Achieving global carbon neutrality targets requires innovative approaches to improve the energy performance and thermal comfort of buildings. This study explores the use of Physics-Informed Neural Networks (PINNs) to predict indoor thermal comfort conditions in residential buildings, combining physical modelling with Machine Learning (ML) techniques. The selected case study is a Multi-Storey Residential Building (MSRB) constructed in the 1980s in Bari, Southern Italy. A dynamic energy simulation of the building is performed using DesignBuilder software, calibrated in accordance with ASHRAE Guideline 14 (2023), and employing real IGDG weather data. The dataset includes hourly data for heating loads, solar gains, occupancy, ventilation, and outdoor air temperatures during the winter season. These variables are used as input features for the PINN, which embeds the first law of thermodynamics into the loss function via a soft constraint defined by a custom partial differential equation (PDE). Hyperparameter optimization is conducted using the Optuna framework, and an ensemble learning approach, based on three independent training runs with different random seeds, is adopted to improve generalization and stability. The model achieves high predictive accuracy in forecasting the Indoor Operative Temperature ( \( {\text{T}}_{{\text{o}}_{\text{p}}} \) ) one hour ahead, with performance metrics of CVRMSE = 0.31%, MBE = −0.003 °C, NME = 0.17%, and NMBE = −0.02%, all within ASHRAE-recommended thresholds. PMV is computed from predicted \( {\text{T}}_{{\text{o}}_{\text{p}}} \) values using ISO 7730 assumptions. The PINN-based approach achieves a comfort compliance rate of 14.12%, showing a relative improvement of +3.37% compared to the traditional PMV method. These results confirm the effectiveness of PINNs for physically consistent, data-driven predictions of indoor comfort, providing a robust framework to support energy-efficient and occupant-centred design strategies in the residential building sector.