基于LSTM-GMM模型的大风多级预警方法

Multi-Level Early Warning Method for Gale Based on LSTM-GMM Model

为保障大风场景下的高铁运行安全,针对风的强随机特性提出一种基于长短期记忆网络和高斯混合模型的多级预警(LSTM-GMM-MELW)方法。首先,通过长短期记忆网络和高斯混合模型(LSTM-GMM)建立风速误差值与风速预测值的联合概率密度,以此确定风速预测值的概率密度;然后,通过多级预警方法计算风速预测值落在高铁限速风速区间的概率值并结合实际情况设置不同阈值,当得到超出阈值的概率时输出阈值对应的预警等级;最后,采用预测区间的覆盖概率、平均宽度和覆盖宽度评价LSTM-GMM方法的概率性预测结果,而采用预警准确率评价多级预警方法的预警效果。依托平潭海峡公铁两用大桥29个风速样本进行实例分析,结果表明:95%置信度下的预测区间的覆盖概率为96%,平均宽度为1.51;第1、第2级别的预警准确率分别高于85%和93%,预警准确率达到100%的风速样本达14个,总体预警准确率高。该方法能有效避免风速在限速分界线附近波动时的误报。

In order to ensure safe operation of high-speed railway in gale scenarios,a multi-level early warning(LSTM-GMM-MELW)method based on long short-term memory network and Gaussian mixture model(LSTM-GMM)is proposed for the strong random characteristics of wind.Firstly,the joint probability density of the wind speed error value and wind speed prediction is established through LSTM-GMM,so that the probability density of wind speed prediction value is determined.Then,the multi-level early warning method is used to calculate the probability value of the wind speed prediction value falling within the limit range of wind speed of the high-speed railway,and different thresholds are set according to the actual situation.When the probability exceeding the threshold is obtained,the early warning level corresponding to the threshold is output.Finally,the coverage probability,average width and coverage width of prediction interval are used to evaluate the probabilistic prediction results of the LSTM-GMM method,while the early warning accuracy is used to evaluate the early warning effect of the multi-level early warning method.The example analysis is carried out based on 29 wind speed samples collected from the Pingtan Strait Rail-cum-Road Bridge.The results show that the coverage probability of the prediction interval is 96%with the average width of 1.51 at a 95%confidence level.The early warning accuracy of the first and the second levels is higher than 85%and 93%,respectively.The number of wind speed samples with 100%early warning accuracy reaches 14,and the overall early warning accuracy is high.Compared with the traditional early warning methods,the proposed method can effectively avoid false alarm when the wind speed fluctuates near the dividing line of speed limit.