Robust Decision and Data Science Application for Reservoir Management : Probabilistic Forecasting, History Matching, Sequential Decision Making and Value-of-Information

Abstract

Reservoir Management (RM) is defined as the utilization of available technology, financial assets, and human resources to maximize the economic recovery of a reservoir. This type of management involves a series of operations and decisions, from the initial stage of the discovery of a reservoir to the final stage of field abandonment. Thus, RM is a decision-oriented activity, for which Decision Analysis (DA) will add value. Indeed, decisions are the only means we have to create value. DA is an approach intended to provide clarity of action on the decisions which we focus our attention. Due to the inherent uncertainty in the outcomes from our decisions, good decisions can have bad or good outcomes. A good decision is logically consistent with alternatives, information, and preferences available at the time the decision is made. The most challenging phenomenon we face in decision making is uncertainty. Uncertainty is inseparable from all significant decisions. Hence, clear thinking about uncertainty is a requirement for making good decisions. Although the oil and gas industry has long been aware of the importance of uncertainty understanding and management, data-driven decision approaches that include consistent uncertainty quantifications are not commonly or comprehensively used. This dissertation address three of the main challenges commonly encountered in reservoir management: production forecasting, uncertainty quantification for history matching problems and sequential decision making. We propose solutions to each problem that employ different algorithms of data-driven decision techniques and model-based forecast that allow a coherent integration of uncertainty and decisions. The first challenge is addressed through illustrating and discussing the implementation of probabilistic Machine Learning (ML) techniques in decline curve analysis. Unlike decline curve models, the ML approach can be regarded as “model-free” (non-parametric) because the pre-determination of decline curve models is not required. However, the main problem of pure ML techniques is lack of stability in long-term forecasts. To solve this, we have combined the decline curve model, a recognized technology in the reservoir engineering community, that provides stable long-term forecasts in an unconventional reservoir, with neural networks to automatically adjust decline curve model’s parameters. We illustrated the incorporation of Neural Ordinary Differential Equations (ODEs) with Bayesian Inference, The No-U-Turn MCMC sampler (NUTS) (Hoffman et al., 2014), which allows the prediction uncertainty of Decline Curve Analysis (DCA) to be quantified. Physics-based neural networks, which are a relatively new technique that makes it possible for physics constraints to be integrated into neural network architecture, are the foundation of this approach. The second challenge is addressed through illustrating and discussing the implementation of two techniques in uncertainty quantification and reservoir model calibration with much-reduced simulation computation time; one relies on ML Dimensionality Reduction (DR) techniques and the other one on Bayesian Evidential Learning (BEL) framework. The ML-DR approach is based on a sequential combination of nonlinear dimensionality reduction techniques: t-Distributed Stochastic Neighbor Embedding (t-SNE), the Gaussian Process Latent Variable Model (GPLVM), and clustering K-means, along with the data assimilation method Ensemble Smoother with Multiple Data Assimilation (ES-MDA). Cluster analysis with t-SNE and GPLVM is used to reduce the number of unknown parameters and select a set of optimal reservoir models that have similar production performance to the observed data from the field. We then apply ES-MDA for data assimilation. BEL is a general data scientific framework used to quantify uncertainty within the decision-making context. BEL relies mainly on data, model, prediction, and decision applying Bayesian inference methodologies. BEL is usually divided into six main stages: (1) Formulation and definition of the decision problem; (2) prior model definition and specification; (3) Monte Carlo simulation and falsification of the prior uncertainty models; (4) Global sensitivity; (5) Uncertainty reduction using data; (6) Posterior falsification and decision making. In step 5, one may opt for classical inversion or direct forecasting (DF). DF utilizes a combination of statistical learning techniques and the Monte-Carlo sampling method to ensure direct relationships between the data and the prediction variables. It should be noted that this method requires no completed explicit model inversions (update the model parameters). This results in it being less computationally expensive when compared to the standard inversion methods. The third challenge is addressed the problem of making optimal decisions while considering the evolution of uncertainty (learning over time). To efficiently account for the impact of future information on optimal decisions, we have used an Approximate Dynamic Programming (ADP) approach, often described as simulation-regression method to address a significant number of decision-making problems related to RM. RM is regarded as sequential problem, as most petroleum engineers and geoscientists are used to consider the gathered information, support their future decision making, and maximize the value created by the reservoirs. However, the models for reservoir management decisions may be computationally prohibitive and intractable if the state-space which involves the number of decisions, the number of alternatives for each decision and the number of uncertainties are included. To solve this issue and provide better good RM decisions, DA is recommended due to its several advantages. Howard (1980) stated that “DA is a systematic procedure for transforming opaque decision problems into transparent decision problems through a series of transparent procedures.” In the context of reservoir management, DA is used as a consistent mean for evaluating different approaches and alternatives to determine the optimal scenario to maximize the profitability of investment of any project e.g., Net Present Value (NPV). In addition, to utilize fewer computational resources, ADP, is a viable technique that can handle complex, large-scale problems and discover a near-optimal solution for intractable sequential decision making. Furthermore, we present and test the performance of several machine learning techniques to quantify geological uncertainty with the reservoir development plan and making sequential decision in the context of the Enhanced Oil Recovery (EOR) process and CO2 storage monitoring. This work presents several examples to demonstrate the value of applying ML and DA techniques in RM. The main contribution of the dissertation is the investigation of how ML methods can contribute to probabilistic forecasts, uncertainty quantification and sequential decision making in RM. To achieve this goal, we: 1. Show how to integrate and apply Bayesian ML for unconventional oil production forecasting to inform and support RM decision-making; 2. Illustrate and discuss how to use ML dimensionality reduction techniques to support history matching and reservoir model calibration with significant reduction in computing time; 3. Illustrate and discuss how to use the BEL framework to quantify uncertainty in the context of CO2 storage monitoring; 4. Show how to apply simulation-regression method for EOR processes and CO2 storage operations to maximize the value and reliability of the reservoir development plan. We believe that this work is relevant and material in demonstrating the benefits and value creation potential of implementing DA and ML methods to support RM decisions. Although the current implementations may be somewhat simplified, they can serve as a guidance for future research attacking the challenges involved in implementations in real-world settings.