Power transformer reliability is fundamental to electricity distribution networks, where unexpected equipment failures disrupt service continuity, increase maintenance expenditure, and compromise infrastructure resilience. Conventional condition assessment methods frequently evaluate dissolved gas concentrations, thermal deterioration, and loading conditions independently, limiting their capacity to capture interacting degradation mechanisms and predict emerging faults. This study proposes a physics informed predictive framework integrating dissolved gas analysis, thermal aging models, and historical load profiles to estimate transformer failure risk and support condition-based maintenance. Dissolved concentrations of hydrogen, methane, ethylene, acetylene, and carbon oxides characterise electrical and thermal fault signatures, while hotspot temperature calculations and insulation aging equations quantify cumulative thermal deterioration. Time varying load profiles capture operational stress, overload frequency, and temperature dependent degradation. Physics constrained machine learning models combine these indicators with transformer operating characteristics to estimate fault probabilities and remaining useful life. Model evaluation incorporates temporal validation, prediction uncertainty, calibration, and comparisons against conventional diagnostic approaches. The proposed framework strengthens interpretable asset health assessment, supports maintenance prioritisation, and provides a foundation for reducing unplanned outages, optimising replacement decisions, and improving long term distribution transformer reliability.