Hydrocarbon reservoir model detection from pressure transient data using coupled artificial neural network—Wavelet transform approach

Highlights

• Various reservoir models have been detected using novel coupled MLP-DWT model.

• Our proposed model detected reservoir models from training set with TCA = 95.37%.

• Our coupled model recognized reservoir models from test set with TCA of 94.34%.

• The proposed model has been validated using synthetic and actual noisy real data.

Abstract

Well testing analysis is performed for detecting oil and gas reservoir model and estimating its associated parameters from pressure transient data which are often recorded by pressure down-hole gauges (PDG). The PDGs can record a huge amount of bottom-hole pressure data, limited computer resources for analysis and handling of these noisy data are some of the challenging problems for the PDGs monitoring. Therefore, reducing the number of the recorded data by PDGs to a manageable size is an important step in well test analysis. In the present study, a discrete wavelet transform (DWT) is employed for reducing the amount of long-term reservoir pressure data obtained for eight different reservoir models. Then, a multi-layer perceptron neural network (MLPNN) is developed to recognize reservoir models using the reduced pressure data. The developed algorithm has four steps: (1) generating pressure over time data (2) converting the generated data to log-log pressure derivative (PD) graphs (3) calculating of the multi-level discrete wavelet coefficient (DWC) of the PD graphs and (4) using the approximate wavelet coefficients as the inputs of a MLPNN classifier. Sensitivity analysis confirms that the most accurate reservoir model predictions are obtained by the MLPNN with 17 hidden neurons. The proposed method has been validated using simulated test data and actual field information. The results show that the suggested algorithm is able to identify the correct reservoir models for training and test data sets with total classification accuracies (TCA) of 95.37% and 94.34% respectively.

Introduction

In the recent years pressure down-hole gauges (PDGs) have been widely installed in oil and gas fields during intelligent completions of the production and injection wells. The main objectives of the PDGs installation are continuous and real-time measurements of pressure, temperature, downhole flow rate, and phase fractions over a long period of time for well monitoring and evaluating performance of reservoirs. During this long period of recording time, the data acquire noises and outliers from different sources, or even it may fail to record data at several occasions. These issues along with limited computer resources for handling of these massive and noisy data are some of the challenging problems that can adversely affect the interpretation process of PDG data. Therefore, it becomes imperative to develop a data processing algorithm that can correctly process the PDGs data for further analysis.

Construction of reliable dynamic model which can predict both current and future transient behavior of reservoir is a crucial stage in the optimization and management of productions policy of oil and gas reservoirs. Although a direct identification of these heterogeneous hydrocarbon reservoirs is almost impossible, some indirect techniques such as seismic, well log and well test have been developed to construct a reliable reservoir model. In spite of static description of reservoirs by well log and seismic techniques, well testing can present a dynamic view of these highly heterogeneous media. Since 1937, the well testing is the most widely used tools in the petroleum engineering for identifying hydrocarbon reservoirs [1]. Well testing is basically conducted by creating a flow disturbance in wellbore and monitoring the pressure response at the bottom-hole. By analyzing the recorded pressure signal over time which is obtained from well testing operations, reservoir model and its boundary (formation model) can be identified [2]. Moreover some reservoir parameters such as initial reservoir pressure, average conductivity of matrix and fracture, storativity ratio, interporosity flow coefficient, value of reservoir damage can be estimated using these signals [2], [3], [4]. It should be noted that prior to start the parameter estimation, decision should be made on the formation model. All of the aforementioned parameters and formation model can be evaluated by knowing both pressure and production/injection flow rate over time.

Pressure derivative plot i.e., the log–log presentation of the rate of pressure change with respect to superposition time function is one of the most widely used techniques for detection of reservoir model and its boundaries [5], [6].

Since the focus of the present study has only been put to detecting of reservoir structure and its boundary model, the type curve matching by pressure derivative plot is effectively employed for the considered task. Once the formation model is detected, their various parameters can be estimated utilizing the specific portion of the pressure derivative plot. Various types of flow regimes have been extensively explained in our previous research [5].

Artificial Neural Network (ANN) is one of branches of the artificial intelligence methodologies which have played important roles in many scientific disciplines for replacing traditional analysis by computer aided ones [5], [6], [7], [8]. MLP and recurrent network are among the most frequently used ANNs for interpretation and recognition of complicated pattern in various fields specially in well testing [5], [8], [9], [10], [11]. In our previous researches two different automated models based on MLPNN and recurrent network were developed for detection of eight different oil reservoir models from synthetic PD patterns which have 33 sample points [5], [9]. In 1995, Athichanagorn and Horne used MLPNN for recognizing characteristic parts and their appearance times in pressure derivative plots of some candidate reservoir models [12]. The obtained values from MLPNN are then used as initial guesses in sequential predictive probability method for diagnosing reservoir model and estimating its parameters [12].

In the last two decades, wavelet transform has appeared several times in the petroleum engineering for detecting changes in flow rate, transient identification, de-noising and up-scaling of reservoir properties [13], [14], [15]. Kikani and He used the wavelet transform for data de-noising and transient detection among synthetic pressure transient data [13]. Athichanagorn et al. developed a wavelet based model for pre-processing and interpretation both long-term simulated as well as actual field data [14]. Olsen and Nordtvedt investigated the wavelet transform ability for filtering and noises removal from production data [15]. They also proposed some empirical rules and automatic methods for threshold approximation and reducing the amount of reservoir production data [15].

In the next section, a brief explanation of well testing, discrete wavelet transform, MLP network, the employed procedures for generating pressure transient data and calculating log–log pressure derivative graphs is presented.