基于特征相关度分析的路面凝冰短时预测方法

Short-term Prediction Method for Road Icing Based on Feature Correlation Analyses

针对高速公路易结冰路段的路面凝冰预测问题,提出了一种基于特征相关度分析的路面凝冰短时预测方法。该方法利用路侧设备的测量数据,包括结冰厚度、相对湿度、风向与风速等,通过ADF(Augment Dickey-Fuller)检验方法分析数据集的平稳性,进而设计出基于长短期记忆网络(Long Short-term Memory,LSTM)的路面凝冰短时预测算法。根据Spearman相关度系数法分析计算上述多种凝冰监测数据的相关度与置信度,并形成基于Spearman特征相关度的数据筛选模型,优化LSTM神经网络中的输入数据集。在此基础上,搭建面向凝冰预测误差的LSTM神经网络模型,并利用筛选后的凝冰数据集训练优化预测算法中的模型参数,提高目标路段路面凝冰预测的效率与精度。最后,通过数值仿真分析比较不同特征相关度下路面凝冰短时预测算法的均方根误差,确定最优预测模型,并于西延高速KM200+918路段进行实地测试。研究结果表明:路侧设备的测量数据中相关度较低的数据对路面凝冰预测算法存在反向作用,并非将所有数据进行组合即可得到最优结果,需对测量数据进行有效筛选,进而优化LSTM神经网络,提高凝冰预测精度。利用所提出的基于特征相关度的数据筛选方法,将相关度较高数据进行组合会得到更精准的预测结果,预测值的均方根误差相比传统方法降低了41%,有效提高了路面凝冰短时预测算法的有效性与精准性。

To improve the driving safety of expressways,a pavement icing prediction method is proposed for solving the problem of pavement icing prediction on icing risk sections of the expressways,in which the long short-term memory(LSTM)model based on data feature correlation analyses is employed to solve the problem.First,different meteorological data,including icing thickness,relative humidity,wind direction,and wind speed,were measured using roadside equipment.After an Augment Dickey-Fuller(ADF)test of the sampled data,an LSTM-based neural network algorithm was developed to estimate the road icing states.According to the Spearman correlation coefficient technique,the correlation and confidence between the meteorological data and the road icing thickness data were calculated.Thus,an input dataset for the LSTM was chosen from the original measurements.Subsequently,the specified LSTM neural network was developed,and the model parameters were obtained by training using the optimized icing dataset.Finally,the average errors of the short-term expressway pavement icing prediction models under different feature correlations were analyzed and compared using numerical simulations.Furthermore,a field test was conducted on the KM200+918 section of the Xiyan Expressway.The results show that data with a low correlation in the sampled dataset have a negative effect on the pavement icing prediction model.It is difficult to obtain an optimal training result by directly using all sampled data.The measurement data must be effectively filtered.Using the proposed feature correlation analysis method,the prediction accuracy can be effectively improved by only using data with a high correlation.Compared with the traditional method without the feature correlation analysis method,the root mean square error(RMSE)between the predicted and actual values was reduced by 41%,which effectively improves the effectiveness and accuracy of the short-term prediction algorithm for pavement icing on sections of the expressways with icing risk.