Probabilistic Load Forecasting with Weather-Aware Neural Networks for Urban Energy Management

Abstract

The growing adoption of renewable energy and distributed electrical infrastructure has increased the uncertainty associated with urban electricity demand. Accurate load forecasting is particularly challenging during extreme weather events, when conventional forecasting models often underestimate rapid demand changes. This paper develops a weather-aware probabilistic neural network for short-term electricity demand prediction. The proposed architecture combines temporal convolutional layers with probabilistic forecasting heads to jointly estimate expected energy consumption and prediction uncertainty. The evaluation covers 24 months of simulated electricity consumption across 35 urban districts, comprising approximately 2.4 million hourly observations. Weather conditions, calendar patterns, and historical demand profiles are incorporated as model inputs. Experimental comparisons show that the proposed method achieves a mean absolute percentage error of 5.9%, compared with 7.8% for a conventional temporal forecasting network. During high-demand periods, peak load prediction error decreases by 21.3%. Additionally, the model achieves 91.6% empirical coverage for nominal 90% prediction intervals, indicating reasonably calibrated uncertainty estimates. These findings demonstrate the potential of probabilistic machine learning for improving demand planning and operational decision-making in intelligent energy systems.