Based on the 10 min observation data from the traffic meteorological stations along the Tianjin expressway during 2018-2023, a long short-term memory network model is used to forecast the pavement temperature in the next 3 h, and the forecasting effect and spatial transfer applicability of the model in the case of extreme pavement temperatures are evaluated and analyzed. The results show that the model forecasting effect is the best when the observation data of the time step 3 h are used as the input. The model can forecast the pavement temperature accurately, and the forecasting ability decreases with the extension of the forecast lead time, with the mean absolute error of 0.38-2.21℃, the root mean square error of 0.63-3.31℃, and the accuracy rate of 76.15%-99.48%. The model can accurately forecast the occurrence time and extreme values of extreme low temperature events on pavements, with the forecasting accuracy being about 90% for the 1 h forecast. For extreme high pavement temperature, the 1 h forecast of the model can also simulate the variation trend and the time for extreme high pavement temperature to occur. The model has a certain spatial transfer ability, and the average accuracy of the model forecast is more than 70% after the transfer, with an mean absolute deviation less than 3℃. Within a certain range, compared with the transfer distance, the forecasting ability of the model at its original station has a greater impact on the forecasting performance of the model after transfer.