Research On Measurement Method Of Oil-well Dynamic Liquid Level Based On Soft Sensor Technology

Oil,as China's main energy pillar industry,plays an important role in the development of the national economy.At present,sucker rod pumping of beam pumping wells is the main oil recovery method in the oilfield,in which the dynamic liquid level data of oil well is an important guiding parameter for the adjustment of the oil recovery process and the optimization of production measures.Generally,artificial echo instrument is used to acquire dynamic liquid level parameter.This method not only is unable to achieve online real-time measurement,but also have problem of large error,high cost,and so on.All of these go against with the development of digitalization of oilfield.In this thesis,the actual production process of oilfield is taken as the research background,and the soft sensor technology is used to realize the online measurement of the dynamic liquid level.The corresponding solutions are proposed for the analysis of the existing problems.The main research contents are as follows:(1)To ensure the accuracy of the modeling data of the multiple condition prediction model of dynamic liquid level,a multiple condition recognition algorithm based on similarity factor analysis is proposed in this thesis.The current production conditions are featured by the characteristics of the window sample data,the similarity factor analysis is used to replace the traditional distance calculation,the improved K-means algorithm is used to cluster the different window sample data,and to correctly classify the sample data of different working conditions.This algorithm overcomes the shortcomings of the traditional clustering method,which is easy to be affected by production fluctuations and abnormal data,resulting in inaccurate data classification and misjudgment of some working conditions,and provides accurate modeling data for the establishment of dynamic liquid level soft sensor models under different working conditions.(2)A static modeling method which based on improved AdaBoost algorithm is proposed in this thesis.This method,according to different types of production conditions,adopted AdaBoost learning algorithm to establish their corresponding integrated models,and finally form a multiple model prediction set.Compared with the conventional BH-LSSVM single model algorithm,the improved AdaBoost integrated learning algorithm can highlight the role of prediction error in modeling sample weights and weak learning machine weight updates,and increase the chances of large error samples entering the training set.At the same time,multiple weak learning machines are used for weighted output to minimize the impact of large individual errors on the overall prediction accuracy of the model,so that the generalization ability of the model is significantly improved and the output is more scientific.(3)In order to improve the adaptability of soft sensor output model to dynamic production process of oilfield,this thesis proposes an adaptive soft sensor modeling method,which based on fuzzy evaluation for dynamic liquid level.A fuzzy expert system for the reasoning trending of liquid change trend is established.The real-time fault diagnosis of dynamic liquid level output is carried out by calculating the goodness of fit between the system prediction value and the actual value.This method is dynamically updated by adopting similar sample data,which enhances the model's ability to adapt to production fluctuations and improves the prediction accuracy of the model.The whole adaptive soft sensor updating modeling strategy includes three stages: off-line modeling,on-line measurement and adaptive updating,and finally constitutes a complete dynamic liquid level measurement system for oil wells.