基于卷积神经网络的汽轮机抗燃油泄漏智能预警技术研究

Research on intelligent early warning technology of steam turbine anti fuel leakage based on convolutional neural network

汽轮机数字电液控制系统依靠抗燃油(EH油)对汽门进行调节,抗燃油一旦发生泄漏,将危及机组的运行安全。为提高EH油泄漏故障的监测可靠性,减少电厂非计划停机次数,基于大数据挖掘与数据驱动技术,构建了多个不同结构的卷积神经网络模型,用于预测监控EH油箱油位。利用机组实际运行积累的大量数据,建立模型训练与验证数据集,从预测准确率、交叉熵损失及运算耗时等方面对模型进行了综合评价。评估结果表明,A结构的模型准确性最高,约为98.92%,交叉熵损失最低,约为0.043 1,而模型运算时长中等,综合性能最优。将A智能模型整合进电厂分散式控制系统中进行实际验证,其监测准确,预警及时,显著减少了监盘和巡检人员的工作量,提高了电力生产的自动化、智能化水平,相关应用经验可供同类型机组参考。

The digital electro-hydraulic control system of steam turbine relies on fire-resistant oil(EH oil)to regulate the steam valve.Once the fire-resistant oil leaks,it will endanger the operation safety of the unit.In order to improve the monitoring reliability of EH oil leakage fault and reduce the number of unplanned shutdown of power plant,several convolution neural network models with different structures are constructed based on big data mining and data-driven technology to predict and monitor the oil level of EH oil tank.Using a large amount of data accumulated in the actual operation of the unit,the model training and verification data set is established,and the model is comprehensively evaluated from the aspects of prediction accuracy,cross entropy loss and operation time.The evaluation results show that the model accuracy of structure a is the highest,about 98.92%,and the cross entropy loss is the lowest,about 0.0431.The operation time of the model is medium,and the comprehensive performance is the best.The intelligent model a is integrated into the distributed control system of the power plant for practical verification.It s monitoring is accurate and early warning is timely,which significantly reduces the workload of supervision and inspection personnel,and improves the automation and intelligence level of power production.The relevant application experience can be used as a reference for similar units.