• Energy Research
• 2016-2025close

In this paper, the load time-series measured from 2016 to 2018 in Qingdao is investigated in order to make a prediction more accurately by using an artificial neural network model combined with some regular and irregular features. The results of spectral analysis show that several periodic variations in diurnal, semidiurnal and weekly frequencies are prominent, and considered as critical parts of predictior variables of the training and test sets in this forecasting model. However, a significant decline of load happens during the national statutory festivals in China, and is more obvious with the longer holidays, which should be taken into account as abnormal conditions. Moreover, both temperatuer and hummity filtered out by mutual information method are also added as two essential weather factors to improve the prediction accuracy, especially in the hot summer. Finally, the comparative results of five different experiments in term of mean absolute percent error show that the forecasting model combined with all periodic and non-periodic factors outperforms the others with one single or a few of factors. This multifactorial model, taking fully account of internal charateristics of load data and external important influences, is proved more suitable to predict the load trend in Qingdao.

There are many factors that affect short‐term load forecasting performance, such as weather and holidays. However, most of the existing load forecasting models lack more detailed considerations for some special days. In this paper, the applicability of the bagged regression trees (BRT) model combined with eight variables is investigated to forecast short‐term load in Qingdao. The comparative experiments show that the accuracy and speed of forecasting have some improvements using the BRT than the artificial neural network (ANN). Then, an indicator variable is newly proposed to capture the abnormal information during special days, which include national statutory holidays, bridging days, and proximity days. The BRT model combined with this indicator variable is tested on the load series measured in 2018. Experiments demonstrate that the improved model generates more accurate predictive results than BRT model combined with previously variables on special days.

Accurate electricity load forecasting is an important prerequisite for stable electricity system operation. In this paper, it is found that daily and weekly variations are prominent by the power spectrum analysis of the historical loads collected hourly in Tai’an, Shandong Province, China. In addition, the influence of the extraneous variables is also very obvious. For example, the load dropped significantly for a long period of time during the Chinese Lunar Spring Festival. Therefore, an artificial neural network model is constructed with six periodic and three nonperiodic factors. The load from January 2016 to August 2018 was divided into two parts in the ratio of 9 : 1 as the training set and the test set, respectively. The experimental results indicate that the daily prediction model with selected factors can achieve higher forecasting accuracy.