Abstract: To realize the load optimization scheduling of power transformers and timely warning of thermal faults to improve the operational reliability of power equipment, this paper proposes a short-term and ultra-short-term multi-timescale prediction method of top oil temperature based on an adaptive extended Kalman filtering algorithm. The method combines the Kalman filtering algorithm and the Susa thermal circuit equivalent model, selects the top oil temperature, oil index, and oil time constant as state variables, ambient temperature, and load current as the inputs. It achieves iterative optimization of the oil index and oil time constant by comparing the estimated and observed values of the top oil temperature to improve the prediction accuracy of the top oil temperature in multiple time scales. In addition, the model utilizes an adaptive noise estimator to correct the noise statistical parameters to automatically optimize the simple noise initial value setting, thus further improving the model's prediction accuracy. Taking two 110kV oil-immersed transformers as examples, the results show that the method has higher prediction accuracy for ultra-short-term and short-term prediction of the top oil temperature compared with the hot circuit equivalent model calculation and the extended Kalman filtering algorithm.