Data Mining Methodologies Enhance Probabilistic Well Forecasting

The assessment of the nation's undiscovered petroleum resources by the US Geological Survey uses the identification of geologically defined plays and the estimation of the size distribution and number of undiscovered fields in each play from historical production data. The Appalachian basin contains six major oil and gas plays, but in contrast to more recently drilled basins, the plays of the mature Appalachian basin lack a complete geologic and production database. Data such as reservoir porosity, permeability, thickness, saturation, drilling depth, and historical production are either unknown, scattered in the literature, deeply buried in company and government files, or exist in incomplete regional government and private commercial data files. In an effort to increase the database and to improve the assessment of the basin, an informal atlas was initiated to test the feasibility of acquiring additional data and the utility of these data in improving the assessment. Critical to the assessment procedure and noticeably incomplete in the basin are the historical cumulative production data necessary for analyzing the distribution of the number of fields and field sizes for the individual plays. Of the Appalachian basin plays, the Oriskany Sandstone play was initially selected and encouraging amount of previously unidentified productionmore » data is available. Achieving the maximum amount of the quality of data for each play in the atlas will require the cooperation of those working in the Appalachian basin.« less

Surface network production flow in the oil and gas industry is a complex process influenced by fluid properties, pipeline configurations, and operational conditions. Traditionally, this estimation is performed using nodal analysis or empirical correlations, both of which rely on physical assumptions and theoretical models. However, these conventional methods often introduce significant assumptions and simplifications that may not fully capture the complexities of real-world production behavior. In this work, a Data-Driven approach was used, where machine-learning models analyze historical production data to learn patterns without relying on physics-based assumptions. Trends and relationships are identified solely from historical data, eliminating the need for complex physical equations or empirical formulas. The objective is to use pure data analytics to build an accurate oil-production prediction model. This study explores the application of machine-learning techniques for oil and gas production forecasting, integrating exploratory data analysis (EDA) and predictive modeling using historical production data from different oil platforms. A Python-Dash dashboard was developed for historical monitoring of key production metrics, featuring dynamic visualization and seamless data integration. Prediction of oil, gas, and water rates was carried out using three regression models, Random Forest, XGBoost, and Support Vector Regression (SVR) with data preprocessing ensuring numerical consistency and null-free values for effective training. Model performance was evaluated using Mean Squared Error (MSE), Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and the R² score, with XGBoost identified as the best-performing model. Predictions on the validation dataset were visualized through line plots, highlighting production trends and fluctuations. The cleaned and forecasted data were exported for further analysis and decision-making. This approach demonstrates the potential of machine learning in optimizing oilfield operations and enhancing resource management through accurate production forecasting.

ProOpter: An advanced platform for production analysis and optimization

We present a robust control system and methodology for physics-informed artificial intelligence (PAI) used to optimize and improve oil recovery, demonstrated in the Yates Field operated by the Kinder Morgan CO2 company. The system consists of a robust control system (referred to as Dynamic Feedback Loop, or DFL) equipped with novel hydrocarbon sensors that measure oil concentration and other parameters continuously and simultaneously on a set of producing wells. The goal of this system is to optimize operational parameters (e.g. choke valve settings, injection rates) to reach specific target metrics of production (e.g. maximizing produced oil while minimizing produced gas).The key element of our approach is the use of a multi-layer artificial neural network (deep neural network, or DNN, to be specific) that extracts physics-based parameters from the real-time measurements and predicts relevant parameters of the DFL control system. DNNs are prone to overfitting in training, making them ineffective in unfamiliar or challenging situations outside the training dataset. To overcome this problem, we have developed a physics-informed robust neural network technique, where the reservoir physics and sensor data are used to train DNN representations of the key physical parameters. Typically, only simplified physical models are developed using available geostatic or historical production data. Also, due to the dynamic nature of these systems, the accuracy of the models often changes over time. To improve predictive capability of the model, we combine the DNN with the system-theoretic robust control concepts based on physics with a model uncertainty formulation. The concept was first validated using a combination of simulations, isolated sensor data and analyses based on sets of historic production data.A study using historic production data on Kinder Morgan's Yates Field Unit (YFU) 4045 Pilot (3 producing wells) indicates application of the DFL system results in an increase in cumulative production of up to 35% per year, compared to what is obtained through a traditional (fixed-point) control system. Currently, the DFL is being field-tested on a different set of wells in the Yates Field, instrumented with the novel hydrocarbon sensors that generate continuous and simultaneous production data.

Understanding how different units of an industrial production plant are operationally related is key to improving production quality and sustainability. Data science has proven indispensable in obtaining such understanding from vast amounts of historical process data. Path modelling is a valuable statistical tool to obtain such information from historical production data. Investigating how relationships within a process are affected by multiple production conditions and their interactions can however provide an even deeper understanding of the plant’s daily operation. We therefore propose conditional path modelling as an approach to obtain such improved understanding, demonstrated for a milk protein powder production plant. For this plant we studied how the relationships between different production units and steps are dependent on factors like production line, different seasons and product quality range. We show how the interaction of such factors can be quantified and interpreted in context of daily plant operation. This analysis revealed an augmented insight into the process that can be readily placed in the context of the plant’s structure and behavior. Such insights can be vital to identify and improve upon shortcomings in current plant-wide monitoring and control routines.

RUA field is classified into heavy oil reservoir type due to the high viscosity value and low API degree . This causes the RUA field can not be produced conventionally. the solution of this problem is to apply steam or thermal injection into reservoir which could reduce the viscosity of the heavy oil (Bera Babadagli, 2015). One of the best EOR methods that has been proven to overcome this issue is using CSS method (Suranto et al., 2020). During the production period, the CSS process can affect the viscosity of the oil by increasing the temperature of the oil in the reservoir. In one production well, cyclic work are applied periodically, its called repeated cyclic (J. J. Sheng, 2013). This is because time of reservoir temperature stays above the baseline temperature reservoir shortly. Even though the cyclic already done repeatedly, there is still a decrease of oil production, different peak reservoir temperatures, and found the possibility of pump damage after the cycle job which led to the need for analysis on these issues. The analysis was performed by looking at the historical production data, historical reservoir temperature data, and production pump work data in the RUA field. After a production history data that reprsentative analyzed, it was found that teh production after cyclic there is increasing, and there is also a decline from the previous cyclic production. Based on the results of the production analysis, it was found that 53.24% of the production wells in the RUA field were already in the ramp down stage and 46.75% were already in the ramp-up stage. Meanwhile, the average HET for regular cyclic jobs is 3-4 months and 5-6 months for long cyclic jobs. And from the pump work data, only 3 wells were damaged. This suggests that cyclic stimulation is completely safe to be performed in this field.

The continuous growth of solar power capacity raises challenges to distribution system operators regarding power quality and security of supply. Network management systems must be enhanced with short-term forecasting functionalities able to predict the solar plants production in the next hours or days. The provision of individual forecasts for each solar plant on the network is often required. To that purpose, historical measurements are needed for tuning the forecasting models. The situation is challenging for new plants for which long history of measurements is not yet available. In that case, models able to provide accurate production forecasts based on few historical production data, are required. In this paper, we investigate the performance of state-of-theart short-term PV forecasting models as a function of the historical data available for tuning. We compare the results with those obtained by a reference model whose utilization does not require more than one day of past production data. Our analysis relies on production data from a 200 kWc solar plant located in the south-east of France. It shows that satisfactory performances can be expected from state-of-the-art models, when calibrated with no more than one or two weeks of training data. (4 pages)

This paper presents a new methodology for constraining 3-D stochastic reservoir models to well and historical production data. Based on the inversion framework, this methodology is mainly composed of a new algorithm for gradually deforming continuous geostatistical models and of an efficient optimization algorithm. The gradual deformation algorithm generates a process of realizations which evolve smoothly while preserving the overall statistical characteristics of the stochastic model. Deformation control by a limited number of parameters improves the efficiency of the process. In comparison with other approaches, the number of iterations is significantly reduced. The deformation procedure can be coupled with an optimization algorithm to obtain automatic history matching capability. The inversion of several model parameters and the gradual selection of a new constrained realization can be performed simultaneously. The analytical gradient method as well as approximated gradients are used to improve the convergence rate of the procedure. The successful application to a 3-D synthetic case, derived from an actual field case, is described in the paper. The shape of the top structure of the reservoir and the activity coefficient of an aquifer are simultaneously constrained to the historical data recorded for 5 wells. The structural uncertainties are described by a geostatistical model. The definition of the top structure is improved by gradual deformation to ensure the geostatistical consistency of the model during the match of production data. An interference well test case is also presented to illustrate the decrease in uncertainties obtained by integrating a geological description and pressure data. P. 221

PediTrack: Blackberry stores historical data and generates reports and pedigrees in an easily understandable format. Other database programs are available commercially, but are often expensive, broadly-based, or unadaptable for use in specific situations. PediTrack: Blackberry allows a PC user with Microsoft Access® version 2000 or higher to use the simple program without charge. Microsoft Access® is widely available and easily adaptable to a variety of breeding program functions. PediTrack: Blackberry is a genus specific database for blackberry cultivars names and parental information, historical production data, and pedigree output.

For the last decade shale gas has made a significant contribution to the oil and gas market by introducing multi-fractured horizontal wells. Forecasting the production for these wells is a challenging task due to reservoir properties and fracture network complexity created by hydraulic fracturing. Two methods ‘Analytical and simulation’ proposed in this paper to forecast the well performance, and quantify the differences one might expect to encounter in a well’s Estimated Ultimate Recovery "EUR". Analytical models are widely used as tools for history matching and forecasting production in unconventional resources due to the advantages of simplicity and efficiency. On the other hand there is an increasing interest in the use of numerical simulation for unconventional reservoirs to predict well performance and Estimated Ultimate Recovery "EUR". The two Methods used to match and predict the production of hydraulically fractured unconventional horizontal wells, and the results comparison was made to check on the efficiency and quality of these two methods. This study will help to quantify the differences one might expect to encounter in a well’s Estimated Ultimate Recovery "EUR" when using analytical model-based Rate Transient Analysis "RTA" versus numerical simulation-based workflows in unconventional reservoirs. This workflow provides multiple history-matched models that are constrained by historical production data. The probabilistic forecast method employed in this work provides P90 (conservative), P50 (most likely), and P10 (optimistic) values for EUR. All the results obtained with the numerical method were compared to the analytical, the findings between these two methods will help to know which most reliable method to be used in predicting the future wells performance to Estimated "EUR". The operators looking for fast and reliable answer to predict the well performance for increasing the success of the models used in the completion, and help for better business planning. This study will help us to know what the most consistent method to use along with hydraulic fractures properties.

The light oil bearing, low permeability, Spearfish formation, in Southwest Manitoba, Canada, has been a recent target of extensive horizontal drilling and fracture stimulation. An extensive completion database, along with historical production data, is available in the public domain. It is beneficial to review what the production trends illustrate about this play, in order to optimize the development of other similar types of low permeability sandstone plays, with regards to horizontal well spacing and fracture design. Decline analysis was performed on over 120 horizontal wells that had sufficient completion information available, to study the effects of varying fracture spacing and size. The emphasis was to review the longer term production histories of the oil wells, as well as the predicted ultimate reserves, and see if there were any discernible trends, when production was compared to various fracture parameters. Both the initial production rates and the ultimate reserves, showed significant data scatter when compared with fracture size, and number of fractures per well. It was difficult to make strong conclusions about initial production rates, being a function of varying fracture parameters. Closer analysis of individual wells, suggests that ultimate reserves appears related to the moveable oil-in-place, within the drainage area that a horizontal well produces from. It appears that well spacing effects and localized reservoir properties, must be taken into consideration, when analyzing the long-term performance of multi-fractured horizontal wells. It seems that caution is required, when drawing long-term conclusions from initial production results. The basic reservoir engineering principle of an oil well draining a defined reservoir volume, are still applicable to low permeability reservoirs. There appears to be a finite number of horizontal wells that can be drilled in any given area, and an optimum number of fracture stimulations that can be placed per well.

Steady-state process models are common in virtual flow meter applications due to low computational complexity, and low model development and maintenance cost. Nevertheless, the prediction performance of steady-state models typically degrades with time due to the inherent nonstationarity of the underlying process being modeled. Few studies have investigated how learning methods can be applied to sustain the prediction accuracy of steady-state virtual flow meters. This paper explores passive learning, where the model is frequently calibrated to new data, as a way to address nonstationarity and improve long-term performance. An advantage with passive learning is that it is compatible with models used in the industry. Two passive learning methods, periodic batch learning and online learning, are applied with varying calibration frequency to train virtual flow meters. Six different model types, ranging from data-driven to first-principles, are trained on historical production data from 10 petroleum wells. The results are two-fold: first, in the presence of frequently arriving measurements, frequent model updating sustains an excellent prediction performance over time; second, in the presence of intermittent and infrequently arriving measurements, frequent updating in addition to the utilization of expert knowledge is essential to increase the performance accuracy. The investigation may be of interest to experts developing soft-sensors for nonstationary processes, such as virtual flow meters.

This year’s edition of the SPE Liquids-Rich Basin Conference in Midland, Texas, merged topics that have dominated recent industry discussions: technologies centered on big data, interwell communication, and Permian Basin production. Given its location in “one of the hottest shale plays in the world right now,” the September event was “more accessible to those most directly working” in the basin, including committee members, presenters, and attendees, said Amber Voithofer, chair of the conference’s program committee. Presentations involved several big Permian operators, top oilfield service firms, and engineering and technology companies. “There is a real focus in the industry to start doing more with all the data we have been collecting,” Voithofer said. “We have all of this data and really have not utilized it to its full potential.” As a result, further discussion in the coming years will revolve around machine learning, artificial intelligence, and predictive analytics. To address these advancements in an unconventional context, the conference was weighted toward the application of data analytics on local data sets and knowledge gained from modeling on themes such as parent-child well interaction, bashing, and sequencing. Special attention was paid to well spacing and well performance analysis methods. Two presentations that sought to provide simpler answers to complex problems generated much interest. In SPE 191777, Schlumberger “challenged the status quo” on how estimated ultimate recovery is calculated, Voithofer noted. “Instead of fitting a curve through all the historic production data and extending it into the future, the paper proposed we match segments of well production, taking into account changes after well interventions.” In SPE 191783, Apache offered a glimpse into how operators are attempting to manage uncertainties around well spacing. “The Permian is hot, and as companies continue to invest in development out here, I expect we will see more novel processes or technologies coming out of the basin,” Voithofer said, adding that there’s another liquids-rich basin to watch going forward: the Powder River.

A real-time decision-making of maintenance using fuzzy agent

Historical production data from November 2021 to February 2022 revealed that polybox plastic containers suffered from a 2.84% defect rate, exceeding the company’s 2% threshold. This study applies the Six Sigma DMAIC (Define, Measure, Analyze, Improve, Control) methodology to enhance product quality. The Define phase utilized SIPOC and Critical to Quality (CTQ) diagrams to outline quality characteristics. In the Measure phase, P and U Control Charts established a baseline performance with a Defect Per Million Opportunities (DPMO) of 8,420 and a sigma level of 3.89. The Analyze phase employed Fishbone diagrams and Failure Mode and Effects Analysis (FMEA) to isolate root causes. High Risk Priority Numbers (RPN) highlighted worker carelessness (294), welding inaccuracies (252), and manual handling issues (210). Consequently, the Improve phase implemented new Standard Operating Procedures (SOPs), welding guidelines, and hand-truck trolleys. Post-implementation monitoring during the Control phase showed immediate improvements: the DPMO dropped to 6,305, and the sigma level increased to 3.99. These findings confirm the efficacy of Six Sigma interventions in reducing manufacturing defects and aligning production with quality standards.