Comparative Analysis of the Accuracy of Traditional and Machine Learning Methods for Predicting Electricity Consumption in Power Grid

This paper investigates the forecasting of the total active power consumed by industrial enterprises from the power grid. Accurate forecasts support operating-mode optimization, maintenance planning, and reliable power supply. The study reviews widely used forecasting methods, including the Holt-Winters method, regression methods, autoregressive integrated moving average models, support vector machines, a technocenological approach, and neural networks. Its objective is to compare forecasting accuracy and identify effective methods for different forecasting conditions. An experiment was conducted using synthetic hourly electricity consumption data covering one year for a mining-and-processing plant and an industrial city, with climatic factors taken into account. Accuracy was assessed using the mean absolute percentage error over four test weeks representing different seasons. Several methods showed comparable performance. For the mining-and- processing plant, the Holt-Winters method and random forest achieved the lowest errors, at 1.32% each. For the industrial city, support vector regression yielded the lowest error, at 2.15%, followed by random forest at 2.17% and the Holt-Winters method at 2.37%. The neural network did not outperform these methods in this experiment. The results indicate that method selection depends on the forecasting conditions and the available dataset. Further validation with real-world data is needed.