Hybrid Long Short-Term Memory and Convolutional Neural Network Architecture for Electric Submersible Pump Condition Prediction and Diagnosis

Abstract

Electric submersible pump (ESP) is one of the common artificial lift technologies in offshore production systems. ESP failures are the main cause of the decline in the production efficiency of oil wells. Early warning and diagnosis of ESP failures are crucial to improve well production efficiency. In this study, a hybrid model of long short-term memory neural network and convolutional neural network (LSTM-CNN) for accurate early warning and diagnosis of ESP faults is proposed, based on electrical data as the basis of analysis. Using hyper-parameters to optimize the LSTM neural network structure and highly fit the field electrical data so that it can be applied to anomaly prediction before ESP faults, the results show that the optimized LSTM model with R2 (test set) = 0.79, root mean square error (RMSE) (test set) = 0.89, which can predict the future electrical data more accurately, and the predicted data are plotted in the polar coordinate system to simulate the ammeter card as the validation set. The ammeter card data set is simulated by collecting electrical data from different working conditions, which is expanded using data expansion, and different CNN models are trained to fine-tune the parameters using transfer learning. The results show that the GoogLeNet model has a significant diagnostic accuracy of 97%, which is 2% and 18% better than VGG16 and ResNet34, respectively. The model evaluation shows that the GoogLeNet model has good recall, F1-Score, and confusion matrix. The trained GoogLeNet model was applied to the validation set, and the hybrid LSTM-GoogLeNet model was found to be well-suited for ESP warning and diagnosis.