Abstract: To address the incomplete information problems commonly existing in the actual operation of integrated energy systems (IES), such as sparse observation and missing measurement data, this paper proposes an adaptive physics-informed neural network (APINN)-driven mechanism-data fusion modeling method. Firstly, a unified mechanism-data fusion network framework is constructed, which embeds multi-energy flow physical mechanisms, including power flow of power grids, fluid dynamics of natural gas pipelines, and thermodynamics of heating pipelines, into the training objective in the form of residuals, and establishes observation consistency constraints with a small number of anchor measurements to alleviate convergence difficulties caused by non-unique solution space and trivial solutions in pure physics-informed training. Secondly, a loss amplitude-aware adaptive weight adjustment mechanism is designed to dynamically optimize gradient weights of multi-physics field learning tasks, effectively resolving model optimization imbalance caused by dimensional mismatch across multiple physics fields. Finally, numerical validation on an electricity-gas-heat coupled IES consisting of an IEEE 24-bus power system, a 20-node natural gas system and a 16-node heating system shows that the proposed method achieves stable global state reconstruction with favorable physical consistency and anti-disturbance robustness.