Forecasting indoor temperature and thermal loads in office buildings using multi-phase trained NARX networks: Toward smart energy applications

Forecasting indoor temperature and thermal loads plays a crucial role in improving energy efficiency, enabling advanced control strategies, and supporting design optimization. However, many data-driven approaches struggle to ensure reliability when applied across different climates and operational scenarios, which limits their robustness and transferability. To overcome this limitation, this study proposes a novel training procedure for nonlinear autoregressive with exogenous input (NARX) neural networks that accounts for multiple heating and cooling setpoint scenarios to maximize prediction reliability. The approach is tested on a representative office building prototype developed by ENEA (Italian national agency for new technologies, energy and sustainable development), typical of central Italy constructions from 1946 to 1970. EnergyPlus is used to perform dynamic simulations across 52 locations covering several Italian climatic zones. The resulting datasets are employed for multi-phase training of NARX networks: first under fixed setpoints, then validated and tested under different conditions. Results show strong predictive performance, with mean square errors between 0.038 and 0.14 °C2 for indoor temperature, and between 10−5 and 0.33 kW2 for thermal loads, and coefficients of determination consistently close to 1. By integrating climate variability and operational scenarios into the training process, the proposed method provides a versatile and accurate forecasting tool, adaptable to other buildings and contexts.