| Popis: |
The variation of road surface temperature along highways is a crucial indicator for traffic meteorological conditions and constitutes a significant focus in the research on meteorological disasters related to transportation. Accurate forecast of pavement temperature, timely issuance of pavement condition warnings, and alerting relevant personnel to take defensive measures are of paramount importance for ensuring the safety of people's lives and property. Observations from 4 expressway meteorological stations along Nanjing City Ring Expressway and the corresponding ERA5-land reanalysis data from 2019 to 2022 are analyzed. Utilizing feature engineering techniques that consider the daily and seasonal temperature variations as well as temperature trends, a long-short-term memory (LSTM) neural network model, incorporating prior knowledge, is established for multi-step pavement temperature forecasting at 10 min intervals for the next 3 hours. The models are validated under different scenarios including extreme high and low pavement temperature conditions. They are further transferred and applied to 5 additional meteorological stations to investigate the model universality. This approach addresses the challenge of pavement temperature forecasting for stations with limited historical data due to new construction or equipment maintenance. Results indicate that the incorporation of prior knowledge facilitates a more comprehensive consideration of environmental influences by maximizing the feature extraction capabilities of LSTM. All forecasting performance metrics of the model exhibit significant improvements, with the accuracy exceeding 85%. As the forecast lead time extends, the enhancement in various forecast metrics becomes more pronounced, reaching a maximum accuracy improvement of 36%. The model accurately predicts the occurrence time and extremities of extreme low temperatures, but it exhibits relatively weaker capabilities in forecasting extreme high temperatures, with approximately 1 h advance in occurrence time and an underestimation of about 4 ℃. Despite this generally lower forecasting efficacy, the model still provides valuable information. When applying models to forecast pavement temperatures at other meteorological stations, the accuracy exceeds 62%. The forecast performance is better for short lead times, with the accuracy surpassing 80%. The underlying surface type plays a crucial role in the selection of different models. The suburban station model performs relatively optimally for urban meteorological stations and suburban meteorological stations, while the rural station model performs relatively optimally for rural meteorological stations. |
