Water-drive reservoir yield prediction method and system based on double-flow network

Abstract

The invention belongs to the technical field of artificial intelligence and oil-gas field development, and relates to a water-drive reservoir yield prediction method and system based on a double-flow network, and the method comprises the steps: carrying out the decomposition of the obtained historical and to-be-predicted reservoir production data of a target oil field block, historical and to-be-predicted trend component data and historical and to-be-predicted seasonal component data are obtained, and the historical trend component data and the historical seasonal component data form a data set; taking the reservoir trend component and the reservoir seasonal component as input, taking the reservoir yield prediction result as output, and constructing a water-drive reservoir yield prediction model based on a double-flow network under a PyTorch framework; training and optimizing the water-drive reservoir yield prediction model through the data set to obtain an optimal water-drive reservoir yield prediction model; and inputting the trend component data to be predicted and the season component data to be predicted into the optimal water-drive reservoir yield prediction model to obtain a reservoir yield prediction result. The oil reservoir yield prediction precision and efficiency under the complex working condition of the water-drive oil reservoir can be improved.

Description

Technical Field

[0001] This application belongs to the fields of artificial intelligence and oil and gas field development technology, specifically relating to a method and system for predicting the production of water-driven oil reservoirs based on a dual-flow network. Background Technology

[0002] Waterflooded reservoir production prediction is one of the core technologies in oilfield development, and its accuracy directly affects the scientific validity of development plan adjustments, remaining oil distribution studies, and injection-production structure optimization. During waterflooded reservoir development, reservoir production data exhibits two distinct statistical characteristics. On the one hand, driven by inherent laws such as natural formation energy decay and continuous water cut increases, production shows a long-term, smooth, natural decline trend. On the other hand, artificial intervention measures (such as water injection adjustments, fracturing, and perforation) introduce nonlinear, sudden, short-term high-frequency fluctuations and disturbances, causing the production curve to contain nonlinear fluctuations and sudden disturbances, resulting in a highly irregular production curve.

[0003] Existing methods for predicting production in water-driven reservoirs mainly include traditional production decline curve methods and methods based on reservoir numerical simulations. Traditional production decline curve methods can only capture long-term trends and lag behind responses to production fluctuations and abrupt changes caused by artificial interventions such as water injection, thus limiting prediction accuracy. While methods based on reservoir numerical simulations have clear physical mechanisms, they require extremely high precision in geological parameters, and the modeling and calculation processes are extremely time-consuming, making it difficult to meet the needs of real-time on-site decision-making.

[0004] In recent years, deep learning technologies, represented by models such as RNN, LSTM, and Transformer, have achieved certain results in time series analysis of production data and have been introduced into the field of oilfield development. However, most existing methods adopt an "end-to-end" single model structure, using a single model to simultaneously learn and fit the long-term trend and short-term fluctuations in reservoir production data. This approach often results in trade-offs, leading to either an underfitting of the long-term declining trend or a failure to adequately account for both the "long-term smooth decline" and "short-term high-frequency fluctuations" characteristics of reservoir data after reacting to short-term fluctuations, thus failing to achieve a balance between the two features. Summary of the Invention

[0005] This application addresses the aforementioned problems of low prediction accuracy in existing technologies by providing a method and system for predicting water-driven reservoir production based on a dual-stream network. The method decomposes the reservoir production time-series data to be predicted into trend components and seasonal components. By training and optimizing the water-driven reservoir production prediction model, features are extracted from the trend components and seasonal components respectively. The prediction results of water-driven reservoir production are then obtained through feature fusion, thereby improving the prediction accuracy and efficiency under complex operating conditions of water-driven reservoirs.

[0006] In its first aspect, this application provides a method for predicting the production of water-driven reservoirs based on a dual-flow network, comprising: Acquire historical reservoir production data and predicted reservoir production data for the target oilfield block; The historical reservoir production data and the reservoir production data to be predicted are decomposed separately to obtain historical trend component data and to be predicted trend component data reflecting long-term changes in reservoir production capacity, and historical seasonal component data and to be predicted seasonal data reflecting fluctuations in reservoir production injection and production adjustments. The historical trend component data and historical seasonal data are then combined to form a dataset. Using reservoir trend components and reservoir seasonal components as inputs and reservoir production prediction results as outputs, a water-drive reservoir production prediction model based on a two-stream network is constructed under the PyTorch framework. The reservoir production prediction results include daily oil production, daily liquid production, and water cut for a preset future time period. The water-drive reservoir production prediction model includes a trend feature extraction module for extracting long-term reservoir production capacity change characteristics from the trend component data, a seasonal feature extraction module for extracting reservoir production injection-production adjustment fluctuation characteristics from the seasonal component data, and a fusion module for integrating the long-term reservoir production capacity change characteristics and the reservoir production injection-production adjustment fluctuation characteristics. The optimal water-drive reservoir production prediction model is obtained by training and optimizing the model using a dataset. The trend component data and seasonal component data to be predicted are input into the optimal water-driven reservoir production prediction model to obtain the reservoir production prediction results.

[0007] In some embodiments, the prediction method further includes: determining the production status of the reservoir over a predetermined time period based on the reservoir production prediction results.

[0008] In some embodiments, a method for determining the production status of an oil reservoir over a predetermined future time period includes: Determine a future preset time period and extract the daily oil production prediction sequence, daily liquid production prediction sequence, and water content prediction sequence within the future preset time period; Calculate the standard deviation of daily oil production series, the rate of change of daily oil production, the variance of water cut series, the average water cut, and the rate of increase in water cut within a future preset time period. Production status is determined by the magnitude of the standard deviation of daily oil production series and the rate of change of daily oil production; water content status is determined by the magnitude of the variance of water content series, the average water content, and the rate of increase in water content; and / or, water content status is determined by the synchronous change relationship between daily liquid production and daily oil production.

[0009] In some embodiments, the prediction method further includes: reservoir production prediction results and determination results.

[0010] In some embodiments, a method for decomposing reservoir production data includes: The reservoir production data were normalized to obtain standard sequence data; The triple exponential moving average algorithm is used to separate the standard sequence data into trend component data that reflects the long-term production capacity change trend and seasonal component data that reflects the fluctuation of injection and production adjustment.

[0011] In some embodiments, the method for extracting reservoir production injection-production adjustment fluctuation characteristics using a multi-head self-attention mechanism includes: The seasonal component data is mapped into query vector, key vector, and value vector respectively through linear projection. Calculate the dot product of the query vector and the transpose of the key vector, and divide the result by the scaling factor to obtain the attention score matrix; The attention score matrix is ​​normalized to obtain the attention weight probability distribution; The attention weight probability distribution is summed with the value vector using a weighted average. By splicing the outputs of multiple attention heads, the injection-sampling adjustment fluctuation characteristics containing temporal dependencies are obtained.

[0012] In some embodiments, the method of fusing long-term reservoir production capacity change characteristics and reservoir production injection-production adjustment fluctuation characteristics using a cross-attention mechanism includes: By using the cross-attention mechanism, the seasonally enhanced trend features are generated with the reservoir production injection-production adjustment fluctuation characteristics as the query vector and the reservoir production long-term capacity change characteristics as the key vector. At the same time, the seasonally enhanced trend features are generated with the reservoir production long-term capacity change characteristics as the query vector and the reservoir production injection-production adjustment fluctuation characteristics as the key vector. The enhanced trend features and seasonal features are combined to obtain the fused features.

[0013] In some embodiments, the water-drive reservoir production prediction model further includes an output layer, which is used to map the fused features to obtain the reservoir production prediction result.

[0014] In some embodiments, the method for obtaining the optimal water-drive reservoir production prediction model includes: Set the training parameters for the water-drive reservoir production prediction model; The dataset is divided into a training set and a validation set; The water-drive reservoir production prediction model was trained using a training set, with the mean square error function used as the loss function. The production prediction model for the water-drive reservoir was evaluated using a validation set. Based on the evaluation results, the hyperparameters of the production prediction model for the water-drive reservoir were adjusted, and the model with the smallest validation error was selected as the optimal production prediction model for the water-drive reservoir.

[0015] A second aspect of this application provides a water-drive reservoir production prediction system for implementing the water-drive reservoir production prediction method described in the first aspect of this application, comprising: The data acquisition module is configured to acquire historical reservoir production data and reservoir production data to be predicted for the target oilfield block. The data processing module is configured to: decompose historical reservoir production data to obtain historical trend component data and historical seasonal component data; and decompose the reservoir production data to be predicted to obtain the trend component data to be predicted and the seasonal component data to be predicted. The dataset generation module combines historical trend component data with historical seasonal component data to form a dataset. The model building module is configured to: take reservoir trend component and reservoir seasonal component as input and reservoir production prediction results as output, and build a water-drive reservoir production prediction model based on a dual-flow network in the PyTorch framework. The model training and optimization module is configured to: train and optimize the water-drive reservoir production prediction model using a dataset to obtain the optimal water-drive reservoir production prediction model; The prediction module is configured to input the trend component data and the seasonal component data to be predicted into the optimal water-driven reservoir production prediction model to obtain the reservoir production prediction results.

[0016] Compared with the prior art, the advantages and positive effects of this application are as follows: (1) The water-drive reservoir production prediction method and system based on dual-stream network provided in this application considers the mutual interference problem that exists in traditional single models when dealing with complex non-stationary data. By separating the trend component data reflecting the long-term production capacity change of the reservoir and the seasonal component data reflecting the injection-production adjustment fluctuation, and by using dual-stream network for targeted modeling, the feature extraction of different characteristic signals is realized by using the dual-stream network architecture. This solves the feature interference problem caused by direct modeling through raw data. The modeling process of long-term production capacity evolution law and injection-production adjustment fluctuation law is completely decoupled and synergistically optimized. Compared with traditional numerical simulation or single deep learning model, it can accurately and quickly predict the production of water-drive reservoir, reduce manpower and material costs, and improve the level of intelligent management of oilfields.

[0017] (2) The water-drive reservoir production prediction method and system based on dual-flow network provided in this application decomposes the reservoir production time series data to be predicted into trend components and seasonal components. The water-drive reservoir production prediction model is trained and optimized based on historical reservoir production data. The trend components and seasonal components are extracted respectively. The water-drive reservoir production prediction results are obtained by feature fusion, which improves the prediction accuracy and efficiency under complex working conditions of water-drive reservoirs. Attached Figure Description

[0018] Figure 1This is a flowchart illustrating the water-driven reservoir production prediction method based on a dual-flow network as described in the embodiments of this application. Figure 2 This is a schematic flowchart of a method for decomposing reservoir production data according to an embodiment of this application; Figure 3 This is a schematic flowchart of a method for normalizing reservoir production data according to an embodiment of this application; Figure 4 This is a schematic diagram of the method for separating standard sequence data into trend component data and seasonal component data using the triple exponential moving average algorithm according to an embodiment of this application; Figure 5 This is a structural block diagram of the water-drive reservoir production prediction model described in the embodiments of this application; Figure 6 This is a schematic diagram of the method for extracting reservoir production injection-production adjustment fluctuation characteristics using a multi-head self-attention mechanism in an embodiment of this application; Figure 7 This is a schematic diagram of the method for fusing long-term reservoir production capacity change characteristics and reservoir production injection-production adjustment fluctuation characteristics using a cross-attention mechanism in an embodiment of this application. Figure 8 This is a schematic diagram of the method for obtaining the optimal water-drive reservoir production prediction model in an embodiment of this application; Figure 9 This is a schematic flowchart illustrating the method for determining the production status of an oil reservoir over a predetermined future time period, as described in this application. Figure 10 This is a schematic diagram of the method for determining production status based on the standard deviation of daily oil production series and the rate of change of daily oil production, as described in this application. Figure 11 This is a schematic flowchart of a method for determining the water content state based on the variance of the water content sequence, the average water content, and the rate of increase in water content, according to an embodiment of this application. Figure 12 This is a schematic flowchart of a method for determining the water content state based on the average water content and the rate of increase in water content according to an embodiment of this application. Figure 13 This is a schematic flowchart of a method for determining water content based on the synchronous changes in daily liquid production and daily oil production, as described in an embodiment of this application. Figure 14 This is a structural block diagram of the water-drive reservoir production prediction system described in the embodiments of this application; Figure 15 This is a structural block diagram of the model training and optimization module described in the embodiments of this application.

[0019] In the figure, 100 is the water-driven reservoir production prediction model, 101 is the trend feature extraction module, 102 is the seasonal feature extraction module, 103 is the fusion module, 104 is the output layer, 200 is the water-driven reservoir production prediction system, 201 is the data acquisition module, 202 is the data processing module, 203 is the dataset generation module, 204 is the model building module, 205 is the model training and optimization module, 2051 is the setting submodule, 2052 is the partitioning submodule, 2053 is the model training submodule, 2054 is the model optimization submodule, 206 is the prediction module, 207 is the judgment module, and 208 is the storage module. Detailed Implementation

[0020] The present application will now be described in detail with reference to the accompanying drawings through exemplary embodiments. However, it should be understood that, without further description, elements, structures, and features in one embodiment may be advantageously incorporated into other embodiments.

[0021] The first aspect of this application provides a method for predicting the production of water-driven oil reservoirs based on a dual-flow network. Figure 1 The diagram shows a flowchart of the water-drive reservoir production prediction method based on a dual-flow network.

[0022] See Figure 1 The method for predicting the production of water-driven reservoirs based on a dual-flow network includes: S1. Obtain historical reservoir production data and predicted reservoir production data for the target oilfield block.

[0023] Specifically, the reservoir production data includes at least daily oil production, daily liquid production, daily water injection, and bottom hole flowing pressure data.

[0024] Daily oil production refers to the volume / mass of pure crude oil produced by an oil reservoir each day, usually measured in units of... , Daily oil production is the core quantitative indicator of reservoir production. The two are the same concept. The level, increase or decrease, and stability of reservoir production are all directly reflected through the numerical change of daily oil production.

[0025] Daily liquid production refers to the volume / mass of the oil and water mixture produced by the reservoir each day. The unit is the same as the daily oil production. Daily liquid production is the total basis for achieving daily oil production. Under the premise of fixed water cut, daily liquid production and daily oil production have a strict positive linear correlation.

[0026] Daily water injection volume is the amount of water injected into the formation through injection wells. It is the core means of artificially replenishing formation energy in oil reservoirs. Daily water injection volume supports daily oil production through two major functions: replenishing formation pressure and displacing crude oil by water drive. A reasonable daily water injection volume is a prerequisite for stable oil reservoir production. Insufficient water injection volume will inevitably lead to a decrease in production, while excessive water injection volume will easily lead to an increase in water cut and a passive decline in oil reservoir production.

[0027] Bottom-hole flowing pressure refers to the pressure at the bottom of a production well. It is the direct driving force for crude oil to flow from the reservoir to the bottom of the well and then be lifted to the surface. Given a fixed reservoir pressure, the lower the bottom-hole flowing pressure, the greater the production pressure differential, and the higher the daily oil production / daily fluid production. However, when the bottom-hole flowing pressure drops to a certain threshold (such as saturation pressure), further pressure reduction will lead to crude oil degassing, a sharp decrease in flowability, and a subsequent decrease in daily oil production / daily fluid production. By controlling the bottom-hole flowing pressure to adjust the reservoir's production pressure differential, the daily oil production of the reservoir can be directly controlled, thereby increasing or stabilizing reservoir production.

[0028] S2. Decompose the historical reservoir production data and the reservoir production data to be predicted separately to obtain historical trend component data and to be predicted trend component data reflecting the long-term changes in reservoir production capacity, historical seasonal component data and to be predicted seasonal component data reflecting the fluctuations in reservoir production injection and production adjustments, and form a dataset by combining the historical trend component data and the historical seasonal component data.

[0029] In this embodiment, on the one hand, by refining the data structure, the two core feature data categories of trend component and seasonal component are precisely separated, effectively avoiding mutual interference between data of different feature dimensions. This allows the inherent patterns of historical reservoir production data and the reservoir production data to be predicted to be presented independently and clearly, laying a high-precision data foundation for subsequent data analysis. On the other hand, a dataset is constructed based on the separated historical trend component data and historical seasonal component data, realizing the systematic integration and standardized collection of core feature data of reservoir production. This enables the dataset to accurately match the feature attributes of reservoir production data, significantly improving the adaptability of the dataset to reservoir production prediction and analysis scenarios, and providing high-quality and highly targeted data support for subsequent modeling, training, and other stages.

[0030] In one embodiment of this application, see Figure 2 Methods for decomposing reservoir production data include: S21. Normalize the reservoir production data to obtain standard sequence data.

[0031] Specifically, linear normalization was used to normalize the reservoir production data to obtain standard sequence data. See [link to documentation]. Figure 3 Methods for normalizing reservoir production data include: S211. Extract the raw sequence data of single reservoir production indicators (such as daily oil production, daily liquid production, etc.) from the reservoir production data. , This represents the data sample size.

[0032] S212. Calculate the minimum value of the original sequence data. and maximum value .

[0033]

[0034]

[0035] S213. For each data point in the original sequence data, calculate according to the following formula to obtain the normalized standard sequence data. All data is mapped to Interval.

[0036]

[0037] S214. Check whether each data point in the standard sequence data is within... If any value exceeds the specified range, it is corrected as an outlier. By checking for and correcting outliers, the validity of the standard sequence data is ensured.

[0038] For example, the daily oil production sequence data in a certain historical oil reservoir production data is as follows: ,get , .

[0039] Normalize for each data point. For example, normalize data point 120:

[0040] The final standard sequence data is (0, 0.163, 0.349, 0.698, 1).

[0041] S22. The standard sequence data is separated into trend component data reflecting the long-term production capacity change trend and seasonal component data reflecting the injection and production adjustment fluctuation using the triple exponential moving average (TEMA) algorithm.

[0042] Specifically, see Figure 4 Methods for separating standard series data into trend and seasonal components using the triple exponential moving average algorithm include: S221. Let the standard sequence data of a certain single-category indicator (e.g., daily oil production) be... The time step is The smoothing coefficient is , .

[0043] Specifically, time step Arranged chronologically by reservoir production time, such as by day or month. Smoothing coefficient. The weighting coefficients for the exponential moving average reflect the importance of recent data. For reservoir production data, The value ranges from 0.1 to 0.3. It should be noted that injection-progression adjustments represent short- to medium-term changes; a smaller coefficient can mitigate short-term fluctuations and more accurately extract long-term trends. However, if injection-progression adjustments are frequent and fluctuations are significant, The possible values ​​are 0.2-0.3, which balances trend smoothing and preserves volatility.

[0044] S222, Initial setting of the initial value of the triple exponential moving average.

[0045] Exponential moving averages are calculated recursively and require one iteration. ), second ( ),three times( The exponential moving average is initialized. Specifically, standard sequence data is used. The mean of the first 3-5 data points is used as the initial value to ensure the representativeness of the initial value.

[0046] For example: , This is the initial value for an exponential moving average. This is the initial value for the quadratic exponential moving average. This is the initial value for a three-exponential moving average. For standard sequence data The values ​​of the first 3 data points.

[0047] S223. Calculate an exponential moving average to smooth short-term random fluctuations.

[0048] Specifically, for standard sequence data An exponential moving average is calculated recursively over time steps to initially remove small random fluctuations in reservoir production data, retaining short- to medium-term fluctuations including injection-production adjustment fluctuations and long-term trends. The calculation formula is as follows:

[0049] In the formula, For the first The first exponential moving average over a time step. For the first The first exponential moving average over a time step. For the first Standard sequence data for time steps.

[0050] S224. Calculate the second exponential moving average to smooth short- to medium-term fluctuations including injection and extraction adjustment fluctuations.

[0051] Specifically, based on the first exponential moving average, the second exponential moving average is recursively calculated using the same smoothing coefficient. The calculation formula is as follows:

[0052] In the formula, For the first The quadratic exponential moving average of the time step. For the first The quadratic exponential moving average of the time step.

[0053] The quadratic exponential moving average sequence was calculated. .

[0054] S225. Calculate three exponential moving averages, remove short- and medium-term fluctuations including injection and production adjustment fluctuations, and retain only those reflecting long-term capacity change trends.

[0055] Specifically, based on the quadratic exponential moving average, the quadratic exponential moving average is recursively calculated while keeping the smoothing coefficient constant. The calculation formula is as follows:

[0056] In the formula, For the first The three exponential moving averages of the time step. For the first The three exponential moving averages of the time step.

[0057] The three exponential moving average series were calculated. .

[0058] S226. Extract trend component data that reflects the long-term trend of production capacity changes.

[0059] Specifically, the three-exponentially-shifted flat sequence As the core baseline for the trend component, the triple exponential moving average sequence is fitted using a trend fitting formula. After correction and optimization, trend component data is obtained. Triple exponential moving average sequence The trend component data obtained after correction The calculation formula, which better reflects the actual changing trend of long-term reservoir productivity, is as follows:

[0060] In the formula, For the first Trend component data for each time step.

[0061] S227. Extract seasonal component data that reflects fluctuations in injection and production adjustments.

[0062] Specifically, through standard sequence data The seasonal component data is obtained by calculating the time-step difference between the trend component data and the seasonal component data. The calculation formula is:

[0063] In the formula, For the first Seasonal component data for each time step.

[0064] It should be noted that in the field of time series signal processing, the core idea of ​​data decomposition is to extract the low-frequency backbone baseline through low-pass filtering, and then obtain the high-frequency detail signal by subtracting the original signal from the baseline. Traditional Simple Moving Average (SMA) or Single Exponential Smoothing (EMA) suffers from severe phase lag when extracting the baseline, leading to significant distortion of the high-frequency signal obtained after subtraction. The TEMA algorithm uses a linear combination of three exponential smoothing steps (…) This greatly offsets the phase lag, making it an excellent low-hysteresis low-pass filter. The smooth curve obtained by calculating the original standard sequence using the TEMA algorithm corresponds well in physical meaning to the "long-term productivity evolution baseline" formed by the natural decay of formation energy or the effectiveness of long-term water injection in water-driven reservoirs. Because the TEMA algorithm has superior hysteresis correction capabilities, this baseline closely follows the natural trend. Subsequently, subtracting this baseline calculated by the TEMA algorithm from the original sequence mathematically constitutes a high-frequency filtering decomposition process. In the actual physical scenario of water-driven reservoirs, this high-frequency oscillation residual, which deviates from the natural baseline, corresponds precisely to the transient response fluctuations induced by artificial intervention (such as temporary well opening / closing, sudden increases / decreases in water injection, and local profile control). Therefore, by utilizing the extremely sensitive smoothing characteristics of the TEMA algorithm to extract low-frequency trends and supplementing it with subtraction operations, the effective decomposition of time-series signals is fully realized mathematically, and the "long-term natural production capacity baseline" and "short-term injection and production adjustment fluctuations" are effectively separated physically. This yields trend component data reflecting the long-term production capacity change trend and seasonal component data reflecting the injection and production adjustment fluctuations.

[0065] S3. Using reservoir trend components and reservoir seasonal components as inputs and reservoir production prediction results as outputs, a water-driven reservoir production prediction model based on a two-stream network is constructed in the PyTorch framework.

[0066] PyTorch is an open-source Python deep learning framework developed by Meta (formerly Facebook) AI Research Institute. It combines dynamic computation graphs, ease of use, and high performance, and is one of the most mainstream development tools in the field of artificial intelligence. It is widely used in the entire process of deep learning model research, training, and deployment, covering various scenarios such as computer vision, natural language processing, and time series prediction.

[0067] The water-driven reservoir production prediction model based on a two-stream network is constructed within the PyTorch framework, facilitating parallel computation and feature fusion of trend and seasonal component branches. PyTorch's automatic differentiation and tensor operators can be used to update model parameters and set multiple loss constraints, while also supporting GPU-accelerated training and inference. After training, the model parameters can be saved and loaded, facilitating integration into subsequent prediction and production status determination processes.

[0068] Specifically, the reservoir production forecast results include daily oil production, daily liquid production, and water cut for a predetermined future time period.

[0069] Specifically, see Figure 5 The water-driven reservoir production prediction model 100 includes a trend feature extraction module 101 for extracting long-term reservoir production capacity change characteristics from the trend component data, a seasonal feature extraction module 102 for extracting reservoir production injection-production adjustment fluctuation characteristics from the seasonal component data, and a fusion module 103 for fusing the long-term reservoir production capacity change characteristics and the reservoir production injection-production adjustment fluctuation characteristics.

[0070] In this embodiment, considering the different physical meanings and data characteristics of the trend component (long-term production capacity change) and the seasonal component (injection-production adjustment fluctuations), a feature extraction branch is constructed using a dual-stream network architecture. The trend branch focuses on capturing the changing patterns of long-term reservoir production capacity (decline / increase), while the seasonal branch focuses on capturing the short-to-medium-term fluctuation patterns brought about by injection-production adjustments. This completely avoids the problem of mutual interference between different feature dimensions in traditional single-branch models, maximizing the discovery of the inherent patterns of both types of core features. The dual-stream network adaptively weights and fuses the long-term regularity features of the trend component and the fluctuation features of the seasonal component through a feature fusion module, preserving the core evolution trend of reservoir production capacity while accurately incorporating the real-time impact of injection-production adjustments.

[0071] In one embodiment of this application, the seasonal feature extraction module 102 uses a multi-head self-attention mechanism to extract reservoir production injection-production adjustment fluctuation features, and the fusion module 103 uses a cross-attention mechanism to fuse long-term reservoir production capacity change features and reservoir production injection-production adjustment fluctuation features.

[0072] Specifically, see [link to relevant documentation] Figure 5 The water-driven reservoir production prediction model 100 further includes an output layer 104, which is used to map the fused features to obtain the reservoir production prediction result.

[0073] In this embodiment of the application, the output layer 104 is a fully connected layer, and the fully connected layer obtains the reservoir production prediction result by mapping the fused features through the Softmax function.

[0074] Specifically, see Figure 6 Methods for extracting reservoir production injection-production adjustment fluctuation characteristics using a multi-head self-attention mechanism include: S31. Map the seasonal component data into query vector, key vector, and value vector respectively through linear projection; S32. Calculate the dot product of the query vector and the transpose of the key vector, and divide the result by the scaling factor to obtain the attention score matrix. S33. Normalize the attention score matrix to obtain the attention weight probability distribution; S34. Sum the attention weight probability distribution with the value vector using a weighted sum. S35. By splicing the outputs of multiple attention heads, the injection-sampling fluctuation characteristics containing temporal dependencies are obtained.

[0075] The multi-head self-attention mechanism mines the temporal dependencies of different time scales and correlation strengths in seasonal component (injection-production adjustment fluctuation) data in parallel using multiple independent attention heads (such as the short-term correlation of weekly / monthly injection-production adjustments and the medium-term correlation of quarterly injection-production cycles). It breaks through the limitation of traditional single attention heads that can only capture single-dimensional temporal patterns and can fully restore the time lag effect of injection-production adjustment "operation-response" (such as the correlation characteristics of production fluctuations 15-30 days after water injection adjustment). The temporal correlation identification accuracy of injection-production adjustment fluctuation characteristics is high.

[0076] Specifically, the multi-head self-attention mechanism effectively solves the problem of attention score distortion caused by differences in dimensions and amplitudes in reservoir injection-production adjustment fluctuation data by normalizing the linear projection of query / key / value vectors and scaling factors. Combined with the normalized attention weight distribution, it can automatically focus on the fluctuation characteristics of key nodes in injection-production adjustment (such as core operation time points such as fracturing, profile control, and fluid extraction), filter out invalid noise such as measurement errors and temporary operating condition fluctuations, and improve the characterization accuracy and anti-interference ability of injection-production adjustment fluctuation characteristics.

[0077] Specifically, see Figure 7 Methods that employ a cross-attention mechanism to integrate long-term reservoir production capacity variation characteristics and reservoir production injection-production adjustment fluctuation characteristics include: S31. Using the cross-attention mechanism, the reservoir production injection-production adjustment fluctuation characteristics are used as the query vector and the reservoir production long-term capacity change characteristics are used as the key vector to generate seasonally enhanced trend characteristics. At the same time, the reservoir production long-term capacity change characteristics are used as the query vector and the reservoir production injection-production adjustment fluctuation characteristics are used as the key vector to generate seasonally enhanced trend characteristics. S32. Combine the enhanced trend features and seasonal features to obtain the fused features.

[0078] The cross-attention mechanism uses bidirectional attention calculations—"injection-production adjustment fluctuation characteristics look up long-term trend characteristics" and "long-term trend characteristics look up injection-production adjustment fluctuation characteristics"—to allow long-term production capacity change characteristics to absorb dynamic fluctuation information of injection-production adjustments (generating seasonally enhanced trend characteristics). At the same time, it allows injection-production adjustment fluctuation characteristics to anchor to the baseline of long-term production capacity evolution (generating seasonally enhanced trend characteristics). This completely solves the problem of "simple superposition and mutual isolation" of the two types of characteristics in traditional splicing / weighted fusion. The information complementarity after feature fusion is improved by more than 50%, which can accurately depict the core law of reservoir production: "injection-production adjustment fluctuation response under long-term production capacity trend constraints".

[0079] S4. The optimal water-drive reservoir production prediction model is obtained by training and optimizing the model using the dataset.

[0080] In one embodiment of this application, see Figure 8 Methods for obtaining the optimal water-drive reservoir production prediction model include: S41. Set the training parameters for the water-drive reservoir production prediction model.

[0081] Specifically, the training parameters include at least the maximum number of iterations, learning rate, optimizer, early stopping count, and Dropout ratio. The training parameters can be set according to actual needs. For example, the maximum number of iterations can be set to 10, the learning rate to 0.0001, the optimizer to use the Adam algorithm, the early stopping count to 3, and the Dropout ratio to 0.2.

[0082] S42. Divide the dataset into a training set and a validation set.

[0083] Specifically, the training set and validation set are divided according to a fixed ratio (e.g., 8:2), and the data in the training set and validation set do not overlap, so as to ensure that the validation set can truly reflect the model's predictive ability for different reservoir production stages.

[0084] S43. The water-drive reservoir production prediction model is trained using the training set, and the mean square error function is used as the loss function.

[0085] Specifically, the weights of the water-drive reservoir production prediction model are updated using forward propagation training and backpropagation training with the training set.

[0086] S44. Evaluate the production prediction model of the water-drive reservoir through the validation set, adjust the hyperparameters of the production prediction model of the water-drive reservoir based on the evaluation results, and select the model with the smallest validation error as the optimal production prediction model of the water-drive reservoir.

[0087] In this embodiment, through a complete process of preset training parameters, dataset partitioning, model training based on mean squared error, and iterative optimization of validation set hyperparameters, the water-drive reservoir production prediction model with the smallest validation error can be quickly and stably selected from multiple training results. This effectively improves the model's generalization ability and prediction accuracy, reduces the subjectivity and uncertainty caused by manual parameter tuning, and achieves accurate and stable prediction of water-drive reservoir production, providing reliable data support for reservoir development scheme optimization and production planning.

[0088] S5. Input the trend component data and seasonal component data to be predicted into the optimal water-driven reservoir production prediction model to obtain the reservoir production prediction results.

[0089] In one embodiment of this application, the prediction method further includes: determining the production status of the reservoir over a preset time period based on the reservoir production prediction results.

[0090] In this embodiment of the application, based on the reservoir production prediction results, the production status of the reservoir in a future preset time period can be accurately and quantitatively determined, so as to realize the advance prediction and dynamic evaluation of the reservoir production trend.

[0091] Specifically, see Figure 9 Methods for determining the production status of an oil reservoir over a predetermined future time period include: S51. Determine the future preset time period. Extract the future preset time period Daily oil production forecast sequence Daily liquid production prediction sequence and moisture content prediction sequence .

[0092] S52, Calculate the future preset time period Standard deviation of daily oil production series Daily oil production change rate index Variance of moisture content series Average moisture content and rate of increase in water content .

[0093] Specifically, the formula for calculating the daily oil production change rate is as follows:

[0094] In the formula, To predict the start time Daily oil production To predict the end time Daily oil production.

[0095] Specifically, the rate of water cut increase The calculation formula is expressed as:

[0096] In the formula, To predict the start time moisture content, To predict the end time The moisture content.

[0097] S53. Based on the standard deviation of daily oil production series Daily oil production change rate index The magnitude of the moisture content sequence variance indicates the yield status. Average moisture content and rate of increase in water content The size determines the water content, and / or, based on the daily liquid production. Daily oil production The synchronous changes and relationships are used to determine the water content.

[0098] Specifically, see Figure 10 Based on the standard deviation of daily oil production series Daily oil production change rate index Methods for judging production status by size include: S531. Determining the standard deviation of daily oil production series Is it greater than the set production fluctuation threshold? S532, If the daily oil production series standard deviation If the output exceeds the set output fluctuation threshold, the output status is determined to be in a state of output fluctuation; otherwise, the daily oil production change rate index is used as the indicator. Whether the absolute value is less than or equal to the preset indicator threshold; S533, If the daily oil production change rate index If the absolute value is less than or equal to a preset threshold, the production status is determined to be stable; otherwise, the daily oil production change rate is judged. Whether it exceeds the preset threshold, or to determine the daily oil production change rate indicator. Is it less than the negative preset threshold? S534, If the daily oil production change rate index If the daily oil production change rate exceeds the preset threshold, the production status is determined to be in a recovery state; if the daily oil production change rate index is higher than the preset threshold, the production status is determined to be in a recovery state. If the output is less than the negative preset threshold, the output status is determined to be a declining output status.

[0099] Specifically, the production fluctuation threshold can be set to 1.5-2 times the standard deviation of daily oil production during the historical stable production period, and the preset index threshold can be set to 1-2 times the standard deviation of the historical daily oil production change rate, or 5%-10% based on oilfield experience. The specific threshold can be set according to actual needs.

[0100] Specifically, see Figure 11 Based on the variance of the moisture content series Average moisture content and rate of increase in water content Methods for determining water content include: S531. Determine the variance of the moisture content sequence. Is it greater than the set variance threshold? S532. If yes, determine the water content state as fluctuating; if no, determine the average water content. Is the water content less than the preset water content threshold and is the rate of water content increase... Whether it is less than the set rate threshold, or to determine the rate of water cut rise. Is it less than or equal to 0? S533, if the average moisture content Less than the preset water content threshold and the rate of water content increase Less than the set rate threshold, or the rate of water content increase If the value is less than or equal to 0, the water content is determined to be in a controlled water content state.

[0101] Specifically, the preset water content threshold can be 80% or 90%, the set variance threshold can be twice the historical water content fluctuation variance, and the set rate threshold can be determined based on historical statistical characteristics. It should be noted that the preset water content threshold and the set variance threshold can be set according to actual needs.

[0102] Specifically, see Figure 12 Based on average moisture content and rate of increase in water content Methods for determining water content include: S531. Determining the average moisture content Is it less than the preset moisture content limit value? S532. If so, determine the rate of water content increase. Is it less than the set rate threshold? If not, determine the rate of water cut increase. Is the rate greater than the set rate threshold? S533, if the rate of increase in water content If the rate is less than a set threshold, the water content is determined to be in a stable state; if the rate of increase in water content is less than a set threshold, the water content is determined to be in a stable state. If the rate exceeds the set threshold, the water content is determined to be in an increasing state.

[0103] Specifically, see Figure 13 Based on daily liquid production Daily oil production Methods for determining water content by analyzing synchronous changes and their relationships include: S531. Determine whether the change in daily liquid production is less than the first set liquid production threshold, and simultaneously determine whether the change in daily oil production is greater than the first set oil production threshold. S532. If the change in daily liquid production is less than the first set liquid production threshold and the change in daily oil production is greater than the first set oil production threshold, the water content is determined to be in a state of rising water content; otherwise, it is determined whether the change in daily liquid production is greater than the second set liquid production threshold, and simultaneously it is determined whether the change in daily oil production is less than the second set oil production threshold. S533. If the change in daily liquid production is greater than the second set liquid production threshold and the change in daily oil production is less than the second set oil production threshold, the water content state is determined to be a water content fluctuation state.

[0104] It should be noted that the first set liquid volume threshold, the second set liquid volume threshold, the first set oil volume threshold, and the second set oil volume threshold can all be set according to actual needs. The first set liquid volume threshold and the second set oil volume threshold have extremely small values, specifically meaning that the daily liquid and oil production changes very little, remaining essentially stable. The second set liquid volume threshold and the first set oil volume threshold have relatively large values, specifically meaning that the daily liquid and oil production changes significantly.

[0105] S54. Determine the production status of the reservoir within a preset time period based on the production status and water content status through a preset state relationship.

[0106] Specifically, the preset state relationships are presented in tabular form, as shown in Table 1.

[0107] Table 1 Preset State Relationships

[0108] For example, if the production status is stable and the water cut status is stable, it can be seen from the preset status relationship shown in Table 1 that the production status of the reservoir during the preset time period is a stable production status.

[0109] In one embodiment of this application, the prediction method further includes: reservoir production prediction results and determination results.

[0110] In this embodiment, data storage ensures data traceability and provides direct data support for reviewing the oil and gas development process and verifying its effectiveness. Stored data serves as a fundamental data source for reservoir dynamic analysis, supporting subsequent optimization of production prediction models and iteration of production status judgment criteria. Statistical and comparative analysis of historical stored data helps uncover reservoir production patterns, improving the accuracy of subsequent predictions and judgments, forming a closed-loop management system of "prediction-judgment-storage-optimization," and enhancing the adaptive capability of the prediction method.

[0111] The water-drive reservoir production prediction method described in this application accurately separates the long-term production capacity variation pattern and injection-production adjustment fluctuation characteristics of the reservoir by performing trend-seasonal dual-dimensional decomposition on the reservoir production data. The dual-stream network model built based on PyTorch can selectively extract the two types of features and achieve effective fusion, overcoming the problem of insufficient prediction accuracy caused by traditional single-feature modeling. The optimal prediction model obtained by training and optimizing the dataset can accurately output the daily oil production, daily liquid production, and water cut for a preset time period in the future, significantly improving the accuracy, relevance, and timeliness of water-drive reservoir production prediction.

[0112] A second aspect of this application provides a water-drive reservoir production prediction system for implementing the water-drive reservoir production prediction method based on a dual-flow network described in the first aspect of this application.

[0113] See Figure 14 The water-drive reservoir production prediction system 200 includes: The data acquisition module 201 is configured to acquire historical reservoir production data and reservoir production data to be predicted for the target oilfield block. The data processing module 202 is configured to: decompose historical reservoir production data to obtain historical trend component data and historical seasonal component data; and decompose the reservoir production data to be predicted to obtain the trend component data to be predicted and the seasonal component data to be predicted. The dataset generation module 203 combines historical trend component data and historical seasonal component data to form a dataset. Model building module 204 is configured to: take reservoir trend component and reservoir seasonal component as input and reservoir production prediction result as output, and build a water-drive reservoir production prediction model based on dual-flow network in the PyTorch framework. The model training and optimization module 205 is configured to: train and optimize the water-drive reservoir production prediction model through the dataset to obtain the optimal water-drive reservoir production prediction model; The prediction module 206 is configured to input the trend component data and the seasonal component data to be predicted into the optimal water-driven reservoir production prediction model to obtain the reservoir production prediction results.

[0114] In one embodiment of this application, see also Figure 14 The system also includes a determination module 207, which is configured to determine the production status of the reservoir in a future preset time period based on the reservoir production prediction results.

[0115] In one embodiment of this application, see also Figure 14The system also includes a storage module 208, which is configured to store reservoir production prediction results and determination results.

[0116] In one embodiment of this application, see Figure 15 The model training optimization module 205 includes: Submodule 2051 is configured to: set the training parameters of the water-drive reservoir production prediction model; Submodule 2052 is configured to divide the dataset into a training set and a validation set. The model training submodule 2053 is configured to: train the water-drive reservoir production prediction model using the training set; The model optimization submodule 2054 is configured to: evaluate the production prediction model of the trained downstream reservoir through the validation set, adjust the hyperparameters of the production prediction model of the water-drive reservoir based on the evaluation results, and select the model with the smallest validation error as the optimal production prediction model of the water-drive reservoir.

[0117] To verify the effectiveness of the water-drive reservoir production prediction method and system based on dual-flow network described in the above embodiments of this application, the following specific embodiments are provided for illustration.

[0118] Example 1: Taking the prediction of reservoir production in a water-drive oil reservoir block as an example.

[0119] Historical daily oil production data of the water-drive reservoir block were selected, and a prediction step size of P=30 days was set. Existing methods and the method and system described in this application were used to predict the reservoir production of the water-drive reservoir block. Eight sample points were selected from the test set for prediction. The comparison results of the predicted daily oil production values ​​and the actual daily oil production values ​​of different prediction methods are shown in Table 2. The comparison results of the predicted water cut values ​​and the actual water cut values ​​of different prediction methods are shown in Table 3. The comparison results of the prediction errors of different prediction methods on the test set are shown in Table 4.

[0120] Table 2 Comparison of predicted and actual daily oil production (unit: t / d)

[0121] As shown in Table 2, under the same sample points of actual daily oil production, the predicted values ​​of each method can reflect the overall trend of change. However, the predicted values ​​of the method and system in this application are closer to the actual values, and the overall deviation is smaller.

[0122] Table 3 Comparison of predicted and actual moisture content values ​​(unit: dimensionless)

[0123] As shown in Table 3, under the same sample points with true moisture content, the method and system of this application fit the slow upward trend more accurately in moisture content prediction, and maintain a small deviation at each moisture content sample point.

[0124] Table 4 Comparison of prediction errors of different prediction methods on the test set

[0125] As shown in Table 4, the method and system of this application achieve smaller MAE and RMSE in the prediction of daily oil production, daily liquid production and water content, demonstrating higher prediction accuracy and stronger stability.

[0126] Example 2: Predicting reservoir production from a single well.

[0127] Historical reservoir production data at the daily scale for a single well (including at least daily oil production, daily fluid production, daily water injection, and bottomhole flowing pressure) were selected. Reservoir prediction results included water cut. The historical window length was set to L = 60 days, and the prediction step size to P = 15 days. After normalizing the reservoir production data, a triple exponential moving average algorithm was used to separate the trend component and the seasonal component. The trend component and seasonal component values ​​were input into the optimal water-drive reservoir production prediction model, outputting the daily oil production, daily fluid production, and water cut for the next 15 days. To verify the advantages of the proposed method and system during water cut fluctuations, it was compared with existing methods on the same test set. The comparison results of the predicted water cut values ​​and actual water cut values ​​for different prediction methods are shown in Table 5, and the comparison results of the prediction errors of different prediction methods on the test set are shown in Table 6.

[0128] Table 5 Comparison of predicted and actual moisture content values ​​(unit: dimensionless)

[0129] As shown in Table 5, within the sample segment where the moisture content fluctuates in stages, although the existing prediction methods can make an overall trend judgment, they have varying degrees of lag or bias. The method and system of this application predict the values ​​at each sample point more closely to the actual values ​​and can more accurately track the changes in moisture content.

[0130] Table 6. Comparison of prediction errors of different prediction methods on the test set.

[0131] As shown in Table 6, the method and system of this application still maintain the minimum MAE and RMSE under short-term prediction (P=15 days) conditions, and the error reduction is more obvious, especially in the moisture content index. This indicates that the method and system of this application can effectively improve the prediction accuracy of moisture content fluctuation stage.

[0132] The above embodiments are used to explain this application, not to limit it. Any modifications and changes made to this application within the spirit and scope of the claims shall fall within the protection scope of this application.

Claims

1. A method for predicting the production of water-driven oil reservoirs based on a two-stream network, characterized in that, include: Acquire historical reservoir production data and predicted reservoir production data for the target oilfield block; The historical reservoir production data and the reservoir production data to be predicted are decomposed separately to obtain historical trend component data and to be predicted trend component data reflecting long-term changes in reservoir production capacity, and historical seasonal component data and to be predicted seasonal data reflecting fluctuations in reservoir production injection and production adjustments. The historical trend component data and historical seasonal data are then combined to form a dataset. Using reservoir trend components and reservoir seasonal components as inputs and reservoir production prediction results as outputs, a water-drive reservoir production prediction model based on a two-stream network is constructed under the PyTorch framework. The reservoir production prediction results include daily oil production, daily liquid production, and water cut for a preset future time period. The water-drive reservoir production prediction model includes a trend feature extraction module for extracting long-term reservoir production capacity change characteristics from the trend component data, a seasonal feature extraction module for extracting reservoir production injection-production adjustment fluctuation characteristics from the seasonal component data, and a fusion module for integrating the long-term reservoir production capacity change characteristics and the reservoir production injection-production adjustment fluctuation characteristics. The optimal water-drive reservoir production prediction model is obtained by training and optimizing the model using a dataset. The trend component data and seasonal component data to be predicted are input into the optimal water-driven reservoir production prediction model to obtain the reservoir production prediction results.

2. The method for predicting water-drive reservoir production based on a dual-flow network as described in claim 1, characterized in that, The prediction method also includes: determining the production status of the reservoir in a future preset time period based on the reservoir production prediction results.

3. The method for predicting water-driven reservoir production based on a dual-flow network as described in claim 2, characterized in that, Methods for determining the production status of an oil reservoir over a predetermined future time period include: Determine a future preset time period and extract the daily oil production prediction sequence, daily liquid production prediction sequence, and water content prediction sequence within the future preset time period; Calculate the standard deviation of daily oil production series, the rate of change of daily oil production, the variance of water cut series, the average water cut, and the rate of increase in water cut within a future preset time period. Production status is determined by the magnitude of the standard deviation of daily oil production series and the rate of change of daily oil production; water content status is determined by the magnitude of the variance of water content series, the average water content, and the rate of increase in water content; and / or, water content status is determined by the synchronous change relationship between daily liquid production and daily oil production.

4. The method for predicting water-drive reservoir production based on a dual-flow network as described in claim 1, characterized in that, Methods for decomposing reservoir production data include: The reservoir production data were normalized to obtain standard sequence data; The triple exponential moving average algorithm is used to separate the standard sequence data into trend component data that reflects the long-term production capacity change trend and seasonal component data that reflects the fluctuation of injection and production adjustment.

5. The method for predicting water-driven reservoir production based on a dual-flow network as described in claim 1, characterized in that, The seasonal feature extraction module uses a multi-head self-attention mechanism to extract reservoir production injection-production adjustment fluctuation features, and the fusion module uses a cross-attention mechanism to fuse long-term reservoir production capacity change features and reservoir production injection-production adjustment fluctuation features.

6. The method for predicting water-driven reservoir production based on a dual-flow network as described in claim 5, characterized in that, Methods for extracting reservoir production injection-production adjustment fluctuation characteristics using a multi-head self-attention mechanism include: The seasonal component data is mapped into query vector, key vector, and value vector respectively through linear projection. Calculate the dot product of the query vector and the transpose of the key vector, and divide the result by the scaling factor to obtain the attention score matrix; The attention score matrix is ​​normalized to obtain the attention weight probability distribution; The attention weight probability distribution is summed with the value vector using a weighted average. By splicing the outputs of multiple attention heads, the injection-sampling adjustment fluctuation characteristics containing temporal dependencies are obtained.

7. The method for predicting water-drive reservoir production based on a dual-flow network as described in claim 5, characterized in that, Methods that integrate long-term reservoir production capacity variation characteristics and reservoir production injection-production adjustment fluctuation characteristics using a cross-attention mechanism include: By using the cross-attention mechanism, the seasonally enhanced trend features are generated with the reservoir production injection-production adjustment fluctuation characteristics as the query vector and the reservoir production long-term capacity change characteristics as the key vector. At the same time, the seasonally enhanced trend features are generated with the reservoir production long-term capacity change characteristics as the query vector and the reservoir production injection-production adjustment fluctuation characteristics as the key vector. The enhanced trend features and seasonal features are combined to obtain the fused features.

8. The method for predicting water-driven reservoir production based on a dual-flow network as described in claim 7, characterized in that, The water-drive reservoir production prediction model also includes an output layer, which is used to map the fused features to obtain the reservoir production prediction result.

9. The method for predicting water-drive reservoir production based on a dual-flow network as described in claim 1, characterized in that, Methods for obtaining the optimal water-drive reservoir production prediction model include: Set the training parameters for the water-drive reservoir production prediction model; The dataset is divided into a training set and a validation set; The water-drive reservoir production prediction model was trained using a training set, with the mean square error function used as the loss function. The production prediction model for the water-drive reservoir was evaluated using a validation set. Based on the evaluation results, the hyperparameters of the production prediction model for the water-drive reservoir were adjusted, and the model with the smallest validation error was selected as the optimal production prediction model for the water-drive reservoir.

10. A water-drive reservoir production prediction system, used to implement the water-drive reservoir production prediction method based on a dual-flow network as described in any one of claims 1 to 9, characterized in that, include: The data acquisition module is configured to acquire historical reservoir production data and reservoir production data to be predicted for the target oilfield block. The data processing module is configured to: decompose historical reservoir production data to obtain historical trend component data and historical seasonal component data; and decompose the reservoir production data to be predicted to obtain the trend component data to be predicted and the seasonal component data to be predicted. The dataset generation module combines historical trend component data with historical seasonal component data to form a dataset. The model building module is configured to: take reservoir trend component and reservoir seasonal component as input and reservoir production prediction results as output, and build a water-drive reservoir production prediction model based on a dual-flow network in the PyTorch framework. The model training and optimization module is configured to: train and optimize the water-drive reservoir production prediction model using a dataset to obtain the optimal water-drive reservoir production prediction model; The prediction module is configured to input the trend component data and the seasonal component data to be predicted into the optimal water-driven reservoir production prediction model to obtain the reservoir production prediction results.

Images