different time-scales but with limited valid datasets. In this paper, taking the work in a city power company of China as an example, an empirical study for this problem is considered and then a robust forecasting system is built. The forecasting models are designed and screened and then constructed a model library for the system. In order to improve the robustness of the designed system, a mean absolute percentage error (MAPE) based dynamic model selection method is proposed and finally the performance of the system is verified on the State Grid Chongqing Company of China. The results indicate that our designed multi-temporal-spatial-scales load forecast system can dynamically select the best model for different demands, which demonstrates that the build system is an effective and reliable system-level solution. In particular, for the annual based load forecast, 80% of the industries in the considered city have a MAPE less than 5%, while the MAPE of the whole city is less than 2.160%; for the monthly based load forecast, 87% of the industries in the considered city have a MAPE less than 10%, while the MAPE for the whole city is less than 3.755%. In addition, for the annual based forecast, GM(1,1) is mostly used for the industry-based forecasts, but the model with the highest average forecast accuracy is LR-ER. While for region-based forecasts, the mostly selected models are GM(1,n) and AR. For the monthly based forecast, GM(1,1)-SI and FOA-HW have a higher average accuracy than other models for both industries- and region-based forecasts, respectively.