In the context of new power systems, the grid exhibits intermittent and fluctuating characteristics, subjecting power transformers to complex working conditions involving multi-physical field coupling. To address the limitations of traditional power transformer condition assessment models, such as their inability to effectively uncover defect and failure mechanisms or to make full use of multi-source heterogeneous data, a power transformer condition assessment based on physics-informed LSTM network is proposed. The approach integrates model-driven and data-driven concepts, offering a new research direction for the refined management and intelligent operation and maintenance of power transformers under new power system conditions. First, the physical model of power transformers under electromagnetic-fluid- temperature multi-physics coupling is analyzed to establish a prior knowledge function for transformer condition assessment. Then, the knowledge function is embedded into the loss function of the LSTM network to penalize “black-box” data that violate physical prior knowledge. This enables effectively discovery of the evolution patterns of power transformers under complex coupled conditions. A Softmax classifier is used to obtain the health index and life prediction model of power transformers. Finally, the effectiveness of the proposed algorithm is verified based on the actual transformer operation data, achieving a comprehensive condition assessment of key performance indicators and their interrelationships.