Prediction Method For Pumping Well System Efficiency Of Data Driven Based On Deep Learning

With the development of Digital Oil Field,a large number of measured data have been collected in the actual operation of Pumping Well,which have not been fully applied to oil field management at present.Therefore,based on deep learning and a large number of measured data,this dissertation predicts the efficiency of the pumping well system,which is close to the actual value.It can effectively improve the production and management ability of the oil field,and guide the optimization of the system operating parameters,and achieve the goal of increasinand g production and saving energy.Based on the simulation model of rod string longitudinal vibration under different working conditions,the system efficiency simulation model of pumping unit is established.Through simulation calculation,a data set containing 23834 records of the relationship between pumping well system efficiency and influencing factors is established,which lays a foundation for the prediction of pumping well system efficiency based on deep learning.Based on the control variable method,the influence of each influencing factor on the system efficiency is analyzed,and the irrelevant influencing factors are removed.Based on principal component analysis,local linear embedding,random forest,reverse feature elimination and other algorithms,the first 6 ～ 9 influencing factors were extracted as the main feature parameters.According to the deep learning theory,taking the main features extracted by PCA,LLE,RF and RFE as input,the deep neural network prediction model of pumping well system efficiency based on LSTM and CNN is designed.Based on 20,000 records of system efficiency training set data,the neural network model parameters were optimized and the neural network prediction model training was completed.The test results show that RFE-CNN and RF-LSTM have high test accuracy.Based on RFE-CNN and RF-LSTM models,the system efficiency of 3834 recorded test sets of Wells was predicted.The prediction results show that the neural network prediction model of RF-LSTM has high prediction accuracy,and the MAE of test set prediction is 1.6650,the MAPE of test set prediction is 13.57%.among which the MAPE of 57.51% test set is within ±5%,and the MAPE of 77.62% test set is within ±10%.The MAPE of 83.99% test sets is within ±15%.The curve of the simulation value and the predicted value of the system efficiency is highly coincident,which can meet the actual demand of engineering.