With the development of horizontal well and volume fracturing technologies as well as the conventional oil reserves depleting, the vast tight oil discovered in the world play a important role in source of energy supply. However, it is challenge to effectively forecast the recovery efficiency and the traditional knowledge and methodologies are not suitable for tight oil reservoirs.
To better forecast the recovery efficiency and exploit the tight oil reservoirs with lower expense, a variable weight combination model of XGBoost-SVR has been proposed based on data mining in this paper. It can be acquired by combining the Limit Gradient Climbing (XGBoost) and Support Vector Regression (SVR) and using the method of residual analysis. This model establishes the relationship between recovery efficiency and essential parameters including geography, formation and engineering by machine learning analysis.
The predictive results have shown that the cluster number of fracture, horizontal length, sand intensity and permeability play a significant role in recovery efficiency. However, the well spacing, liquid intensity and reservoir temperature have an weak effect on recovery efficiency. The accuracy in predicting recovery efficiency for combination model of XGBoost-SVR can reach as high as 94.63%, which is higher than those of single XGBoost model and SVR model. Therefor, the predictive model based on XGBoost-SVR can be used as a feasible tool for economic evaluation.
The methodology, predictive model and predictive results demonstrated in this paper will have a great and novel effect on development of tight oil. This study gives an insight on the production mechanism for tight oil reservoir from a big data mining perspective, as well as a feasible and accurate method to forecast recovery efficiency. The methodology and model established in this paper can be easily applied to other unconventional oil reservoirs.

In the past decade, machine learning and deep learning have been popular tools for well production forecasting since these black-box approaches can bypass the incomplete understanding of physics while obtaining satisfactory prediction results given a considerable amount of data. However, due to their large data requirements, inability to produce physically consistent results, and their lack of generalizability to out-of-sample scenarios, pure machine learning approaches cannot meet the requirements for predictive performance and generalization ability in complex production forecasting problems.
Especially in tight oil and shale gas fields, the post-fracture production mechanisms of oil and gas wells have been challenging because of the application of horizontal drilling and hydraulic fracturing techniques. Typical deep learning models are generally scenario-specific and they may fail to capture relationships behind post-fracture production directly from limited observation data, leading to their failure to generalize to scenarios not encountered in training data.
In this work, we propose a new workflow for post-fracture production profile forecasting that could constrain production prediction profiles within known physics laws for more robust results, even beyond the training set. To this end, we use a state-of-the-art deep learning method called Physics-informed Long Short-Term Memory (PI-LSTM) networks. In this approach, PI-LSTM models are trained with additional production physics constraints of decline curve analysis incorporated into the loss function.
Our goal is to impose appropriate physics constraints on the LSTM networks and also explore the mathematical constraint patterns. In this way, we can take advantage of the complementary strengths of machine learning and corresponding physics equations to obtain more reliable and robust prediction results. Furthermore, this grey-box approach may provide us with more insights into the production behavior of fractured wells that cannot be simply described by simplified physics or empirical equations.
We conduct comparison experiments on two cases to show the forecasting ability of the proposed PI-LSTM approach, conventional decline curve analysis methods and deep learning models (i.e. RNN and LSTM). The results show that the PI-LSTM model has the smallest prediction errors both in the simulated case and in the actual field case.

The application of different methods of machine learning in the oil and gas industry is becoming relevant. The data–driven approach makes it possible to build excellent oil prediction models to increase oil recovery. This article discusses machine learning algorithms for solving the problem of polymer injection into an oil reservoir. The problem of supervised learning, which is one of the classes of machine learning problems, is considered. Our problem belongs to the class of regression problems in terms of machine learning methods. The considered generated data from the mathematical model were used for the training and test set. To build a machine learning model, such parameters of the oil production problem as porosity, viscosity of the oil phase, polymer injection ratio, absolute rock permeability and oil recovery factor were used. The value of the oil recovery factor was chosen as the output parameter of the regression model. Over 350 thousand generated data were applicated to implement multiple regression methods. Different regression algorithms were developed and tested, and it was also found that for our synthetic data, the considered models train quite well and predict the value of the oil recovery factor. An artificial neural network with multiple hidden layers with optimally selected hyperparameters was built. The hyperparameters of the artificial neural network were optimally selected. To prevent overfitting, the early stopping function was used, where training stops at the number of epochs without improvement. The tensorflow deep learning library was used to implement regression algorithms for predicting the oil recovery factor. In this way, it is supposed that the data-based different regression algorithms reviewed in the article can be valuable for predicting the oil recovery factor using engineering data from diffrerent oil fields during processing stages.

In recent years, Physics Informed Neural Networks (PINNs) have generated considerable interest in the scientific computing community as an alternative, or potential contender, to traditional numerical discretization methods such as finite- difference, volume, and element methods, for solving partial differential equations (PDE). In PINNs, the governing equations are incorporated into a loss function as a regularization term to guide the neural network so that the solution respects the underlying physics, hence the name ‘physics informed’. In this work, we investigate the performance of PINNs in solving the highly anisotropic diffusion equation that models fluid flow in subsurface porous media. Several levels of permeability anisotropy are tested. The results show that PINNs have excellent performance when the solution is smooth regardless of the strength of permeability anisotropy. However, PINNs struggle to give adequate results when the solution has large gradients, for example, when the solution is induced by a concentrated source term. The problem is exacerbated by higher levels of permeability anisotropy. Our results highlight a few limitations in the current implementation of physics-informed neural networks for fluid flow in porous media and show that we still have ways to go before it can compete with traditional numerical methods.

Reduced-order modeling (ROM) has been used to simulate subsurface flow in porous media for decades. With recent advances in machine learning and deep learning methods, new ROMs have been presented in the literature. In this work, we present extensions to the embed to control-based (E2C) models limited to two-dimensional (2D) reservoir models to three-dimensional (3D) reservoir models.
E2C-based models are built to mimic the ROM known as proper orthogonal decomposition trajectory piecewise linearization (POD-TPWL) by using blocks of neural networks, where two of the blocks; namely encoder and decoder, are used to transform back and forth from the space of system states to a low-dimensional space and a transition block that predicts the evolution of system states linearly in the null-space. This framework predicts the system state variables such as pressure and saturation across the reservoir, and system outputs such as rate and pressure at production or injection wells are computed by using the predicted state variables in explicit well model equations for the E2C model. The other E2C-based proxy model, which is referred to as embed to control and observe (E2CO), can predict system outputs directly by using another network block called transition output and does not require explicit well-model equations. We upgraded the existing E2C and E2CO models by using 3D convolution layers and modified the loss functions to address the 3D flux conditions.
The proposed methods are tested using a small and a large portion of the SPE10 benchmark reservoir model with channelized heterogeneous permeability for a waterflooding scenario. The small-scale model contains 14,400 cells and 8 wells whereas the large-scale model contains 528,000 cells and 53 wells spread across the reservoir in a 5-spot pattern. 300 simulations from a commercial high-fidelity simulator (HFS) are generated to train the proxy models. Both the E2C and E2CO provide accurate estimates of the state variables with acceptable errors when compared with the test data obtained from HFS for both the small-scale and large-scale reservoir models. It is observed that the well output predictions made by the E2CO are more accurate than the predictions of the E2C. Compared with HFS, these proxy models result in several orders of magnitude faster forward predictions.