A method for predicting tight oil fracture cluster production contribution

Abstract

This invention discloses a method for predicting the production contribution of fracturing clusters in tight oil, belonging to the field of tight oil fracturing production prediction technology. The method includes: constructing a stage-aware gating model based on conventional geological, engineering, and production data; integrating multi-dimensional feature encoding and gating modulation to accurately predict single-well production; identifying effective fracturing clusters using a LightGBM classification model; quantifying the contribution using an integrated regression model; and finally allocating production through mass conservation constraints and normalization. This method eliminates reliance on expensive downhole monitoring, addresses the problem of poor physical interpretability through production pattern embedding and mass conservation constraints, and overcomes the pain points of sample imbalance and weak generalization ability by employing cost-sensitive learning, leave-one-well cross-validation, and ensemble strategies. It outputs a high-resolution dynamic production matrix, thereby achieving the prediction of fracturing cluster production contribution without downhole monitoring, providing data support for fracturing optimization and development adjustments.

Description

Technical Field

[0001] This invention relates to the field of tight oil fracturing production prediction technology, and in particular to a method for predicting the production contribution of tight oil fracturing clusters. Background Technology

[0002] In the development of multi-stage, multi-cluster fracturing in horizontal wells of tight oil reservoirs, accurate assessment of the production contribution of fracturing clusters is a core aspect of fracturing design optimization, fracturing effect evaluation, production enhancement measures formulation, and remaining oil potential tapping. Currently, the industry mainly relies on downhole direct monitoring technologies such as FSI (Fluid Scanning Imaging) logging to obtain cluster-level production contribution data. However, these technologies suffer from high equipment investment costs, high construction risks, and complex operation procedures, making it difficult to achieve large-scale application in oilfields. The vast majority of oil wells do not meet the conditions for implementing FSI logging, resulting in a long-term lack of crucial data support for refined evaluation of fracturing effects and subsequent optimization adjustments.

[0003] The core challenges of existing technologies lie primarily in three aspects. First, fracturing cluster-level samples are extremely scarce and their distribution is severely imbalanced. Accurate identification of effective fracturing clusters relies entirely on high-cost, small-sample FSI logging data, resulting in a very limited number of positive samples (effective clusters) for supervised model training. Furthermore, the proportion of effective clusters within a single well is extremely low, creating a demanding training condition of "small sample size and high imbalance." This makes conventional machine learning models prone to overfitting or excessive bias towards ineffective clusters, leading to severely insufficient model generalization ability. Second, the strong nonlinearity and time-varying characteristics of production dynamics are difficult to model accurately. Tight oil well production exhibits a three-stage nonlinear evolution pattern: "high initial production – rapid decline – long-term stable production." Traditional time-series models and single deep learning networks lack an adaptive perception mechanism for production stages, failing to dynamically adjust prediction strategies at the inflection points of stage transitions. This results in a significant decrease in prediction accuracy at critical points, making it difficult to reflect true physical laws. Finally, in well-cluster cross-scale prediction, existing methods often treat total quantity prediction and inter-cluster allocation as independent tasks. The prediction results generally violate the law of conservation of mass, that is, the sum of the predicted production of each cluster often does not match the actual (or predicted) total production of the whole well. This physical inconsistency makes it difficult to directly use the prediction results for engineering decision-making and optimization, even if the statistical indicators are good, due to the lack of physical credibility.

[0004] Therefore, there is an urgent need for a mathematical algorithm based on conventional data that does not rely on expensive downhole monitoring equipment, but can replace (or partially replace) expensive FSI logging, and calculate the production of each cluster at low cost. Summary of the Invention

[0005] In order to overcome the shortcomings and deficiencies of the existing technology, the present invention provides a method for predicting the production contribution of tight oil fracturing clusters.

[0006] The technical solution adopted in this invention is a method for predicting the production contribution of tight oil fracturing clusters, comprising the following steps: S1, acquiring geological, engineering, production dynamics, and historical fluid scanning imaging (FSI) logging data, constructing a standardized input dataset, and defining a single-well production prediction task; S2, constructing a stage-aware gated production prediction model, using temporal feature encoding, static feature encoding, and production stage encoding, combined with a stage gating mechanism and feature modulation, to perform adaptive production prediction for the planned time period; S3, using historical well fracturing clusters as samples, utilizing conventional geological and engineering data as features, and fluid scanning imaging interpretation results as labels, training a Light algorithm for fracturing cluster effectiveness recognition. S4. After extracting features from the target well data, the effective fracturing clusters are identified through the classification model, and the initial production contribution of the effective clusters is quantified using the ensemble regression model. S5. The production prediction results of a single well are used as the total baseline, and the initial contribution of each cluster is normalized at the wellbore level to obtain the weight. The daily predicted production of a single well is allocated to each effective cluster according to the weight, and the production of invalid clusters is set to zero. S6. The dynamic production matrix at the fracturing cluster level is output. Each row in this matrix represents the predicted production curve of a fracturing cluster, and each column represents the distribution of production among clusters of the entire well on a certain day.

[0007] Furthermore, the temporal feature encoding employs a bidirectional long short-term memory network, and the forward computation satisfies: Backward computation satisfies: After merging, we get: in, For the hidden layer dimension, The input feature vector is the value at time step t. This is the set of parameters for the LSTM model, including the weight matrix and bias terms. This is a hidden state for aggregating forward and backward context information.

[0008] Furthermore, the static feature encoding is calculated using a multilayer perceptron and satisfies: Production-stage coding is computed through a learnable embedding layer to meet [the requirements]. ;in, This is the embedding vector for static geological features. This is a static geological feature vector. For the set of parameters of the MLP model, This is the weight matrix. For bias vectors, For the production stage embedding vector at time step t, To embed the lookup table, The stage label for a single time step. For the embedded vector dimension.

[0009] Furthermore, the stage gating mechanism satisfies: and Phase-aware feature modulation satisfies: and in, For the Sigmoid function, The gated signal vector, For the gated weight matrix, This is the gated bias vector. For element-wise multiplication, This represents the feature representation after stage filtering. For modulation signal, For the modulation weight matrix, This is the modulation bias vector. For feature cascade weighting operations, For learnable scaling parameters, This is a post-modulation enhancement feature.

[0010] Furthermore, the method also includes fusion prediction, satisfying: The loss function satisfies: in, ;in, For planning time period Production forecast at each time step For output layer weights, To fuse the weight matrix, To fuse the bias vector, For output bias terms, and For hyperparameters, Mean square error, To aid in the perception of stage-specific losses, For the number of stages, This is the sample set corresponding to the k-th stage. For the model's true value, These are the model's predicted values. This is the smoothing constant.

[0011] Furthermore, the output contribution prediction of the ensemble regression model satisfies: Normalized weights satisfy: Effective cluster dynamic daily output satisfies: in, The predicted value of the contribution of fracturing clusters. To integrate the weighting coefficients, The prediction results are from the LightGBM regression model. The prediction results of the ElasticNet regression model. For feature vectors, To contribute weight to normalized output, This is the original contribution prediction value. For an effective cluster set, Let i be the yield of cluster i on day t. Let T be the predicted daily fluid production of the target well T on day t.

[0012] Further, S2 includes the following sub-steps: S21, constructing a time-series encoder using a bidirectional long short-term memory network to encode the time-series production feature matrix from both forward and reverse directions, aggregating contextual information implied by historical trends and planned time-period trends; S22, performing a nonlinear transformation on the static geological feature vector using a multilayer perceptron to map it to the same semantic space as the dynamic time-series features, obtaining the embedding vector of the static features; S23, converting each stage label in the expert-annotated production stage sequence into a corresponding embedding vector through a learnable embedding layer, performing quantitative representation of production stage information; S24, generating a gating signal vector based on the production stage embedding, filtering the time-series hidden states, fusing the production stage embedding and static feature embedding to generate a modulation signal, performing residual modulation on the gating features, and then outputting the production prediction value through a fully connected network.

[0013] Further, S3 includes the following sub-steps: S31, taking each independent fracturing cluster in each horizontal well with fluid scanning imaging data as a basic sample, extracting features from the cluster's geological static parameters and cluster-level fracturing engineering parameters to form a feature vector, and using the fluid scanning imaging interpretation results to construct an effectiveness label and a normalized relative contribution label for each effective cluster; S32, constructing a binary classification model using the LightGBM algorithm, implementing a cost-sensitive learning strategy to address the class imbalance problem, training and evaluating the model through a leave-one-well cross-validation strategy, calculating multiple classification performance indicators and optimizing hyperparameters to obtain the final classification model; S33, selecting effective fracturing cluster samples and their normalized contribution labels to form a regression training set, training the LightGBM regression model and the ElasticNet regression model respectively, determining the fusion weight coefficients through cross-validation, and integrating the prediction results of the two base learners using a linear weighted fusion method to form an ensemble regression model.

[0014] Further, step S4 includes the following sub-steps: S41, according to the specifications established in S1, feature extraction is performed on the original data of the target well to ensure that the features of the target well and the model training data are consistent in feature definition, forming a normalized feature vector corresponding to each fracturing cluster; S42, the normalized feature vector is input into the LightGBM classification model optimized by leaving one well for cross-validation, and the model outputs the posterior probability that each fracturing cluster is an effective production cluster, and the effectiveness status of the fracturing cluster is determined based on this probability; S43, the set of fracturing clusters predicted to be effective is selected, and the feature vector subset corresponding to each effective cluster in the set is extracted and input into the trained ensemble regression model in batches; S44, the ensemble regression model performs independent forward calculation on each input feature vector, outputs the relative weight of the production contribution capability of each effective cluster relative to the calibrated effective clusters in the target well, and obtains a preliminary contribution prediction value.

[0015] Further, S5 includes the following sub-steps: S51, calling the stage-aware gated production prediction model trained and optimized in S2, inputting the historical production dynamics, static geological characteristics and production stage sequence of the target well, and outputting the daily production prediction value of the target well within the specified prediction period of the planning time period, forming a time series as the total benchmark; S52, performing wellbore-level normalization operation on the preliminary contribution prediction value of the effective clusters obtained in S4, eliminating the potential cumulative prediction bias of the regression model at the well scale, and obtaining the relative proportion weights that meet the physical allocation requirements; S53, for each effective cluster and each specific date within the prediction period, allocating the total production of the whole well on that day according to the corresponding normalized weights, and calculating the production of the effective cluster on that day; S54, for fracturing clusters determined to be invalid by the classification model, setting their daily production within the entire prediction period based on physical priors, and ensuring that the sum of the production of each cluster at any prediction time is consistent with the total production of the whole well through spatiotemporal coupling synthesis.

[0016] Beneficial Effects: This invention proposes a method for predicting the production contribution of fracturing clusters in tight oil reservoirs. This method does not rely on expensive downhole monitoring technologies such as fluid scanning imaging logging, but only uses conventional geological, engineering, and production dynamic data for prediction, significantly reducing operating costs and risks, while supporting large-scale application. By incorporating the dynamic evolution of reservoir production into the model construction through a stage-aware gating mechanism, and combining dynamic production synthesis under mass conservation constraints, it ensures that the sum of the production of each fracturing cluster is consistent with the total well production, completely solving the problems of poor physical interpretability and violation of mass conservation in purely data-driven models. Addressing the challenges of sample scarcity and class imbalance, cost-sensitive learning, clean sample training, and ensemble regression strategies are employed, coupled with cross-validation with one well to ensure the model's cross-well generalization ability, effectively improving the accuracy and predictive stability of effective fracturing cluster identification. The final output fracturing cluster-level dynamic production matrix clearly presents the changes in the production contribution of each cluster throughout its entire life cycle, providing high spatiotemporal resolution decision support for fracturing scheme optimization, production adjustment, and remaining oil potential tapping, comprehensively improving the intelligence and precision of tight oil reservoir fracturing development. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating the overall process of the method of the present invention. Figure 2 This is a flowchart of method step S2 of the present invention; Figure 3 This is a flowchart of method step S3 of the present invention; Figure 4 This is a flowchart of method step S4 of the present invention; Figure 5 This is a flowchart of step S5 of the method of the present invention. Detailed Implementation

[0019] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. The application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] like Figure 1As shown, a method for predicting the production contribution of tight oil fracturing clusters includes the following steps: S1, acquiring geological, engineering, production dynamics, and historical fluid scanning imaging (FSI) logging data, constructing a normalized input dataset, and defining a single-well production prediction task; S2, constructing a stage-aware gated production prediction model, using temporal feature encoding, static feature encoding, and production stage encoding, combined with a stage gating mechanism and feature modulation, to adaptively predict production for the planned time period; S3, using historical well fracturing clusters as samples, utilizing conventional geological and engineering data as features, and fluid scanning imaging interpretation results as labels, training a LightGBM classification system for fracturing cluster effectiveness recognition. The model and effective cluster production contribution quantification LightGBM-ElasticNet ensemble regression model; S4, after feature extraction of target well data, effective fracturing clusters are identified through a classification model, and the initial production contribution of effective clusters is quantified using the ensemble regression model; S5, the production prediction results of single wells are used as the total baseline, and the initial contribution of each cluster is normalized at the wellbore level to obtain weights. The daily predicted production of single wells is allocated to each effective cluster according to the weights, and the production of ineffective clusters is set to zero; S6, the fracturing cluster-level dynamic production matrix is ​​output. Each row in this matrix represents the predicted production curve of a fracturing cluster, and each column represents the inter-cluster distribution of production of the entire well on a certain day.

[0021] Step S1 is the data foundation building stage of the entire prediction method. Through multi-source data integration and standardization, it provides high-quality input for subsequent model training and prediction, while clarifying the core task boundary of single-well production prediction. In specific implementation, four types of key data are first comprehensively collected: geological data, including parameters directly reflecting reservoir characteristics such as porosity, permeability, oil saturation, effective thickness, reservoir pressure, and temperature; engineering data, including key parameters of fracturing operations such as horizontal section length, number of fracturing sections, number of clusters, total fluid volume, total sand volume, sand addition intensity, and construction displacement; production dynamic data, involving real-time production indicators such as daily oil production, daily fluid production, water cut, and bottom hole flowing pressure that change over time; and historical fluid scanning imaging logging data, which serves as a label source for model training, providing a verified cluster-level production contribution reference. After collection, missing values ​​in the original data are supplemented using specific techniques, outlier samples are identified and removed using the 3-standard deviation criterion, and then the data dimensions are unified through extreme value normalization or Z-score standardization to construct a structured and formatted input dataset. Based on this, the task of single-well production prediction is clearly defined, and single-well data is represented as a triple including a time-series production feature matrix, a static geological feature vector, and a production stage sequence. The time-series production feature matrix includes multiple time steps, each of which includes multiple dynamic features such as water cut and flowing pressure. The static geological feature vector includes multiple reservoir attributes that do not change over time, such as porosity and permeability. The production stage sequence is labeled with 0, 1, and 2 to correspond to the three stages of initial high production, rapid decline, and stable production, respectively. The task objective is to learn a mapping function based on historical data from multiple time steps to predict the production sequence of a specified time step in the planned period, providing a clear objective for subsequent model training.

[0022] Step S2 performs time-series prediction of daily production from a single well. By constructing a stage-aware gated production prediction model and integrating multi-dimensional feature encoding and intelligent modulation mechanisms, it achieves accurate and adaptive prediction of production during the planned period. Its core value lies in integrating expert knowledge from the production stage into the model, improving the physical rationality and dynamic adaptability of the prediction. In the specific implementation process, time-series feature encoding is first performed. A bidirectional long short-term memory network is used to construct a time-series encoder, which encodes the time-series production feature matrix from both forward and reverse directions. Forward encoding generates the current hidden state based on the input features of the current time step and the hidden state of the previous time step, while reverse encoding generates the current hidden state based on the input features of the current time step and the hidden state of the next time step. Subsequently, the hidden states obtained from forward and reverse encoding are merged to form an aggregated feature that includes twice the hidden layer dimension information, comprehensively capturing the bidirectional time-series dependency relationship implied by historical production trends and changes in the planned period. Next, static feature encoding is performed. A two-layer fully connected multilayer perceptron is used to nonlinearly transform the static geological feature vectors. After processing with a modified linear unit activation function, the static features are mapped to the same semantic space as the dynamic temporal features, resulting in dense static feature embedding vectors. Then, production stage encoding is performed. A learnable embedding layer converts expert-annotated production stage labels into fixed-dimensional embedding vectors, achieving a quantitative representation of production stage information. A stage gating mechanism is then activated. Based on the production stage embedding vectors, a gating signal vector with each element taking values ​​between 0 and 1 is generated. This selectively filters the corresponding dimensions of the temporal hidden state, amplifying features strongly correlated with the current production stage and suppressing interference from irrelevant features. Finally, stage-aware feature modulation is used to fuse the production stage embeddings and static feature embeddings to generate a modulated signal. Residual modulation is employed to enhance the gating features, preserving the original temporal feature core information while dynamically incorporating the influence of geological attributes and production stages. Finally, the modulated features are input into a fully connected network and trained end-to-end using a multi-objective weighted loss function. This loss function combines mean squared error and stage-aware auxiliary loss to ensure that the model maintains balanced performance in the three sub-stages of high initial production, rapid decline, and stable production while ensuring overall prediction accuracy. Ultimately, this enables the prediction of production sequence at multiple time steps during the planning period, providing a reliable total benchmark for subsequent cluster-level production allocation.

[0023] Step S3 constructs an intelligent prediction model system for fracturing clusters with cross-well generalization capabilities, including an effectiveness identification model and a production contribution quantification model. Through scientific sample processing, model design, and training strategies, it ensures that the model can still output reliable results even in the absence of fluid scanning imaging monitoring data. In specific implementation, the sample set is first constructed and processed. Each independent fracturing cluster in each horizontal well with fluid scanning imaging data is used as the basic sample. Features are extracted from the cluster's geological static parameters (such as porosity and permeability) and cluster / segment-level fracturing engineering parameters (such as injected fluid volume, injected sand volume, and cluster spacing), and combined into a feature vector. This feature vector only includes conventionally obtainable parameters and does not involve the fluid scanning imaging monitoring data itself. At the same time, a supervision label is constructed using the fluid scanning imaging interpretation results. A threshold of 0.5% of the cluster production as the total well daily production is set. Clusters above this threshold are marked as valid clusters and assigned a value of 1, while those below are marked as invalid clusters and assigned a value of 0. The normalized relative contribution of the valid cluster samples is calculated as a regression label. Subsequently, a fracturing cluster effectiveness classification model was trained. The LightGBM algorithm was used to construct a binary classification model. To address the class imbalance problem where the number of effective cluster samples is far less than that of ineffective clusters, a cost-sensitive learning strategy was implemented to improve the model's ability to identify effective clusters. To ensure cross-well generalization ability, a leave-one-well cross-validation strategy was adopted. M historical wells were used as independent test sets in sequence, and the remaining M-1 wells were used as training sets. M rounds of training and evaluation were carried out, and key performance indicators such as accuracy, recall, precision, F1 score, and AUC were calculated. The model hyperparameters were optimized based on the average values ​​of the indicators. Finally, the final classification model was obtained by training with all historical well data. Finally, an ensemble regression model for the contribution of effective cluster output was trained. A pure sample strategy was adopted, selecting only effective cluster samples to form the regression training set. LightGBM regression model and ElasticNet regression model were trained separately. The LightGBM regression model was responsible for capturing the complex nonlinear relationship between geological and engineering features and cluster contribution. The ElasticNet regression model improved model stability by combining L1 and L2 regularization to achieve feature selection and multicollinearity mitigation. The fusion weight coefficients were determined through cross-validation, and the prediction results of the two base learners were integrated using a linear weighted fusion method to form an ensemble regression model with complementary advantages, ensuring that high-precision and stable contribution measurement can still be achieved even with a limited number of effective cluster samples.

[0024] Step S4 is the model application stage. By applying the intelligent model with strong generalization ability trained in Step S3 to the target well, the effectiveness of the fracturing clusters in the target well is automatically identified and the initial quantification of their production contribution is achieved, providing accurate intermediate input for subsequent production synthesis under the constraint of quality conservation. In the specific implementation process, the target well data is first subjected to rigorous preprocessing and feature extraction, strictly following the multi-source data fusion and standardized processing specifications established in Step S1, ensuring that the target well data is completely consistent with the model training data in terms of feature definition and data processing method, eliminating prediction bias caused by data differences. On this basis, each fracturing cluster in the target well is taken as an independent unit, and features such as porosity, permeability, injected fluid volume, injected sand volume, and cluster spacing are extracted from the conventional geological and engineering data of the target well according to the feature engineering standards defined in Step S3. These features are combined to form a standardized feature vector corresponding to each fracturing cluster. For a target well with N fracturing clusters, a set of N feature vectors is finally formed as the model input. Subsequently, an automatic identification process for the effectiveness of fracturing clusters is initiated. Normalized feature vectors are input into the LightGBM classification model, optimized through cross-validation with one well retained. The model outputs a posterior probability value between 0 and 1 for each input feature vector, representing the model's confidence level in classifying the fracturing cluster as an effective production cluster. Based on a preset confidence threshold, the effectiveness status of each fracturing cluster is automatically determined, replacing the "production segment identification" function of fluid scanning imaging logging. Finally, the production contribution of effective clusters is initially quantified, and a set of fracturing clusters deemed effective is selected. A subset of feature vectors corresponding to each effective cluster in this set is extracted and batch-input into the ensemble regression model trained in step S3. The ensemble regression model performs independent forward computation on each input feature vector, outputting the relative weight of each effective cluster's production contribution capacity relative to other effective clusters in the target well. This yields a preliminary, non-normalized contribution prediction value, which directly reflects the relative production potential of each effective cluster, providing a core basis for subsequent production allocation.

[0025] Step S5, as the synthesis step of the entire prediction method, organically integrates the single-well-level production prediction results with the cluster-level contribution prediction results through a spatiotemporal coupling mechanism under the constraint of mass conservation. This generates a cluster-level dynamic production profile that conforms to the law of mass conservation, solving the key problem of physical inconsistency in pure data-driven models. In specific implementation, firstly, dynamic constraints on single-well production are applied. The stage-aware gated production prediction model trained and optimized in step S2 is called, and the historical production dynamic data, static geological feature vector, and production stage sequence of the target well are input. The model outputs the daily production prediction value of the target well within the specified prediction period of the planned time period, forming a time series sequence of length m. This sequence serves as the total benchmark for cluster-level production allocation, ensuring that the sum of cluster-level production is consistent with the total well production. Subsequently, a normalization calculation of cluster-level contribution weights is performed. Addressing the issue that the sum of the preliminary contribution predictions of effective clusters obtained in step S4 is not 1, a wellbore-level normalization operation is executed. Using the sum of the preliminary contribution predictions of all effective clusters in the target well as the denominator and the preliminary contribution prediction of each effective cluster as the numerator, the normalized production contribution weight of each effective cluster is calculated. This weight eliminates the potential cumulative prediction bias of the regression model at the well scale, retaining only reliable information on the relative contribution relationships between clusters, thus meeting physical allocation requirements. Next, spatiotemporal coupling and dynamic production calculation are performed. For each effective cluster and each specific date within the prediction period, the predicted total well production for that day is multiplied by the normalized weight of that effective cluster to obtain the dynamic daily production of that effective cluster on that day. For all fracturing clusters determined as invalid by the classification model, based on the prior physical knowledge that "invalid clusters contribute no to the overall well production," their daily production is directly set to zero throughout the entire prediction period, simplifying the calculation while ensuring physical rationality. Finally, physical consistency verification is performed. Through the above synthesis mechanism, it is ensured that at any prediction time, the sum of the daily production of all fracturing clusters is equal to the predicted total production of the whole well on that day, strictly following the law of conservation of mass, and providing a core guarantee for the reliability of the final output results.

[0026] Step S6 constructs a structured fracturing cluster-level dynamic production matrix, presenting the results of previous model predictions and synthesis in an intuitive and engineering-oriented form. This provides high spatiotemporal resolution decision support data for subsequent fracturing effect evaluation, production optimization, and remaining oil potential tapping. In specific implementation, based on the cluster-level dynamic production calculation results completed in Step S5, a fracturing cluster-level dynamic production matrix with dimensions of N rows and m columns is constructed, where N is the total number of fracturing clusters in the target well, and m is the number of production days in the prediction period. Each row of the matrix corresponds to a fracturing cluster, fully recording the daily production data of that cluster throughout the entire prediction period, forming the full life-cycle predicted production curve of the fracturing cluster. This clearly shows the production change trend from initial high production, rapid decline to stable production, intuitively reflecting the fluid production capacity and decay characteristics of the cluster. Each column of the matrix corresponds to a specific production day within the prediction period, recording the production distribution of all fracturing clusters in the target well on that day, clearly presenting the production contribution ratio of each fracturing cluster on that day, providing direct evidence for analyzing the fluid production efficiency of different clusters within a single day. The output format of this production matrix not only meets the needs of engineering applications for data structuring and visualization, but also includes core information with high spatiotemporal resolution. It can accurately support the analysis of the distribution law of effective clusters in the process of fracturing scheme optimization, the key control of high-yield clusters in the process of production adjustment, and the transformation decision of low-contribution clusters in the process of tapping the potential of remaining oil. It realizes the seamless connection from data prediction to engineering application and gives full play to the engineering value of the entire prediction method.

[0027] Preferably, the temporal feature encoding employs a bidirectional long short-term memory network, and the forward computation satisfies: Backward computation satisfies: After merging, we get: in, For the hidden layer dimension, The input feature vector is the value at time step t. This is the set of parameters for the LSTM model, including the weight matrix and bias terms. This is a hidden state for aggregating forward and backward context information.

[0028] Specifically, the time-series feature encoding uses a bidirectional long short-term memory network as the core encoding tool. Through the collaborative operation of forward and reverse bidirectional encoding and feature merging, it comprehensively captures the bidirectional dependencies in production time-series data. The implementation is as follows: during forward encoding, the hidden state of the current time step is calculated and generated jointly by the input feature vector of that time step and the hidden state of the previous time step, focusing on mining the influence of historical production data on the current state. Conversely, reverse encoding calculates the hidden state of the current time step by combining the input feature vector of that time step and the hidden state of the next time step, emphasizing the implications of the planned period trend for the current state. After completing bidirectional encoding, the hidden states obtained from forward and reverse encoding are merged to form an aggregated feature vector with a dimension twice that of the hidden layer. The hidden layer dimension can be set according to the actual data complexity and prediction accuracy requirements, typically ranging from 64 to 512. The input feature vector values ​​at each time step include real-time production data such as daily oil production, pressure, and water content, ensuring that the encoding process fully utilizes dynamic production information. The model parameter set of the Long Short-Term Memory Network includes a weight matrix and bias terms, specifically including parameters related to forget gate, input gate, output gate, and cell state update. These parameters are adaptively optimized through model training. The resulting aggregated hidden state can simultaneously integrate historical trends and planning time period trends, providing comprehensive and accurate temporal feature support for subsequent gating mechanisms and feature modulation, and significantly improving the temporal correlation and accuracy of single-well production prediction.

[0029] Preferably, the static feature encoding is calculated using a multilayer perceptron and satisfies: Production-stage coding is computed through a learnable embedding layer, satisfying the following: in, This is the embedding vector for static geological features. This is a static geological feature vector. For the set of parameters of the MLP model, This is the weight matrix. For bias vectors, For the production stage embedding vector at time step t, To embed the lookup table, The stage label for a single time step. For the embedded vector dimension.

[0030] Specifically, static feature encoding is accomplished using a multilayer perceptron. During implementation, static geological feature vectors, including reservoir inherent properties such as porosity, permeability, and oil saturation, are taken as input. These vectors undergo a nonlinear transformation using two fully connected layers and a modified linear unit activation function, mapping the static features to the same semantic space as the dynamic temporal features. The multilayer perceptron's model parameter set includes weight matrices and bias vectors corresponding to the two fully connected layers. The weight matrices implement linear transformations of the features, while the bias vectors provide translational degrees of freedom for each layer. The modified linear unit activation function effectively introduces nonlinear features, mitigating the gradient vanishing problem during model training. The final output static feature embedding vector is a dense vector, ensuring fusion compatibility with dynamic temporal features. Production stage encoding is implemented through learnable embedding layers. Expert-labeled production stage tags (0 for initial high production, 1 for rapid decline, and 2 for stable production) are converted into fixed-dimensional embedding vectors. The embedding vector dimension is typically set between 16 and 64. The embedding lookup table is continuously optimized through model training, enabling production stage information to participate in model calculations in a quantified form. The two encoding methods respectively realize the semantic mapping of static geological features and the quantitative representation of production stage information, providing multi-dimensional feature inputs for subsequent stage gating mechanisms and feature modulation, ensuring that the model can simultaneously consider the influence of reservoir inherent properties and dynamic changes in the production stage.

[0031] Preferably, the stage gating mechanism satisfies: and Phase-aware feature modulation satisfies: and in, For the Sigmoid function, The gated signal vector, For the gated weight matrix, This is the gated bias vector. For element-wise multiplication, This represents the feature representation after stage filtering. For modulation signal, For the modulation weight matrix, This is the modulation bias vector. For feature cascade weighting operations, For learnable scaling parameters, This is a post-modulation enhancement feature.

[0032] Specifically, the implementation process of the stage-based gating mechanism is as follows: First, based on the production stage embedding vector, a gating signal vector is generated through a fully connected layer and a sigmoid activation function. The value of each element in this vector is strictly controlled between 0 and 1, and each element corresponds to a dimension in the temporal hidden state. The closer the element value is to 1, the higher the degree to which the corresponding dimension's features are "enabled," and the closer it is to 0, the higher the degree to which they are "suppressed." The gating weight matrix and bias vector are learned through model training to ensure that the gating signal can accurately match the feature requirements of different production stages. Then, the gating signal vector is multiplied element-wise with the temporal hidden state to obtain the feature representation after stage selection, enabling the model to focus on key features in different production stages. In the stage-aware feature modulation stage, the production stage embedding vector and the static feature embedding vector are first cascaded and fused. A modulation signal is generated through a neural network layer and a hyperbolic tangent activation function. The modulation weight matrix and bias vector are optimized through training to ensure that the modulation signal can effectively integrate production stage information and inherent reservoir properties. Subsequently, residual modulation was employed to enhance the gated features. A learnable scaling parameter (ranging from 0 to 1) controlled the degree of influence of the modulated signal. This parameter was adaptively adjusted during training to ensure that the core information of the original temporal features was preserved while incorporating the dynamic influence of geological attributes. The two mechanisms worked synergistically, enabling the model to amplify features related to fracturing conductivity and initial production capacity in the initial high-production stage, focus on features such as pressure reduction rate and recovery rate in the rapid decline stage, and focus on features such as matrix seepage velocity and long-term trends in the stable production stage, significantly improving the model's stage-adaptive predictive capability.

[0033] Preferably, the method further includes fusion prediction, satisfying: The loss function satisfies: in, ;in, For planning time period Production forecast at each time step For output layer weights, To fuse the weight matrix, To fuse the bias vector, For output bias terms, and For hyperparameters, Mean square error, To aid in the perception of stage-specific losses, For the number of stages, This is the sample set corresponding to the k-th stage. For the model's true value, These are the model's predicted values. This is the smoothing constant.

[0034] Specifically, the implementation process of fusion prediction is as follows: Feature vectors rich in multi-source information, obtained after stage gating and feature modulation, are input into a fully connected network. The fully connected network achieves high-dimensional mapping of features through a fusion weight matrix. After processing by a modified linear unit activation function, the predicted output value for a specified time step in the planning period is output through a linear transformation of the output layer weights and output bias terms. Prediction can be performed in a multi-step rolling manner to generate a complete production sequence for the planning period. A multi-objective weighted loss function is used, balancing the importance of the main loss and stage-aware auxiliary loss through hyperparameters. The hyperparameter values ​​range from 0 to 1 and can be adjusted according to actual prediction needs. The main loss uses mean squared error to measure the deviation between the overall predicted value and the true value, ensuring overall prediction accuracy. The stage-aware auxiliary loss calculates the average relative error of the predicted value in each production stage. A smoothing constant (10 to the power of -6) is introduced to prevent the denominator from being zero. The number of stages is fixed at 3 (corresponding to the initial high production, rapid decline, and stable production stages). The final auxiliary loss is obtained by calculating the mean relative error within the sample set of each stage and then averaging the mean values ​​of all stages. This loss function design forces the model not only to ensure overall prediction accuracy, but also to maintain balanced prediction performance at each production stage. This avoids model bias caused by imbalanced sample size, enabling the model to learn a more generalizable stage-aware representation and improving the prediction stability at different production stages.

[0035] Preferably, the output contribution prediction of the integrated regression model satisfies: Normalized weights satisfy: Effective cluster dynamic daily output satisfies: in, The predicted value of the contribution of fracturing clusters. To integrate the weighting coefficients, The prediction results are from the LightGBM regression model. The prediction results of the ElasticNet regression model. For feature vectors, To contribute weight to normalized output, This is the original contribution prediction value. For an effective cluster set, Let i be the yield of cluster i on day t. Let T be the predicted daily fluid production of the target well T on day t.

[0036] Specifically, the production contribution prediction of the ensemble regression model adopts a linear weighted fusion method, integrating the prediction results of the LightGBM regression model and the ElasticNet regression model. The fusion weight coefficients range from 0 to 1, and the optimal values ​​are determined by performing cross-validation on the training set. Mean squared error or coefficient of determination is typically used as the evaluation metric to ensure that the fused model achieves the best balance between prediction accuracy and stability. The LightGBM regression model is responsible for capturing the complex nonlinear relationship between geological and engineering features and cluster contribution, while the ElasticNet regression model, through a combination of L1 and L2 regularization, achieves feature selection and multicollinearity mitigation. The complementary advantages of both models improve prediction reliability. When calculating the normalized weights, the sum of the original contribution predictions of all effective clusters in the target well is used as the denominator, and the original contribution prediction of each effective cluster is used as the numerator. The normalized production contribution weight is obtained through division. This operation effectively eliminates the potential cumulative prediction bias of the regression model at the well scale, ensuring that the weights only reflect the relative contribution relationship between clusters. In the calculation of the dynamic daily production of effective clusters, for each specific date within the forecast period, the predicted total well production value for that day is multiplied by the normalized weight of the corresponding effective cluster to obtain the production of that effective cluster on that day. These three calculation steps are progressively advanced, respectively achieving precise quantification of the contribution of effective clusters, standardization of inter-cluster contribution weights, and spatiotemporal allocation of dynamic production. This ensures that the final output of cluster-level production conforms to the law of conservation of mass, providing reliable data support for engineering applications.

[0037] Preferred, such as Figure 2 As shown, S2 includes the following sub-steps: S21, constructing a time-series encoder using a bidirectional long short-term memory network to encode the time-series production feature matrix from both the forward and reverse directions, aggregating contextual information implied by historical trends and planned time-period trends; S22, performing a nonlinear transformation on the static geological feature vector using a multilayer perceptron to map it to the same semantic space as the dynamic time-series features, obtaining the embedding vector of the static features; S23, converting each stage label in the expert-annotated production stage sequence into a corresponding embedding vector through a learnable embedding layer, performing quantitative representation of production stage information; S24, generating a gating signal vector based on the production stage embedding, filtering the time-series hidden states, fusing the production stage embedding and static feature embedding to generate a modulation signal, performing residual modulation on the gating features, and then outputting the production prediction value through a fully connected network.

[0038] Specifically, the implementation process of step S2 clarifies the construction and operation logic of the stage-aware gating production prediction model through four sub-steps, ensuring effective fusion of multi-source features and accurate production prediction. S21, as the core of temporal feature extraction, employs a bidirectional long short-term memory network to build a temporal encoder. It mines the influence of historical production trends by forward traversing the temporal production feature matrix and captures the changing trends during the planning period by backward traversal. The bidirectional encoding results are aggregated to form a feature representation including complete contextual information, providing comprehensive temporal data support for subsequent analysis. S22 focuses on static feature processing, using a multilayer perceptron to perform nonlinear transformations on static geological feature vectors such as porosity and permeability. Through multilayer network mapping, static features are transformed into a semantic space matching dynamic temporal features, generating dense static feature embedding vectors and achieving the fusion of static and dynamic features. S23 is responsible for quantifying production stage information. Through a learnable embedding layer, the expert-annotated three stages—initial high yield, rapid decline, and stable production—are converted into fixed-dimensional embedding vectors, enabling this abstract information of the production stage to participate in model calculations in numerical form, providing a basis for stage adaptive adjustment. S24, as the core computational component of the model, first generates a gated signal vector based on the production stage embedding vector, selectively filters the temporal hidden states, and retains features strongly correlated with the current stage; then it merges the production stage embedding and static feature embedding to generate a modulated signal, and enhances the targeting of the gated features through residual modulation; finally, it inputs the modulated comprehensive features into a fully connected network to complete the production prediction for the planned period. The four sub-steps are progressively advanced, realizing the deep fusion and adaptive prediction of multi-dimensional features.

[0039] Preferred, such as Figure 3 As shown, S3 includes the following sub-steps: S31, taking each independent fracturing cluster in each horizontal well with fluid scanning imaging data as a basic sample, extracting features from the cluster's geological static parameters and cluster-level fracturing engineering parameters to form a feature vector, and using the fluid scanning imaging interpretation results to construct an effectiveness label and a normalized relative contribution label for each effective cluster; S32, using the LightGBM algorithm to construct a binary classification model, implementing a cost-sensitive learning strategy to address the class imbalance problem, training and evaluating the model through a leave-one-well cross-validation strategy, calculating multiple classification performance indicators and optimizing hyperparameters to obtain the final classification model; S33, selecting effective fracturing cluster samples and their normalized contribution labels to form a regression training set, training the LightGBM regression model and the ElasticNet regression model respectively, determining the fusion weight coefficients through cross-validation, and integrating the prediction results of the two base learners using a linear weighted fusion method to form an ensemble regression model.

[0040] Specifically, the construction process of the intelligent prediction model for fracturing clusters in step S3 involves three sub-steps: sample processing, classification model training, and integrated regression model construction, ensuring the model has a high recognition rate and cross-well generalization ability. S31 lays the data foundation for model training. Each independent fracturing cluster in each horizontal well with fluid scanning imaging data is used as the basic sample. Features such as porosity, permeability, injected fluid volume, and cluster spacing are extracted from the cluster's geological static parameters and cluster segment-level fracturing engineering parameters and combined into a feature vector. At the same time, the results are interpreted using fluid scanning imaging. An effectiveness label is constructed with the cluster production accounting for 0.5% of the total well daily production as the threshold. The normalized relative contribution of the effective clusters is calculated as the regression label, forming a complete supervised training sample set. S32 focuses on training an effective cluster identification model, employing the LightGBM algorithm to construct a binary classification model. Addressing the class imbalance problem caused by the scarcity of effective cluster samples, a cost-sensitive learning strategy is implemented to improve the model's ability to identify effective clusters. To ensure cross-well generalization, a leave-one-well cross-validation strategy is adopted, using multiple historical wells sequentially as independent test sets and the remaining wells as training sets for multiple rounds of training and evaluation. Key metrics such as accuracy, recall, and AUC are calculated, and hyperparameters are optimized, ultimately resulting in a stable classification model. S33 is dedicated to building a production contribution quantification model. Effective cluster samples are selected to form a pure regression training set, training both a LightGBM regression model and an ElasticNet regression model. The former captures complex nonlinear relationships, while the latter improves stability and interpretability through double regularization. Cross-validation is used to determine the fusion weight coefficients, and a linear weighted fusion method is employed to integrate the prediction results of the two base learners, forming a complementary ensemble regression model to ensure the accuracy and reliability of contribution prediction.

[0041] Preferred, such as Figure 4 As shown, S4 includes the following sub-steps: S41, according to the specifications established in S1, feature extraction is performed on the original data of the target well to ensure that the features of the target well and the model training data are consistent in feature definition, forming a normalized feature vector corresponding to each fracturing cluster; S42, the normalized feature vector is input into the LightGBM classification model optimized by leaving one well for cross-validation, and the model outputs the posterior probability that each fracturing cluster is an effective production cluster, and the effectiveness status of the fracturing cluster is determined based on this probability; S43, the set of fracturing clusters predicted to be effective is selected, and the feature vector subset corresponding to each effective cluster in the set is extracted and input into the trained ensemble regression model in batches; S44, the ensemble regression model performs independent forward calculation on each input feature vector, outputs the relative weight of the production contribution capability of each effective cluster relative to the calibrated effective clusters in the target well, and obtains the preliminary contribution prediction value.

[0042] Specifically, the model application process in step S4 involves four sub-steps to identify the effectiveness of the target well fracturing clusters and initially quantify their contribution, providing accurate input for subsequent quality conservation constraint synthesis. Step S41 focuses on ensuring data consistency, strictly adhering to the multi-source data fusion and standardized processing specifications established in step S1. Feature extraction is performed on the original geological and engineering data of the target well to ensure that the target well features are completely consistent with the model training data in terms of definition, dimensions, and processing methods, eliminating prediction bias caused by data differences. For target wells with multiple fracturing clusters, feature parameters for each cluster are extracted one by one to form a standardized feature vector set. Step S42 performs automatic identification of effective clusters, inputting the standardized feature vectors into the LightGBM classification model optimized through cross-validation with one well retained. The model outputs a posterior probability between 0 and 1 for each feature vector, representing the confidence level that the cluster is an effective production cluster. Based on a pre-set confidence threshold, the effectiveness status of each fracturing cluster is automatically determined, efficiently completing the identification of the production segment. S43 performs effective cluster feature screening, selecting the effective cluster set from all fracturing clusters, and extracting the feature vector subset corresponding to each effective cluster in the set to ensure that subsequent contribution prediction is only based on effective samples and avoids interference from invalid data. S44 performs preliminary contribution quantification, inputting the batch of effective cluster feature vector subsets into the ensemble regression model trained in step S3. The model performs independent forward computation on each input vector, outputting the relative weight of each effective cluster's production contribution capability relative to other effective clusters in the target well, obtaining preliminary, non-normalized contribution prediction values, providing core data support for subsequent production allocation. The four sub-steps are sequentially connected, realizing the complete application of the model from data input to intermediate result output.

[0043] Preferred, such as Figure 5 As shown, S5 includes the following sub-steps: S51, calling the stage-aware gated production prediction model trained and optimized in S2, inputting the historical production dynamics, static geological characteristics and production stage sequence of the target well, and outputting the daily production prediction value of the target well within the specified prediction period of the planning time period, forming a time series as the total benchmark; S52, performing wellbore-level normalization operation on the preliminary contribution prediction value of the effective clusters obtained in S4, eliminating the potential cumulative prediction bias of the regression model at the well scale, and obtaining the relative proportion weights that meet the physical allocation requirements; S53, for each effective cluster and each specific date within the prediction period, allocating the total production of the whole well on that day according to the corresponding normalized weights, and calculating the production of the effective cluster on that day; S54, for fracturing clusters that are determined to be invalid by the classification model, setting their daily production within the entire prediction period based on physical priors, and ensuring that the sum of the production of each cluster at any prediction time is consistent with the total production of the whole well through spatiotemporal coupling synthesis.

[0044] Specifically, the mass conservation constraint synthesis process in step S5 achieves spatiotemporal coupling between single-well production and cluster-level contribution through four sub-steps, generating cluster-level dynamic production that conforms to the law of mass conservation. S51 establishes a total baseline by calling the stage-aware gated production prediction model trained and optimized in step S2. The model inputs historical production dynamic data, static geological characteristics, and production stage sequences of the target well. The model outputs the daily production prediction value for the specified prediction period within the target well's planned timeframe, forming a continuous time-series production sequence. This sequence serves as the total constraint for cluster-level production allocation, ensuring that the sum of cluster-level production is consistent with the total well production. S52 addresses the contribution normalization problem. For the issue that the sum of the preliminary contribution prediction values ​​of effective clusters obtained in step S4 is not 1, a wellbore-level normalization operation is performed. The sum of the preliminary contribution prediction values ​​of all effective clusters in the target well is used as the denominator, and the preliminary contribution prediction value of each effective cluster is used as the numerator. This calculates the normalized production contribution weight, eliminating the cumulative prediction bias of the regression model and ensuring that the weight only reflects the relative contribution relationship between clusters. S53 completes dynamic production calculation. For each valid cluster and each specific date within the prediction period, the predicted total well production value for that day is multiplied by the normalized weight of that valid cluster to accurately obtain the dynamic daily production of that valid cluster on that day, realizing the spatiotemporal allocation of production. S54 processes invalid cluster production and ensures physical consistency. For fracturing clusters determined to be invalid by the classification model, their daily production is set to be zero throughout the entire prediction period based on physical priors. Through the above spatiotemporal coupling synthesis mechanism, it is ensured that the sum of the production of all fracturing clusters at any prediction time is equal to the total well production for that day, strictly following the law of conservation of mass. The four sub-steps are progressively advanced, ultimately generating physically interpretable cluster-level dynamic production data.

[0045] A method for predicting the production contribution of fracturing clusters in tight oil reservoirs utilizes only conventional geological, engineering, and production dynamic data, eliminating the need for expensive downhole monitoring technologies such as fluid scanning imaging logging. This significantly reduces operating costs and implementation barriers, enabling large-scale application. By constructing a stage-aware gating model that integrates multi-dimensional features from temporal, static, and production stages, and combining gating mechanisms with feature modulation, the method accurately captures the dynamic patterns of reservoir production. Furthermore, by employing an integrated learning architecture and various optimization strategies, the method improves the accuracy of effective fracturing cluster identification and the stability of production contribution prediction. The output fracturing cluster-level dynamic production matrix combines high spatiotemporal resolution with engineering practicality.

[0046] This method addresses the issue of insufficient physical interpretability by integrating the evolutionary laws of reservoir production stages into model construction. Through a dynamic production synthesis mechanism under mass conservation constraints, it enforces consistency between the sum of production from each cluster and the total well production, overcoming the shortcomings of purely data-driven models that violate mass conservation. To address the weak generalization ability caused by scarce samples and class imbalance, a cost-sensitive learning strategy is employed to improve the effective cluster identification rate. Cross-validation with one well left over simulates new well prediction scenarios to avoid overfitting. Combined with pure sample training and an ensemble regression model, it enhances prediction stability under small sample conditions, ensuring reliable results under various geological and engineering conditions, thus completely overcoming the application limitations of existing technologies.

[0047] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for predicting the production contribution of tight oil fracturing clusters, characterized in that, Includes the following steps: S1 acquires geological, engineering, production dynamics, and historical fluid scanning imaging (FSI) logging data, constructs a standardized input dataset, and defines a single-well production prediction task. S2, construct a phase-aware gating output prediction model, and use time-series feature encoding, static feature encoding, production phase encoding, combined with phase gating mechanism and feature modulation to perform adaptive output prediction for the planning period. S3 uses historical well fracturing clusters as samples, conventional geological and engineering data as features, and fluid scanning imaging interpretation results as labels to train a LightGBM classification model for fracturing cluster effectiveness identification and a LightGBM-ElasticNet ensemble regression model for measuring the contribution of effective cluster production. S4. After extracting features from the target well data, effective fracturing clusters are identified through a classification model, and the initial production contribution of effective clusters is quantified using an integrated regression model. S5, call the single well production prediction results as the total benchmark, perform wellbore-level normalization on the preliminary contribution of each cluster to obtain the weight, and allocate the daily predicted production of the single well to each effective cluster according to the weight, and set the production of the invalid cluster to zero. S6 outputs the fracturing cluster-level dynamic production matrix. Each row in this matrix represents the predicted production curve of a fracturing cluster, and each column represents the inter-cluster distribution of production for the entire well on a given day.

2. The method for predicting the production contribution of a tight oil fracturing cluster according to claim 1, characterized in that, The temporal feature encoding employs a bidirectional long short-term memory network, and the forward computation satisfies: Backward computation satisfies: After merging, we get: in, For the hidden layer dimension, The input feature vector is the value at time step t. This is the set of parameters for the LSTM model, including the weight matrix and bias terms. This is a hidden state for aggregating forward and backward context information.

3. The method for predicting the production contribution of tight oil fracturing clusters according to claim 1, characterized in that, The static feature encoding is calculated using a multilayer perceptron and satisfies: Production-stage coding is computed through a learnable embedding layer, satisfying the following: in, This is the embedding vector for static geological features. This is a static geological feature vector. For the set of parameters of the MLP model, This is the weight matrix. For bias vectors, For the production stage embedding vector at time step t, To embed the lookup table, The stage label for a single time step. For the embedded vector dimension.

4. The method for predicting the production contribution of tight oil fracturing clusters according to claim 1, characterized in that, The stage gating mechanism satisfies: ， Phase-aware feature modulation satisfies: in, For the Sigmoid function, The gated signal vector, For the gated weight matrix, This is the gated bias vector. For element-wise multiplication, This represents the feature representation after stage filtering. For modulation signal, For the modulation weight matrix, This is the modulation bias vector. For feature cascade weighting operations, For learnable scaling parameters, This is a post-modulation enhancement feature.

5. The method for predicting the production contribution of a tight oil fracturing cluster according to claim 1, characterized in that, This method also includes fusion prediction, satisfying: The loss function satisfies: in, in, For planning time period Production forecast at each time step For output layer weights, To fuse the weight matrix, To fuse the bias vector, For output bias terms, and For hyperparameters, Mean square error, To aid in the perception of stage-specific losses, For the number of stages, This is the sample set corresponding to the k-th stage. For the model's true value, These are the model's predicted values. This is the smoothing constant.

6. The method for predicting the production contribution of a tight oil fracturing cluster according to claim 1, characterized in that, The output contribution prediction of the ensemble regression model satisfies: Normalized weights satisfy: Effective cluster dynamic daily output satisfies: in, The predicted value of the contribution of fracturing clusters. To integrate the weighting coefficients, The prediction results are from the LightGBM regression model. The prediction results of the ElasticNet regression model. For feature vectors, To contribute weight to normalized output, This is the original contribution prediction value. For an effective cluster set, Let i be the yield of cluster i on day t. Let T be the predicted daily fluid production of the target well T on day t.

7. The method for predicting the production contribution of tight oil fracturing clusters according to claim 1, characterized in that, S2 includes the following steps: S21 uses a bidirectional long short-term memory network to construct a time-series encoder, which encodes the time-series production feature matrix from both the forward and reverse directions, and aggregates the contextual information implied by historical trends and planned time period trends. S22, the static geological feature vector is nonlinearly transformed by a multilayer perceptron and mapped to the same semantic space as the dynamic temporal features to obtain the embedding vector of the static features; S23, through a learnable embedding layer, converts each stage label in the expert-annotated production stage sequence into a corresponding embedding vector to quantitatively represent the production stage information. S24 generates a gated signal vector based on the production stage embedding, filters the temporal hidden state, integrates the production stage embedding and static feature embedding to generate a modulation signal, performs residual modulation on the gated features, and then outputs the production prediction value through a fully connected network.

8. The method for predicting the production contribution of tight oil fracturing clusters according to claim 1, characterized in that, S3 includes the following steps: S31. Taking each independent fracturing cluster in each horizontal well with fluid scanning imaging data as the basic sample, features are extracted from the geological static parameters of the cluster and the fracturing engineering parameters of the cluster segment to form a feature vector. The interpretation results of fluid scanning imaging are used to construct an effectiveness label and a normalized relative contribution label of the effective cluster for each sample. S32 uses the LightGBM algorithm to build a binary classification model, implements a cost-sensitive learning strategy to deal with the class imbalance problem, trains and evaluates the model through leave-one-well cross-validation strategy, calculates multiple classification performance indicators and optimizes hyperparameters to obtain the final classification model; S33. Select effective fracturing cluster samples and their normalized contribution labels to form a regression training set. Train the LightGBM regression model and the ElasticNet regression model respectively. Determine the fusion weight coefficients through cross-validation. Use a linear weighted fusion method to integrate the prediction results of the two base learners to form an ensemble regression model.

9. The method for predicting the production contribution of a tight oil fracturing cluster according to claim 1, characterized in that, S4 includes the following steps: S41. Following the specifications established in S1, feature extraction is performed on the original data of the target well to ensure that the features of the target well are consistent with the feature definitions of the model training data, thus forming a standardized feature vector corresponding to each fracturing cluster. S42, input the normalized feature vector into the LightGBM classification model after cross-validation optimization with one well left, the model outputs the posterior probability of each fracturing cluster being an effective fluid-producing cluster, and the effectiveness status of the fracturing cluster is determined based on this probability; S43, select the set of fracturing clusters that are predicted to be effective, extract the feature vector subset corresponding to each effective cluster in the set, and input them in batches into the trained ensemble regression model; S44, the ensemble regression model performs independent forward computation on each input feature vector, outputs the relative weight of the production contribution capability of each effective cluster relative to the calibrated effective cluster in the target well, and obtains the preliminary contribution prediction value.

10. The method for predicting the production contribution of a tight oil fracturing cluster according to claim 1, characterized in that, S5 includes the following steps: S51 calls the stage-aware gated production prediction model trained and optimized by S2, inputs the historical production dynamics, static geological characteristics and production stage sequence of the target well, and outputs the daily production prediction value of the target well within the specified prediction period of the planning period, forming a time series as the total benchmark. S52, for the preliminary contribution prediction values ​​of the effective clusters obtained in S4, perform wellbore-level normalization to eliminate the potential cumulative prediction bias of the regression model at the well scale and obtain the relative proportional weights that meet the physical allocation requirements. S53, for each effective cluster and each specific date within the prediction period, the total well production for that day is allocated according to the corresponding normalized weight, and the production of that effective cluster on that day is calculated. S54. For fracturing clusters that are determined to be invalid by the classification model, their daily production rate is set based on physical priors throughout the entire prediction period. Spatiotemporal coupling synthesis is used to ensure that the sum of the production rates of each cluster at any prediction time is consistent with the total well production rate.

Images