传统基于热力学的混凝土结构温度场分析方法存在着假设过多、参数取值困难、计算能耗过大的缺点． 为研究轨道板竖向温度梯度分布规律，结合轨道板温度场的长期观测数据，建立误差反向传播的多层人工神经网络，选用易于取得的气象参数作为训练样本，对轨道板竖向温度梯度进行预测，并采用实测数据验证其准确性． 在此基础上研究了日温差、日照时数和风速对轨道板竖向温度梯度的影响规律． 研究表明:采用日温差、日平均风速和日照时数 3 种气象参数作为训练样本，所建立的 4-16-1 结构人工神经网络预测结果最大误差为 2． 0℃，平均相对误差为 0． 38%，可准确预测轨道板竖向温度梯度，且具有较好的鲁棒性;各气象参数与轨道板竖向温差之间存在着复杂的非线性映射关系，总体而言，日照越强，风速越高，轨道板竖向温度梯度越大;对我国中部地区而言，轨道板竖向温度梯度为 －2 ～10℃．

The traditional thermodynamics-based analysis methods of the temperature fields of concrete structuresare characterized by too many assumptions and excessive energy consumption in calculation,and with these methodsit is difficult to obtain parameter values.In order to know more about the vertical temperature gradient distributionin track plate,a multi-layer artificial neural network based on error back propagation is established by using thelong-term observation data of the temperature field of track plate.Then,the meteorological parameters easy to beobtained are used as the training samples to predict the vertical temperature gradient of track plate,and the predic-tion accuracy is verified by the measured data.On this basis,the influences of the daily temperature difference,thesunshine hours and the wind speed on the vertical temperature gradient of track plate are discussed.The resultsshow that (1) when the artificial neural network of a 4-16-1 structure is established with the daily temperaturedifference,the daily average wind speed and the sunshine hours as the training samples,the network has a strongrobustness and can accurately predict the vertical temperature gradient of track plate with a maximum error of 2.0℃and an average relative error of 0.38%; (2) there is a complex nonlinear relationship between each meteorologicalparameter and the vertical temperature difference of track plate; (3) generally speaking,the stronger the sunshineis and the higher the wind speed is,and the greater the vertical temperature gradient of track plate will be; and(4) the vertical temperature gradient of track plate is －2 ～10℃ in the central region of China.