沥青路面温度预测的长短记忆循环神经网络模型研究

Research on Long Short Term Memory Recurrent Neural Network Model for Asphalt Pavement Temperature Prediction

摘要: 沥青路面的力学特征演化、环境老化与温度直接相关，路面温度的预测是对沥青路面性能评价、研究的基础。首先依路面热传导方程将路面温度场离散为表面热交换、内部热传导两个部分，路面表面热流可以近似表示为气象环境的线性组合，路面温度为路面历史温度与边界热流的线性组合。然后建立与离散路面温度场耦合的双层长短记忆(LSTM)深度神经网络模型，模型采用时间步迭代方式运行，迭代步内神经元能够模拟表面热交换、内部热传导过程。模型时间序列长度48 h，训练时前24小时的数据主要用于模型参数更新，后24小时用于验证训练效果；验证时全部时间序列用于检验模型效果。然后，实测、收集了武汉、广州、唐山、苏州的温度场数据集、气象数据，并将路面温度数据归一化成统一的深度，对路面温度存在锯齿形不光滑的数据做标记剔除短时降雨的影响。然后用武汉、广州训练集训练模型，比选模型的输入气象因素、神经网络神经元个数与层数、学习率、训练时段，遴选出模型架构为4 + 5的双层LSTM神经网络。用唐山、苏州数据评价模型在0、2、4、6、8 cm深度的预测标准差为2.98、2.32、1.95、1.68、1.5℃，剔除降雨标记数据时模型预测标准差为2.0、1.71、1.54、1.25、1.08℃，路面深度越大模型预测效果越好。模型的因素分析表明，气温升高对路面温度的提升效应最明显，太阳辐射提高对白天路面温度有提高效应，气温实际是非独立的路面与环境交互作用的体现者。

Abstract: The evolution of mechanical characteristics and environmental aging of asphalt pavement are directly related to temperature. The prediction of pavement temperature is the basis of performance evaluation and research of asphalt pavement. Firstly, according to the pavement heat conduction equation, the pavement temperature field is divided into two parts: surface heat exchange and internal heat conduction. The pavement surface heat flow can be approximately expressed as a linear combination of meteorological environment, and the pavement temperature is a linear combination of historical temperature and boundary heat flow. Then, a two-layer long short term memory (LSTM) deep recurrent neural network model coupled with a discrete pavement temperature field is established. The model operates in a time step iteration mode. In the iteration step, neurons can learn the function relationship between input and output of surface heat exchange and internal heat conduction respectively. The length of the model time series is 48 hours. The data for the first 24 hours is mainly used to update the model parameters, and the data for the last 24 hours is used to validate the model. Then, the temperature field data sets and meteorological data of Wuhan, Guangzhou, Tangshan and Suzhou are measured and collected, and the pavement temperature data are normalized to a unified depth, and the data with zigzag unsmooth pavement temperature are marked to eliminate the impact of short-term rainfall. Then, Wuhan and Guangzhou training sets are used to train the model. The input meteorological factors, the number and layers of neural network neurons, learning rate and training period of the model are compared, and the double-layer LSTM recurrent neural network with model architecture of 4 + 5 is selected. Using Tangshan and Suzhou data evaluation model, the prediction standard deviation of 0, 2, 4, 6, 8 cm depth is 2.98, 2.32, 1.95, 1.68 and 1.5˚C, and the model prediction standard deviation is 2.0, 1.71, 1.54, 1.25, 1.08˚C when excluding the rainfall marker data, and the greater the pavement depth, the better the prediction effect of the model. The factor analysis of the model shows that the increase of temperature has the most obvious effect on the road surface temperature, and the increase of solar radiation has an effect on the road surface temperature in the daytime. The temperature is actually the embodiment of the interaction between the road surface and the environment.