Abstract: In the construction of the new power system, the volatility of renewable energy and the introduction of power electronic equipment have transformed transformer loss from a traditionally relatively fixed parameter into a complex dynamic variable. To achieve accurate prediction of transformer loss, this paper integrates multi-source parameters such as the operating voltage, load current, load rate, and winding temperature of the transformer, and proposes a short-term transformer loss prediction method based on an LSTM-GRU hybrid neural network with a residual structure. Verified by the measured data of a 220kV transformer in a substation, the proposed method achieves a prediction accuracy of 98.64% and outperforms the single LSTM and GRU network models. The research results can provide important references for power grid energy conservation and transformer operation optimization.