Soft Sensor Method Research And Implementation For Dynamic Fluid Level Of Pumping Well

To deal with complex measurement tasks and achieve online estimation of hard-to-measure variables,the soft sensor technology is considered to be an effective method.The oil dynamic fluid levels are key parameters in the oilfield production.As it locates thousands of meters underground,it is difficult to be measured,which is considered as a typical hard-to-measure variable.It has great significances for timely updates of the effective oilfield production measurements and improvement of the oil recovery technology by realizing online soft sensor for dynamic fluid level of pumping well.In this paper,we use the practical oilfield production as an application background and introduce t,he soft sensor theories into the measurement of the oil dynamic fluid levels;we have deep researches on perdiction methods of the oil dynamic fluid levels based on the soft sensor modeling,and have some further discussion and experiments when they are applied into the practical application in actual production.Our main contributions can be summarized as follows:(1)In order to ensure the accuracy of the collected data,a sensor fault diagnosis method based on the iterative principal component analysis and squared weighted prediction error is proposed in this paper.The soft sensor modeling uses the auxiliary variables to deduce the main variables;so the data accuracy of the auxiliary variables is important to guarantee the model prediction ability.The oilfield production fields mainly locate in the wilderness with the bad working environments,so the sensors have high failure rate.The proposed method combines the iterative principal component analysis algorithm and the fault identification with SWE index.Compared to the traditional PCA method,this method can overcome the problems of false negative and miscalculation when the production model has dynamic changes.It can improve the sensitivity of the system for the failure through the analysis of faults in different residual spaces.(2)The BH-LSSVM soft sensor modeling method for the dynamic fluid levels of pumping well is proposed.In order to improve the modeling training effectiveness,we analyze the relevant parameters of the oil dynamic fluid levels according to the practical production process to select rational auxiliary variables.The least squares support vector machine method based on the statistical theory is used to establish the soft sensor model,which needs fewer samples and has strong generalization abilities;and the black hole algorithm is used to optimally select the best regularization constant C and kernel parameter ?.This optimal algorithm has the characteristics of high optimization precision,fast convergence rate,unease to fall into local optimum and no premature and deception phenomena.And more importantly,compared with particle swarm optimization algorithm,genetic algorithm,the harmony search algorithm and the other optimal algorithms,this algorithm needs not to set the subjective parameters,which can reduce the bad impacts from the optimization algorithm itself during modeling process.(3)The just-in-time learning strategy based on the subspace similarity is used to update the soft sensor model.As the oil production is a dynamic process,so the selection of the modeling data is very important to aim at the model update in the modeling process.Hence,it is necessary for the soft sensor model to update itself in its working process.The abnormal data in the production process can make the soft sensor model have poor prediction performance as the oil process is complex and affected by various factors.In order to prevent the disturbances from the abnormal data to the statistical properties of the soft sensor model,through the analysis of previous modeling selection methods,we use a period of production data to calculate the similarity of subspaces to improve the accuracy of the modeling samples selection.Two memory parameters are used to judge the similarity degree between the current production condition and the model calculation conditions;and then samples with similar subspace are selected to update the soft sensor model.(4)In order to improve the applicability of the soft sensor output,an adaptive soft sensor method is proposed to correct the model.In order to adapt to the dynamic process of the oilfield production,the empirical mode decomposition method is used to decompose the auxiliary variables to two parts:tendency value and fluctuation value;the tendency value is used to predict the dynamic fluid levels,and the fluctuation value is used to predict the compensating errors,then both of them are added to get the final dynamic fluid levels.The output of the soft sensor model is considered as the output of the "soft" sensor,which is evaluated by the SWE index.The model is dynamically updated by an adaptive mode according to the evaluation result,which can follow the dynamic production process.(5)According to the research works in this paper,the design and field tests of a soft sensor system for oil dynamic fluid levels are completed.The system concludes two main parts:soft system and hardware system.The soft system concludes:data preprocessing module,sensor fault diagnosis module,dynamic fluid levels soft sensor algorithm module,model updating module,etc.;it can achieve the following functions:data backup,sensor fault detection,output of the dynamic fluid levels,real-time monitoring,etc..The hardware system concludes:computers for monitoring and soft sensor,detection devices and process control network,etc.;it can achieve the integration of the electrical,instrumentation and computer system,etc..