Recurrent and Ensemble Models for Short-Term Load Forecasting of Coal Mining Companies

Abstract

For open cast mining enterprises, electricity costs significantly affect the self-cost of production. To reduce the electricity tariff, an enterprise should improve the accuracy of short-term power consumption forecasting (day-ahead). Forecasting the power consumption of a mining enterprise is a difficult task due to the influence of many factors: technological, geological, metrological, and administrative. Therefore, it is necessary to use artificial intelligence methods based on machine learning, such as artificial neural networks and ensemble models. They show high efficiency in forecasting the daily curve of electricity consumption of large power supply systems, households, and industrial enterprises. At the same time, at present, there are practically no studies of modern machine learning methods concerning the short-term power consumption forecasting of mining enterprises. It is largely due to the lack of open access data on mining enterprises' power consumption. Research and verification of the results require the data on various enterprises for several years. In this work, the authors' data on four enterprises in Yakutia operating in the open cast coal mining and processing for four years are used. A study of two different classes of machine learning methods has been carried out. The first one is processing retrospective power consumption data as a time series using recurrent neural networks. The second one is selecting the most significant features and applying ensemble models based on decision trees. The advantages and disadvantages of these approaches are shown; the obtained forecast accuracy for four enterprises that differ in their technological processes are given.