冰雪天气下基于LSTM的跑道温度数据-机理联合预测

Runway temperature data mechanism joint prediction based on LSTM under ice and snow

温度是跑道结冰的重要因素,针对跑道除冰运行的跑道热特性参数瞬态变化问题和温度周期序列缓慢变化特性,建立冰雪天气下基于长短时记忆(LSTM)的跑道温度数据-机理联合预测模型。通过最大信息系数法选择数据模型的输入特征变量,采用动态时间弯曲法进行跑道温度数据聚类划分,建立基于LSTM的数据预测模型;通过跑道热力学知识获取跑道温度预测机理模型,采用最小误差赋权法建立跑道温度数据-机理联合预测模型。仿真预测显示,预测时长为20 min、残差阈值为±0.5℃时,数据-机理联合预测模型优于单独的数据预测模型和机理模型,预测准确率可达99.34%;横向对比显示,在相同边界条件下,数据-机理联合预测模型优于BP神经网络、多元回归模型和支持向量机模型,平均准确率提高26.11%。研究表明,基于LSTM的跑道温度数据-机理联合预测模型契合冰雪天气下跑道除冰运行实际,可获得较好的跑道温度短时预测结果。

Runway temperature is an important factor in runway icing.Fully considering the transient characteristics of the temperature mechanism model and the time sequence of temperature multivariate time series data,the paper has developed a joint model based on the long short term memory(LSTM) neural network and temperature mechanism model.Firstly,the influencing elements with a greater correlation with runway temperature were selected by the study using the maximum information coefficient approach to serve as the model's input.Secondly,the paper uses the dynamic time warping method to cluster temperature data under different snowfall conditions,and then develops an LSTM model adapted to different snowfall or icing situations.Finally,to solve the disadvantage of LSTM which can not be characterized by the runway parameters that change irregularly and frequently,the paper developed a joint model based on the LSTM neural network and temperature mechanism model by using the minimum error method.The joint model's degree of accuracy is 99.34%,which is superior than both the data model and the mechanism model,when the prediction time step is 20 minutes and the residual threshold is±0.5°C,according to the simulation's result based on the ice and snowfall weather condition data.With the same condition,the joint model has better accuracy than the BP model,the regression model and the support vector machine model.Average accuracy increased by 26.11%.It proved the joint model based on the LSTM neural network and temperature mechanism model has better accuracy according to the transient characteristics of the mechanism model and the periodic time sequence of the multivariate time series of pavement temperature.