Residential Net-Load Forecasting with Statistical and Learning-Based Models

The increasing adoption of behind-the-meter (BTM) solar photovoltaic systems has introduced greater uncertainty in residential demand patterns, making accurate net-load forecasting essential for reliable grid operation. This paper presents a unified comparison of statistical, machine-learning, and deep-learning models for residential net-load forecasting using net-load and weather data from Austin, Texas. A systematic feature-engineering framework is proposed to improve forecasting performance and facilitate a fair comparison between statistical and learning-based approaches. Model performance is evaluated using deterministic and probabilistic metrics. Results show that feature engineering significantly improves nonlinear models, with the feature-engineered Temporal Fusion Transformer (TFT) achieving the best overall performance, with an R-squared score of 0.923 and a weighted CRPS of 2.99. The findings demonstrate that combining domain-specific feature engineering with advanced deep learning architectures enables accurate, uncertainty-aware residential net load forecasting for renewable-integrated power systems.

Recommended citation: Imam, T., Sifat, A. I., Chowdhury, M. R. & Ferdaus, F. (2026, October). "Residential Net-Load Forecasting with Statistical and Learning-Based Models." In 2026 IEEE 58th North American Power Symposium (NAPS). IEEE.