Soft-sensor Modeling For Dynamic Fluid Level Of Sucker-rod Pumping Process

Sucker rod pumping of beam pumping wells is the most common form of artificial lift in oilfield of the world.Fluid level of oil well is an important working parameters in the production process,which is difficult to obtain in real time but very important for optimization and control of the production process.Up to now,there is no applied online sensor can detect the parameters directly,which is mostly obtained from artificial temporarily installed echometer in oil field.The above-described method has brought great difficulties for fluid level real-time monitoring and optimal control of oil well operation.Therefore,under the existing technical conditions,how to achieve the pumping well fluid level online testing is a serious problem in oil production.Soft sensor technology is a new method to on-line estimate the parameters which is difficult to measure on line,based on readily available parameters.In this paper,based on the research background of oilfield production process,we introduced soft sensor technology into the on-line detection for fluid level of pumping oil wells,deeply studied on the methods to detect fluid level of pumping oil wells on line and practiced the relevant application in the oil field.The main contents are as follows:(1)According to the basic principles of rod pump production,we briefly analyze the rod pump production process and the significance of fluid level on-line detection for the production process.Then we create a dynamic equation to describe the process of oil leaking into the wellbore in rod pump production process;build a mechanism model described how the three-phase fluid flows into the pump in the working process of deep well pump,which determines the boundary conditions underground to solving spring-mass-damper distributed parameter systems for sucker rod longitudinal vibration;develop the coupled dynamic model for the load of beam pumping unit suspension point;studied the variation of dynamic fluid level in annular during rod pump production process;analyze the integrated complex characteristics of the well production key parameters and production process;find out the factors,which is related to the characteristics or effect dynamic fluid level underground directly;reveal the viable theories and technical approaches for soft sensor of dynamic fluid level;define the auxiliary variables of soft sensor model,which laid a solid foundation for soft sensor modeling.(2)Due to our overall low level of current oil production automation and information,it is hard to get a lot of historical data for soft sensor modeling based on data driven in advance.We put forward engineering methods that take dynamic fluid level observation feature point(or area)obtained directly from dynamometer card on the ground and well working status feature information strongly related to dynamic fluid level as input variables of soft sensor model through the in-depth research and analysis of rod pump oil recovery mechanism.Furthermore,we have established a new dynamic fluid level soft sensor mechanistic model.(3)Considering the development of digitalization construction in Chinese major oil fields gradually,we put forward soft sensor methods based on online Gaussian regression process which is a data-driven soft sensor method applicable to pumping well liquid level real-time detection with higher measuring accuracy.Aiming at the disadvantages such as depend on initial value,easy to fall into local optimum in using conjugate gradient method to calculate the Gaussian process regression hyper-parameters,simulated annealing algorithm is proposed for optimizing the hyper-parameters in the process of soft sensor modeling.In view of the dynamic and time-varying characteristics in sucker rod pump oil recovery process and the requirement for online updating soft sensor model according to the changes of working conditions,we judge whether add the new data to basic training set by calculating the variation of the model in each new input sample.Then we propose to delete old data by calculating the data information gain with forgetting function.In addition,we put forward a kind of incremental learning algorithm to update Gaussian process regression model on line,which makes the model adapted to the changes of well working status in the actual production process and improves the prediction accuracy and generalization ability of the soft sensor model.(4)Further,a dynamic fluid level soft sensor model integrated with multi-source information characteristics is established.Building dynamic fluid level soft sensor model as a single model will lead to long period of training,poor process feature matching and inadequate information fusion,low prediction accuracy and generalization because of the sucker rod pump oil recovery process multi-operation features and demand for complete historical data samples.Moreover,horsehead load,vibration,dynamometer card,electrical parameters and wellhead pressure data contain rich oil well condition information in periodic operation process of beam pumping unit,but there is integrated complex features between the above data such as multi-source,multi-scale,relevance,complementary and redundancy.For these reasons,we establish Gaussian process regression working fluid level integrated soft sensor model through extracting and selecting oil wells multi-source characteristics in frequency with the kernel principal component analysis and least squares support vector machine(SVM)method,realizing the training sample concentration with the improved fuzzy ISODATA algorithm based on the evidence weight optimization fusion method and the fish algorithm optimization combination kernel function.(5)According to actual production engineering and samples from oil field,applying the wireless communication technology,automatic detection and control technology in the real-time online monitoring of oil field production,digital oilfield production monitoring system platform setting the key production parameters automatic acquisition and remote wireless transmission,data storage and sharing,the well condition diagnosis and closed-loop optimization control into an organic whole is built.Finally,we embed the dynamic fluid level soft sensor method in above-mentioned digital oilfield production monitoring system platform,that meets the requirements of oilfield engineering application in practice.