Method for predicting yield of tight oil reservoir based on machine learning

Abstract

As an important unconventional resource in China, the compact oil and gas reservoir is widely distributed and abundant in reserves, and has important significance for improving the national energy safety. However, due to the complex geologic features, a traditional model is difficult to capture the complex nonlinear relation between reservoir parameters and the yield, and then generalization and the poor prediction effect occur. In order to improve the prediction precision, the invention provides a compact reservoir yield prediction method based on machine learning, and a hybrid model of an improved grey wolf optimization algorithm (GWO) and a bidirectional long-short-term memory network-attention mechanism is used. According to the improved GWO, the global search capability is remarkably improved by fusing a group disturbance strategy and an adaptive weighting strategy. Experimental results show that the model has a good prediction effect and shows high robustness to reservoir parameter fluctuation.

Description

Technical Field

[0001] This invention relates to the field of machine learning, specifically to the research on algorithms for predicting the production of tight oil reservoirs in oil fields. Background Technology

[0002] With the widespread application of artificial intelligence (AI) technology across various industries, object detection and recognition have become core tasks in computer vision. The essence of this technology is to accurately locate and identify targets of interest from images or videos, typically defining their positions with bounding boxes and assigning them category labels. With the evolution of deep learning models, especially the breakthroughs in speed and accuracy achieved by detection frameworks such as YOLO and Faster R-CNN, this technology has rapidly moved from laboratory research to large-scale industrial deployment. Today, it is not only widely used for pedestrian and vehicle perception in autonomous driving, defect location in industrial quality inspection, and suspicious target screening in security systems, but also plays a crucial role in emerging fields such as smart retail, medical image analysis, and drone inspection, becoming an important technological foundation for achieving environmental perception and intelligent decision-making.

[0003] In the field of tight oil reservoir production prediction, traditional methods mainly rely on physical models and numerical simulations based on simplified geological assumptions, as well as statistical methods based on empirical formulas. These methods struggle to accurately characterize the complex nonlinear relationship between reservoir parameters and production. With the development of machine learning technology, shallow models such as Artificial Neural Networks (ANNs) and Support Vector Machines (SVRs) have been introduced, but limitations remain, including hyperparameter sensitivity, susceptibility to local convergence, and weak temporal modeling capabilities. In recent years, deep learning models such as LSTM and BiLSTM have better captured long-term dependencies, while attention mechanisms can dynamically focus on key features, and optimization algorithms such as the Grey Wolf Algorithm (GWO) have been used for hyperparameter tuning. However, existing research often fails to systematically integrate these technologies, still facing problems such as insufficient model generalization ability, low optimization efficiency, and unclear feature importance allocation. Therefore, a hybrid intelligent method that integrates optimization algorithms, temporal modeling, and feature enhancement is urgently needed to achieve more accurate and robust production prediction for tight oil reservoirs. Summary of the Invention

[0004] To address the challenges of complex geological parameters, strong nonlinear coupling, low prediction accuracy, and weak generalization ability of traditional models in tight oil reservoir production prediction, this invention proposes a hybrid intelligent prediction model integrating the improved Grey Wolf Optimization Algorithm (IGWO), Bidirectional Long Short-Term Memory Network (BiLSTM), and an attention mechanism. Tight oil reservoirs are characterized by low porosity, low permeability, and strong heterogeneity. Their production is influenced by a complex interplay of factors such as porosity, permeability, formation pressure, temperature, and mineral content. Traditional physical models rely on excessive simplification assumptions, while conventional machine learning methods such as ANN and SVR are prone to getting trapped in local optima and are sensitive to hyperparameters, making it difficult to capture their deep nonlinear dynamic relationships. Therefore, this invention constructs a three-layer collaborative modeling framework: IGW0 is responsible for global hyperparameter optimization, BiLSTM extracts temporal features, and the attention mechanism adaptively weights key geological parameters. This model not only overcomes the limitations of traditional methods in complex oil reservoir systems but also achieves high-precision mapping from multi-source data to production output through end-to-end training, providing reliable data-driven support for the efficient development and dynamic adjustment of tight oil reservoirs.

[0005] The algorithm for predicting the production of tight oil reservoirs in oilfields proposed in this invention includes:

[0006] An Improved Gray Wolf Optimization Algorithm (IGWO): The improved IGWO algorithm, based on the standard GWO, introduces a population perturbation strategy and an adaptive weighting strategy, significantly enhancing global search capability and convergence efficiency. The standard GWO is prone to premature convergence and getting trapped in local optima when optimizing high-dimensional, multi-modal functions. This invention first adds a random perturbation term to the position update, controlling population diversity through a dynamic perturbation factor λ to avoid algorithm stagnation; secondly, it designs an adaptive weight allocation mechanism, dynamically adjusting the contribution weight of an individual in the position update based on its fitness, making the guiding role of superior individuals stronger and accelerating the convergence process. IGWO significantly improves the generalization ability and stability of the BiLSTM-Attention model by iteratively optimizing key hyperparameters (such as learning rate, number of hidden layer neurons, batch size, etc.). Experiments show that IGWO outperforms traditional optimization algorithms in both optimization efficiency and accuracy, providing a solid parameter foundation for subsequent deep learning modules.

[0007] Bidirectional Long Short-Term Memory (BiLSTM) networks serve as the core temporal modeling component, capturing long-term dependencies and dynamic evolution patterns in tight oil reservoir production data. BiLSTM effectively addresses the gradient vanishing / exploding problem inherent in traditional RNNs by capturing the contextual information of time series data through dual-path forward and backward LSTM layers. In tight oil reservoir production prediction, production changes are often influenced by multiple stages of factors, such as historical production status and injection-production strategies, exhibiting significant time lag effects. BiLSTM can memorize key historical states and selectively transmit information through a gating mechanism, thereby achieving accurate modeling of production time-series trends. This invention uses multi-dimensional geological parameters such as porosity, permeability, and pressure as temporal inputs. After extracting high-order features through BiLSTM, it outputs time-aware hidden states, providing rich semantic representations for subsequent attention-weighted processing.

[0008] An attention mechanism is introduced to enhance the model's focus on key geological parameters, enabling adaptive feature selection and weight allocation. Among the extracted temporal features, the influence of different geological parameters on yield varies with time and block, making it difficult for traditional models to distinguish their importance. The attention mechanism calculates the weight coefficients of features at each time step, allowing the model to focus on the feature dimensions that contribute most to the current prediction and suppress interference from irrelevant or noisy information. This invention uses a weighted summation method to reconstruct the BiLSTM output, generating a context vector, which is then mapped to the final yield prediction value through a fully connected layer. This mechanism not only improves the model's interpretability but also maintains strong robustness to changes in data distribution, adapting to the highly heterogeneous characteristics of tight oil reservoirs.

[0009] An intelligent prediction framework with adaptive tuning capabilities is constructed by globally optimizing hyperparameters such as learning rate, number of hidden layer units, and batch size of the BiLSTM-Attention model using the IGWO algorithm. Traditional deep learning models rely on empirical settings or grid search for parameter tuning, which is inefficient and prone to getting trapped in local optima. IGWO uses prediction error as the fitness function, gradually approaching the optimal parameter combination in continuous iterations, so that the model structure achieves the best match with its actual data distribution. The optimized model converges faster during training, significantly improves prediction accuracy, and exhibits good transferability across different blocks. This framework is not only applicable to tight oil reservoirs but can also be extended to other oil and gas reservoir types or time series prediction tasks, demonstrating high engineering applicability and promotional value.

[0010] The experimental part of this invention is based on actual production data from multiple blocks in the Tarim Oilfield, including RMSE, MAE, and R... 2Multiple metrics have verified the superiority of this invention. Compared with mainstream models such as CNN, ANN, and CNN-LSTM, IGWO-BiLSTM-Attention achieved the lowest error and the highest goodness of fit in all three test blocks, with R... 2 The highest value reached 0.9788. Ablation experiments further demonstrated that the contributions of IGWO optimization, the Attention mechanism, and the BiLSTM structure are all irreplaceable, and their synergy significantly improved the model performance. This model can not only accurately predict production trends but also effectively reflect the impact of geological parameter fluctuations on production capacity, providing a reliable technical tool for dynamic analysis of oil and gas fields, adjustment of development plans, and economic benefit assessment, and has significant engineering application prospects. Attached Figure Description

[0011] Figure 1 A flowchart of the GWO algorithm is shown.

[0012] Figure 2 A flowchart of the present invention is shown.

[0013] Figure 3 A flowchart of the improved version of the present invention is shown. Detailed Implementation

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] according to Figure 1 The GWO algorithm mimics the social hierarchy and hunting behavior of gray wolf packs, guiding the entire population towards the optimal solution region through three leader wolves: α, β, and δ. In each iteration, individuals update their own positions based on the positions of these three wolves, and combine a perturbation strategy with an adaptive weighting mechanism to balance global exploration and local exploitation capabilities, gradually approaching the optimal solution. The algorithm stops when a preset termination condition is met (such as reaching the maximum number of iterations or fitness convergence), and outputs the found optimal solution and its fitness value.

[0016] Example 1: A method for predicting tight oil reservoir production based on machine learning is proposed, mainly including the following steps: Data acquisition and preprocessing: Historical production data and related geological and engineering parameters of tight oil reservoirs are collected, including porosity, permeability, formation pressure, temperature, and mineral content. The data is cleaned and normalized, and a time-series sample set is constructed in chronological order. The sample set is randomly divided into training, validation, and test sets according to a predetermined ratio. Model construction and training: A dual-channel IGWO-BiLSTM-Attention prediction model is constructed. This model uses BiLSTM as its core and fully extracts the temporal dependencies in the production sequence through forward and backward loop structures. Based on this, a dual-channel attention mechanism is introduced to guide the cross-attention of deep and shallow features respectively: deep features are corrected by the weights of shallow channels, and shallow features are enhanced by the weights of deep channels, thereby achieving complementarity and enhancement between features and improving the model's ability to represent key geological parameters. Optimization Strategy: An improved Grey Wolf Optimization (IGWO) algorithm is used to globally optimize the network structure parameters of BiLSTM (such as the number of hidden layer units, learning rate, batch size, etc.). IGWO enhances the algorithm's ability to escape local optima by introducing a population perturbation strategy and an adaptive weight update mechanism, thereby improving parameter search efficiency and model generalization performance. Model Validation and Testing: The trained model is tuned using a validation set, and its predictive performance is evaluated using a test set. Root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) are selected. 2 Using this method as an evaluation metric, it is compared with traditional models such as CNN, ANN, and CNN-LSTM to verify its superiority in tight oil reservoir production prediction. Results and Applications: The model ultimately outputs production prediction curves and error analysis results for each well, providing data support for development decisions and production optimization in tight oil reservoirs. This method exhibits strong robustness and adaptability and can be further extended to production prediction for other unconventional oil and gas resources.

[0017] according to Figure 2 Data enters the BiLSTM layer from the input layer. The BiLSTM extracts time-series features through a bidirectional recurrent structure, capturing the long-term dependence of tight oil reservoir production. Subsequently, the feature data is passed to the Attention layer, which calculates feature weights, enhancing the influence of key geological parameters and improving the model's attention distribution. Finally, the optimized features enter the output layer to generate predicted oil reservoir production values. The IGWO layer (improved gray wolf optimization), as a hyperparameter optimization module, does not directly participate in the data flow but optimizes the BiLSTM's hyperparameters (such as the learning rate and the number of hidden layer neurons), improving the model's generalization ability and prediction accuracy.

[0018] Example 2: A method for predicting tight oil reservoir production based on machine learning is proposed, mainly including the following steps: Data acquisition and preprocessing: Historical production data and related geological and engineering parameters of tight oil reservoirs are collected, including porosity, permeability, formation pressure, temperature, and mineral content. The data is cleaned and normalized, and a time-series sample set is constructed in chronological order. The sample set is randomly divided into training, validation, and test sets according to a predetermined ratio. Model construction and training: A dual-channel IGWO-BiLSTM-Attention prediction model is constructed. This model uses BiLSTM as its core and fully extracts the temporal dependencies in the production sequence through forward and backward loop structures. Based on this, a dual-channel attention mechanism is introduced to guide the cross-attention of deep and shallow features respectively: deep features are corrected by the weights of shallow channels, and shallow features are enhanced by the weights of deep channels, thereby achieving complementarity and enhancement between features and improving the model's ability to represent key geological parameters. Optimization Strategy: An improved Grey Wolf Optimization (IGWO) algorithm is used to globally optimize the network structure parameters of BiLSTM (such as the number of hidden layer units, learning rate, batch size, etc.). IGWO enhances the algorithm's ability to escape local optima by introducing a population perturbation strategy and an adaptive weight update mechanism, thereby improving parameter search efficiency and model generalization performance. Model Validation and Testing: The trained model is tuned using a validation set, and its predictive performance is evaluated using a test set. Root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) are selected. 2 Using this method as an evaluation metric, it is compared with traditional models such as CNN, ANN, and CNN-LSTM to verify its superiority in tight oil reservoir production prediction. Results and Applications: The model ultimately outputs production prediction curves and error analysis results for each well, providing data support for development decisions and production optimization in tight oil reservoirs. This method exhibits strong robustness and adaptability and can be further extended to production prediction for other unconventional oil and gas resources.

[0019] according to Figure 3This invention proposes an improved gray wolf optimization algorithm (IGWO) that is tightly integrated with a BiLSTM-Attention model, forming a two-stage "optimization-training" framework. First, the network structure and hyperparameter search space are initialized, and a gray wolf population is randomly generated. In the optimization phase, a leader individual is selected by calculating fitness, and a population perturbation strategy is introduced to avoid premature convergence. Adaptive weight updates are combined to improve search efficiency, and the process iterates until the termination condition is met. The optimal hyperparameter combination (such as learning rate and number of hidden units) is output during the optimization phase. Then, the training phase begins, where the optimized hyperparameters are used to fully train the BiLSTM-Attention model, allowing it to fully learn the complex relationship between production time-series features and geological parameters. Model performance is continuously validated during training until convergence or the maximum number of iterations is reached. Finally, the optimal prediction model and corresponding performance metrics are output. This process enhances global optimization capabilities by improving the GWO mechanism, while ensuring efficient and stable model training, significantly improving the accuracy and generalization ability of tight oil reservoir production prediction.

[0020] Example 3: A method for predicting tight oil reservoir production based on machine learning is proposed, mainly including the following steps: Data acquisition and preprocessing: Historical production data and related geological and engineering parameters of tight oil reservoirs are collected, including porosity, permeability, formation pressure, temperature, and mineral content. The data is cleaned and normalized, and a time-series sample set is constructed in chronological order. The sample set is randomly divided into training, validation, and test sets according to a predetermined ratio. Model construction and training: A dual-channel IGWO-BiLSTM-Attention prediction model is constructed. This model uses BiLSTM as its core and fully extracts the temporal dependencies in the production sequence through forward and backward loop structures. Based on this, a dual-channel attention mechanism is introduced to guide the cross-attention of deep and shallow features respectively: deep features are corrected by the weights of shallow channels, and shallow features are enhanced by the weights of deep channels, thereby achieving complementarity and enhancement between features and improving the model's ability to represent key geological parameters. Optimization Strategy: An improved Grey Wolf Optimization (IGWO) algorithm is used to globally optimize the network structure parameters of BiLSTM (such as the number of hidden layer units, learning rate, batch size, etc.). IGWO enhances the algorithm's ability to escape local optima by introducing a population perturbation strategy and an adaptive weight update mechanism, thereby improving parameter search efficiency and model generalization performance. Model Validation and Testing: The trained model is tuned using a validation set, and its predictive performance is evaluated using a test set. Root mean square error (RMSE), mean absolute error (MAE), and coefficient of determination (R²) are selected. 2Using this method as an evaluation metric, it is compared with traditional models such as CNN, ANN, and CNN-LSTM to verify its superiority in tight oil reservoir production prediction. Results and Applications: The model ultimately outputs production prediction curves and error analysis results for each well, providing data support for development decisions and production optimization in tight oil reservoirs. This method exhibits strong robustness and adaptability and can be further extended to production prediction for other unconventional oil and gas resources.

Claims

1. A method for predicting the production of tight oil reservoirs based on machine learning, characterized in that, Includes the following steps: S1: Data collection and preprocessing, using Spearman's rank correlation coefficient to screen features that are significantly related to yield and remove irrelevant variables (such as fluid viscosity); S2: Feature engineering and input construction, model input, construction of supervised learning samples; S3: Model building and training, BiLSTM-Attention model construction and IGWO hyperparameter optimization; S4: Model training and validation, independent testing on three different blocks (A, B, C) to evaluate generalization ability; S5: Comparison with benchmark models such as CNN, ANN, and CNN-LSTM.

2. The method for predicting tight oil reservoir production based on machine learning according to claim 1, characterized in that, In S1, production data from a tight reservoir at a station in the Tarim Oilfield were used, including porosity, permeability, pressure, temperature, mineral content, and fluid viscosity. Missing and outlier values ​​were handled to ensure data quality. Spearman's rank correlation coefficient was used to screen features significantly related to production and to remove irrelevant variables.

3. The method for predicting tight oil reservoir production based on machine learning according to claim 1, characterized in that, In the S2 process, porosity, permeability, pressure, temperature, and mineral content are selected as model inputs, and time series data are divided into windows to construct supervised learning samples.

4. The method for predicting tight oil reservoir production based on machine learning according to claim 1, characterized in that, In step S3, a BiLSTM-Attention model is built, constructing a bidirectional LSTM layer to capture forward and backward temporal dependencies. An Attention mechanism is introduced to assign higher weights to key geological parameters, and the output layer is used for regression prediction (yield). IGWO hyperparameter optimization uses an improved gray wolf optimization algorithm to optimize the hyperparameters of the BiLSTM-Attention model (such as learning rate, number of hidden layer neurons, batch size, etc.). A gray wolf population is initialized, and a fitness function (such as the validation set RMSE) is defined. Individual positions are updated through a population perturbation strategy and adaptive weights to avoid local optima. Iterative optimization continues until convergence, yielding the optimal hyperparameter combination.

5. The method for predicting tight oil reservoir production based on machine learning according to claim 1, characterized in that, In the S4 process, the training set, validation set, and test set are divided. The BiLSTM-Attention model is trained using optimized hyperparameters and tested independently on three different blocks (A, B, C) to evaluate its generalization ability.

6. The method for predicting tight oil reservoir production based on machine learning according to claim 1, characterized in that, In the S5 process, compared with benchmark models such as CNN, ANN, and CNN-LSTM, ablation experiments were conducted: IGWO and the Attention mechanism were removed respectively to verify the contribution of each component, using RMSE, MAE, and R... 2 The accuracy of predictions is evaluated using various indicators, the prediction curves are visualized, and the model's fitting effect and robustness are analyzed.

Images