Recent Development of Smart Field Deployment for Mature Waterflood Reservoirs

Abstract

In the petroleum industry, artificial intelligence has been applied in seismic and logging interpretation, accurate modeling, optimized drilling operations, well dynamics prediction, safety warning, etc. However, field-scale application and deployment remain a challenge due to the lack of sufficiently powerful algorithms for the integration of multi-granularity data in the time and space domain, for the construction of a deep-learning network able to represent the evolution of well and reservoir dynamics, and finally the lack of investment in surveillance data acquisition. This paper offers a concise review of smart field deployment for mature waterflood reservoirs, including the current status of data foundation construction, and the research progress for applied AI algorithms, as well as application scenarios and overall deployment. With respect to data, the domestic and international oil and gas industry has completed or at least started the large-scale construction and deployment of lake data. However, the data isolation phenomenon is serious in China. Preparation for the integration of new monitoring data for the overall research of reservoirs is insufficient. With respect to algorithms, data-based and model-based AI algorithms have been emerging recently, but the development of the overall proxy model for rapid prediction and automatic model calibration is still in the preliminary period. For application scenarios, relatively simple and independent applications related to geophysical interpretation and production engineering are continuing to emerge, while large-scale reservoir and field application require substantial investment in data acquisition, game-changing algorithms with cloud-based computing architecture, and top-down deployment.

1. Introduction

According to statistical analysis of the Information Energy Agency (IEA), up until 2050, global primary energy will still be in high demand, and the leading status of oil and natural gas will not fundamentally change. As new, easily recoverable oil and gas reserves are becoming scarce, the vast majority of global oil and gas production is sourced from tapping the potential of old oil and gas fields or the development of complex unconventional oil and gas fields. However, traditional reservoir engineering methods have encountered challenges in many aspects of reservoir analysis. For example, processing, cleaning, and analyzing raw field data can be laborious and computationally intensive. In most cases, reservoir engineers must follow a trial-and-error protocol to translate seismic survey data, well logs, core analysis, production history, and other data into a more practical engineering format [1]. With the development of artificial intelligence and big data, the application of artificial intelligence has allowed the development of traditional oil and gas fields to gradually enter the era of intelligent, digital oil and gas fields, which greatly improves the intelligence degree of these fields, not only helping to reduce the cost of oil and gas field development, but also improving the quality and efficiency. Big data and artificial intelligence will certainly be more widely used in the development of oil and gas fields. In terms of tapping the potential of mature oilfields, as development progresses, many oilfields in China are in the “double high period” of high recovery degree and high water cut (the average water content is 87.7% and the average recovery percentage of recoverable reserves is 74.7%), and major oilfields such as Daqing and Shengli are even in the ultra-high water cut period, as shown in Figure 1. Optimization production in mature fields is a complex problem that is difficult to solve with current subsurface modeling solutions. On the one hand, the traditional 3D numerical reservoir solver has high fidelity for the physical of multiphase porous media flows and has the advantage of long-term prediction accuracy. However, complex and time-consuming workflows mean that models and predictions are immediately out of date. In addition, the computational cost of simulation means that it is virtually impossible to run thousands of scenarios to quantify the uncertainty [2]. The main challenges include: (i) the highly dispersed distribution of the remaining oil; (ii) the complex relationship between oil and water; and (iii) the difficulty of stabilizing oil production and controlling water [3]. In terms of development and production, on the one hand, the development of long well sections leads to large differences in production between vertical layers; on the other hand, the casings of old wells are seriously damaged and deformed, making it difficult to form a complete injection–production pattern on the plane [4]. In addition, due to the high heterogeneity of oilfields, after long-term water injection and flushing, a preferential seepage channel forms, which results in ineffective injection of water and inefficient circulation. As a result, these effects restrict oilfield development and further increase the difficulty of exploiting recoverable reserves. There are many methods to improve the production of oilfields by evaluating and optimizing the water injection efficiency, such as using the CRM (capacitance resistance model) to characterize the dynamic connectivity of the injection and production wells and analyze the influence of injection and production imbalance on oil production, changing the flow direction of oil, subdividing the injection and production unit, and carrying out water shut-off to improve the areal sweeping efficiency [5,6,7,8]. For more complex injection–production relationships, with frequent changes in displacement performance and ineffective water circulation, it is of great significance to carry out fine-detail and intelligent reservoir analysis to improve the development efficiency and oil recovery of mature oilfields. Considering that current mature oilfields in the high water-cut period urgently require new technology to promote layer water injection performance, the fourth-generation separated-layer water injection technology has been developed for mature oilfields. The core of this technology is to monitor and adjust parameters automatically in the whole separated–layered injection process [9].

For unconventional oil and gas resources, the US accomplished the efficient development of large-scale tight and shale oil in the early 2000s. Shale oil and gas investment in the United States far exceeded conventional oil and gas investment during the last 10 years, and has become the investment focus of large and medium-sized oil companies. In 2021, the United States announced a total of USD 66 billion worth of shale oil and gas asset transactions. In January 2022, Chesapeake announced that it would spend USD 2.6 billion to acquire Chief Oil and Gas Company, which is engaged in shale oil and gas production [10]. In recent years, China has also achieved large-scale commercial production of unconventional oil and gas in the Sichuan basin and the Mahu sag in Xinjiang. By the end of 2020, China’s tight and shale oil had developed rapidly, with a cumulative productivity of 450 × 104 t/a, and an annual output of about 200 × 104 t, with many tight and shale oil development demonstration projects having been implemented in Junggar, Songliao, Ordos and other basins, and a number of horizontal wells having achieved high production, demonstrating good resources and development prospects [11]. With the increase in the number of wells, on the one hand, a large amount of geophysical, geological interpretation and engineering data were rapidly accumulated; on the other hand, unconventional reservoirs are highly heterogeneous, making single-well production more sensitive to fracturing operation conditions and engineering designed parameters. Furthermore, considering the high complexity of the flow in the nanoscale media and the coupled process of rock mechanics, reliable reservoir simulation and accurate dynamic prediction are extremely difficult. Therefore, through the new generation of intelligent algorithms, using the massive data generated in the unconventional oil and gas exploration and development process to achieve rapid and accurate analysis and prediction is a significant undertaking, with the aim of reducing the costs and improving the production of unconventional oil and gas.

With global industrial technology developing rapidly in the direction of digitalization and information, the application of artificial intelligence (AI) in oil and gas exploration and development has become a hot topic and development trend, and has been applied in almost every aspect of the petroleum industry in recent years, including well testing, exploration, drilling and completion, reservoir engineering, and surface engineering [12].

With respect to sedimentary reservoirs, some scholars have begun to use intelligent analysis of core images to realize precise quantitative research, which has achieved a certain effect in industrial application and shown great application potential. Eberhart and Kennedy proposed the particle swarm optimization (PSO) technique based on the natural model of flocking birds or fish. The PSO algorithm starts from a randomly generated population, uses a fitness function value to evaluate the population, and uses random technology to update populations and search. Recently, PSO has been successfully applied in many fields, such as continuous function optimization, neural network training [13], static function optimization, dynamic function optimization, multimodal optimization and data clustering. However, its application in the petroleum industry can be traced back to about six years ago. To optimize the recovery factor of heavy oil reservoirs, Wang and Qiu showed the convergence behavior and performance of three different PSO algorithms [14]. The results show that the conventional PSO algorithm produced the highest objective function. Humphries et al. linked PSO and generalized pattern search (GPS) in a synchronous and sequential manner to optimize well layout and control problems [15].

In logging, some petroleum enterprises and research institutes have adopted AI technologies such as machine learning and deep learning in curve reconstruction, lithology identification, reservoir parameter prediction, oil, gas and water layer identification, intelligent stratification, imaging logging and other aspects of exploration, research, and preliminary application [16,17,18,19]. Alireza Moazzeni et al. proposed that real-time drilling data can be used as input data of 3-layer neural networks to predict formation type and lithology [20]. J. Lowell et al. proposed that the AI technology can be used to accurately delineate and interpret the fault network in oil and gas reservoirs [21]. In geophysical exploration, seismic data processing and interpretation, such as structural interpretation, seismic phase recognition, seismic wave field forward modeling, seismic inversion, initial arrival pickup, seismic data reconstruction and interpolation, and seismic attribute analysis, are realized using computer vision technology such as target detection, image segmentation and image classification [22,23,24,25,26,27]. In drilling and completion, the application of AI is mainly reflected in the intelligent optimization of wellbore trajectory, intelligent steering drilling, intelligent optimization of drilling rate, etc. [28,29,30]. In reservoir engineering, the data of fine separated-layer water injection [31] are used to realize the intelligent water injection optimization of oil and gas wells [32], which greatly improves the recovery efficiency. In addition, the application of production forecasting and intelligent optimization of production measures based on cyclic neural networks has also achieved preliminary results.

With respect to waterflooding, guiding operational decisions about injection location and injection volume can be quite challenging. Understanding the sub-surface connections between injectors and productors is what makes the process challenging, and improving the efficiency between them is even more so [33]. Shahvali et al. proposed an alternative simplified approach to determine inter-well connectivity by solving the steady-state tracer equation. This method uses the finite volume method to calculate the flux support allocation from the injector to the producer, but returns information similar to the streamline method [34]. However, the amount of parameters involved in the simulation and the large degree of freedom in the reservoir model result in the high computational cost of the whole reservoir simulation process [35,36,37]. As a mathematical or statistical regression function, the surrogate model based on artificial intelligence can replicate the output of the simulation model with selected input parameters with high computational efficiency and accuracy, and is widely used in numerical simulation [38]. The primary application of this proxy model is the uncertainty qualification of performance forecasts of reservoir production, history matching, production optimization, well schedule optimization, and probabilistic forecasting of oil recovery [39,40,41]. Shams et al. used the artificial neural network (ANN) to predict the recovery performance of a layered reservoir waterflooded by five-spot pattern [42].

Petroleum engineering is acquiring more and diverse data faster than ever before, which, in addition to data collected by means of tens of thousands of sensors, also include a large amount of semi-structured and unstructured data. Oil and gas exploration and development targets consist of underground rock and fluid, so petroleum engineering is more dependent on data [43]. Compared with traditional information technology, big data technology can analyze and process massive data more quickly and efficiently, improve the accuracy, timeliness and comprehensiveness of decision making, and play an important role in promoting the increases in oil and gas storage, production, cost reduction and efficiency [44]. Shell’s Smart Field focuses on the collaborative work environment, intelligent well control, optical fiber monitoring, real-time optimization of production, intelligent water injection, and closed-loop reservoir management. The “Field of the Future” proposed by British Petroleum (BP) focuses on the application of real-time information systems to optimize operation [45]. Total signed an agreement with Google to jointly explore intelligent solutions for oil and gas exploration and development, focusing on intelligent processing and interpretation of underground imaging, especially seismic data processing and interpretation, in order to improve the development efficiency of oil and gas fields [46]. Sinopec began to explore intelligent manufacturing in 2012 [47]. It has successively started the planning, design, and construction of smart factories, intelligent oilfields, and intelligent research institutes, and built an intelligent cloud industrial platform for oilfields, which profoundly integrates the new generation of information technology with enterprise business, promoting the digital transformation and upgrade of enterprises.

Against the backdrop of the new era of intelligence, the intelligent conducting of petroleum exploration and development is a core and highly important strategic research goal of the energy industry [48]. To overcome problems in periods of high recovery and high water cut, Liu et al. developed fourth-generation separated-layer water injection technology, which realizes the digital real-time monitoring of single well layering pressure and water injection volume for injector, enhancing the water flooding effect. In addition, extensive data have been produced by mature oilfields that have experienced long-term production. It is significant to integrate the massive production data obtained from mature oilfields and dynamic observation data from reservoirs generated in the fourth-generation separated-layer water injection process to conduct intelligent reservoir analysis on water flooding, thus improving the development efficiency and recovery rate of mature oilfields [49]. Meanwhile, substantial data are generated in development of unconventional oil and gas reservoirs, which can be used in an intelligent algorithm for the purposes of analysis and accurate prediction, reducing the development cost and increasing the production of unconventional resources.

The outbreak of COVID-19 directly affected the global energy supply and demand pattern in the short term and further lowered crude oil prices, thus requiring higher demand in order to achieve the “cost reduction and efficiency enhancement” of oil exploration and development in China. In recent years, AI has gradually been employed in various industries, such as manufacturing, medicine, finance, and transportation. Machine learning and deep learning algorithms rely on powerful computing capability to mine the correlation between feature and target on the basis of big data, and to make accurate predictions and quick decisions. In the petroleum industry, driven by the integration of geological engineering and full data, the application of AI algorithms, and the development of intelligent oilfields, rapid and accurate prediction of single well and reservoir performance for mature conventional oilfields and unconventional oil and gas has become possible, which will certainly facilitate the development of engineering optimization plans, improve intelligent construction and promote field development efficiency.

In this study, the research on AI in petroleum exploration and development is extensively investigated, especially with respect to smart field technology for mature waterflood reservoirs. Combined with the actual demand for oil and gas development and related work regarding digitization, automation and integration of separated-layer water injection, the research progress and application of AI are discussed and prospected in three fields, including the current status of data foundation construction, the research progress of applied AI algorithms, application scenarios, and overall deployment expounded, and future directions in artificial intelligence are discussed and prospected, with a focus on their application and development trends in energy.

Current Status of Data Foundation Construction

Digitalization in the oil and gas industry has almost completed rescaled deployment. The new generation of monitoring and controlling technologies, such as fourth-generation separate layer water injection, has provided a large amount of previously difficult-to-obtain layered measurement data. However, there has been insufficient preparation for the integration of whole reservoir construction and application data with the new generation of monitoring data.

From a global perspective, oil and gas exploration and development companies recognized the value of massive data and digitalization at a relatively early point in time. In recent years, an increasing number of upstream companies have focused on automatic collection and equipment and processing control of big data. According to Gartner, 75% of international oil companies’ investment in informatization is related to big data. According to industry evaluation, oil companies can increase oil recovery by at least 6% and production by 8% through big data analysis and optimization. Big data technology has received approval from the industry and has become the key to the smart oilfield.

In the past three years, there have been continuous reports about cooperation between oil companies or oilfield service companies, such as BP, Shell, Eni, Saudi Aramco, and Schlumberger, and IT service companies such as Amazon, Google, and Microsoft, on establishing data centers and industry big data solutions. For example, a single fiberoptic cable from an oil well can generate more than 1 TB of data per day according to Chevron’s data collection and analysis system. Chevron has deployed thousands of optical fiber wells, and the data collected and analyzed have expanded dramatically over the years. Using big data technology provided by Microsoft, Chevron has increased the drilling efficiency of some oilfields in the Gulf of Mexico by more than 20%. Through the optimization of exploration and development on the basis of big data, Chevron has dramatically improved the number and quality of the newly discovered oil and gas fields.

Another example is the offshore unmanned platform system advocated by Equinor (Statoil ASA, Equinor ASA is a Norwegian state-owned multinational energy company headquartered in Stavanger), which has integrated the data for the whole system, including geophysics, geology, drilling and completion engineering, oil production engineering, pipe network, and platform measurement. On the basis of this platform, Equinor used the intelligent algorithm to quickly predict system status and abnormal risks, provide corresponding solutions, and finally finish implementation using intelligent sensor and control equipment. Figure 2 presents the data type and volume of the Volve field in the North Sea, which is a part of a joint study between Equinor and Stanford University.

China started establishing the digital oilfields early in the century, with PetroChina as the leader. At present, the A series data system surrounding upstream exploration and development service has been preliminarily completed, and covers exploration and production data, well production data, pipeline and geographic information, oil production and surface engineering, etc. At the same time, an integrated platform for collaborative research and applications related to exploration and development is in the continuous construction stage, and serves as a good foundation for the construction of a big data platform for reservoirs.

For example, since 2011, in order to meet the needs of large-scale construction in oil and gas fields, Changqing Oilfield has widely adopted advanced information technology, actively integrated into “Internet +”, and led the research trend of enterprises changing from the traditional “artificial + computer” to “digital and intelligent” big data. At present, by establishing a “data service center” with massive data and multiple data chains, Changqing Oilfield has built not only a scene-based system with the theme of “standardization real-time, and visualization”, but also digital services that extend underground. In addition, Changqing Oilfield has integrated more than 200 million pieces of basic data in 19 categories, including drilling, mud logging, well-logging, oil and gas testing, analysis and testing, and oil and gas production. Digital and intelligent technologies have been employed in front-line duty, early warning with respect to on-site safety, remote training for employees, offices with the use of handheld mobile tools, and intelligent lithology.

On the other hand, in addition to managing the massive data generated with respect to geophysics, geology, reservoir, and engineering throughout the whole process of petroleum exploration and development, the layered dynamic monitoring data of single wells and reservoirs, generated using new measurement methods and sensing equipment, has become an extremely important supplement to big reservoir data. Different from geophysical interpretation data, this type of observational data is more direct and reliable, and can be collected more frequently and intensively, supplementing data for comprehensively sensing the dynamics of subsurface pressure and fluid distribution. AI technology combines this massive engineering production history data to dynamically adjust reservoir properties. First of all, the production history data is passed through the inverse history matching model to preliminarily predict the reservoir physical properties. The reservoir properties predicted by the high-fidelity numerical model were used to test the matching quality. If a good match is obtained, the result is saved. Otherwise, prospective data will be used to further adjust reservoir properties. In this case, a global optimization algorithm can be combined with an intelligent system to minimize historical matching errors. Figure 3 illustrates an AI-assisted history-matching workflow for reservoir properties.

This progress is typified by fourth-generation separated-layer water injection technology. After more than 60 years of development, three generations of separated-layer water injection technologies have formed: fixed type, wire casting and fishing, and cable measurement and adjustment. The adjustment method evolved from starting and lowering the pipe string to casting and fishing nozzles. Data recording developed from card marking to electronic storage and performing direct readings on the ground. In addition, the method for obtaining corresponding parameters changed from recording a single parameter to recording multiple parameters at the same time. The separated-layer water injection method has developed from fixed layer [50], to movable layer, conventional eccentric layer, concentric integrated layer and bridge-type eccentric layer water injection [51]. The supporting measurement and adjustment technology has developed from steel wire casting to a direct reading from the steel pipe cable [52,53,54,55]. With the refinements in oilfield developments, there are limitations in the application of the current layer water injection technology, including the failure of stratification parameters in long-term continuous monitoring, resulting in inaccurate water dispensing plans and an inability to meet fine water distribution requirements; the use of point-like discontinuous measurement and adjustment not being able to maintain the rate of water injection at a high level for a long time; the fact that, as the layer division at the injection side becomes finer, the displacement performance changes more and more frequently, and as a result, the workload of measurement and adjustment workload is multiplied; the fact that the production end mostly produces oil in general form, and the liquid production in the high-permeability interval rises rapidly during the high water-cut period, further intensifying the vertical contradiction between layers; the low monitoring rate of the liquid production profile, the unclear liquid production structure, the unclear distribution of the remaining oil, and the difficulty of adjusting the injection-production matching (Figure 4). Hence, China has carried out research into fourth-generation separated layer water injection technology [56], including: real-time monitoring of single-well layer pressure and water injection volume in water injection wells; dynamic monitoring of water injection in blocks and reservoirs; design and optimization of water injection schemes that are integrated into reservoirs and engineering; and the adjustment of downhole layer water injection, which can greatly improve the efficiency of reservoir development and refinement and reduce management costs.

In recent years, breakthroughs have been made in core technologies such as downhole permanent interval adjustment, layered measurement and wellbore two-way communication, and real-time monitoring and automatic control technologies have been developed for separated-layer water injection. The digitalization of real-time monitoring of well layer pressure and flux in injection wells and the informatization of reservoir performance monitoring have been realized, promoting the development of layered water injection technology towards digitalization, automation and integration, and resolving the contradiction presented by the exponential increase in production cost caused by increases in measurement and adjustment workload, and personnel and equipment investment following the increasingly refined interval division of a single well [57]. At the same time, the level of management and oil reservoir analysis have been promoted with respect to their degree of refinement and intelligence. However, traditional water-flooding reservoir analysis methods have limitations such as long duration, reliance on engineering experience and limited optimization schemes. Therefore, integrating a large number of dynamic injection–production data will further increase the computational cost of numerical simulation. Meanwhile, highly efficient, precise, and intelligent reservoir optimization simulation technology has become the key to future reservoir management. With the development of fourth-generation separated-layer water injection technology, some key technologies, such as interval flow detection, allocation, injection volume adjustment, and big data-driven water injection optimization algorithms for old oil reservoirs have been integrated with reservoir and oil production engineering, enabling the realization of the design and optimization of water injection schemes and the real-time adjustment of downhole layer water injection.

In general, a certain foundation has been developed for the construction of “digital oilfields”. However, there remain some challenges in applying big data algorithms to reservoirs with respect to data measurement and interpretation, with some specific examples including: (1) the inhomogeneous space and time distribution of interpretation and observation reservoir data; (2) the high degree of uncertainty in the understanding of the geological characteristics and physical properties of reservoirs; and (3) the scarcity of measurement data regarding reservoir performance, especially the data used to make decisions, such as time-dependent layer water injection volume and oil/water production. These challenges lead directly to machine learning algorithms being unable to obtain sufficient and adequate training samples [58]. Although massive production data can be obtained during the process of development, effective samples for the development of big data algorithms are scarce. As a result, it is urgently necessary to build a “whole data” (generated from explorations to development) platform that can effectively assist with conducting reservoir research.

2. Research Progress of the Application of AI Algorithms

In recent years, AI algorithms have been widely applied in the petroleum industry, and are essential for industrial decision making. In this section, we provide a detailed review of the research progress of the AI algorithms in the petroleum industry.

Data models or analysis algorithms for geophysical interpretation, oil production engineering, and single well analysis are relatively mature. For the research into entire oil reservoirs, in recent years there have been many automatic history-matching algorithms based on dynamic data or physical models combined with AI algorithms. However, an overall surrogate model that is able to achieve fast prediction of reservoir performance and automatic matching of single wells is still in its preliminary stage.

History matching is an extremely important part of the reservoir simulation process. Artificial history matching requires reservoir engineers to have a rich body of experience from which to draw, and to spend large amounts of time repeatedly adjusting parameters. Hence, the automatic history matching method has become a hot research topic in oil reservoir engineering in recent years. Automatic history matching methods can be roughly divided into model space-based inversion and data space-based inversion.

Research progress of model space-based inversion algorithms: Oliver and Chen [59] produced various model based data assimilation algorithms through optimization. The random maximum likelihood (RML) method [60,61] is a sampling method that, based on the Bayesian framework, randomly perturbates the examples and then generates different models by minimizing specific objective functions [62,63]. Sampling-based methods such as reject sampling and Markov chain Monte Carlo [64,65,66] are extremely time consuming, and as a result, they are not suitable for large-scale problems. Ensemble-based data assimilation algorithms, such as the Ensemble Kalman Filter (EnKF), have been used widely in many fields [67,68,69]. EnKF is based on the Bayesian framework, and requires multiple restart operations during fitting [70]. EnKF assumes a linear relationship between the model parameters and the flow response during each update; however, for nonlinear problems, an error may occur. Ehrendorfer [71] analyzed and summarized the sampling error of the EnKF algorithm. Van Leeuwen and Evensen [72] proposed a new ensemble based ensemble smoothing algorithm (ES). In contrast to EnKF, ES does not require the simulator to be restarted and the model parameters to be updated globally at single time, and thus it runs faster than EnKF [73]. However, Emerick and Reynolds [74] found that ES only updates model parameters through one iteration, resulting in insufficient reservoir history matching. Emerick and Reynolds [75] proposed an ensemble-smoothing algorithm under multiple data assimilation (ES-MDA), which functions well in reservoir history matching. Jenni et al. [76] carried out history matching research for naturally fractured reservoirs; however, their method only generated a single posterior model, which was not able to effectively reduce the uncertainty of the reservoir prediction index.

Research progress of data space-based inversion algorithm: the SCERF research group led by Professor Jef Caers and the SFC research group led by Louis Durlofsky at Stanford University carried out pioneering work. Scheidt et al. [77] proposed an uncertainty quantification method based on prediction focused analysis (PFA) to predict pollutant concentration data. On this basis, Satija and Caers [78] proposed the standard function component analysis method to reduce the dimensions of the data and the prediction variables, thus solving the problem of PFA [79]. This method has been applied to reservoir development and pollutant concentration prediction [80]. Scheidt et al. [77] proposed a Bayesian Evidence Learning (BEL) process to reduce the uncertainty inherent in oilfield development index prediction, and finally to effectively guide oilfield development management and make decisions. He et al. [81] established an ensemble-variance analysis method to carry out quantitative analysis of uncertainty in subsurface flow problems. Sun and Durlofsky [82] proposed a new data space inversion (DSI) algorithm based on the Bayesian framework to perform a rapid quantitative analysis of the uncertainty of reservoir development index prediction. Jiang and Durlofsky [83] further proposed the DSIVC algorithm based on the DSI algorithm to avoid the problem of repeated calculation and adjustment of prior models under specific well control conditions.

The above automatic history matching method was able to effectively overcome the defects arising from the dependence on engineers’ experience and the long research cycle; however, they are not satisfactory applied in large-scale reservoirs. Methods based on model-space inversion require an a posteriori model that conforms to the observation data, and the process for generating this is complex and time consuming; the methods based on data space inversion may lead to dimension disasters when they are used to deal with high-dimensional data problems, thus decreasing the model’s prediction ability. The practicability and prediction ability of traditional automatic history matching methods have certain limitations. Hence, it is urgent to carry out research on intelligent history matching methods based on traditional models and data, cut down the research cycle time, improve the prediction accuracy of the dynamic data, obtain dynamic predictions that are more consistent with the actual production, and provide reliable models for scheme adjustment and optimization.

There has been extensive research on the development of surrogate models for subsurface flow simulation predictions, and these can be divided into two categories: physics-based methods and data-driven methods. Physics-based methods often ignore or simplify some physical or numerical aspects of the problem, such as through simplifying physics modeling, coarsening the mesh modeling, or employing reduced-order modeling based on proper orthogonal decomposition (POD) [84,85,86,87,88]. However, applying these methods in solving inverse problems presents some limitations [89,90].

Data-driven methods rely entirely on data generated by a solver to train the model, in order to approximate the mapping relationship between input and output. Artificial neural network-based methods have also found broad applications in the petroleum industry, and usually require massive data in order to train the model. Yilmaz et al. (2002) described a model based on backpropagation ANN and fractal geostatistics to solve the problem of selecting the best bit in terms of actual rock bit data, gamma ray, and acoustic logging data of several wells in carbonate oilfields. Huang et al. proposed an ANN model to predict the MMP of pure and impure CO2 and oil systems [91], which uses the molecular weight of the C5+ component, the reservoir temperature, and the oil concentrations of volatile and intermediate components. Ahmadi used a hybrid model combining feedforward neural networks and the imperialist competitive algorithm (ICA) to predict asphaltene deposits in reservoirs. He evaluated the performance of the ICA-ANN model compared with the scale model and the traditional ANN model and showed the effectiveness of the former [92]. In history matching, Hamdi et al. [93] applied a Gaussian process in surrogate modeling for an unconventional gas reservoir system including 20 parameters. Bazargan et al. [94] constructed a polynomial chaotic expansion surrogate model for a 40-parameter history matching study of channelized reservoirs. Costa et al. [95] applied ANN to build a 16-parameter surrogate model to assist in the history-matching problem of oil-water systems. Although these data-driven surrogate models have advantages and disadvantages, they all share a similar limitation: they are only suitable for low-dimensional spaces.

The recent rapid development of deep neural networks and their successful applications in high-dimensional data regression problems such as image recognition [96,97,98] and natural language processing [99,100] have inspired research on deep learning-based proxy modeling for high-dimensional nonlinear systems. Unlike shallow artificial neural networks, well-designed deep neural networks are able to capture complex high-dimensional nonlinear relationships and avoid overfitting [101,102]. Zhu and Zabaras [103,104] first introduced a fully convolutional encoder–decoder network to approximate flow simulation, considering single-phase steady-state flow in a model characterized by a Gaussian permeability field, and demonstrated that their deep convolutional neural network can accurately predict high-dimensional pressure field maps. They then conducted more applied research, including predicting CO₂ saturation and groundwater contaminant concentrations under uncertainty quantification and inverse problem solution [105]. Jin et al. [106] proposed a deep learning-based E2C (Embed-to-Control) method, a surrogate model capable of providing very fast production performance predictions for two-phase flow systems under different well control conditions. These studies demonstrate the ability of deep convolutional neural networks to extract high-dimensional nonlinear relationships in subsurface flow systems.

With the development of AI technology, foreign scholars have already established data-driven reservoir simulation TDM methods [107], which, to a certain extent, are able to overcome some of the weaknesses of the traditional reservoir engineering and reservoir numerical simulation methods. This is essential for the development evaluation of mature oilfields. However, due to the traditional fully connected neural network structure, the current method is not able to deal with time series data, such as reservoir production performance well. In consideration of the advantages of the recurrent neural network (RNN) in processing time series data, some scholars have established data-driven reservoir modeling methods using the recurrent neural network to carry out reservoir production prediction [108]. Chaki et al. constructed two types of neural networks (DNN and RNN) to predict oil and water production in 10 years [109]. Compared with the physical flow simulator, both neural network models are able to reduce the calculation cost by 100 times without sacrificing the prediction accuracy. They concluded that DNN is more suitable for short-term production forecasting, while RNN is more suitable for long-term production forecasting [110]. The results show that the sensitivity analysis of super parameters is significant for the accurate prediction of data-driven models. However, this method is limited in application scenarios such as long-term prediction and parameter inversion. Currently, research methods for oil and gas development mainly comprise reservoir engineering, reservoir numerical simulation, and data-driven reservoir simulation methods. The conventional reservoir engineering method is mainly based on the statistics of block production rules, and its accuracy is not high; the traditional reservoir numerical simulation method, which has limitations with respect to grid shape and resolution, takes a long time to perform dynamic simulation, which is detrimental to subsequent history matching; the future development direction is to use the data-driven method to solve reservoir simulation problems. Nevertheless, there are some problems, such as the inherent shortage of reservoir data source granularity, that make it impossible to obtain massive sample data. Therefore, it is urgent to establish a “data physics” model that is consistent with the dynamic data, and that integrates the petroleum physical model with the real-time data of layer water injection. This model, driven by dynamic and static big data, would be automatically updated, improving layer water injection over that of conventional reservoir engineering methods and being faster than numerical simulation methods. As a result, it would have a revolutionary impact on reservoir influence and reservoir development evaluation methods.

The method based on long short-term memory (LSTM) is widely used in the production prediction of time-varying single well time-varying problems, because the special gate structure is able to learn long-term correlation and deal with complex problems involving time series. Many studies have applied the LSTM algorithm to predict the production performance of single wells, especially for the purposes of developing oilfield time series databases. Sun et al. implemented the LSTM algorithms and tradition physical to predict the dynamic production rates for single and multiple wells. It was shown that LSTM is able to capture subtle changes in the dynamic production curve, while DCA can only describe the overall trend. In addition, the LSTM model provides more accurate results with a lower calculation cost under three-phase conditions (oil, gas, and water) or in complex operation scenarios [111]. Similar to character translation in natural language processing, Biswis et al. pointed out that LSTM has been proven to be able to evaluate the production of adjacent wells with variable history and more flexible capabilities. However, with increasing amounts of data, LSTM may exhibit a more robust learning ability and thus perform better [112]. Because LSTM has firm nonlinear fitting and time memory capabilities, as well as effective performance in time series tasks, LSTM was successfully applied to monthly hydrocarbon production prediction (for example, production in the last six months and 12 months). Temizel conducted a comparative study between a physical simulation model and LSTM for a reservoir [113]. The results showed that the prediction ability may have been better due to the limited amount of reservoir data based on the physical simulation model due to the strict physical laws of fluid flow in porous media. Therefore, it is necessary to use the results of traditional numerical simulation to obtain more input data containing different operating conditions and wells for data-driven model learning.

3. Application Scenarios and Overall Deployment

Applications in data analysis, oil production engineering, automatic control, geophysical interpretation and other relatively isolated and independent scenarios are increasing. For reservoir prediction, the relevant research has been focused on automatic history matching, 3D reservoir modeling, flow simulation, and data physical modeling, etc.

Since the dramatic decrease in international oil prices in 2014, people in petroleum engineering have tried to accomplish sustainable development by means of data analysis, real-time monitoring and automation in order to improve their competitiveness and anti-risk capabilities. AI technology is widely used in various areas in petroleum engineering, such as in software, intelligent equipment, operation platforms and special services. The application range covers all aspects, from management through to exploration, development, and construction at site, thus improving the automation and intelligence of petroleum engineering.

Shell cooperates with HP to deploy numerous sensors and optical fiber cables in the oilfields and refineries, and transmits massive data to the private cloud maintained by Amazon. Through the construction of the database and the application of a large number of data analysis methods, Shell is able to achieve a more accurate understanding of the reservoir conditions. Data from any one field can be compared with other data sourced from many oilfields, potentially thousands of miles away, in order to identify the best drilling targets. Shell uses SAS predictive asset maintenance software to extend the service life and operation time of equipment, and the increased production of oil and gas generates benefits in the form of tens of millions of dollars in company revenue. Big data applies to every aspect of equipment operation, and through the use of data analysis technology based on this, Shell maintains the lead worldwide. Shell also eliminates the uncertainty in the device operation process with the use of SAS software. When SAS sends an early warning signal, engineers are able to quickly diagnose and prevent or mitigate serious problems.

Baker Hughes uses the Digital twin technology to integrate physical machinery with the analytical technology. By means of the deep-learning model stored on the Predix industrial Internet platform, it is possible to automatically detect equipment defects and abnormal conditions, provide early warning of potential breakdown, develop detection plans on the basis of risk, and avoid unnecessary routine periodic detection and maintenance. By means of digital simulation, Baker Hughes aims to implement “synchronous simulation reproduction” of all elements and the whole process of oilfield production, and finally achieve full visualization, predictability and control of the production system. In 2016–2017, Baker Hughes developed a performance management scheme, based on digital twin technology, for the top drive, draw works, thrusters, main engines, and other key components of 10 Maersk Drilling rigs. To date, Baker Hughes has created digital twins for more than 5000 instruments, and is investigating the construction of the digital twins of wells. By means of sensors in the wellbore, Baker Hughes is able to obtain information on tools and equipment in the wellbore and the reservoir status, and is able to combine the well status, equipment operation status, and well production status. SparkCognition, a software package developed by Flowserve, a pump manufacturer in the oil and gas industry, can also provide similar services. The software provides automatic modeling, establishing a reliable model, and performing predictive maintenance functions when the fault data required for model training are missing.

In China, the application of big data in oilfield exploration and development is still restricted to local domains, such as seismic data processing, real-time analysis of well sites, oilfield production analysis, and equipment and facility maintenance. However, complete big data applications oriented toward geological and reservoir research, including the establishment of multidisciplinary data and business analysis models, are still under exploration and construction. Some typical application scenarios include: seismic stacking velocity quality control and modeling based on the fully connected neural network, seismic first break picking and modeling based on convolutional neural networks, sedimentary facies planar interpolation based on deep neural network, seismic profile characterization and interpolation based on deep neural networks, 3D property modeling based on deep neural networks, and logging curve generation method based on the recurrent neural network [114,115].

In terms of overall reservoir research, using AI algorithms, scholars have conducted research to explore 3D reservoir modeling, flow simulation, and physical data modeling in recent years. In reservoir modeling, Liu et al. [116] used principal component analysis and a convolutional neural network algorithm to propose a new low-dimensional parameterization method to characterize non-Gaussian models; Klie and Florez [117] applied the Latin hypercube sampling algorithm to study the uncertainty of fracture distribution in shale reservoirs; based on training images, Laloy et al. [118] proposed a generative adversarial network model to solve inversion problems; combined with multiple data assimilation methods and variational self-encoder, Canchumuni et al. [119] researched the river equivalent complex geological model and proposed a new parameterization method; Temirchev et al. [120] constructed a reservoir down-order simulation process based on a variational autoencoder and recurrent neural network, significantly improving the simulation speed; in porous media flow simulation, Zhu and Zabaras [103] proposed a Bayesian method based on a convolutional self-encoder; based on Zhu and Zabaras’s method, Mo et al. [121] established a proxy model for the quantification of multiphase flow uncertainty. In physical data modeling, De Bézenac et al. [122] built a physics-informed neural network to predict sea surface temperature accurately; Seo and Liu [123] proposed a network architecture of differentiable maps, which significantly improved the weather prediction ability; Grigo and Koutsourelakis [124] integrated self-encoder and internal physical models and carried out small-data research; and Zhu et al. [104] added the physical model equation to the loss function and trained a neural network based on the generative model of conditional flow.

With respect to reservoir dynamic analysis, Jia et al. used traditional numerical simulation combined with optimization algorithms to study the flow relationship between injection and production wells [125]. Sarma et al. proposed a closed production optimization method for water injection reservoirs, which was based on the Sequential Quadratic Programming (SQP) solution [126]. This closed production optimization method also combines the Karhunen–Loeve transformation and the polynomial accurate replacement approach to obtain more reliable uncertain parameters such as reservoir permeability and porosity. Yan et al. also proposed a new closed-loop production optimization method, which was based on the Ensemble Kalman filter (EnKF) [127]. The main feature of this method is to optimize the expected net present value (NPV) based on model updating. This method is relatively stable, and avoids the necessity of solving redundant equations during the solution process. Patrick et al. studied the closed-loop production optimization problem and used a method that is famous in control theory—the nonlinear model predictive control (NMPC) method to deal with the serious nonlinear problems in the reservoir [128]. They used the numerical simulation software Eclipse to establish the reservoir model. This algorithm can be used to obtain the optimal flow control valve (ICV) setting parameters during the optimization process. A.I. Hassan proposed a discrete optimization method for studying the production of intelligent horizontal wells, which distributes the flow volume of injection and production wells by optimizing the parameters of the flow control valves [129]. After applying the discrete optimization method, the production of oil ring reservoirs improved dramatically. In addition to the above optimization model, many studies have focused on accelerating the numerical simulation of large dynamical systems with high computational costs. One such technique is the Model Order Reduction (MOR) method. Eduardo Gildin tried to reduce the computational costs in reservoir water injection optimization problems by combining different optimization parameterization methods and model order reduction techniques. H. Zalavadia proposed a method based on MOR as a solution to the well location optimization problem, which requires a large number of high-fidelity simulation runs [130].

For production forecasting, Wu et al. successfully introduced a feedforward neural network to predict oilfield production for the first time in China [131]. Wang et al. used five models to predict well production: the integration method, linear regression, support vector machine, regression tree, Gaussian process regression, and LSTM [132]. They found that the prediction results of the LSTM model were the most accurate. Lee et al. used the LSTM model to predict the monthly production of shale gas wells in Alberta, Canada, and demonstrated the advantages of LSTM over decline curve analysis. Because the LSTM model has the benefits of high accuracy in production prediction, and only requires a small amount of data, it is more suitable for unconventional oil and gas production. Scholars have continuously optimized the LSTM model [133]. For example, Kocoglu et al. used the Bi LSTM model based on Bayesian optimization to predict the production of horizontal wells. This model demonstrated improved prediction accuracy compared with the traditional LSTM model [134]. Zhang et al. combined two LSTM sub-models with different characteristics using integrated methods, significantly improving the prediction accuracy of the model [135]. Qiu set seven characteristic quantities, including tubing pressure, casing pressure, and reservoir temperature, and used the LSTM model to predict the gas wells’ production data in the Ordos Basin for 1–4 years, achieving high accuracy. Although data-driven approaches have grown in popularity over the past two decades due to the increase in data collection with oilfield development [136], the extrapolation ability of purely data-driven approaches is limited, and a sufficient volume of training data is required. Training the neural network prediction model with limited data will lead to overfitting and poor prediction performance. Unlike statistical models, physics-based models can provide reliable predictions across a wide range of inputs. Jodel et al. proposed a model that was able to embed the physics flow function to the observed data in order to improve production forecasting [137].

The traditional water injection optimization method is based on numerical simulation. Firstly, a geological model is established according to the dynamic and static data of the reservoir, and after numerical simulation of the reservoir, artificial historical fitting is carried out according to the experience of reservoir engineers. Finally, a better scheme is selected. This engineering approach is time-consuming, depends on reservoir engineer experience, and has limited optimization options. The whole process of water injection optimization simulation based on AI technology is divided into two steps. In the first step, the data assimilation algorithm is incorporated into the reservoir model, so that the correction of the geological model parameter field can be carried out automatically under continuous data drive. In the second step, the machine learning algorithm is used to quantify and evaluate the effect of multi-well stratified water injection, analyze the direction of water injection adjustment, and finally determine the water injection adjustment plan through the big data intelligent optimization algorithm. Figure 5 represents the optimization workflow of the water injection scheme by traditional and AI technologies.

In addition, water composition can significantly affect oil recovery in mature fields. Saudi Aramco, through its upstream research division (EXPEC Advanced Research Center), launched a research program called Smart Water Flood to explore the potential for enhancing oil recovery by adjusting the properties of the injected water (with respect to parameters such as salinity, ion composition, interfacial tension, viscosity, etc.). Leonardo Fonseca Reginato used a machine learning method to predict the relative permeability curve affected by changes in wettability under the conditions of different concentrations of the ion components in water injection [138]. Shams presented a novel empirical correlation based on a feed-forward neural network to predict the recovery efficiency of low-salinity waterflooding (LSWF) in a heterogeneous reservoir at and beyond water breakthrough [139]. Rashida analyzed the effect of oil/brine parameters on LSWF performance using multivariable linear regression [140]. Low-salinity water flooding, as a novel enhanced oil recovery method, has gained increased attention, especially during the past decade. Akin to other methods, low salinity water flooding carries some risks, which could lead to failure in the target reservoir. To reduce the potential risk of incompatibility between injected water and initial oil/brine/rock system, the composition of injected water needs to be optimized to prevent formation damage due to incompatibility problems. Moreover, careful selection of the composition of injected water would profoundly influence the performance of LSWF. Z Negahdari et al. proposed a detailed workflow to resolve problems in terms of maximizing wettability alteration and minimizing compatibility in a core model of carbonate rock. The proposed framework can be utilized to plan an effective injection during LSWF [141]. For the simulation of low-salinity water flooding, the method proposed in our previous work considering time-dependent wettability alteration was applied. Furthermore, a genetic algorithm (GA) was utilized to optimize the injected water composition.

In general, with respect to data, the digitalization of the oil and gas industry is complete, or has at least commenced large-scale deployment; however, the phenomenon of data isolation is serious in China. There has been insufficient preparation for the integration of new monitoring in data construction and application scenarios in order to perform overall studies on reservoirs; in recent years, data-based and model-based automatic history matching algorithms have emerged globally, representing major progress. Tracy Energy Technologies released its digital twin software “Timeline” for oil and gas reservoirs in 2022. It features accurate and super-fast surrogate modeling of reservoir dynamics with varying well controls, and it has been successfully applied in 15 oil and gas fields in China. However, the development of the overall proxy model for rapid prediction and automatic history matching of each individual well and overall reservoir performance is still in the preliminary period; for application scenarios, relatively simple and independent applications continue to emerge, such as data analysis, oil production engineering, automatic control, and geophysical interpretation. However, in terms of reservoir prediction, related studies have mostly focused on automatic history matching, 3D reservoir modeling, flow simulation, data physics modeling, etc. Large-scale reservoir and field applications require substantial investments in data acquisition, game-changing algorithms with cloud-based computing architecture, and top-down deployment.

4. Conclusions

This study provided a comprehensive investigation on applications of AI in petroleum engineering, especially in the smart deployment for mature waterflood reservoirs, with respect to three aspects: the current status of the construction of the data foundation, the research progress on applied AI algorithms, and the application scenarios and overall deployment for mature waterflood reservoirs. In addition, potential challenges and limitations associated with well-known AI algorithms were also discussed. To date, the application of AI in petroleum exploration and development has mainly occurred in logging processing and interpretation (e.g., lithology identification, and curve reconstruction), seismic processing and interpretation (e.g., first arrival wave pickup, and fault identification), real-time regulation and control of water flooding development, and production prediction. The application of AI algorithms improves the intelligence level of integrated analysis software. However, since the underground conditions of reservoirs are complex and variable, exploration and development faces uncertainties and small samples, and the full application of AI still presents a challenge.

The application of AI algorithms to achieve fast and accurate analysis and forecasting using the vast amount of data accumulated in the development of mature waterflood fields is important for reducing costs and increasing production. A certain foundation has been laid for the construction of “digital oilfields”; however, there are some specific challenges regarding the application of big data algorithms to reservoirs with respect to obtaining sufficient and adequate training samples, including the inhomogeneity of the space and time distributions of reservoir data, high uncertainty in the understanding of geological characteristics and physical properties, and the scarcity of fine measurement data representative of reservoir performance. With regard to the application of algorithms, many studies have been performed in which the data models or analysis algorithms described are now relatively mature, especially with respect to the fields of geophysical interpretation, oil production engineering, and single well analysis. In terms of overall reservoir research, using AI algorithms, scholars have conducted research on 3D reservoir modeling, flow simulation, and data physical modeling. However, an overall surrogate model that can achieve fast prediction of reservoir performance and automatic matching of single wells is still in the preliminary stages, while a major direction of future development will be to use data-driven methods to solve reservoir simulation problems. In China, the application of big data in oilfield exploration and development is still restricted to local domains, such as seismic data processing, real-time analysis of well sites, oilfield production analysis, and the maintenance of equipment and facilities. However, the phenomenon of data isolation is serious in China, and the complete application of big data oriented toward geological research and reservoir research, including the establishment of multidisciplinary data and business analysis models, is still being explored and developed. Large-scale reservoir and field applications require a substantial investment in data acquisition, game-changing algorithms with cloud-based computing architecture, and top-down deployment.