Forecasting Disaggregated Electricity Data with General Regression Neural Networks

Electrical load forecasting is crucial for the efficient planning of electricity companies. Load forecasting traditionally refers to forecasting the expected electricity demand at aggregated levels. More recently, with the rise of high granularity data, load forecasting has shifted towards a disaggregated approach. However, forecasting disaggregated electrical loads such as residential or commercial buildings is a challenge due to high variability in load curves influenced directed by individuals. This study proposes using general regression neural networks to forecast disaggregated electrical loads from an electric power distribution system in São Paulo for four consumption profiles: Residential, Commercial, Municipal Public Power and Public Water Service. In general, the Mean Absolute Percentage Errors (MAPE) obtained in the forecast of the Commercial, Municipal Public Power, and Public Water Services profiles using the GRNN were consistent with the results found in the literature. These MAPEs were, on average, 12,97%, 14,39% and 9,11%, respectively. When using the modified GRNN, lower or equal MAPEs were obtained with a lower training and validation time. The MAPEs obtained in the forecast of the residential profile with the GRNN, on the other hand, averaged 144.13% and demonstrated that the forecasting system was not able to generalize the behavior for these loads.