Research On Evidence Driven Condition Early Warning Method With Applications In Power Plant

It is essential and challenging to monitor the equipment or processes in power plants and thus make an early warning for abnormal conditions and diagnosis the possible warning reason,so as to ensure the healthy and stable operation of thermal power units as well as the power grid.However,the inaccuracy and uncertainty involved in equipment's operating condition,the difficulty in modeling complex equipment,and the scarcity of fault samples have made the monitoring and early warning tasks more challenging.In view of these problems,this paper investigated on the evidence-driven condition monitoring,early warning and diagnosis method for thermal power units.The content consists of typical condition mining and representation,condition monitoring and early warning,inversion analysis of the warning reason for potential abnormality.Besides,the practical applications of the condition early warning method on the Yunnan power generation monitoring and early warning platform are also introduced in this paper.The main contributions of this paper can be concluded as follows:A typical condition mining method based on density peaks clustering(DPC)is proposed to find the typical operating conditions of the equipment,and the typical conditions are then described using mass functions under the framework of the Evidence Theory,so as to establish the condition library of the equipment.The clustering method DPC can find the typical operating conditions of the equipment reasonably according to the density distribution characteristics of the historical operating data,and does not need to predetermine the number of clusters.Describing the operating conditions in the form of evidence(mass function)can fully express the inaccuracy and uncertainty involved in the operating condition.However,performing clustering analysis with DPC directly on massive operating data of the equipment usually takes a long time and could be a challenge for ordinary computers.As a consequence,a dynamic density biased sampling(DBS)algorithm is proposed to solve this problem.A typical sample set is firstly sampled from the massive operating data by dynamic DBS.Then the typical operating conditions are found by DPC from the typical sample set.The dynamic DBS algorithm has solved the problem that there is always a deviation between the actual and the expected sampling size in the traditional DBS.In addition,an index P_S that can evaluate the performance of the dynamic DBS algorithm is proposed based on the Adjusted Rand Index(ARI).On the basic of the performance index P_S,some parameters in dynamic DBS algorithm are studied.In view of the scarcity of fault samples,a condition monitoring and early warning method called CMEW-EKNN is proposed based on the modified evidential KNN rule(EKNN).The distance reject option in the EKNN rule is utilized,such that only normal operating data is needed to construct the monitoring and early warning model.To improve the accuracy and robustness of CMEW-EKNN,an adaptive discounting factor?is proposed to make the early warning boundary adaptive to local distribution characteristics of the training samples.Besides,the leave-one-out cross-validation method is adopted to optimize the value of the discounting factor without damaging its adaptive ability,so as to further improve the performance of the early warning method.The CMEW-EKNN method is able to present the evolution processes of the equipment's operating conditions and raise the alarm timely at the origin of a possible abnormality to prevent it from further deterioration.A new distance calculation method for accurate search of k-nearest neighbors is proposed to make the searching process of k-nearest neighbors more accurate,especially when the equipment is in abnormal condition.In the EKNN based condition monitoring and early warning method,the searching process of the k-nearest neighbors(KNN)is determined according to the distances between the sample to be monitored and the typical condition samples of the equipment in the condition library.Consequently,the abnormality of one or more operating variables will lead to deviations in the KNN searching process.In fact,the operating variables of the equipment can usually be divided into two types,namely input variables and output variables.And the real abnormal condition of the equipment is usually reflected on the abnormal deviation of one or more output variables,while has almost no influence on those input variables.Based on this,a new distance,calculated as the weighted sum of the Euclid distances of the input variables and the output variables,is utilized to search the k-nearest neighbors,which makes the KNN searching process more accurate especially when some abnormality happens.At the same time,the KNN searching deviation caused by inertia and delay between the input and output of the equipment can also be reduced.Based on k-nearest neighbor(KNN)residuals,an inversion analysis method regarding early warning reasons for abnormal conditions of the equipment is proposed.According to the core idea of KNN,the typical operating samples in the condition library of the equipment are treated as the reference set to generate KNN residuals.In this case,the difficulty of mechanism modeling in the model-based residual generation method can be avoided.Then the generated KNN residuals are combined through evidence combination rule,so as to locate the abnormal operating variables of the equipment.Based on the located abnormal variables,we can determine whether a real abnormal condition or a new condition is going to happen when the warning is triggered.For real abnormalities,the possible causes can be analyzed through evaluating the developing direction of the KNN residuals of the abnormal variables,in combination with expert knowledge or experience.The causes obtained can provide guidance for the maintenance of the equipment.The simulation data regarding various types of abnormal conditions including abrupt and slowly abnormalities,such as leakage,fouling,etc.of a high pressure heater and a condenser is used to verify the effectiveness of the proposed condition early warning method in this paper.Though some actual field anomaly data has already been used to test the performance of the condition warning method,it is still not enough.In view of the scarcity of abnormal and fault samples,dynamic simulation models of a high pressure heater as well as a condenser from Taizhou No.2 Power Plant are established.Based on the simulation models,full load normal operating conditions and different kinds of abnormal conditions of the two objects are simulated to get the test data.Then,typical condition library of the equipment is established using the normal operating data.Based on the condition library,the proposed condition early warning method is used to detect and diagnosis the simulated faults to see if all the faults can be detected and diagnosed accurately.The results have verified the ability of the evidence driven condition early warning method in detecting and diagnosing different types of abnormal conditions.