Remote Monitoring of Well Production Performance Based on Machine Learning

Abstract

Abstract A set of dynamic remote monitoring method of production performance based on Machine Learning is proposed for the production process of electric submersible pump (ESP) well with multi-dimensional parameters. Aiming at dealing with the characteristics of multi-dimensional parameters in the complex production system, to implement dynamic monitoring of production performance. Helping the engineers at data centers to find the anomaly remotely and make response in a timely manner. It puts forward a procedure for large amount, high dimension and low information density production data in complex production system, using the dimensionality reduction algorithm to reduce the dimensionality into one comprehensive parameter changing over time, time series analysis algorithm for the production anomaly detection and prediction based on Machine Learning. The Principal Component Analysis (PCA) is used to reduce the dimensionality and extract the crucial information. The Autoregressive Integrated Moving Average (ARIMA) model is used to conduct timing anomaly detection, and fbProphet model is used to analyze the dimensionality reduced data to provide prediction of the production. With the dimensionality reduction, time series comprehensive parameter analysis and anomaly detection method based on Machine Learning, more than 40 ESP wells with 15 dimensions production daily parameters up to 1,000 days were analyzed, which realized the comprehensive description of ESP wells with multiparameter. Although the PCA retained only 47.73% of the information in the first principal component, which may be related with the low information density of industrial big data, the subsequent analysis proved the effectiveness. The time series analysis realized many times anomaly detection during the life period of each ESP well, and visualized the production data and the anomalous events. More than 100 anomalous events were detected in advance and which were robust corresponding to the subsequence real production events, among which 95% agreement rate is achieved. The procedure proposed reported the anomaly events with high confidence up to 90%, and low misstatement rate and omission rate, realized the production perception and abnormal detection in a timely manner. Based on this algorithm, the best time for the well intervention is determined, so that the loss of production is avoided and the revenue is maximized. The novelty of the procedure of Machine Learning using the multiple production data is in the ability to provide a solution of dealing with the low information density and high noise in the complex multi- dimensional production data of production wells, realize the comprehensive description, analysis and prediction of the production. It is helpful for engineers find the abnormalities in time, and will support the decision making of production, optimization and well intervention for the production.