A Deep Learning Wag Injection Method for Co2 Recovery Optimization

CO2 has some critical technical and economic reasons for its use as an injection gas for oil recovery. CO2 is very soluble in crude oil at reservoir pressures; it contributes to sweep efficiency enhancement as it swells the oil and significantly reduces its viscosity. Although the mechanism of CO2 flooding is the same as that for other gases, CO2 is easier to handle, it is cheaper, and it is an environmentally better candidate than other gases. Formation evaluation and reservoir engineering have been major areas in the oil and gas industry that are heavily influenced by technology advances, to increase efficiency, improve hydrocarbon recovery and allow real-time reservoir monitoring. Water flooding for increasing oil recovery has been amongst the oldest production mechanisms widely utilized since the end of the 19th century to maintain pressure levels in the reservoir and push hydrocarbons accumulations towards the producing wellbore locations (Satter, Iqbal, & Buchwalter, 2008). Produced water from the reservoir formation was reinjected in order to maintain pressure levels, as well as seawater and aquifer water injection have also taken a strong mandate. With the advent of technology and processing plants this injection process was further refined, allowing salinity control of the injected water as well as monitor the injection and distribution of the water levels in near real time (Boussa, Bencherif, Hamza, & Khodja, 2005). Formation evaluation has seen an even greater penetration of technology in its area with the quest to achieve real-time formation evaluation during the drilling process. Conventional formation evaluation is conducted utilizing wireline logging technology, which is deployed after the drilling of the well and allows to analyze the reservoir formation. Given the significant advancement of logging technologies, acquiring the measurements during the drilling process (LWD) has been at the forefront of interest, allowing improved well placement and geosteering as well as real-time formation evaluation to optimize well completion strategies (Hill, 2017). Amongst the technologies recently deployed, surfaced logging and advanced mud and logging allow to determine on cuttings in real time mostly any of the properties previously possible only on direct measurements on cores (Santarelli, Marsala, Brignoli, Rossi, & Bona, 1998; Katterbauer & Marsala, A Novel Sparsity Deploying Reinforcement Deep Learning Algorithm for Saturation Mapping of Oil and Gas Reservoirs, 2021; Katterbauer, Marsala, Schoepf, & Donzier, 2021). With advances in AI, reservoir characterization is now moving towards real-time or near real-time analysis at the rig site. For near real-time analysis, the main physical source of data is drill cuttings as it guides the drilling operation by determining important depth point such as formation tops, coring intervals. Traditionally, the description of these cuttings is done manually by geologists at the well site. The accuracy of these descriptions can be variable depending on the geologist's experience and indeed their mental state and tiredness level. Cores is another source of data. New techniques and older techniques imbued with AI components new allow for greater automation, efficiency, and consistency. The use of AI on traditional images are of great interest in the oil and gas community as they are: 1) fast to acquire, and 2) do not typically require expensive hardware. For example, Arnesen and Wade used convolutional neural networks; specifically, an inception-v3 inspired architecture, to predict lithological variations in cuttings (Arnesen & Wade, 2018). In their study, each sample is related to one lithology. Buscombe used a customized convolutional neural network to predict the granulometry of sediments, specifically the grain size distribution (Buscombe, 2019). Similarly, automated core description systems (e.g., (Kanagandran; de Lima, Bonar, Coronado, Marfurt, & Nicholson, 2019; de Lima, Marfurt, Coronado, & Bonar, 2019) and microfossil identification systems (e.g., (de Lima, Bonar, Coronado, Marfurt, & Nicholson, 2019)) are also being explored using neural networks with varying degree of success. A comprehensive review on the state of usage of rock images for reservoir characterization presented by de Lima et al. (de Lima, Marfurt, Coronado, & Bonar, 2019). In addition, the community is also recognizing the potential of improving older techniques by integrating artificial intelligence into their workflow. In reservoir characterization, chemostratigraphic analysis X-ray fluorescence is a prime example for this especially with the difficulties encountered when analyzing mudrocks in shale plays using traditional methods. The rise of XRF measurement was also fueled by the introduction of highly portable XRF devices that take 10s of seconds to measure one sample. The use of artificial intelligence techniques is being studied. For example, fully connected neural networks are applied on XRF data to predict total organic carbon (Lawal, Mahmoud, Alade, & Abdulraheem, 2019; Alnahwi & Loucks, 2019). In addition to the traditional elemental to mineralogical inversion methods such as constrained optimization, neural networks are being utilized (Alnahwi & Loucks, 2019). The integration between XRF, X-ray diffraction (XRD) measurements (Marsala, Loermans, Shen, Scheibe, & Zereik, 2012), and well logs using traditional statistical methods and neural network methods is also being explored (Al Ibrahim, Mukerji, & Hosford Scheirer, 2019). The integration between artificial intelligence systems and automated robotic scanning systems (e.g., (Croudace, Rindby, & Rothwell, 2006)) is key in introducing these technologies into the daily rig operations. The low density of CO2 relative to the reservoir fluid (oil and water) results in gravity override whereby the injected CO2 gravitates towards the top of the reservoir, leaving the bulk of the reservoir uncontacted. This may lead to poor sweep efficiency and poor oil recovery; this criticality can be minimized by alternating CO2 injection with water or similar chase fluids. This process is known as Water Alternating Gas (WAG). A major challenge in the optimization of the WAG process is to determine the cycle periods and the injection levels to optimize recovery and production ranges. In this work we present a data-driven approach to optimizing the WAG process for CO2 Enhanced Oil Recovery (EOR). The framework integrates a deep learning technique for estimating the producer wells’ output levels from the injection parameters set at the injector wells. The deep learning technique is incorporated into a stochastic nonlinear optimization framework for optimizing the overall oil production over various WAG cycle patterns and injection levels. The framework was examined on a realistic synthetic field test case with several producer and injection wells. The results were promising, allowing to efficiently optimize various injection scenarios. The results outline a process to optimize CO2-EOR from the reservoir formation via the utilization of CO2 as compared to sole water injection. The novel framework presents a data-driven approach to the WAG injection cycle optimization for CO2-EOR. The framework can be easily implemented and assists in the pre-selection of various injection scenarios to validate their impact with a full feature reservoir simulation. A similar process may be tailored for other Improved Oil Recovery (IOR) mechanisms.