Enhancing Road Surface Temperature Prediction: A Novel Approach Integrating Transfer Learning with Long Short-Term Memory Neural Networks

Timely and accurate prediction of winter road surface temperature is crucial for the effective operation of a road weather information system (RWIS), which is essential to road traffic safety. A major challenge in achieving high-precision predictions is the lack of extensive data, particularly in newly established road weather stations. To address this challenge, this study proposes a transfer learning and long short-term memory network-based (TL-LSTM) model for road surface temperature prediction. This model is designed to overcome the accuracy limitation typically encountered in small sample modeling. First, the pretrained model containing the long short-term memory (LSTM) network feature extraction module and prediction module is constructed, which learn the pattern in road temperature time series using the long-term data from the established road weather station. Subsequently, the pretrained model is transferred to the target road weather station data set with a small sample for fine-tuning weights to determine the optimal transfer strategy. The results show that the best prediction performance is achieved when freezing the LSTM feature extraction module and the first two fully connected layers of the predictor module. In the case of small samples, the TL-LSTM model improves accuracy by 30% compared to the baseline model, achieving a mean absolute error (MAE) of 0.673, a mean square error (MSE) of 1.314, and a mean absolute percentage error (MAPE) of 12.8%. Notably, the model performs particularly well in the low-temperature range (-5 degrees C to 5 degrees C). It adeptly identifies the periodic fluctuations and uncertainties in road surface temperature. During both cloudy and sunny conditions, its forecasts align closely with the observed values, demonstrating the model's robust reliability.