Comparing various neural network methods for temperature prediction of CRTS II slab track on transition sections

Based on the meteorological measuring data, the prediction of the temperature field is worthy of thermal performance evaluation of CRTS II slab tracks in bridge-subgrade transition sections. To find the best temperature prediction method, this present study shows a comparison of internal temperature predictions in CRTS II slab track by using three typical neural network methods (ANN, CNN, LSTM) subjected to different meteorological factors. Firstly, the distribution characteristics of four meteorological factors (e.g. ambient temperature, solar radiation, wind speed, and humidity) and internal temperature for the CRTS II slab track on three different foundations are analyzed. Moreover, temperature prediction effects of track slab and base plate on three foundations under five meteorological testing cases are compared by using three neural network models, respectively. The results show that the ambient temperature ranging from 15°C and 25°C accounts for about 7% and the solar radiation during daytime mainly ranges from 100 W/m2 to 1100 W/m2. The solar radiation has more effect on the temperature gradients of the CRTS II slab track on bridge and transition zone than that of the ambient temperature, and Case 5 with five different input variables has the best prediction accuracy for three predict models among five testing cases. Although the LSTM model has the best prediction accuracy among the three prediction models with R2 values of about 0.85, it costs the longest calculation time of about 180 s. In addition, the track slab on bridge has the worst prediction accuracy for the ANN and CNN models among the three foundations with RMSE values of 4.5 and 2.5 for Case 2, and the base plate on transition zone has the best prediction accuracy both for the CNN and LSTM models among the three foundations with RMSE values of 3 for Case 5.