Improvement Of Extreme Learning Machine Algorithm And Its Application Research In Pumping Well

With the rapid development of Chinese economy, oil productionstatus is becoming more and more important. It is very important significance for thequality of oil of the entire oil field, the safety production of oil pumping operation,and the comprehensive benefits of oil field to grasp the changes in the productionand the operational status of pumping well. However, for now, the application of oilwell dynamic data is single and the changes in oil production cannot be mastered inadvance in the oil production, which leads to we cannot develop production planseffectively. Suck rod pumping system plays a dominant role in the oil exploitation,but the working conditions of underground is complex and harsh, which leads to thehigher failure rate. Although there are already a lot of pumping well yield predictionand fault diagnosis, such as improvement of BP neural network by LM algorithm(LMBP)、improvement of BP neural network by genetic algorithm (GABP)、RBFneural network、support vector machine (SVM) and so on, which cannot accomplishyield prediction and fault diagnosis very well. Therefore, the oil production forecastmethod based on gray relational analysis combined with improved of extremelearning machine and the rod pumping well fault diagnosis method based on waveletpacket combined with improved extreme learning machine are proposed in this paper.The two methods can quickly and accurately forecast and diagnose monthlyproduction of individual well and the breakdown of rod pumping well.This paper mainly focuses on the following five aspects to carry out research.(1) This section thinks the oil production forecast and its influence factors as theresearch objects to analyze the influence factors of oil production, then using grayrelational analysis to extract the main factors, and giving the detailed steps.(2) Research on fault diagnosis principle of rod pumping well, then making ananalysis the formation of the indicator diagram theory and the working principle ofsucker rod pumping well, then giving the detailed description of the graphical features fault indicator diagram and reasons. And on this basis, we propose thecharacteristic energy extraction method which uses wavelet packet, then giving thedetailed steps which uses three layer wavelet packet to decompose and constructindicator diagram signal.(3) Discussing the extreme learning machine theory, to analyze its origin,advantages and disadvantages. Using wavelet substitutes network’s commonactivation function, structural risk minimization principle is integrated into the model,we propose the improved extreme learning machine (RWELM) theory and specificimplementation methods.(4) Building predictive model with monthly output of single well and faultdiagnosis model of pumping well, then study the selection method of hidden layerneurons and fine tuning parameter.(5) The simulation is carried out based on MATLAB R2010b. Using the actualproduction data to verify monthly output of single well predictive model, andcompared with the forecast results of ELM, LMBP, GABP. Using the collected dataof measured dynamometer to verify the pumping well fault model, and comparedwith diagnostic results of ELM, SVM, RBF.The relative error of single well monthly production forecast is only1.1204%,the fault diagnostic accuracy of the pumping well is96.667%, the experimentalresults show that the improved extreme learning machine has some feasibility onpumping well oil production forecast and fault diagnosis.