The invention relates to a multi-step prediction method for early warning of overtemperature of the tube wall of the heating surface of an ultra-supercritical boiler, comprising: step 1, collecting training data in the training stage, and constructing a time series network model after preprocessing the collected training data And carry out offline training; step 2, collect the input data in the use stage as the input of the time series network model, and obtain the prediction result of the metal wall temperature data. The beneficial effects of the present invention are: the temperature of the tube wall on the heating surface can be predicted in advance, and guidance can be provided for over-temperature control; longer time-series characteristic information can be captured, and the prediction accuracy can be improved; by sinusoidal encoding of time, the transition of each day can be smoothed, It is more in line with the actual operating conditions; it can not only solve the problem that the sensor cannot work for a long time by direct measurement, but also realize the prediction of wall temperature at different positions; it can easily add the latest data and retrain the model, and has good anti-interference.