Research On Hybrid Probabilistic Modeling For Flow Rate Characteristics Of Reciprocating Multiphase Pumps

In the multiphase transportation conditions,the interaction is intricate between the multiphase flows of the pump cavity and irregular movements of the check valves.The flow rates of the reciprocating multiphase pumps show complex properties,including quick and strong nonlinearity,dynamic,time-varying and multi-stage process characteristics.Moreover,the resulting strong flow pulsation will cause serious pressure fluctuation and vibration to decrease the pump stability greatly.Therefore,reducing flow pulsation is one of the problems to be solved urgently.However,the research on the multiphase flow mechanism of the valve gap is still in the exploratory stage.Additionally,the accurate mechanism and numerical models for describing the relationship between the flow rates and multiphase transportation conditions are not available.Consequently,the theoretical and technical basis cannot be provided in a direct and simple manner,for the design of the multiphase pumps to reduce the flow pulsation effectively.Data-driven modeling methods have been widely used to model the time-varying and nonlinear systems.However,limited to the existing measurement techniques of multiphase flow parameters,it is very difficult to improve the prediction performance of the data-driven empirical model by obtaining useful information from a large number of experimental data.Additionally,for complicated multiphase transportation processes,the direct application of a single empirical model may not be enough,mainly because it is difficult to quickly capture important information with limited modeling samples.Therefore,this thesis aims to explore efficient,simple,practical hybrid probabilistic modeling methods using the Gaussian process regression(GPR)model and its probabilistic property,for modeling and prediction of flow rate characteristics of reciprocating multiphase pumps in different multiphase transportation conditions.The main contributions of this thesis are as follows:(1)The superiority of computational fluid dynamics(CFD)transient model and GPR empirical model is integrated to develop a hybrid probabilistic model for the prediction of the flow rate characteristics.The experimental reSsults show that,instead of the time-consuming CFD design process,a better CFD transient model can be selected with a simple and effective method to reduce the reliance on designers'experience.Moreover,a suitable GPR model is constructed using the training data from the selected CFD transient model,for online prediction of a new condition(2)Based on two probabilistic indices,a self-adaption hybrid probabilistic model is proposed to automatically select a suitable model from several local GPR and just-in-time GPR models for each new sample.The experimental results show that,using the essential properties of various GPR models,the proposed method can better describe the complex characteristics of different stages and transitions between stages,for a discharge flow rate curve with li,mited modeling samples(3)The GPR probabilistic information and process knowledge are integrated to construct a hybrid probabilistic model,for individual modeling of a discharge flow rate characteristics.First,the probabilistic information is used to divide the flow rate curve into several stages.Additionally,the process knowledge is integrated for individual modeling with different methods.The experimental results show that,compared with the self-adaption hybrid probabilistic model,the proposed model can better handle the characteristics of the complicated process in a simpler and more efficient manner.(4)Two active learning methods are proposed to sequentially design of a few and significant training data to enhance the prediction performance of a GPR-based flow rate characteristics model.The experimental results show that,compared with the traditional GPR learning method using random selection of the samples,the variance-based and relative variance-based methods can obtain better learning and prediction performance with much less training data.Additionally,the proposed relative variance-based approach explores the dynamical characteristics between the samples and has significant advantages in the feasibility and simplicity for the engineering applications.(5)A performance test system for reciprocating multiphase pumps is designed.It provides reliable modeling data for the proposed hybrid probabilistic model,and verifies the validity of the proposed modeling metlhod.Additionally,the test system can also be used for in-plant performance test and factory inspection of reciprocating multiphase pumps.