Averaged Errors as a Risk Factor for Intelligent Forecasting Systems Operation in the Power Industry

Abstract

The paper discusses the operational risk in intelligent systems for forecasting time series. Typically, when developing and testing regression models based on machine learning, their accuracy is calculated over a long time interval, from several months to several years, and then is averaged. However, in the real-life operation of such systems, the customer is likely to draw a conclusion about the system efficiency based on the results of the first 2–4 weeks of operation. If one or several large errors appear on this short interval, they will not be averaged as it happens over a long one. As a result, there is a risk of failure in the intelligent forecasting system implementation due to the discrepancy between the calculated mean error and that obtained over a short time period at the start of operation. This study considers the problem of solar power plant generation short-term forecasting, analyzes the distribution of errors over short time periods, and substantiates the need to use more detailed accuracy metrics of machine learning models than the error values averaged over a long interval.