基于特征提取的输油管道泄漏系数预测*

为准确预测管道泄漏系数，估计管道泄漏量，以基于瞬变流方法的模拟数据为例，建立多个管道泄漏系数预测模型(多层感知机、长短期记忆网络、随机森林、支持向量机以及K近邻回归)，综合考虑管道流量和压力数据特点，提出序列提取法和均值提取法2种管道时序数据预处理方法，模型评价指标为相关系数（R2）和平均绝对百分比误差（MAPE）。研究结果表明:随机森林和多层感知机的抗噪性较强，在5%的噪声影响下，模型准确度下降幅度较小；均值提取法去噪功能较好，可在一定程度上降低噪声影响；基于均值提取法的多层感知机模型效果相对较好，R2为0.997 5，MAPE为1.599%，研究结果可为准确预测管道泄漏系数、估计泄漏量提供指导。

Prediction on leakage coefficient of oil pipeline based on feature extraction

In order to accurately predict the leakage coefficient of pipeline and estimate the leakage amount of pipeline,taking the simulation data based on transient flow method as an example,multiple prediction models for the leakage coefficient of pipeline (multi-layer perceptron,long short-term memory network,random forest,support vector machine,K nearest neighbor regression) were established.Considering the data characteristics of pipeline flow rate and pressure comprehensively,two preprocessing methods of pipeline time series data,including the sequence extraction method and mean extraction method,were proposed.The model evaluation indicators were the correlation coefficient (R2) and the mean absolute percentage error (MAPE).The results showed that the random forest and multi-layer perceptron were more resistant to the noise,and the accuracy of models did not decrease much under the influence of 5% noise.The mean extraction method was powerful in denoising,which could reduce the influence brought by noise to a certain extent.The multi-layer perceptron model based on the mean extraction method worked the best with R2 of 0.997 5 and MAPE of 1.599%.The results have the guidance significance for the accurate prediction of pipeline leakage coefficient and the estimation of leakage amount.