COMPARISON OF CLASSIC AND NEURAL NETWORK MODELS FOR FORECASTING ELECTRICITY CONSUMPTION OF NON-HOUSEHOLD CONSUMER

Abstract

The study addresses the problem of forecasting the electrical load of an energy facility under conditions of high consumption variability. A comparative analysis of forecasting model performance is carried out for horizons of 1 and 24 hours. For the first case, the SSA, Holt–Winters methods, as well as LSTM and Transformer neural network architectures, were examined. For the second case, models with prior decomposition based on the Hilbert–Huang transform (model M1) and polynomial regression (model M2) were additionally considered. The quality of the models was evaluated using four metrics: mean error (ME), mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). The results show that, for the 1-hour horizon, the Transformer model achieved the lowest MAE and MAPE values (2.54 kW and 4.95%, respectively), indicating high accuracy. LSTM demonstrated similar accuracy, with the smallest forecast bias. The SSA and Holt–Winters models were significantly less accurate, though they showed better stability in avoiding large errors. For the 24-hour horizon, the Transformer model achieved the best results in both accuracy and stability (MAE = 3.61 kW). The M1 model, based on the Hilbert–Huang transform, showed balanced performance across all metrics, while LSTM achieved high absolute accuracy. Additional analysis of mean error distribution frequencies showed that Transformer and LSTM provide high densities of accurate forecasts within narrow error intervals, unlike SSA and Holt–Winters, which are characterized by systematic biases. The conclusions have practical significance for energy management tasks in microgrid conditions, particularly for operational load planning, loss reduction, and optimization of backup power sources. References 16, figures 2, tables 3.