Oil and gas field integrated dynamic simulation construction method and system based on artificial intelligence

Abstract

The invention discloses an oil and gas field integrated dynamic simulation construction method and system based on artificial intelligence, and the method comprises the steps: S1, carrying out the space-time alignment and standardized cleaning of a multi-source heterogeneous data signal, and generating a fusion data signal with unified space-time; s2, a physical information embedded deep learning model architecture is constructed, the architecture receives the fusion data signal, and a mixed training target signal fusing data driving and mechanism driving is generated; s3, training the deep learning model by using a fusion data signal of a historical time sequence, and finally obtaining a convergent intelligent agent model; and S4, feeding the current oil and gas reservoir state parameters and the production control parameters into the intelligent agent model as input signals, and directly outputting a dynamic prediction signal. The oil and gas field integrated dynamic simulation construction method and system based on artificial intelligence can solve the problem that traditional oil and gas field numerical simulation calculation is low in efficiency.

Description

Technical Field

[0001] This invention relates to the field of oil and gas field development simulation and artificial intelligence, specifically to an integrated dynamic simulation construction method and system for oil and gas fields based on artificial intelligence. Background Technology

[0002] Oil and gas field development is a complex systems engineering project, and its lifecycle decisions heavily rely on accurate predictions of underground oil, gas, and water flow dynamics. Traditionally, this task has been primarily undertaken by numerical simulation techniques based on physical equations. This technique predicts the future dynamics of oil and gas fields by solving a set of partial differential equations describing the flow of multiphase fluids in porous media. Although this method has a solid physical foundation and high predictive reliability, its computational process is extremely intensive; building a detailed reservoir model and completing a single simulation prediction often takes hours or even days. This high time cost severely limits its application in scenarios requiring rapid response, such as real-time production optimization, rapid comparison of development schemes, and uncertainty assessment.

[0003] In recent years, with the rise of big data and artificial intelligence technologies, data-driven proxy model technologies have emerged. These technologies learn mapping relationships in historical data through machine learning algorithms, achieving millisecond-level rapid predictions and significantly improving computational efficiency. However, these purely data-driven models are essentially "black boxes," heavily reliant on the coverage of training data. For new operating conditions or development strategies not covered by the data, their predictions often lack physical plausibility, potentially exhibiting violations of fundamental physical laws such as matter and energy conservation, thus greatly diminishing their decision-making guidance and credibility in industrial applications. Therefore, current dynamic simulation technology for oil and gas fields faces a dilemma: traditional numerical simulations offer high accuracy but are slow, failing to meet real-time requirements; artificial intelligence models are fast but lack physical consistency, resulting in unreliable predictions. The industry urgently needs a new simulation method that balances computational efficiency and physical accuracy to establish a rapid channel from data to decision-making, enabling intelligent and refined management of oil and gas field development. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a method and system for constructing integrated dynamic simulation of oil and gas fields based on artificial intelligence, to solve the problems of low computational efficiency in traditional numerical simulation of oil and gas fields, and the lack of physical logic and unreliable prediction results in pure artificial intelligence models. This invention integrates geological statics, production dynamics, and real-time monitoring data to form a standardized signal, which drives a deep learning model embedded with physical information. The core of this model lies in its loss function, which not only includes data fitting terms but also innovatively introduces physical equations characterizing oil and gas seepage as constraint terms, thus deeply embedding physical conservation laws into the learning process of artificial intelligence. The intelligent agent model trained in this way possesses both the rapid computational capabilities of artificial intelligence and the physical logic and reliability of extrapolation prediction in traditional numerical simulation, ultimately achieving rapid and high-precision simulation and prediction of oil and gas field production dynamics.

[0005] This invention provides an artificial intelligence-based method for constructing integrated dynamic simulations of oil and gas fields, comprising: S1: Collect geological static data signals, historical and real-time production dynamic data signals, and downhole monitoring data signals of the target oil and gas field. Perform spatiotemporal alignment and standardization cleaning on the multi-source heterogeneous data signals to generate a spatiotemporally unified fused data signal. S2: Construct a deep learning model architecture with embedded physical information. This architecture receives fused data signals and generates a hybrid training target signal that combines data-driven and mechanism-driven approaches by introducing physical equation constraint terms that characterize the seepage law of oil and gas into the model loss function. S3: The deep learning model is trained using fused data signals from historical time series. The prediction signal is generated by forward propagation, and the model parameters are optimized by backpropagation with the criterion of minimizing the mixed training target signal. Finally, a converged intelligent agent model is obtained. S4: The current reservoir status parameters and production control parameters are used as input signals and fed into the intelligent agent model. Through one forward calculation of the model, the dynamic prediction signals of the pressure field, saturation field distribution of the entire reservoir and the production capacity of a single well in the future prediction period are directly output.

[0006] In one embodiment of the present invention, the spatiotemporal alignment and standardization cleaning of multi-source heterogeneous data signals in step S1 specifically includes: performing spatial discretization interpolation processing on the geological static data signals according to the oil and gas field geological grid model to generate static parameter field signals that match the grid model; performing time series resampling and alignment on the production dynamic data signals and downhole monitoring data signals to unify them to a common timestamp sequence, and using a standardization algorithm to eliminate the influence of different physical dimensions to generate a spatiotemporally unified fused data signal, which is directly used as the standardization input of the deep learning model architecture for embedding physical information in step S2.

[0007] In one embodiment of the present invention, the deep learning model architecture for embedding physical information in step S2 is a composite network structure, which includes a multilayer perceptron or recurrent neural network for encoding spatiotemporal features in the fused data signal, and a parallel physical residual calculation module. The physical residual calculation module calculates the physical conservation residual on the grid cell based on Darcy's law and the mass balance equation in oil and gas seepage mechanics, and uses this residual as a constraint term of the physical equation and the data fitting error term to form a mixed training target signal.

[0008] In one embodiment of the present invention, the physical residual calculation module further introduces specific processing for heterogeneous oil and gas reservoirs when calculating the physical conservation residuals. It dynamically adjusts the coefficients of the physical equations at different spatial locations by fusing the permeability field and porosity field signals in the data signal, so that the constraint terms of the physical equations can accurately characterize the non-uniform seepage characteristics of the actual oil and gas reservoirs. This generates a spatially adaptive hybrid training target signal that is highly consistent with the real physical process, thereby improving the simulation accuracy of the model under complex geological conditions.

[0009] In one embodiment of the present invention, the training process in step S3 adopts a phased optimization strategy. First, the data fitting ability of the deep learning model is optimized separately using the fused data signal of historical time series. Then, while keeping the backbone parameters of the model unchanged, the mixed training target signal is started for joint fine-tuning. This strategy aims to balance the convergence stability of the model in the early stage with the consistency of physical laws in the later stage, and finally obtain a converged intelligent surrogate model that conforms to the trend of historical data and strictly follows the physical conservation law.

[0010] In one embodiment of the present invention, step S3 introduces an uncertainty-based weighting mechanism during the training process. This mechanism dynamically adjusts the weight coefficients of the data-driven and physical-driven terms in the total loss function by monitoring the relative changes of the data-driven and physical-driven terms in the mixed training target signal during the training process. This results in an emphasis on quickly fitting the data in the early stage of training and on satisfying physical constraints in the later stage of training. This adaptive strategy ensures that the training process of the intelligent agent model is smoother and more efficient.

[0011] In one embodiment of the present invention, the current reservoir state parameters and production control parameters in step S4 are used as input signals, specifically including bottom hole flowing pressure, single well injection and production, and the latest known pressure field and saturation field distribution extracted from the fused data signals. The intelligent proxy model simulates the continuous changes of the input signals in a preset future time period in a forward calculation through its internally encoded spatiotemporal evolution law, and directly outputs high-resolution full reservoir pressure field and saturation field evolution signals, as well as dynamic prediction signals of oil production, gas production and water production of each wellhead derived therefrom.

[0012] In one embodiment of the present invention, after the method outputs the dynamic prediction signal in step S4, it further includes an online model self-update step. This step compares the newly acquired actual production monitoring data signal from the oil and gas field with the dynamic prediction signal previously output by the model to generate a model prediction deviation signal. This deviation signal and the corresponding new data signal are then used to incrementally train or fine-tune the parameters of the intelligent agent model, so that the model can continuously adapt to the dynamic changes of the oil and gas reservoir and maintain its long-term prediction accuracy.

[0013] In one embodiment of the present invention, the deep learning model architecture for embedding physical information constructed in step S2 adopts a meta-learning framework. This framework enables the model to learn the general dynamic laws of oil and gas reservoirs by pre-training on a large number of virtual oil and gas reservoir scenarios with different geological and development conditions. When applied to a target oil and gas field, it only needs to be quickly fine-tuned using a small amount of fused data signals from the oil and gas field to adapt the model and form a high-precision intelligent proxy model suitable for the specific oil and gas field.

[0014] This invention also includes an artificial intelligence-based integrated dynamic simulation construction system for oil and gas fields, comprising: The acquisition module collects geological static data signals, historical and real-time production dynamic data signals, and downhole monitoring data signals of the target oil and gas field. It performs spatiotemporal alignment and standardization cleaning on the multi-source heterogeneous data signals to generate a spatiotemporally unified fused data signal. The coordination module constructs a deep learning model architecture that embeds physical information. This architecture receives fused data signals and generates a hybrid training target signal that integrates data-driven and mechanism-driven approaches by introducing physical equation constraint terms that characterize the seepage law of oil and gas into the model loss function. The training module uses fused historical time series data signals to train the deep learning model. It generates prediction signals through forward propagation and optimizes model parameters through backpropagation with the criterion of minimizing the mixed training target signal, ultimately obtaining a converged intelligent agent model. The analysis module takes the current reservoir status parameters and production control parameters as input signals and feeds them into the intelligent agent model. Through one forward calculation of the model, it directly outputs the dynamic prediction signals of the pressure field, saturation field distribution of the entire reservoir and the production capacity of a single well in the future prediction period.

[0015] This invention provides an AI-based integrated dynamic simulation construction method and system for oil and gas fields. It integrates static geological data, dynamic production data, and real-time monitoring data to form a standardized signal, which drives a deep learning model embedded with physical information. The core of this model lies in its loss function, which not only includes data fitting terms but also innovatively introduces physical equations characterizing oil and gas seepage as constraint terms, thus deeply embedding physical conservation laws into the AI ​​learning process. The intelligent agent model trained in this way possesses both the rapid computational power of AI and the physical logic and reliability of extrapolation prediction inherent in traditional numerical simulation, ultimately achieving rapid and high-precision simulation and prediction of oil and gas field production dynamics. Attached Figure Description

[0016] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the 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.

[0017] Figure 1 A flowchart illustrating the method for constructing an integrated dynamic simulation of oil and gas fields based on artificial intelligence; Figure 2 System architecture diagram for constructing an integrated dynamic simulation system for oil and gas fields based on artificial intelligence. Detailed Implementation

[0018] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, unless otherwise specified, the following embodiments and features described therein can be combined with each other.

[0019] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Therefore, the drawings only show the components related to the present invention and are not drawn according to the actual number, shape and size of the components in the actual implementation. In the actual implementation, the form, quantity and proportion of each component can be arbitrarily changed, and the layout of the components may also be more complex.

[0020] In the following description, numerous details are explored to provide a more thorough explanation of embodiments of the invention. However, it will be apparent to those skilled in the art that embodiments of the invention may be practiced without these specific details. In other embodiments, well-known structures and devices are shown in block diagram form rather than in detail to avoid obscuring embodiments of the invention.

[0021] Please see Figure 1-2 The figure shows the method and system for constructing integrated dynamic simulation of oil and gas fields based on artificial intelligence according to the present invention. The AI-based dynamic simulation construction method for integrated oil and gas fields of this invention includes: S1: Collecting geological static data signals, historical and real-time production dynamic data signals, and downhole monitoring data signals of the target oil and gas field; performing spatiotemporal alignment and standardization cleaning on the multi-source heterogeneous data signals to generate a spatiotemporally unified fused data signal; S2: Constructing a deep learning model architecture with embedded physical information. This architecture receives the fused data signal and generates a hybrid training target signal driven by fused data and mechanism by introducing physical equation constraint terms representing the oil and gas seepage law into the model loss function; S3: Training the deep learning model using the fused data signal of historical time series; generating a prediction signal through forward propagation; and optimizing the model parameters through backpropagation with minimizing the hybrid training target signal as the criterion, ultimately obtaining a converged intelligent proxy model; S4: Feeding the current oil and gas reservoir state parameters and production control parameters as input signals into the intelligent proxy model; and directly outputting dynamic prediction signals of the pressure field, saturation field distribution, and single-well production capacity of the entire oil and gas reservoir during the future prediction period through one forward calculation of the model.

[0022] like Figure 1As shown, the method involved in this invention begins with the comprehensive acquisition and in-depth processing of multi-source heterogeneous data from the target oil and gas field. This initial step is the cornerstone for building all subsequent high-precision simulations. Specifically, the data signals to be acquired cover three core dimensions: First, geological static data signals, which include reservoir structures, faults, sequence stratigraphy, and key petrophysical parameters such as the three-dimensional spatial distribution of permeability field, porosity field, and net-to-gross ratio, obtained through geological studies, geophysical logging, and seismic interpretation. Second, historical and real-time production dynamic data signals, which originate from the oilfield production database, specifically including the daily oil production, gas production, water production, and water cut of each production well, as well as the daily water injection or gas injection volume of injection wells. These data form long-term time series on a daily or shorter time scale. Finally, downhole monitoring data signals, which come from permanent sensors installed downhole, such as downhole pressure gauges, thermometers, and distributed fiber optic sensing systems, which can provide more detailed and real-time information on pressure and temperature changes within the wellbore and even the reservoir. These data differ significantly in source, format, and spatiotemporal resolution, constituting a typical "heterogeneous" characteristic. Geological models are static grids in three-dimensional space, production data are time series from single well locations, while monitoring data may be a combination of one-dimensional spatial distribution along the wellbore and a time series. Therefore, spatiotemporal alignment and standardization cleaning are crucial and indispensable preprocessing steps. Spatiotemporal alignment means unifying all data under the same reference frame. Spatially, all well point data, and even some monitoring data, need to be mapped onto a unified three-dimensional geological grid model through grid interpolation or allocation algorithms, ensuring that each grid cell possesses its corresponding static attributes and time-varying dynamic attributes. Temporally, all dynamic data signals need to be unified to a common timestamp sequence. For data with inconsistent sampling frequencies, resampling is required. For example, high-frequency downhole pressure data can be aggregated to a time step consistent with the daily production data, or low-frequency data can be supplemented into higher-frequency time series through interpolation methods to ensure that the data input to the model is completely synchronized in the temporal dimension at each simulation time step.

[0023] Furthermore, standardization cleaning follows, aiming to eliminate the negative impact of different physical dimensions and numerical magnitudes on model training. For example, permeability values ​​may range from a few millidarcy to several thousand millidarcy, while porosity is expressed as a decimal or percentage, and pressure values ​​may reach tens of megapascals. Directly inputting these vastly different data into the model without processing will lead to instability in the optimization process; the model will tend to focus on variables with large values ​​while ignoring variables with smaller values ​​but equally important physical meaning. Therefore, standardization algorithms, such as min-max scaling or Z-score standardization, are needed to transform the values ​​of each data dimension into a unified, dimensionless distribution range, such as the [0, 1] interval or a standard normal distribution with a mean of 0 and a variance of 1. After this series of precise data preprocessing steps, a well-formatted, spatiotemporally unified, and numerically standardized fused data signal is generated. This signal constitutes a high-quality direct input that the subsequent intelligent model can understand and process. After obtaining the high-quality fused data signal, the core innovation of this method lies in the constructed deep learning model architecture that embeds physical information. This architecture is not a standard, off-the-shelf neural network, but a composite network structure tailored for dynamic simulation of oil and gas fields. Its design aims to deeply couple data-driven learning with physical constraints within the model. The architecture first includes a backbone network for encoding complex spatiotemporal features in the fused data signal. This backbone network can be a deep multilayer perceptron, responsible for learning complex nonlinear mappings from input parameters to output states; or, to better capture the dynamic characteristics of time series, a recurrent neural network such as a long short-term memory network or its variants can be used to explicitly model the dependencies of reservoir state evolution over time. However, the real innovation lies in a physical residual calculation module set up in parallel with this backbone network. This module is the physical engine for embedding physical knowledge into the artificial intelligence model. Its working principle is as follows: during each forward propagation of the model training, the backbone network predicts the reservoir state (e.g., the predicted pressure field and saturation field) for the next moment based on the current input signals (e.g., the current pressure, saturation distribution, and production control parameters). At this point, the physical residual calculation module begins to operate. Instead of relying on data, it is based on classic, industry-validated principles of oil and gas flow mechanics, specifically Darcy's law describing the velocity-pressure gradient relationship and the mass conservation law described by the mass balance equation. This module uses these physical equations, with the reservoir state predicted by forward propagation as new input, to recalculate the physical relationships that should be satisfied on each grid cell.

[0024] In one embodiment of the present invention, for a given grid cell, according to the mass balance equation, the difference between the inflow and outflow of fluid should be equal to the change in fluid reserves within that cell. The physical residual calculation module calculates the flow rate based on the predicted pressure field and compares it with the change in reserves calculated based on the predicted change in saturation. If the model's prediction perfectly conforms to physical laws, then this difference, i.e., the physical residual, should be zero. However, in the early stages of training, the prediction is inevitably inaccurate, and this residual is a non-zero value. This non-zero physical residual is precisely calculated and used as a strong physical constraint term. Together with the data fitting error term formed by the difference between the backbone network's prediction results and the actual historical data, it is combined into the hybrid training objective signal defined by this method. This hybrid training objective signal, thus, acts as a "compass" for model training. It not only guides the model to imitate historical data but also mandates that its learning results must comply with basic physical conservation laws, thereby fundamentally improving the physical rationality and extrapolation prediction ability of the model. To address the strong heterogeneity commonly found in actual oil and gas reservoirs, the functionality of the aforementioned physical residual calculation module has been further enhanced and refined. An ideal, homogeneous oil and gas reservoir model is virtually nonexistent in reality; real reservoirs exhibit dramatic and complex variations in permeability and porosity in three-dimensional space. This heterogeneity directly controls the dominant channels and seepage resistance for fluid flow. If the physical residual calculation module uses a single, average physical equation coefficient, it cannot accurately describe these spatially varying seepage characteristics, leading to deviations between the embedded physical constraints and actual conditions. Therefore, the physical residual calculation module in this invention is designed to dynamically incorporate and respond to this spatial variation when performing its core calculations. It dynamically and individually adjusts the coefficients used in the physical equations of each grid cell by directly reading key field data—primarily permeability and porosity field signals—extracted from the fused data signal that characterize reservoir heterogeneity. For example, when calculating the fluid flow rate between adjacent grid cells according to Darcy's law, the conductivity coefficient used will be determined by the permeability, porosity, and geometric relationship between the two grid cells themselves. This means that in high-permeability strips, the physical equations will allow for faster fluid flow, while in low-permeability regions, greater flow resistance will be imposed. This highly customized processing transforms the physical equation constraints from a globally uniform, rigid mathematical formula into a spatially adaptive intelligent physical constraint capable of accurately characterizing the non-uniform seepage characteristics within actual oil and gas reservoirs. The resulting hybrid training target signal thus becomes a more refined and accurate training guide that closely matches real underground physical processes. In this way, the model is not only required to obey macroscopic physical laws during training, but also to learn the specific manifestations of these laws in heterogeneous reservoirs.This greatly improves the simulation accuracy and prediction reliability of the intelligent agent model when facing complex geological conditions, such as the presence of faults, interlayers, and high-permeability channels, thus evolving it from an ideal "theoretical model" into an "industrial-grade model" capable of coping with the complexities of the real world.

[0025] Furthermore, the training process of the method of this invention employs a carefully designed phased optimization strategy. The core purpose of this strategy is to address the training instability and convergence problems that may arise when simultaneously optimizing data fitting and physical constraints. In the initial stage, the training process focuses on enabling the model to grasp the basic mapping relationship from the input signal to the output signal. At this time, the optimization algorithm uses only the difference between historical data and the model's predicted values ​​as the objective function; this stage is called the data fitting pre-training period. Through this stage of training, the parameters of the model's backbone network are initially adjusted, enabling it to capture the main dynamic trends and statistical regularities contained in the production history, forming an initial model with basic predictive capabilities. However, at this time, the model may not have fully internalized the physical laws, and its predictions may have flaws in terms of physical conservation. Subsequently, the training enters the crucial joint fine-tuning stage. In this stage, the optimization objective switches from a single data fitting term to a complete mixed training objective signal, that is, simultaneously including data fitting errors and physical equation constraint terms. At this time, most of the parameters of the model's backbone network are frozen or subjected to a small learning rate, while the parameters of the parts of the model dedicated to physical constraint calculation or the final output layer are finely adjusted using a relatively large learning rate. This phased, differentiated optimization strategy offers the advantage of providing a smooth transition path for model training. It first ensures the model can stably find a solution space region that conforms to historical data, and then guides this initial solution to evolve and refine in a direction that simultaneously satisfies physical laws. This method effectively avoids gradient conflicts or optimization oscillations that may arise from facing complex mixed target signals at the beginning of training, ultimately resulting in a highly convergent and reliable intelligent surrogate model that not only performs well in matching historical data but also whose internal logic and deduction process strictly adhere to the physical conservation laws of oil and gas seepage. To further improve the intelligence and efficiency of the training process, this invention also introduces an adaptive weighting mechanism based on uncertainty. The core idea of ​​this mechanism is to recognize that the reliability and importance of the learning signals provided by data-driven and physics-driven terms change dynamically at different stages of model training. In the early stages of training, due to the random initialization of model parameters, its physical predictions may be extremely inaccurate, resulting in huge physical residuals and unstable gradient directions; at this time, direct data fitting targets can provide relatively clear and stable optimization directions. As training progresses, the model's fit to the data gradually improves, the descent space of the data-driven term decreases, and its learning of physical laws may enter a deeper stage. This weighting mechanism dynamically calculates and assigns the weight coefficients of the data-driven and physics-driven components in the overall loss function by monitoring their numerical magnitude, change amplitude, or gradient characteristics in recent training periods within the mixed training target signal in real time.For example, it can be designed to assign high weights to data fitting terms at the start of training, allowing the model to quickly approximate historical observations. Subsequently, as training progresses, the system automatically and smoothly reduces the weights of data terms while correspondingly increasing the weights of physical constraint terms, forcing the model to adjust its predictions to a trajectory more consistent with physical laws. This dynamic balancing strategy ensures that the training process can adaptively adjust its learning focus, avoiding the problems that might arise under a single fixed weight, such as one factor dominating and prematurely falling into a local optimum, or another factor being completely ignored. It makes the entire training process smoother, more efficient, and more robust, guiding the intelligent agent model to ultimately reach an optimal state with both excellent data fit and physical consistency.

[0026] like Figure 1As shown, the precise definition of the inputs and outputs is crucial for the specific application of the intelligent agent model. The current reservoir state parameters and production control parameters mentioned in step four, as input signals, constitute a highly condensed and information-rich set. The current reservoir state parameters are essentially a snapshot of the distribution of the three-dimensional pressure field and fluid saturation field of the entire reservoir at the latest known time point, extracted from the fused data signals. This represents the initial conditions under which the model begins prediction. The production control parameters refer to the operating conditions set by engineers or given by optimization algorithms during the future prediction period. These mainly include the bottomhole flowing pressure control targets or single-well fluid / gas production indicators for each production well, as well as the water injection or gas injection volume indicators for each injection well. These input signals are collectively fed into the fully trained intelligent agent model. This model internally encodes complex spatiotemporal evolution mapping relationships learned from historical data and physical laws. Unlike traditional numerical simulations that require iterative solutions to partial differential equations, this method directly simulates the continuous dynamic changes of input conditions over a predetermined future timeframe through a single, efficient forward computation—a neural network propagation from the input layer to the output layer. Its output is comprehensive and detailed, including high-resolution, time-series-varying three-dimensional pressure and saturation field evolution signals for the entire oil and gas reservoir, visually demonstrating the migration of underground fluids and the transmission of pressure waves; it also includes dynamic prediction signals for key production indicators derived from this, directly observable and measurable at the wellhead—namely, the oil, gas, and water production of each well. This achieves a complete prediction closed loop from macroscopic field distribution to microscopic single-well indicators. To ensure the intelligent agent model maintains its predictive accuracy throughout the long development lifecycle of an oil and gas field, this invention integrates a crucial online self-updating step after the core prediction process. An oil and gas reservoir is a dynamically changing system whose characteristics slowly evolve with development, and may encounter new production conditions not previously learned by the model. This self-updating step constructs a continuous "learning-prediction-correction" cycle. The specific process is as follows: The system continuously collects newly generated actual production monitoring data signals from the oil and gas field, such as wellhead production, water cut, and downhole pressure data for the new week or month. Then, the system automatically compares this fresh, real data with the model's predictions for the corresponding time period, calculating the difference between the two, thus generating a model prediction bias signal. This bias signal quantifies the gap between the model's current performance and the actual situation. Next, the system uses this bias signal and the corresponding newly collected data signals to perform incremental training or parameter fine-tuning on the deployed intelligent agent model. This incremental training is not a complete retraining from scratch, but rather a small-scale, iterative optimization using new, small-batch data based on the existing model parameters, resulting in minor, adaptive adjustments to the model's parameters.Through this continuous and automated process, the intelligent agent model can learn from new operational data and correct any drift in its predictions in a timely manner. This allows it to dynamically adapt to the slow changes in reservoir characteristics and new development strategies, ensuring that it can provide high-precision and reliable dynamic prediction signals throughout the entire oilfield lifecycle. It truly becomes a living digital twin that evolves in sync with the physical entity.

[0027] Specifically, this involves employing a meta-learning framework to endow intelligent agent models with powerful transfer and rapid adaptation capabilities. Traditional machine learning models are typically trained for a single, specific oil and gas field. When applied to a new work area, even with similar geological and development conditions, it requires collecting a large amount of data from scratch and executing a complete training process, which consumes significant computational resources and time. The introduction of the meta-learning framework aims to fundamentally change this paradigm, with its core idea being "learning how to learn." Specifically, in the preparation phase, the framework first utilizes numerical simulation technology or existing massive historical data to construct a vast virtual training library. This library contains tens of thousands of virtual oil and gas reservoir scenarios with different geological structures (such as anticlines, fault blocks, and lithological traps), different reservoir properties (such as high, medium, and low permeability distributions), different fluid properties (such as light oil, heavy oil, and condensate gas), and different development methods (such as natural energy extraction, water drive, and gas drive). The model is then pre-trained on this vast and diverse set of scenarios. The goal of this pre-training process is not to make the model memorize the details of every specific scenario, but to force it to extract and summarize the universal physical laws and statistical characteristics that control the dynamic evolution of oil and gas reservoirs, such as the basic patterns of pressure propagation, the commonalities of water breakthrough patterns, and the general responses of different injection and production strategies. The advantages of the pre-trained model become apparent when applied to a new and specific target oil and gas field. At this point, we no longer need the massive data accumulation of the target oil and gas field over several years, but only its relatively small amount of initial fused data signals. The model based on the meta-learning framework, with its powerful prior knowledge and generalization ability acquired during pre-training, can quickly integrate its general dynamic understanding with the personalized data of the specific oil and gas field through a process called "rapid fine-tuning." This fine-tuning process is extremely efficient; only a few iterations of optimization are needed for the model parameters to be subtly adjusted, allowing its predictive behavior to quickly match the actual dynamic response of the target work area. Ultimately, a high-precision intelligent proxy model suitable for the specific oil and gas field can be formed in a very short time with limited data support. This significantly lowers the threshold and timeline for building new oilfield models, providing unprecedented technical support for rapid evaluation and early development decisions.

[0028] like Figure 2As shown, this invention also includes an integrated dynamic simulation construction system for oil and gas fields based on artificial intelligence, comprising: an acquisition module, which acquires geological static data signals, historical and real-time production dynamic data signals, and downhole monitoring data signals of the target oil and gas field, performs spatiotemporal alignment and standardization cleaning on the multi-source heterogeneous data signals, and generates a spatiotemporally unified fused data signal; a coordination module, which constructs a deep learning model architecture with embedded physical information, which receives the fused data signal and generates a hybrid training target signal driven by fused data and mechanism by introducing physical equation constraint terms representing the oil and gas seepage law into the model loss function; a training module, which trains the deep learning model using the fused data signal of historical time series, generates a prediction signal through forward propagation, and optimizes the model parameters through backpropagation with minimizing the hybrid training target signal as the criterion, and finally obtains a converged intelligent proxy model; and an analysis module, which feeds the current oil and gas reservoir state parameters and production control parameters as input signals into the intelligent proxy model, and directly outputs the dynamic prediction signals of the pressure field, saturation field distribution, and single-well production capacity of the entire oil and gas reservoir in the future prediction period through one forward calculation of the model.

[0029] Furthermore, the value chain of this invention extends further downstream, entering the core decision-making stage of production optimization. This method seamlessly integrates the dynamic prediction signals into a dedicated production optimization control module, thus achieving a leap from "predicting the future" to "optimizing the future." The core task of this optimization module is to automatically find the optimal production strategy that maximizes economic benefits (usually represented by net present value) or other management objectives (such as improving oil recovery) based on future predictions. It typically employs two types of advanced algorithms to achieve this goal: one is reinforcement learning algorithms, which model the production optimization problem as a sequential decision-making process of continuous interaction between an agent and its environment (i.e., the oil and gas reservoir). The intelligent agent model acts as a virtual environment, capable of quickly evaluating the long-term consequences of any given strategy; the other is gradient optimization algorithms, which calculate the gradient of objective functions such as net present value relative to production control parameters (such as bottomhole flowing pressure and injection-production rates), and gradually adjusts these parameters along the direction of gradient ascent. Its workflow is an intensive, automated "hypothesis-simulation-evaluation" loop. The optimization module first generates a set of initial or random future production control parameter signals, and then feeds these signals to the intelligent agent model. The model performs rapid forward calculations based on this signal, outputting dynamic prediction signals for the entire prediction period, including production, water cut, and the final cumulative oil production and water injection costs. The optimization module calculates the objective function value (e.g., net present value) based on these outputs. Subsequently, the optimization algorithm generates a new set of improved production control parameter signals based on its inherent logic (e.g., strategy gradient, genetic algorithm, or quasi-Newton method), and calls the intelligent agent model for simulation again. This process is repeated hundreds or thousands of times. Because the intelligent agent model has a simulation speed of milliseconds or seconds, this intensive computation, unimaginable in traditional numerical simulation, is achieved. After extensive exploration and evaluation, the optimization module is finally able to select the optimal bottomhole flowing pressure and injection-production control parameter signals that bring the greatest economic benefits throughout the entire prediction period. This directly transforms rapid simulation capabilities into concrete, quantitative, and optimal production decision instructions. To address the complex phase change issues encountered in actual oil and gas reservoir development, particularly in enhanced oil recovery through gas injection and the development of volatile reservoirs, this invention significantly expands the depth and breadth of physical information embedding. Building upon fundamental physical constraints, the physical equation constraint terms in step two are designed to not only encompass Darcy's law and the mass conservation equation describing macroscopic fluid flow, but also to further and more profoundly introduce physical laws that precisely describe the microscopic phase behavior of fluids.This mainly includes the fluid volume coefficient equation, which describes the conversion relationship between the underground oil and gas volume and the volume under standard surface conditions. This relationship is a function of pressure and directly affects the expansion and contraction of the fluid; and the relative permeability curve relationship, which describes the nonlinear characteristics of the effective permeability of each phase fluid changing with saturation when underground multiphase fluids coexist, directly controlling the water and gas breakthrough patterns of the well.

[0030] This invention presents an AI-based integrated dynamic simulation method and system for oil and gas fields. It integrates static geological data, dynamic production data, and real-time monitoring data to form a standardized signal, which drives a deep learning model embedded with physical information. The core of this model lies in its loss function, which not only includes data fitting terms but also innovatively introduces physical equations characterizing oil and gas seepage as constraint terms, thus deeply embedding physical conservation laws into the AI ​​learning process. The intelligent agent model trained in this way possesses both the rapid computational capabilities of AI and the physical logic and reliability of extrapolation prediction inherent in traditional numerical simulation, ultimately achieving rapid and high-precision simulation and prediction of oil and gas field production dynamics.

[0031] Therefore, the method and system for constructing integrated dynamic simulation of oil and gas fields based on artificial intelligence of the present invention can solve the problems of low computational efficiency of traditional numerical simulation of oil and gas fields, as well as the lack of physical logic and unreliable prediction results of pure artificial intelligence models.

[0032] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the invention. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in the present invention should still be covered by the claims of the present invention.

Claims

1. A method for constructing an integrated dynamic simulation of oil and gas fields based on artificial intelligence, characterized in that, include: S1: Collect geological static data signals, historical and real-time production dynamic data signals, and downhole monitoring data signals of the target oil and gas field. Perform spatiotemporal alignment and standardization cleaning on the multi-source heterogeneous data signals to generate a spatiotemporally unified fused data signal. S2: Construct a deep learning model architecture with embedded physical information. This architecture receives the fused data signal and generates a hybrid training target signal that combines data-driven and mechanism-driven approaches by introducing physical equation constraint terms that characterize the seepage law of oil and gas into the model loss function. S3: The deep learning model is trained using the fused data signal of the historical time series. A prediction signal is generated by forward propagation, and the model parameters are optimized by backpropagation with the criterion of minimizing the mixed training target signal, and finally a converged intelligent agent model is obtained. S4: The current oil and gas reservoir state parameters and production control parameters are used as input signals and fed into the intelligent agent model. Through one forward calculation of the model, the dynamic prediction signals of the pressure field, saturation field distribution of the entire oil and gas reservoir and the production capacity of a single well in the future prediction period are directly output.

2. The method for constructing an integrated dynamic simulation of oil and gas fields based on artificial intelligence according to claim 1, characterized in that, Step S1, which involves spatiotemporal alignment and standardization cleaning of the multi-source heterogeneous data signals, specifically includes: performing spatial discretization interpolation processing on the geological static data signals based on the oil and gas field geological grid model to generate static parameter field signals that match the grid model; performing time series resampling and alignment on the production dynamic data signals and downhole monitoring data signals to unify them to a common timestamp sequence, and using a standardization algorithm to eliminate the influence of different physical dimensions to generate the spatiotemporally unified fused data signal, which is directly used as the standardized input of the deep learning model architecture for embedding physical information in step S2.

3. The method for constructing an integrated dynamic simulation of oil and gas fields based on artificial intelligence according to claim 1, characterized in that, The deep learning model architecture for embedding physical information in step S2 is as follows: A composite network structure includes a multilayer perceptron or recurrent neural network for encoding the spatiotemporal features in the fused data signal, and a parallel physical residual calculation module. The physical residual calculation module calculates the physical conservation residual on the grid cell based on Darcy's law and the mass balance equation in oil and gas seepage mechanics, and uses this residual as a constraint term of the physical equation and a data fitting error term to constitute the hybrid training target signal.

4. The method for constructing an integrated dynamic simulation of oil and gas fields based on artificial intelligence according to claim 3, characterized in that, The physical residual calculation module further introduces specific processing for heterogeneous oil and gas reservoirs when calculating the physical conservation residuals. It dynamically adjusts the coefficients of the physical equations at different spatial locations by using the permeability field and porosity field signals in the fused data signal. This allows the constraint terms of the physical equations to accurately characterize the non-uniform seepage characteristics of the actual oil and gas reservoir, thereby generating a spatially adaptive hybrid training target signal that is highly consistent with the real physical process, so as to improve the simulation accuracy of the model under complex geological conditions.

5. The method for constructing an integrated dynamic simulation of oil and gas fields based on artificial intelligence according to claim 1, characterized in that, The training process described in step S3 employs a phased optimization strategy. First, the data fitting ability of the deep learning model is optimized separately using the fused data signal of the historical time series. Then, while keeping the core parameters of the model unchanged, the mixed training target signal is activated for joint fine-tuning. This strategy aims to balance the convergence stability of the model in the early stage with the consistency of physical laws in the later stage, and finally obtain the converged intelligent surrogate model that conforms to the trend of historical data and strictly abides by the physical conservation law.

6. The method for constructing an integrated dynamic simulation of oil and gas fields based on artificial intelligence according to claim 1, characterized in that, In step S3, an uncertainty-based weighting mechanism is introduced during the training process. This mechanism monitors the relative changes of the data-driven and physical-driven terms in the mixed training target signal during the training process and dynamically adjusts their weight coefficients in the total loss function. This results in an emphasis on quickly fitting the data in the early stage of training and on satisfying physical constraints in the later stage of training. This adaptive strategy ensures that the training process of the intelligent agent model is smoother and more efficient.

7. The method for constructing an integrated dynamic simulation of oil and gas fields based on artificial intelligence according to claim 1, characterized in that, The current reservoir state parameters and production control parameters mentioned in step S4 are used as input signals, specifically including bottom hole flowing pressure, single well injection and production volume, and the latest known pressure field and saturation field distribution extracted from the fused data signals. The intelligent agent model simulates the continuous changes of the input signals in a preset future time period in a forward calculation through its internally encoded spatiotemporal evolution law, and directly outputs high-resolution full reservoir pressure field and saturation field evolution signals, as well as dynamic prediction signals of oil production, gas production and water production of each wellhead derived therefrom.

8. The method for constructing an integrated dynamic simulation of oil and gas fields based on artificial intelligence according to claim 1, characterized in that, After outputting the dynamic prediction signal in step S4, the method further includes an online model self-update step. This step compares the newly acquired actual production monitoring data signal from the oil and gas field with the previously output dynamic prediction signal from the model to generate a model prediction deviation signal. This deviation signal and the corresponding new data signal are then used to incrementally train or fine-tune the parameters of the intelligent agent model, so that the model can continuously adapt to the dynamic changes of the oil and gas reservoir and maintain its long-term prediction accuracy.

9. The method for constructing an integrated dynamic simulation of oil and gas fields based on artificial intelligence according to claim 1, characterized in that, The deep learning model architecture for embedding physical information constructed in step S2 adopts a meta-learning framework. This framework enables the model to learn the general dynamic laws of oil and gas reservoirs by pre-training on a large number of virtual oil and gas reservoir scenarios with different geological and development conditions. When applied to the target oil and gas field, it only needs to be quickly fine-tuned using a small amount of fused data signals from the oil and gas field to adapt the model and form a high-precision intelligent proxy model suitable for the oil and gas field.

10. A system using the AI-based integrated dynamic simulation construction method for oil and gas fields according to any one of claims 1-9, characterized in that, include: The acquisition module acquires geological static data signals, historical and real-time production dynamic data signals, and downhole monitoring data signals of the target oil and gas field. It performs spatiotemporal alignment and standardization cleaning on the multi-source heterogeneous data signals to generate a spatiotemporally unified fused data signal. The coordination module constructs a deep learning model architecture with embedded physical information. This architecture receives the fused data signal and generates a hybrid training target signal that combines data-driven and mechanism-driven approaches by introducing physical equation constraint terms that characterize the seepage law of oil and gas into the model loss function. The training module uses the fused data signal of historical time series to train the deep learning model, generates a prediction signal through forward propagation, and optimizes the model parameters through backpropagation with the criterion of minimizing the mixed training target signal, and finally obtains a converged intelligent agent model. The analysis module takes the current oil and gas reservoir state parameters and production control parameters as input signals and feeds them into the intelligent agent model. Through one forward calculation of the model, it directly outputs the dynamic prediction signals of the pressure field, saturation field distribution of the entire oil and gas reservoir and the production capacity of a single well in the future prediction period.

Images