基于LSTM和先验知识的高速公路路面温度预报

Expressway Pavement Temperature Forecast Based on LSTM and Prior Knowledge

为了精准预报高速公路路面温度,为车辆安全行驶提供气象保障,采用2019—2022年南京市绕城高速公路上9个交通气象站及ERA5-land再分析数据,通过构建时间序列特征工程、引入物理机制相关数据两类方法结合先验知识,运用长短期记忆神经网络模型建立研究区域内4个交通气象站未来3 h逐10 min路面温度多步预报模型并进行验证;在此基础上,将已建立的模型应用于其他交通气象站,探究模型的适用性。结果表明:结合先验知识后,模型预报性能明显提高,准确率在85%以上,且随着预报时效的延长,性能提升更为明显,准确率最高提升36%;模型能较为准确地预报路面极端低温发生的时间和极值,且在预报时效较短时对路面极端高温的预报也具有一定参考价值;利用已建立的模型对其他交通气象站的路面温度进行预报时,准确率在62%以上,在预报时效较短时效果较好,准确率在80%以上,且交通气象站所处的下垫面背景类型对模型的选择起关键作用。

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.