Reconstructing hourly power profiles from monthly billing data: A neural network framework with two-phase validation

Electrical grid planning requires accurate hourly power consumption profiles, yet utilities typically possess only monthly billing data. This study presents a neural network framework for reconstructing detailed hourly power profiles from aggregated monthly consumption features. Feature engineering transforms hourly consumption into 46 monthly aggregated features, including tariff-based totals and distribution ratios. Principal Component Analysis and K-means clustering identify 14 distinct user behavioral patterns. Three neural network architectures are systematically compared: Multi-Layer Perceptron, Long Short-Term Memory, and Gated Recurrent Unit networks. The methodology employs temporally separated validation, using 2022 data for training and 2023 data for validation, thereby assessing robustness to inter-annual variations in weather, economic conditions, and consumer behavior. Among the evaluated models, the Gated Recurrent Unit achieved the best overall performance with an R2 of 0.87 and a 40% reduction in mean squared error compared to XGBoost. For peak load estimation, which is critical for grid capacity planning, the proposed approach achieves a peak error of 18.3% for high-consumption users. Clustering stability analysis and evaluation across extreme user segments (high-consumption, high-volatility, and low-consumption) further confirm the robustness of the proposed methodology.