Digital oilfield

Machine-Learning-Based Early-Warning System Maintains Stable Production

This paper describes an accurate, three-step, machine-learning-based early warning system that has been used to monitor production and guide strategy in the Shengli field.

jpt-2020-03-197365hero-526701021.jpg

After long-term waterflooding, a number of mature oil fields in China have entered the high-water-cut stage, and abnormal production decline has become the primary problem for stable production. This paper describes an accurate, three-step, machine-learning-based early warning system (EWS) that has been used to monitor production and guide strategy in the Shengli field. Adding artificial samples to the training process improved the system’s prediction accuracy greatly (Fig. 1).

2020-03-197365f1.jpg
Fig. 1—The work flow to build an EWS model based on machine learning.

Introduction

For conventional Chinese oil fields that have entered the high-water-cut stage after decades of waterflooding, stabilizing production has become increasingly difficult. After stimulation treatments throughout the field’s history, abnormal decline rates—that is, exceeding 5%—occurred more frequently. Production declined dramatically in 2004 and has not been maintained since then.

To prevent future abnormal production declines, an effective EWS was needed that could release a production alarm to enable engineers to take preventive measures in advance. The complete paper includes a discussion of various early-warning models and their limitations.

The paper discusses an EWS based on a neural network method using a previously established data set. Factors that can affect the abnormal decline were selected. The index sets of production composition and injected and produced water obtained from practical statistics were considered as the main assessment indicators. Grey relational analysis was used to evaluate the importance of the different indicators and to eliminate redundant parameters.

Machine learning was adopted to build the EWS. Using the degree of deviation from normal as the input data for the prediction model provided the highest accuracy. However, the basic machine-learning model contains many input parameters that cannot be obtained easily. The number of input parameters was optimized on the basis of the variation of accuracy under different input parameter numbers. To improve prediction accuracy, artificial samples were added to the training process.

The prediction accuracy of the final optimization model can reach 92%. The result reveals the possibility of the occurrence of anomalous decline in different reservoirs, which can guide oilfield production strategy effectively. The EWS was verified by oilfield production.

Fundamentals of Neural-Network Classification

Neural networks use existing data to determine the implicit model between input and output data. Classification and prediction are the most common applications. Neural networks are used in the industry to solve classification and regression problems.

The back-propagation (BP) neural network is a multilayer feed-forward network that corrects the connection weight of each layer from back to front according to the difference between the actual output and the expected output. The basic model consists of one input layer, one output layer, and several hidden layers. It can depict the nonlinear relationship between input variables and output variables. By training multiple samples, it obtains the connection weight and other information of each layer as knowledge storage to predict new samples. The principle of applying BP neural-network technology to the comprehensive evaluation of certain problems is to take the evaluation index system as the input vector and the quantity value representing the corresponding comprehensive evaluation as the output vector. First, the network is trained by using some samples that have achieved success through traditional comprehensive evaluation. After training and learning, the weight of each indicator can be expressed correctly, and the trained neural network can be used as an effective tool for comprehensive evaluation.

For the three-step EWS described in this paper, the warning alert is divided into three levels on the basis of a decline-rate variation. A decline rate of less than 3% is considered normal condition. A decline rate between 3 and 5% signals a medium-warning alert. A decline rate of more than 5% issues a high-warning alert. The complete paper illustrates a neural-network model.

Establishment of an Early-Warning Indicator System

The complete paper presents a detailed discussion of the basic indicator system and the main features used in the model. According to the authors, selecting these features appropriately is the most important step in ensuring the effectiveness of the machine-learning method. For the paper, the authors extracted 213 sets of historical data from 12 units in the Shengli oil field from 1996 to 2013, and considered 10 parameters. Grey relational analysis was used to identify the correlation between decline rate and the input parameter.

Compared with traditional multifactor correlation and regression-analysis methods, grey correlational analysis requires fewer data and calculations, making it easier to use widely. The paper outlines the steps in the analysis process and includes numerous calculations. Correlation degree and rank among production decline and various influencing factors are presented in tabular form.

Maximizing Prediction Accuracy and Field Verification

The paper details the steps taken to optimize data preprocessing to maximize the model’s prediction accuracy. The discussion addresses preprocessing of input features and input data by typical normalization methods, by degree of membership, and by degree of ­deviation from normal, using charts, tables, and equations.

The paper also addresses feature numbers optimization and the use of reservoir simulation to add artificial samples to the machine-training process.

Finally, the authors summarize verification of the EWS forecast results in two blocks in the Shengli field.

Conclusions

  • The main factor that causes abnormal oil production decline is the production composition. To obtain stable production each year, the oil production proportion of new wells and stimulation measures should be in a reasonable range.
  • For the machine-learning method, the appropriate data input form has a great influence on final prediction accuracy. In the EWS model, the preprocess method of the degree of deviation from normal is the most suitable method.
  • The number of features input into the model can be optimized to reduce the training time and avoid the overfitting problems without lower prediction accuracy. The grey relational analysis could provide guidance for the correlation degree between target and features.
  • The sample size of different warning levels has a strong influence on accuracy. For prediction problems, adding artificial samples can be effective in overcoming the problem and achieving high accuracy.

This article, written by JPT Technology Editor Judy Feder, contains highlights of paper SPE 197365, “A Novel Early Warning System of Oil Production Based on Machine Learning,” by Kang Ma, Hanqiao Jiang, and Junjian Li, China University of Petroleum-Beijing, et al., prepared for the 2019 Abu Dhabi International Petroleum Exhibition and Conference, Abu Dhabi, 11–14 November. The paper has not been peer reviewed.