In this study, we aim to demonstrate how machine learning can empower computational models that can predict the flow rate of a given well. Given current real-time data and periodic well tests, this new method computes flow rates using data-driven model. The computational model is based on analyzing the relations and trends in historical data. Relational databases include huge amounts of data that have been accumulated throughout decades. In addition, there is a large number of incoming operational data points every second that gives a lot of insight about the current status, performance, and health of many wells. The project aims to utilize this data to predict the flow rate of a given well.
A variety of well attributes serve as inputs to the computational models that find the current flow rate. Artificial Neural Networks (ANN) were used in order to build these computational models. In addition, a grid search algorithm was used to fine-tune the parameters for the ANN for every single well. Building a single unique model for every well yielded the most accurate results. Wells that are data-rich performed better than wells with insufficient data. To further enhance the accuracy of the models, models are retrained after every incoming patch of real-time data. This retraining calibrates the models to constantly represent the true well performance and predict better. In practice, Flow rate prediction is used by production engineers to analyze the performance of a given well and to accelerate the process of well test verification. One of the main challenges in building unique models for every well is fine-tuning the parameters for the artificial neural networks, which can be a computationally intensive task. Parameter fine-tuning hasn't been discussed in previous literature regarding flow rate prediction. Therefore, our unique approach addresses the individuality of every well and builds models accordingly. This high-level of customization addresses the problem of under-fitting in ANN well models.
Machine learning has attracted the attention of geoscientists over the years. In particular, image analysis via machine learning has promise for application to exploration and production technologies. Demands have grown for the automation of carbonate lithology identification to shorten the delivery time of work and to enable unspecialized engineers to conduct it. The image analysis of carbonate thin sections is time consuming and requires expert knowledge of carbonate sedimentology. In this study, the authors propose an image analysis technique based on deep neural network for carbonate lithology identification of a thin section, which is an important image analysis process required for oil and gas exploration. In addition, the authors consider that porosity and permeability variations in the same facies are controlled by the grain, cement, pore, and limemud contents. If the contents are accurately measured, the porosity and permeability can be determined more accurately than by using traditional methods such as point counting. The elucidation of the complex relation of porosity and permeability is the objective of automation of carbonate lithology identification. To perform image analysis of the thin section, the authors prepared a data set mainly comprising pictures of the Pleistocene Ryukyu Group, which were composed of reef complex deposits distributed in southern Japan. The data set contains 306 thin section pictures and annotation data labeled by a carbonate sedimentologist. The rock components was divided into four types (grain, cement, pore, and limemud). A convolution neural network (CNN) was utilized to train the model. After training the neural network, each of the four categories was interpreted by the trained model automatically. Resultantly, the accuracy of automatic Dunham classification was 90.6% and the mean average test accuracy of category identification was 83.9%. The interpretation seems highly consistent between human vision and machine vision in both the overview and pixelwise scales. This result indicates that it has sufficient potential to assist geologists and become a basic tool for practical applications. However, the accuracy of category identification is still insufficient. The authors believe that the model requires higher quality supervised data and a greater number of supervised data.
Han, Chao (China University of Petroleum East China) | Guan, Zhichuan (China University of Petroleum East China) | Li, Jingjiao (China University of Petroleum East China) | Hu, Huaigang (China University of Petroleum East China) | Xu, Yuqiang (China University of Petroleum East China)
A hybrid methodology is presented to predict equivalent circulating density (ECD), which combines autoregressive integrated moving average (ARIMA) and back propagation neural network (BPNN) models. Research results are compared to previously published ECD prediction method that is based on theoretical calculation of hydraulic parameters.
The hybrid methodology is based on data analysis theory. It uses ARIMA model to capture the linear trends of ECD, and then the BP neural network is used to predict the nonlinear and stochastic change law of ECD. Finally BP neural network prediction results are used to correct the prediction error of ARIMA to get the ECD prediction results. With a deepwater well in the South China Sea, a simulation experiment is carried out to verify the comprehensiveness and accuracy of the method presented in this paper.
The prediction results are similar to those of the traditional hydraulics model, which considers the effect of temperature and pressure on drilling fluid density and rheological parameters. Comparisons are also provided for three classical time series prediction models, including the support vector machine, multiple linear regression and grey prediction. The root mean square error (RMSE) and the mean absolute deviation (MAD) were selected as evaluation indicators. The comparison results show that the hybrid model can reflect the variation law of ECD more accurately. Therefore, relative to the traditional methods of prediction, the hybrid methodology has the advantages of advanced modeling thought and simple operation, and it can be selected for prediction of ECD.
Because of the effects of high temperature, high pressure and uncertain factors, the accurate prediction of ECD is very difficult. This paper provides a novel idea for accurately predicting ECD by analyzing the implicit relationship of ECD series data through data mining.
Ma, Kang (China University of Petroleum-Beijing) | Jiang, Hanqiao (China University of Petroleum-Beijing) | Li, Junjian (China University of Petroleum-Beijing) | Zhang, Rongda (China University of Petroleum-Beijing) | Zhang, Lufeng (China University of Petroleum-Beijing) | Fang, Wenchao (Sinopec Petroleum Exploration and Production Research Institute) | Shen, Kangqi (China University of Petroleum-Beijing) | Dong, Rencheng (University of Texas at Austin)
For mature oilfields which have entered into the high water cut stage, many stimulation measures are adopted in order to maintain the oil production. Those measures include drilling new wells, general measures, and strengthened measures. Even though the oil production increase when the measures conducted, it will cause different degrees of production decline in the next year. Due to the irrational composition of oil production in the matured field, abnormal production decline is becoming the primary problem for stable production. Establish an effective early warning system (EWS) is important to release production alarm and take necessary measures in advance. In this paper, the factors that can affect the abnormal decline are selected and the influence degree of different factors are compared by grey relational analysis. The machine learning was adopted to build the EWS. Three distinct forms of input data are considered to improve the prediction accuracy. By using the degree of deviation from normal as the input data for the prediction model have the highest accuracy. However basic machine learning model contains many input parameters which can't obtain easily. The number of input parameter is optimization based on the variation of accuracy under different input parameter number. In order to improve the prediction accuracy the artificial samples are added into the training process. The prediction accuracy of the final optimization model can reach 92%. According to the EWS the production condition of different reservoir is evaluated. The result reveals the possibility of the occurrence of anomalous decline in different reservoir which can effectively guide the oilfield production strategy. The EWS can be an effective tool in the oil production monitor in the mature oil field.
Shoeibi Omrani, Pejman (TNO) | Vecchia, Adrian Luciano (Wintershall Noordzee B.V.) | Dobrovolschi, Iulian (TNO) | Van Baalen, Thijs (Wintershall Noordzee B.V.) | Poort, Jonah (TNO) | Octaviano, Ryvo (TNO) | Binn-Tahir, Huzaifah (Binn-Tahir Consulting) | Muñoz, Esteban (Wintershall Noordzee B.V.)
Reliable forecasting of production rates from mature hydrocarbon fields is crucial both in optimizing their operation (via short-term forecasts) and in making reliable reserves estimations (via long-term forecasts). Several approaches may be employed for production forecasting from the industry standard decline curve analysis, to new technologies such as machine learning. The goal of this study is to assess the potential of utilizing deep learning and hybrid modelling approaches for production rate forecasting. Several methods were developed and assessed for both short-term and long-term forecasts, such as: firstprinciple physics-based approaches, decline curve analysis, deep learning models and hybrid models (which combine first-principle and deep learning models). These methods were tested on data from a variety of gas assets for different forecasting horizons, ranging from 6 weeks to several years. The results suggest that each model can be beneficial for production forecasting, depending on the complexity of the production behavior, the forecasting horizon and the availability and accuracy of the data used. The performances of both hybrid and physical models were dependent on the quality of the calibration (history matching) of the models employed. Deep learning models were found to be more accurate in capturing the dynamic effects observed during production - this was especially true for mature fields with frequent shut-ins and interventions. For long-term production forecasting, in some cases, the hybrid model produced a greater accuracy due to its consideration of the long-term reservoir depletion process provided by the incorporated material balance model.
A methodical rock texture characterization of core samples and cuttings can provide powerful information that can be used reliably and cost-effectively to optimize fracture stimulation designs by placing fracture stages based on rock characteristics and drill bit selection. This paper presents a method to quantify rock texture using a developed integrated digital image processing to the neural network segmentaion for the analysis of the core-plug thin section.
The standard routine uses the binarization process that converts the colored (RGB; Red Green Blue) core-plug thin section image into a set of binary images corresponding to different mineralogy content using the R, G, and B pixel values for thresholding. The idea is that the threshold values must accommodate all different colors that correlate to the rock texture and pore and help in creating a different mask (binary image) for each content type based on the threshold values used. The collection of the binary images for the different mineralogy will help in the rock texture interpretation. However, the neural network method used the data for multiple thin section images to learn based on the pattern of the color difference between pores and matrix. All results are validated with x-ray deflection (XRD).
The proposed integrated system results show that, the thin section used in this work from an oilfield in the southern of Iraq is mainly quartz with minor Kaolinite and other components, where the mineral color code and weight percentages are recorded and validated with XRD. The comparison between the proposed image processing approach, neural network, and the XRD analysis show that good agreement with an acceptable percentage of error. These variations can help in understanding the rock behavior, optimize fracture stimulation designs and selecting the appropriate drill bits for drilling applications. The variations can also help in smart completion designs.
A new method based on deep learning enables the extraction of formation compressional and shear slownesses from raw waveforms acquired by an acoustic tool regardless of its conveyance system or of its hardware configuration (number of axial receivers, waveform sampling rate, or number of time samples). The proposed approach is very fast, fully automated, and suitable for real-time processing workflows at the wellsite.
Over the years, a large collection of acoustic waveforms has been recorded and processed by experts in a variety of environments. In the proposed method, we apply a convolutional neural network (also known as ConvNet or CNN) to learn from previously processed data to estimate acoustic slownesses from raw waveforms. Because we use an algorithm that is originally designed for visual recognition, we transform the raw waveforms into images with enhanced characteristics that are directly associated with the acoustic slownesses that we are aiming to predict. For monopole waveforms, we were able to improve the prediction results using a short-term average/long-term average (STA/LTA) technique that enhances the main arrivals. We then train a CNN model with both the input images and the expected outputs (i.e., slowness values) on a large variety of data covering the main rock environments of interest. The CNN-trained model is subsequently used to estimate the slowness values from unprocessed waveforms never seen by the CNN model.
To test our method on real data, we gathered a collection of acoustic waveforms recorded by several acoustic tools in 20 wells, drilled in different fields across the world. The wells were drilled with different bit sizes (varying from 6 to 17.5 in.), and the compressional slownesses were ranging from 50 to 165 μs/ft and shear slowness ranging from 80 to 600 μs/ft. In total, we used 96,011 data points where each data point consists of a pair of a waveform array and the associated slowness value calculated by an acoustics expert. We then applied the CNN-trained model to a set of waveforms from
In these wells, previously unseen by the ConvNet model, the average absolute error between the slowness estimated using the CNN-trained model and the slowness calculated by an expert was less than 3 μs/ft, which is comparable to the error associated with state-of-the-art processing techniques for slowness estimation. We also discuss how our method can be extended to estimate shear slowness from dipole data using the same ConvNet. The deep-learning-based technique for slowness estimation that we describe can run extremely fast after the ConvNet training process is completed and provides good slowness results without prior information about the tool configuration or the environment in which waveforms are recorded. Because our technique is fully automated, it can also be used as an automatic quality control (QC) flag for wellsite processing and real-time operations.
There are various factors that contribute to the Well planning. Be it cost, complexity of the reservoir, safety measures, isolating problem areas, engineering, etc., The whole process of well analysis takes a lot of time and efforts, as the method used is mostly manual. In this abstract, we explain the Big Data approach used to perform Well Planning which significantly reduces the drilling engineer time. Future, possible suggestions to the existing system are also discussed in this paper.
Over the past few years, the concept of Big Data has gradually matured. Drilling Engineers, Data Management and the Solution Development Team have jointly discussed the potential and the functionality if this concept to optimize the Drilling Engineer performance and timeframe for preparing and delivering drilling programs for the new development wells. EDM (Enterprise Drilling Management System), drilling suite of applications, In-House drilling data extraction tool with Web based interface, and many different types of excel sheets, were considered as the input sources. Well planning process was automated for offset and related data acquisition, results and knowledge were captured in a central repository, which will be used in future endaovers.
Despite tremendous efforts, AI-based predictive maintenance is still not fully exploited in Oil&Gas plants worldwide. The reason mainly relies on the fact that predictive maintenance algorithms need many examples of failures to be trained on, and this is not always the case. For this reason, we developed an efficient unsupervised approach for predictive maintenance, based on deep learning algorithms and applied successfully to predict and anticipate the failures of a coalescer of an Eni's offshore plant.
Our method is based on a Recurrent Neural Network (RNN) autoencoder architecture, coupled with clustering algorithms. The RNN is based on a combination of two algorithmic steps, respectively called encoder and decoder. The encoder reads multivariate chunks of data and summarizes them in a vector of fixed length, named context vector. Then, the decoder brings this context vector and reconstructs the input signals. Once the reconstruction error is minimized, we cluster context vector by choosing an optimal number of clusters and associating them to the operating conditions of the equipment, in particular by distinguishing ‘healthy’ from ‘faulty’ states.
We applied the aforementioned workflow to distinguish the operating conditions of a small equipment in an Eni's offshore plant. This equipment, an electroastic coalescer, suffered repeated troubles during the first phases of plant start-up. We picked up all the sensor measurements available for the coalescer (pressures, levels, temperatures) with very tight sampling (10 seconds resolution) and trained the RNN architecture on 9 months of data. After the application of a suitable clustering method on the context vector minimizing reconstruction error, we were then able to detect up to 5 different operating conditions of the coalescer, associating them to healthy and faulty states of it. In particular, the method was able to authomatically cluster the failures periods of the coalescer, with an advance of around 4 hours before the failures occurred.
‘Effective unsupervised learning – learning without labelled data – remains a holy grail of AI’ (Andrew Ng, WIPO Technology Trends 2019, Artificial Intelligence). We tried to do a step forward in the application of unsupervised approaches to predictive maintenance of industrial equipments by developing an innovative deep learning based method and applying it to a coalescer of an Oil&Gas plant, getting results that are very promising for massive, large scale application in real production settings.
Low rates of penetration (ROP) were experienced in an area with well-known lithology. The vast drilling experience and similarity of drilling conditions in the area, provided the operator with enough data to improve the well schedule and cost performance through the use of machine learning.
Machine learning, specifically artificial neural networks (ANN), is a statistical tool to find relations between multiple inputs. Details that would have been missed or considered outliers by a mathematical model can be accounted for and explained in the ANN model. The ANN was trained on thousands of real time data points recorded from selected wells in a specific depth interval. Typical drilling parameters such as weight on bit, rotary speed, bit hydraulics, lithological properties, and dogleg severity were the input parameters chosen in the model to generate ROP. Once the model was calibrated to historical data, it was used to find the best parameters to maximize ROP.
R squared factors were 0.729 and 0.675 for 12.25 in. and 17.5 in. sections repectively. This was achieved with an ANN structure of 2 hidden layers consisting of 5 nodes each. Sensitivity analysis identified bit hydraulics, weight on bit, and rotary speed as the major parameters impacting ROP. The ROP model was used to conduct a "virtual drill-off test" to identify drilling parameters that maximize ROP. ROP dependency on weight on bit and lithological analysis suggests bit design can be further improved. Bit hydraulics showed that higher flow rate was needed in sections with higher overbalance. Optimum drilling parameters were tested on four wells and resulted in more than 50% higher ROP compared to original field data.
In an industry increasingly dominated by big data, separating the clean data from the "noise" will be a vital topic. This paper aims to provide a blueprint for the use machine learning to optimize ROP in a manner that is simple and easily replicated.