Search

Summary In this paper we present a new data-driven methodology for a drilling bit position and direction determination. The model is based on machine learning approach and trained on a data collected in a real-time or near real-time: mechanical parameters of drilling, tool-face data, MWD/LWD data, etc. The proposed methodology might be an interest for directional drilling service companies, operator companies that develop low-thickness productive strata. One of the main advantages of the proposed approach is economic efficiency which it provides due to absence of additional costs associated with payments for additional man hours for precise trajectory and direction monitoring. Methodology allows to predict trajectory at any time of drilling. The methodology is illustrated on the historical data of drilling of one oilfield. At the current stage, the results of the testing show good quality. Blind test on 154 independent sliding episodes shows that median absolute error (MedAE) of depth, inclination and azimuth are 0.26 m, 0.25° and 0.42°. These errors will decrease after adding more wells and steps, which are described in future plans.

In this work we present a methodology for optimal management of brownfields that is illustrated on a real field. The approach does not depend on the particular reservoir flow simulator used although streamline-derived information is leveraged to accelerate the optimization. The method allows one to include (nonlinear) constraints (e.g., recovery factor larger than a given baseline value), which are very often challenging to address with optimization tools. We rely on robust (derivative-free) optimization combined with the filter method for nonlinear constraints. It should be noted that the approach yields not only a feasible optimized solution but also a set of alternative infeasible solutions that could be considered in case the constraints can be relaxed. The whole procedure is accelerated using streamline-derived information. Performance in terms of wall-clock time can be improved further if distributed-computing resources are available (the method is amenable to parallel implementation). The methodology is showcased using a real field in West Siberia where net present value (NPV) is maximized subject to a constraint for the recovery factor (RF). The optimization variables represent a discrete time series for well bottomhole pressure over a fraction of the production time frame. An increase in NPV of 7.9% is obtained with respect to an existing baseline. The optimization methods studied include local optimization algorithms (e.g., Generalized Pattern Search) and global search procedures (e.g., Particle Swarm Optimization). We provide solutions with different levels of approximation and computational efficiency. Without the acceleration achieved through streamline-derived information, the method, while effective, could be prohibitive in many practical scenarios. It is worthwhile noting that part of the solution determined in this work has been tested out on the real field. Optimal management of brownfields is typically addressed using bottomhole pressure values or rates as well control variables. Well controls given as bottomhole pressure values, although not directly implementable in the real field, are often much easier to put into practice than if they are given as rates. However, optimization algorithms that deal with well rates as control variables can be in many cases computationally faster than methods based on bottomhole pressure values. In this work we combine the two aforementioned desirable features for the optimal management of mature fields: well controls are given as bottomhole pressure values for a more practical implementation, and these values are also determined efficiently using concepts borrowed from optimization via well rates.

Summary In this paper, we presented an approach to building a machine learning model for predicting well productivity index. The proposed approach is based mainly on LWD data and well log interpretation results, based on the petrophysical model of the oilfield and digital signal processing approaches. The proposed approach was tested on historical data from the Novoportovskoye oilfield. The model was tested based on the LOOCV cross-validation process. As a result, the median relative error over wells is less than 20%.

Summary A lot of anomalies can occur and lead to failures during drilling process. It is crucial to detect these deviations from normal process as soon as possible, so engineers can analyse and decide what activities to take in order to prevent potential NPT. In this work we propose a new machine learning based approach for detection abnormal drilling behaviour in an online manner. The idea is to cluster drilling data, which is preprocessed in a very special way. Our aproach allows using all available data for training as it does not need any labeled data and incorporates both raw drilling parameters and expert knowledge, thus enhancing prediction results.

During the directional drilling, a bit may sometimes go to a nonproductive rock layer due to the gap about 20m between the bit and high-fidelity rock type sensors. The only way to detect the lithotype changes in time is the usage of Measurements While Drilling (MWD) data. However, there are no general mathematical modeling approaches that both well reconstruct the rock type based on MWD data and correspond to specifics of the oil and gas industry. In this article, we present a data-driven procedure that utilizes MWD data for quick detection of changes in rock type. We propose the approach that combines traditional machine learning based on the solution of the rock type classification problem with change detection procedures rarely used before in the Oil\&Gas industry. The data come from a newly developed oilfield in the north of western Siberia. The results suggest that we can detect a significant part of changes in rock type reducing the change detection delay from $20$ to $1.8$ meters and the number of false-positive alarms from $43$ to $6$ per well.

This paper is devoted to the testing of of automatic well logs interpretation testing, developed on the basis of machine learning methods. The basis of the method presented in the paper is recurrent artificial neural networks. For their training and adjustment, log curves and their corresponding interpretation of different years are used. The set of well data is divided into training, validation and test samples. The resulting tool is set up on training and validation samples, and then used on a test sample of wells for which the interpretatoin was hidden, in order to automatically predict net pays and compare the results with the interpretation performed by an expert petrophysicist. For the test sample, traditional machine learning metrics metrics and special geological were calculated to assess the quality of the algorithm. During the work a number of experiments were carried out, in which the dependence of the forecast quality was estimated not only on the different architecture and settings of the artificial neural network, but also on the amount of input data. The iterative approach in the research allowed to determine the best parameters for the solution of the task. For each well of the test set, a forecast of reservoir intervals distribution is made. The resulting interpretation shows high accuracy, both in terms of different mathematical metrics, and the results of analysis and evaluation of the expert petrophysics. Also, during the experiments, an important conclusion was made about the generalizing ability of the proposed methodology. The use of several variants of interpretation of well log data, performed by different specialists at different times and on the basis of different petrophysical models, allows to generalize and use all the accumulated experience of well logs interpretation, thereby improving the quality of the forecast. The main conclusion of the study can be considered a statement about the efficient applicability of machine learning algorithms for automatic well logs interpretation.

Summary In this paper, we present a methodology for determining lithological difference at the bottom of the well during drilling operations. Our approach is based on the analysis of mechanical parameters of drilling. These parameters are receiving as real-time time-series data. The central part of the methodology is a model based on the machine learning approach. Our model and the whole methodology can be applied to real drilling cases. The set of parameters that are required for the methodology can be collected from the typical mud logging station. The main use case for the methodology is an optimization of the geosteering process. The most modern geosteering approaches are based on the LWD data. It is the main restriction of common approaches for the adjustment of the direction of drilling. The problem is that the LWD sensors are placed for a few decimals meters before the bit in a typical Bottom Hole Assembly (BHA) design. As a result, these a few tens of meters are drilling in a "blind window". The methodology is illustrated on the historical data of drilling of the Novoportovskoe oilfield. At the current stage, the results of the testing show that suggested methodology can correctly classify two out of three cases of changes of lithotypes while drilling.

Extra net thickness may bring a huge impact on projects NPV, especially in case of brownfields with vast production wells stock and maintained surface infrastructure. Reservoir beds with sand may be misinterpreted by petrophysicist within a well and miscorrelated spatially. We propose statistical learning methods to identify missed reservoir beds and therefore extra net thickness by predictions of supervised model. Robustness analysis of such identification is the main purpose of our paper. Methodology is tested on 3 brownfields in Western Siberia along with computational experiments with digital outcrop model, representing complex fluvial facies sedimentology. All the three brownfields represent different geological environment and have significant production history. Digital outcrop model is used primarily as a benchmark for different statistical learning algorithms. The main idea behind extra net thickness identification within vertical scale is to train the model on manual interpretation (reservoir/non-reservoir, binary classification) and perform predictions on validation wells. False positives errors give potential reservoir intervals, which were not identified in manual interpretation. Such candidates are evaluated by an expert and validated on production data through perforation. Recurrent neural network is chosen as the baseline algorithm for the methodology. The choice was made according to benchmark testing of different approaches (including Bayesian networks, support vector machines and others) and according to sensitivity analysis of training error for different size of training set (amount of wells). Although RNN gives high accuracy of prediction, this approach still need improvements in term of interpretability and generalization for brownfields covering regions with high variation of geological properties. Feature engineering includes augmentation and creating synthetic curves in case of absence of some significant well log. Missing or noisy well logs were reconstructed based on logs not only from a particular well but also on logs from its neighbor wells. Using of data from neighbor wells as additional features showed dramatic improvement of synthetic log quality. Robustness of a spatial forecast examined in the presented paper was dependent on a number of neighbor wells taken as features and search window size within a particular well. Evaluation of forecast accuracy was done not only by statistical but also by geological metrics such as compartmentalization and net-to-gross ratio. According to the experiments presented in this paper the optimal vertical window is around 1 meter thick, collected from 5 neighbor wells.

Directional oil well drilling requires high precision of the wellbore positioning inside the productive area. However, due to specifics of engineering design, sensors that explicitly determine the type of the drilled rock are located farther than 15m from the drilling bit. As a result, the target area runaways can be detected only after this distance, which in turn, leads to a loss in well productivity and the risk of the need for an expensive re-boring operation. We present a novel approach for identifying rock type at the drilling bit based on machine learning classification methods and data mining on sensors readings. We compare various machine-learning algorithms, examine extra features coming from mathematical modeling of drilling mechanics, and show that the real-time rock type classification error can be reduced from 13.5% to 9%. The approach is applicable for precise directional drilling in relatively thin target intervals of complex shapes and generalizes appropriately to new wells that are different from the ones used for training the machine learning model.

In this paper we present an enhanced and new optimization algorithm based on Particle Swarm Optimization (PSO) that can be used in reservoir characterization applications to determine multiple solutions that reproduce measurements within satisfactory accuracy. The method is illustrated on a history-matching problem where significant speed-up is obtained with respect to a standard PSO implementation. PSO is a global-search iterative technique where the solution space is explored using a swarm of particles. In every PSO iteration each particle is individually attracted to: the best point (in terms of objective function) in the swarm during last iteration, the best point visited by that particle in all iterations, and the best point found by any particle in all iterations. In the PSO variant proposed here, rather than single points, collections of points carefully selected are used as attractors. The algorithm ensures by a series of mechanisms that the number of attractors is adequate and that the solutions progressively determined are sufficiently different from one another. A series of experiments were performed with a synthetic model based on a real oil field for which nine years of historical data were generated. Both standard PSO and the PSO variant were tested to obtain several models that reproduced well rates within acceptable accuracy. In order to determine multiple solutions via standard PSO (which usually returns only one solution), a number of independent runs with different initial conditions were considered (i.e., standard PSO was run in multi-start fashion). In our experiments, which had a budget of 3,072 simulations in the two cases, the number of solutions found on average by standard PSO and by our PSO variant was 8 and 837, respectively. That is, multi-start standard PSO may require, on average, around 100 times more simulations than our PSO variant to compute a comparable number of models (possibly) distributed in similar manner in the solution space. These results may be explained by noticing that the search for multiple solutions may be more efficient when all information is considered simultaneously than separately by means of a number of independent runs. Standard PSO, when applied to many uncertainty quantification problems, may provide only one solution because the particles in the swarm generally tend to converge to essentially a single point. The Ensemble Kalman Filter (EnKF), which has lately become a very popular history-matching algorithm, presents a similar problem because the ensemble sometimes collapses into one solution. The PSO variant introduced in this work can be an efficient alternative to these algorithms in reservoir characterization since it has been designed to provide multiple and different solutions in a single run.

In this paper we present a methodology of automated optimal (in terms of NPV maximization) greenfield development strategy construction that takes into account geological uncertainties of the reservoir. The algorithm is based on ensemble of hydrodynamic models and uses the theory of probabilistic graphical models. The proposed methodology allows making dynamic (in the process of field development) decisions about well grid pattern change on the basis of information obtained from the studies on the oil reservoir carried out with the help of the appraisal wells: appraisal boreholes of the development wells and pre-drilling wells (drilling from current pad to the next one for refining geological targets). The decision of the wells pattern option change does not require any additional calculations. The proposed methodology was applied for creating optimal development strategy for one of the new assets (on production stage) of JSC "Gazprom neft". The reservoir development is complicated by the set of uncertain parameters related to the complex heterogeneous structure of the collector. The parameters have very high influence on the optimal decision choice. As result expected NPV is increased by 15%.

The paper describes an approach to selection of horizontal well placement. Given a field dynamic model we use advanced optimization techniques to select horizontal well length, well placement, well control that improve the field economics (as measured by the net present value, NPV) and increase field recoveries. Well-known problems of using optimization algorithms for field development are: (1) big number of variables, simplest well description assumes 5 variables: position, lateral length, orientation, bottom-hole pressure; (2) computational complexity of hydrodynamic simulation. In order to deal with high dimensionality and computation complexity we propose multi-layer approach. First, we decompose optimal well placement and control task of high dimension into a number of optimization problems of lower dimension: selection of optimal well pattern, local well placement optimization, selection of well control. That allows significantly decrease a number of simulation runs. Second, we use chain of simulators and dynamic models (from analytic models to fine-scale hydrodynamic model). The optimal solution for well placement obtained on simple model is used as input (baseline) solution for more complicated models. Thus we reduce number of complicated model runs. The proposed approach was implemented in experimental software that automatically optimize horizontal well placement and well control using Eclipse dynamic model. We use this software to optimize FDP of a JSC "Gazprom neft" greenfield. The optimization resulted in significant increase of the field NPV and increase of expected recovery factor: 14% and 1.7% respectively. We estimate that optimization techniques contribution is 9% of NPV and 0.4% of RF.

We demonstrate dramatic performance improvements for 3D finite difference RTM using domain partitioning over thousands of compute nodes, processing a shot for a 512x512x512 models with about 3000 time steps in under 30 seconds. We have tested this methodology on velocity models of up to 2048x2048x2048 elements in size. In addition to performance advantages, the demonstrated method benefits from substantially easier programmability because it runs on a homogenous supercomputer using standard C and MPI. It therefore is much easier to maintain, port, tune and extend to novel algorithms codes developed for heterogenous systems such as CPU GPUs. This approach has a high affinity for existing methods, such as reservoir modeling, which are already processed today using domain partitioning and therefore makes merging seismic imaging and reservoir modeling on a single system much easier. And finally, since this methods runs entire velocity models instead of subsets for each shot, it can be easily extended to perform multisource RTM, leading to further performance benefits.