The rapid development of machine learning algorithms and the massive accumulation of well data from continuous monitoring has enabled new applications in the oil and gas industries. Data gathered from well sensors are a foundation of the oilfield digitization and data-driven analysis. Here, we describe a deep learning approach to predict the long-term well performance based on a moderate duration of well monitoring data.
In this study, we first developed the data processing procedures for oilfield time series data and determined the proper selection of data sampling frequency, parameter combinations and data structures for deep learning models. Then we explored how Deep Learning (DL) models can be employed for well data analysis and how can we combine physics and DL models. Recurrent Neural Network (RNN) is a type of sequential DL model, which can be utilized for time series data analysis. This approach preserves preceding information and yields current response with memory of prior well behavior. Two candidate RNN models were tried to determine how well they were able to improve the accuracy and stability of well performance estimates. These two methods are Gated Recurrent Unit (GRU) and Long Short Term Memory (LSTM). In addition, a novel combination of RNN with Convolutional Neural Networks (CNNs), Long- and Short-term Time-series network (LSTNet), was also investigated.
These various models were tested and compared based on the public production datasets from Volve Field. Both GRU and LSTM achieved higher accuracy in performance prediction compared to the simple RNN. In the case of frequent well shut-in and opening, the failure in capturing fast pressure responses and the extreme fluctuations with the simple RNN ultimately leads to high error. In contrast, LSTNet is more stable to frequent or significant well variations. With advanced deep learning structures, engineers can interpret long-term reservoir performance information from responses estimated by deep learning models, instead of performing costly well tests or shut-ins.
For most of the mature fields, the oil well operation and maintenance expenditures continue to put financial pressure on the operators in the low oil price period. Digital oilfields and artificial intelligent technology are the major areas invested to fight for declining oil production and increasing cost. This paper provides a novel artificial intelligent method to monitoring and diagnose the sucker-rod pumping wells using deep learning algorithms.
Traditional method using load and displacement sensors to measure the dynamometer card needs large investment on the equipment installation and maintenance. We build a general model that generates the dynamometer card from electrical parameter using state-of-art deep learning algorithms. The deep learning algorithms can analyze the relationships between the electrical data and corresponding dynamometer card in different conditions, which is very hard for human being to detect. In addition, we build another automated diagnosis deep learning model from thousands of dynamometer cards labeled with different classifications.
We have already tested these newly developed artificial intelligent models on hundreds of sucker rod pumped wells in different oilfields in PetroChina. The field test results show that the dynamometer cards generated from electrical data have above 90% similarity compared to the real dynamometer, which meet the requirement for well diagnosis. The card generation model is stable and prevents the disturbance of hostile environment change and sensor failures. The automated diagnosis model also proved to be a good substitute to the conventional software, with above 95% prediction accuracy. The automated diagnosis model reduces the liability and uncertainty of traditional diagnosis software and can integrated with the former dynamometer card generation model to fulfill well monitoring and diagnosis automatically without any physical model based calculations.
These models developed with artificial intelligent technology will be important components in the "Intelligent Fields". They can also be embedded in the IIoT edge computing machines for automatic diagnosis and control. For ultra-low production wells and the newly producing wells utilized this method, operator can save expenditure and human resources tremendously.
In this paper, we tackle an old problem - production forecast - using techniques that are relatively new to the reservoir engineer toolbox. The problem at hand consists of forecasting oil production in a mature onshore field simultaneously driven by water and steam injection. However, instead of turning to traditional methods, we deploy machine-learning algorithms which will feed on a plethora of historical data to extract hidden patterns and underlying relationships with a view to forecasting oil rate. No geological model and/or numerical reservoir simulators will be needed, only 3 sets of time-series: injection history, production history and number of producers. Two Machine-Learning algorithms are used: linear-regression and recurrent neural networks.
Downhole fluid sampling is ubiquitous during exploration and appraisal because formation fluid properties have a strong impact on field development decisions. Efficient planning of sampling operations and interpretation of obtained data require a model-based approach. We present a framework for forward and inverse modeling of filtrate contamination cleanup during fluid sampling. The framework consists of a deep learning (DL) proxy forward model coupled with a Markov Chain Monte Carlo (MCMC) approach for the inverse model.
The DL forward model is trained using precomputed numerical simulations of immiscible filtrate cleanup over a wide range of in situ conditions. The forward model consists of a multilayer neural network with both recurrent and linear layers, where inputs are defined by a combination of reservoir and fluid properties. A model training and selection process is presented, including network depth and layer size impact assessment. The inverse framework consists of an MCMC algorithm that stochastically explores the solution space using the likelihood of the observed data computed as the mismatch between the observations and the model predictions.
The developed DL forward model achieved up to 50% increased accuracy compared with prior proxy models based on Gaussian process regression. Additionally, the new approach reduced the memory footprint by a factor of ten. The same model architecture and training process proved applicable to multiple sampling probe geometries without compromising performance. These attributes, combined with the speed of the model, enabled its use in real-time inversion applications. Furthermore, the DL forward model is amendable to incremental improvements if new training data becomes available.
Flowline measurements acquired during cleanup and sampling hold valuable information about formation and fluid properties that may be uncovered through an inversion process. Using measurements of water cut and pressure, the MCMC inverse model achieved 93% less calls to the forward model compared to conventional gradient-based optimization along with comparable quality of history matches. Moreover, by obtaining estimates of the full posterior parameter distributions, the presented model enables more robust uncertainty quantification.
Gupta, Harshit (Indian Institute of Technology Delhi) | Pradhan, Siddhant (Indian Institute of Technology Delhi) | Gogia, Rahul (Indian Institute of Technology Delhi) | Srirangarajan, Seshan (Indian Institute of Technology Delhi) | Phirani, Jyoti (Indian Institute of Technology Delhi) | Ranu, Sayan (Indian Institute of Technology Delhi)
Horizons in a seismic image are geologically signficant surfaces that can be used for understanding geological structures and stratigraphy models. However, horizon tracking in seismic data is a time consuming and challenging task. Saving geologist's time from this seismic interpretation task is essential given the time constraints for the decision making in the oil & gas industry. We take advantage of the deep convolutional neural networks (CNN) to track the horizons directly from the seismic images. We propose a novel automatic seismic horizon tracking method that can reduce the time needed for interpretation, as well as increase the accuracy for the geologists. We show the performance comparison of the proposed CNN model for different training data set sizes and different methods of balancing the classes.
Krutko, Vladislav (Gazpromneft Scientific and Technical Center) | Belozerov, Boris (Gazpromneft Scientific and Technical Center) | Budennyy, Semyon (Moscow Institute of Physics and Technology, Center for Engineering and Technology) | Sadikhov, Emin (Moscow Institute of Physics and Technology, Center for Engineering and Technology) | Kuzmina, Olga (Moscow Institute of Physics and Technology, Center for Engineering and Technology) | Orlov, Denis (Skolkovo Institute of Science and Technology) | Muravleva, Ekaterina (Skolkovo Institute of Science and Technology) | Koroteev, Dmitri (Skolkovo Institute of Science and Technology)
A framework for porous media topology reconstruction from petrographic thin sections for clastic rocks is proposed. The framework is based on two sequential stages: segmentation of thin sections imagesinto grains, porous media, cement (with further mineralogical classification of segmented elements) and reconstructing a three-dimensional voxel model of rock at pore scale.
The framework exploits machine learning algorithms in order to segment2D thin section images, perform structural and mineralogical classification of grains, cement, pore space, and reconstruct 3D models of porous media. Segmentation of petrographic thin section images and mineral classification of the segmented objects are performed by the means of combination of image processing methods and Convolutional Neural Networks (CNNs). The 3D porous media reconstruction is done by means of the Generative Adversarial Networks (GANs) are applied to the segmented and classified 2D images of thin sections.
As the criteria of the reconstruction quality, the following metrics were numerically calculated and compared for original and reconstructed synthetic 3D models of porous rocks: Minkowski functionals (porosity, surface area, mean breadth, Euler characteristic) and absolute permeability. Absolute permeability was calculated using pore network model. The 3D reconstruction framework was tested on a set of thin sections and CT tomograms of the clastic samples from the Achimovskiy formation (Western Siberia). The results showed the validity of the goodness-of-fit metrics based on Minkowski functionals for reconstruction the topology of porous media. The combined usage of CNN and GAN allowed to create a robust 3D topology reconstruction framework. The calculated poroperm characteristics are correlated with laboratory measurements of porosity and permeability.
The developed algorithms of automatic feature extraction from petrographic thin sections and 3D reconstruction based on these features allow to achieve the following goals. First is the reduction of the amount of the routine work done by an expert during petrographic analysis. Second leads to the reduction of the number of expensive and time-consuming CT scannings required for each physical sample in order to perform further absolute and relative permeability calculations. The proposed method can bring the petrographic thin section and CT data analysis to a new level and significantly change traditional core experiments workflow in terms of speed, data integration and rock sample preparation.
Severe drilling dynamics of a bottomhole assembly (BHA) causes energy to dissipate into vibrations which undermines drilling efficiency. Dangerous dynamics modes, such as backward whirling and high frequency torsional oscillation, could cause downhole drilling tools to fail prematurely. To mitigate the risk of failure due to these dangerous conditions, it is critical to identify the damaging dynamics modes by interpreting the drilling data. Based on a deep learning approach, a novel method was proposed to automatically identify severe drilling dynamics modes directly from the time-series data.
The drilling dynamics data can be obtained from either a downhole sensor measurement or transient dynamics simulation. First, a deep neural network, which is composed of convolutional and fully connected layers, is employed to explore patterns in the data by generating a feature map of drilling dynamics. Knowledge of drilling dynamics physics can be used to facilitate data clustering in the feature map. Each data cluster can be tagged with the corresponding drilling dynamics mode. Using the tagged dataset, a machine learning classification model can be trained to automatically identify the dynamics modes based on the input of time-series drilling data.
The deep learning approach can be implemented to recognize a collection of dynamics modes of BHA, such as various whirling patterns and high frequency torsional resonance. The most commonly available drilling dynamics data channels, accelerations and collar RPM, were used as the model inputs. The deep neural network was trained to predict the next data sample based on the previous time-series data. One of the hidden layers of the neural network was employed to generate the feature map, in which the dataset forms several clusters. The orbits of BHA movement were plotted on top of the clusters for pattern visualization. After this practice, the simple polygon boundary was drawn between whirling and stable cases, and the dataset was tagged automatically. With the tagged dataset, the classification model was trained to identify various whirling patterns and the stable drilling state. Similar processes can be readily applied to interpret other dynamics modes. Interpreting the drilling dynamics modes provided a high-level description of the data, which offered clues on how to optimize BHA design and drilling practices to improve efficiency.
The automatic interpretation of drilling dynamics data can significantly improve the consistency and efficiency of the existing manual interpretation workflow. The generated feature map enables the exploration of new motion patterns and new vibration modes. This approach eliminates the need to manually tag the data. With minimum human interactions, the dataset can be automatically tagged. The model employs only the raw time series data of basic dynamics channels as inputs, which makes the algorithm universally applicable for various data sources.
A new real-time machine learning model has been developed based on the deep recurrent neural network (DRNN) model for performing step-down analysis during the hydraulic fracturing process. During a stage of the stimulation process, fluids are inserted at the top of the wellhead, while the flow is primarily driven by the difference between the bottomhole pressure (BHP) and reservoir pressure. The major physics and engineering aspects involved are complex and, quite often, there is a high level of uncertainty related to the accuracy of the measured data, as well as intrinsic noise. Consequently, using a machine learning-based method that can resolve both the temporal and spatial non-linear variations has advantages over a pure engineering model.
The approach followed provides a long short-term memory (LSTM) network-based methodology to predict BHP and temperature in a fracturing job, considering all commonly known surface variables. The surface pumping data consists of real-time data captured within each stage, such as surface treating pressure, fluid pumping rate, and proppant rate. The accurate prediction of a response variable, such as BHP, is important because it provides the basis for decisions made in several well treatment applications, such as hydraulic fracturing and matrix acidizing, to ensure success.
Limitations of the currently available modeling methods include low resolution BHP predictions and an inability to properly capture non-linear effects in the BHP/temperature time series relationship with other variables, including surface pressure, flow rate, and proppant rate. In addition, current methods are further limited by lack of accuracy in the models for fluid properties; the response of the important sub-surface variables strongly depends on the modeled fluid properties.
The novel model presented in this paper uses a deep learning neural network model to predict the BHP and temperature, based on surface pressure, flow rate, and proppant rate. This is the first attempt to predict response variables, such as BHP and temperature, in real time during a pumping stage, using a memory-preserving recurrent neural network (RNN) variant, such as LSTM. The results show that the LSTM can successfully model the BHP and temperature in a hydraulic fracturing process. The BHP and temperature predictions obtained were within 5% relative error. The current effort to model BHP can be used for step-down analysis in real time, thereby providing an accurate representation of the subsurface conditions in the wellbore and in the reservoir. The new method described in this paper avoids the need to manage the complex physics of the present methods; it provides a robust, stable, and accurate numerical solution throughout the pumping stages. The method described in this paper is extended to manage step-down analysis using surface-measured variables to predict perforation and tortuosity friction.
The course introduction will attempt to answer the question: How will A.I. change the way we work in the Oil and Gas industry in the coming years? Looking at what is underway in other industries and guessing what type of projects are under development in R&D departments in our industry will help answer that question. Oil and Gas examples will be presented corresponding to each of the terms A.I., Machine Learning, and Deep Learning, allowing participants to reach a clear understanding on how they differ. The course will then focus on Deep Learning (DL) and address all key aspects of developing and applying the technology to Oil and Gas projects. What is DL and how different is it from traditional neural networks?
At times, it may seem that machine learning can be performed without a sound statistical background, but this does not take in to account many difficult nuances. Code written to make machine learning easier does not negate the need for an in-depth understanding of the problem. Nicknamed “warshipping,” the hacking technique allows remote infiltration of corporate networks by hiding a remote-controlled scanning device designed to penetrate a wireless network inside a package.