A distributed medicine injection control system and method based on LSTM dynamic compensation

Abstract

The application relates to the technical field of medicine injection control systems, and particularly discloses a distributed medicine injection control system and method based on LSTM dynamic compensation, which adopts a multi-time scale nested full-closed loop control architecture and comprises a multi-modal data sensing unit, wherein the multi-modal data sensing unit comprises a sparse in-situ online repair effect monitoring network, is used for collecting physical parameters, chemical parameters and acoustic parameters of an injection well site, and collects in-situ pollutant concentration data through the sparse in-situ online repair effect monitoring network. According to the technical scheme, the multi-time scale nested full-closed loop architecture combining edge control and central remote cooperation is constructed, and in the actual operation process of underground medicine injection control, continuous self-adaptive regulation and control of multi-source formation state data can be realized.

Description

Technical Field

[0001] This invention relates to the field of drug injection control system technology, and in particular to a distributed drug injection control system and method based on LSTM dynamic compensation. Background Technology

[0002] Traditional methods for controlling the injection of remediation agents for underground contamination mainly rely on periodic offline manual sampling and analysis. In actual operations, the control mechanisms of multiple injection devices on site are relatively simple, typically using a uniformly set fixed flow rate for continuous or open-circuit injection, lacking technical means to make local dynamic adjustments based on real-time changes in the formation and fluids.

[0003] Due to the heterogeneity of underground geological structures and the dynamic evolution of contamination plume distribution, existing fixed injection control mechanisms are ill-suited to the complex near-wellbore fluid migration. In multi-well collaborative injection scenarios, traditional systems cannot synchronously correlate the physical characteristics of the fluids in the field, such as pressure, temperature, and flow rate, as well as chemical characteristics such as pH and acoustic characteristics. This results in a lack of overall coordination among the injection wells, which can easily lead to ineffective accumulation of reagents in non-target areas, abnormal disturbances in local formations, or reagent injection volumes deviating from the actual degradation density of contaminants, thus reducing the overall accuracy and operational stability of in-situ remediation. Summary of the Invention

[0004] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, the objective of this invention is to propose a distributed drug injection control system and method based on LSTM dynamic compensation to achieve precise distributed drug injection.

[0005] To achieve the above objectives, a first aspect of the present invention proposes a distributed drug injection control system based on LSTM dynamic compensation. The system employs a multi-timescale nested fully closed-loop control architecture, comprising:

[0006] The multimodal data sensing unit includes a sparse in-situ online remediation effect monitoring network, which is used to collect physical, chemical and acoustic parameters at the injection well site, and to collect in-situ pollutant concentration data through the sparse in-situ online remediation effect monitoring network.

[0007] Multiple edge single-well controllers are deployed at the corresponding injection well sites to receive physical, chemical, and acoustic parameters. They have built-in structured long short-term memory network state identification modules and hierarchical risk response modules to output single-well flow control commands according to the first control cycle and extract formation disturbance index.

[0008] The central control platform, deployed in the remote monitoring center, receives the physical parameters and formation disturbance index uploaded by the edge single-well controllers, as well as the in-situ pollutant concentration data. It has a built-in generative geological model dynamic update module and a remediation effect-oriented multi-well collaborative optimization module. It is used to update the geological model based on the physical parameters and formation disturbance index according to the second control cycle, and to optimize the global parameters based on the in-situ pollutant concentration data according to the third control cycle, and output the optimal target flow rate and optimal target concentration to the edge single-well controllers.

[0009] The distributed execution unit is used to receive the single-well flow control command issued by the edge single-well controller based on the optimal target flow rate and the optimal target concentration, and execute the reagent injection action of the corresponding injection well.

[0010] To achieve the above objectives, a second aspect of this invention proposes a distributed drug injection control method based on LSTM dynamic compensation, applied to the aforementioned system. The method is executed based on a multi-timescale nested fully closed-loop control architecture and includes the following steps:

[0011] The physical, chemical, and acoustic parameters at the injection well site are collected by the multimodal data sensing unit, and the in-situ pollutant concentration data are collected by the sparse in-situ online remediation effect monitoring network in the multimodal data sensing unit.

[0012] Multiple edge single-well controllers receive the corresponding physical parameters, chemical parameters and acoustic parameters respectively, and use the built-in structured long short-term memory network state identification module and hierarchical risk response module to extract the formation disturbance index according to the first control cycle and output the single-well flow control command.

[0013] The central control platform receives the physical parameters and formation disturbance index uploaded by the edge single-well controller, as well as the in-situ pollutant concentration data. It uses the built-in generative geological model dynamic update module to update the geological model based on the physical parameters and formation disturbance index according to the second control cycle. It also uses the built-in repair effect-oriented multi-well collaborative optimization module to optimize global parameters based on the in-situ pollutant concentration data according to the third control cycle, and outputs the optimal target flow rate and optimal target concentration to the corresponding edge single-well controller.

[0014] The distributed execution unit receives the single-well flow control command issued by the edge single-well controller based on the optimal target flow rate and the optimal target concentration, and executes the corresponding injection well's reagent injection action.

[0015] To achieve the above objectives, a third aspect of the present invention provides an electronic device including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described distributed drug injection control method based on LSTM dynamic compensation.

[0016] This invention's technical solution constructs a multi-timescale nested closed-loop architecture combining edge control and central remote collaboration. During actual underground chemical injection control operations, it enables continuous adaptive regulation of multi-source formation state data. Through distributed multimodal data sensing units, it synchronously acquires on-site physical, chemical, acoustic, and in-situ concentration parameters. This allows edge single-well controllers to promptly optimize single-well flow control commands based on local fluid conditions within a short timescale. Simultaneously, the central control platform can integrate global data over a longer timescale for dynamic geological model updates and multi-well collaborative optimization of global parameters. This structure changes the traditional single-fixed injection mode, enabling the chemical flow rate and concentration of injection wells in different areas to match the dynamic degradation state of the underground contaminant plume. While effectively coordinating the collaborative operation of each injection well, it reduces the risk of chemical flow deviation or formation disturbance caused by formation heterogeneity, achieving a dynamic balance between precise distributed chemical injection and formation safety protection. Attached Figure Description

[0017] Figure 1 This is a schematic diagram illustrating the implementation of the distributed drug injection control system based on LSTM dynamic compensation provided by the present invention.

[0018] Figure 2 This is a convergence trend data graph of multimodal feature fusion and LSTM state identification in a distributed drug injection control system based on LSTM dynamic compensation provided by the present invention.

[0019] Figure 3 This invention provides a three-dimensional response surface plot of the system comprehensive state exponential nonlinear amplification effect in a distributed drug injection control system based on LSTM dynamic compensation.

[0020] Figure 4 This is a slice bitmap of the anisotropic three-dimensional concentration distribution field based on Kriging interpolation in the distributed drug injection control system based on LSTM dynamic compensation provided by the present invention.

[0021] Figure 5 This is a scatter plot of Pareto front data for adaptive particle swarm multi-objective optimization in a distributed drug injection control system based on LSTM dynamic compensation provided by the present invention.

[0022] Figure 6This is a coupling characteristic diagram of high-frequency acoustic spectrum distortion and pressure drop funnel in the crossflow state of the distributed drug injection control system based on LSTM dynamic compensation provided by the present invention.

[0023] Figure 7 This is a timing waveform diagram of the dynamic adjustment of duty cycle and compensation frequency in the frequency conversion pulse injection mode of the distributed drug injection control system based on LSTM dynamic compensation provided by the present invention.

[0024] Figure 8 This is a multi-parameter dynamic response curve of interface fouling evolution and chemical-mechanical joint defogging in the distributed agent injection control system based on LSTM dynamic compensation provided by the present invention.

[0025] Figure 9 This is a flowchart illustrating the distributed drug injection control method based on LSTM dynamic compensation provided by the present invention.

[0026] Figure 10 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0027] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0028] The following describes, with reference to the accompanying drawings, an embodiment of the distributed drug injection control system, method, and electronic equipment based on LSTM dynamic compensation of the present invention.

[0029] Example 1:

[0030] like Figure 1 As shown, this embodiment discloses a distributed drug injection control system based on LSTM dynamic compensation. The system adopts a multi-timescale nested fully closed-loop control architecture, including the following:

[0031] The multimodal data sensing unit includes a sparse in-situ online remediation effect monitoring network, which is used to collect physical, chemical and acoustic parameters at the injection well site, and to collect in-situ pollutant concentration data through the sparse in-situ online remediation effect monitoring network.

[0032] Multiple edge single-well controllers are deployed at the corresponding injection well sites to receive physical, chemical, and acoustic parameters. They have built-in structured long short-term memory network state identification modules and hierarchical risk response modules to output single-well flow control commands according to the first control cycle and extract formation disturbance index.

[0033] The central control platform, deployed in the remote monitoring center, receives the physical parameters and formation disturbance index uploaded by the edge single-well controllers, as well as the in-situ pollutant concentration data. It has a built-in generative geological model dynamic update module and a remediation effect-oriented multi-well collaborative optimization module. It is used to update the geological model based on the physical parameters and formation disturbance index according to the second control cycle, and to optimize the global parameters based on the in-situ pollutant concentration data according to the third control cycle, and output the optimal target flow rate and optimal target concentration to the edge single-well controllers.

[0034] The distributed execution unit is used to receive the single-well flow control command issued by the edge single-well controller based on the optimal target flow rate and the optimal target concentration, and execute the reagent injection action of the corresponding injection well.

[0035] Specifically, to achieve comprehensive and high-precision perception of complex underground environments, the multimodal data sensing unit further includes a distributed physical parameter monitoring subunit, comprising an array of fiber Bragg grating distributed pressure sensors deployed along the injection well casing, and a high-precision electromagnetic flowmeter, temperature sensor, and density sensor integrated at the wellhead. In actual in-situ remediation engineering applications, underground aquifers often exhibit severe heterogeneity, and traditional single-point pressure monitoring cannot effectively reflect the permeation resistance of the agent at different depths. The fiber Bragg grating distributed pressure sensor used in this system, based on the wavelength drift principle of fiber Bragg gratings, can acquire the stress changes exerted by the formation fluid on the inner wall of the well casing at specific intervals along the longitudinal direction of the well casing. This arrayed method of acquiring physical parameters allows the system to clearly characterize the permeability profile along the longitudinal direction of the injection well; while the high-precision electromagnetic flowmeter at the wellhead, working in conjunction with the Coriolis density sensor, can monitor the mass flow rate and solution concentration fluctuations of the injected agent in real time, providing the basic physical boundary conditions for subsequent mass conservation calculations.

[0036] For example, the chemical parameter monitoring subunit of the multimodal data sensing unit includes an online pH sensor, a redox potential sensor, and an ion-selective electrode integrated into the wellhead reagent mixing pipeline. During in-situ chemical oxidation or reduction remediation, the activity of the reagents is largely limited by the pH and redox potential of the groundwater.

[0037] For example, in the persulfate injection process, if the online pH sensor detects that the system pH value has dropped to the non-reactive range, it means that the currently injected activator has failed to effectively maintain the alkaline environment. At this time, the data of the redox potential sensor will usually also show a synchronous abnormal decay. Ion selective electrodes are specifically designed to continuously track specific target pollutant degradation byproducts (such as chloride ion release) or heavy metal ions (such as the concentration of trivalent chromium ions after the reduction of hexavalent chromium), thereby providing a basic chemical state benchmark before the agent enters the formation.

[0038] Optionally, the acoustic parameter monitoring subunit of the multimodal data sensing unit includes high-precision hydrophones deployed at the bottom and middle of the well casing. In groundwater flow dynamics, when reagents are forced into porous media under high pressure or undergo microscopic flow in rock and soil fissures, weak acoustic emission phenomena in specific frequency bands are generated. High-precision hydrophones can pick up these acoustic parameters caused by fluid shearing, solid particle stripping, or bubble collapse, which provides key implicit characteristic support for determining whether physical channel changes have occurred within the formation. Meanwhile, the sparse in-situ online remediation effect monitoring network includes in-situ online monitoring wells deployed in the core, boundary, and background zones of the pollution plume, and is equipped with online gas chromatographs and fiber optic Raman spectroscopy sensors.

[0039] Due to the high cost of constructing underground monitoring wells, dense well deployment across the entire site is not feasible in actual operations; therefore, a sparse deployment strategy is adopted. An online gas chromatograph can periodically separate and quantify the concentration of volatile organic pollutants in groundwater samples, while a fiber optic Raman spectroscopy sensor utilizes the Raman scattering effect to identify characteristic molecular bond vibrations of pollutants in situ online. This high-dimensional concentration data forms the cornerstone for the system's global remediation effectiveness assessment.

[0040] It is important to note that the distributed execution unit includes a variable frequency speed-regulating injection pump, a proportional regulating solenoid valve, an automatic dosing device, and a high-pressure clean water flushing pump. The edge single-well controller physically executes the single-well flow control command by issuing frequency conversion commands to the variable frequency speed-regulating injection pump and opening adjustment commands to the proportional regulating solenoid valve. At the execution level, to prevent water hammer effects from damaging the well casing and formation structure, the variable frequency speed-regulating injection pump integrates a torque control algorithm, which can smoothly increase and decrease the motor speed according to the received frequency conversion commands. The proportional regulating solenoid valve is used to eliminate static pressure hysteresis in the fluid system. Its internal valve core displacement has a strict linear mapping relationship with the input current signal, ensuring that the actual injection flow rate can accurately track the target value issued by the edge single-well controller. The automatic dosing device is responsible for automatically adjusting the mixing ratio of the concentrate and dilution water according to the optimal target concentration, while the high-pressure clean water flushing pump serves as a backup unit, providing high-pressure kinetic energy when the system needs pipeline maintenance or to clear local blockages near the well.

[0041] After acquiring the aforementioned multi-source heterogeneous data, real-time processing is required at the edge. Due to the significant differences in the dimensions and numerical ranges of various sensors, the system first performs standardization processing on the multimodal sensing data. The standardization conversion algorithm formula is as follows:

[0042] ;

[0043] in, for Standardized eigenvalues ​​at time t. for The original multimodal sensing sample values ​​at time 10:00. The mean of features within a set historical time window, The standard deviation of features within the set historical time window.

[0044] Through this standardized calculation, various parameters (such as pH data ranging from 0 to 14 and acoustic signals ranging from microvolts) are uniformly mapped to a standard normal distribution space with zero mean and unit variance, reducing the problem of neural network gradient update offset caused by differences in data scale.

[0045] Specifically, the structured long short-term memory network state identification module adopts a three-layer stacked long short-term memory network structure. The input layer receives a standardized multimodal feature vector matrix composed of the physical, chemical, and acoustic parameters from multiple past time points. Temporal features are extracted through the long short-term memory units of the hidden layer. The output layer outputs the device degradation index and the formation disturbance index. Wherein:

[0046] The first layer of the Long Short-Term Memory network mainly serves as a low-level feature extractor. It uses its internal forget gate and input gate mechanisms to filter high-frequency noise in the time series and retain transient physical features such as pressure changes and flow oscillations.

[0047] The second layer of the long short-term memory network further performs nonlinear combination of these transient features to extract cross-modal correlation information, such as the synchronous correlation between the hydrophone acoustic band shift and the rise in pressure sensor readings.

[0048] The third layer, Long Short-Term Memory (LSTM), acts as an abstract state mapping layer, outputting two key dimensionless health assessment indicators: the equipment degradation index and the formation disturbance index. The equipment degradation index primarily reflects the decline in mechanical efficiency caused by factors such as impeller wear in variable frequency pumps and pipe scaling; while the formation disturbance index characterizes the degree of pathological evolution of underground microstructures caused by agent injection, such as formation pore blockage, dominant channel development, or formation fracturing.

[0049] like Figure 2This figure displays a convergence trend data graph of multimodal feature fusion and long short-term memory network state identification. The horizontal axis represents runtime in seconds, and the vertical axis represents the state identification index in dimensionless units. The legend indicates the blue equipment degradation index curve and the red formation disturbance index curve.

[0050] The blue equipment degradation index curve shows a slowly rising waveform from an initial value of around 0.2 over time, reaching approximately 0.45 at 500 seconds. This truly reflects the gradual performance decline of actuators such as variable frequency speed control injection pumps due to prolonged operation.

[0051] Meanwhile, the blue curve shows continuous small high-frequency oscillations during its ascent, indicating that background noise from some complex downhole environments still exists when processing physical, chemical, and acoustic parameters at the front end. This demonstrates that the wavelet multi-scale filtering algorithm of the data sensing unit still has room for further upgrades and optimizations in dealing with non-steady-state noise.

[0052] The red formation disturbance index curve remained stable at around 0.1 in the initial stage, but at the 200th second, due to a sudden flow or pore blockage of the simulated underground medium, the curve showed a step increase and eventually converged to the steady-state baseline of 0.6.

[0053] The dynamic details of the waveform transformation show that after the Long Short-Term Memory network captured the step change signal, it generated a tracking delay of about 15 seconds. During the convergence process, an overshoot peak of about 0.65 appeared before gradually falling back and stabilizing.

[0054] The dynamic response curve, which exhibits short-term delay and slight overshoot, objectively reflects the state mapping and risk identification capabilities of the three-layer stacked network structure in handling multimodal transient distortion characteristics. However, it also reveals the system's insufficient transient control smoothness when dealing with extreme hydrodynamic changes, suggesting that the transient robustness of the overall control system can be improved by optimizing the forget gate parameters of the neural network in the future.

[0055] For example, the hierarchical risk response module is used to classify three levels of operational risk based on the system comprehensive state index and execute corresponding control strategies; the formula for calculating the system comprehensive state index is:

[0056] ;

[0057] in, This is the system's overall state index. The degradation index of the equipment is [value missing]. The term refers to the formation disturbance index. It is an exponential function with the natural constant as its base.

[0058] In this formula, an exponential function is used instead of a simple linear weighting because in complex industrial injection systems, equipment degradation and formation disturbance often have a nonlinear mutual amplification effect.

[0059] For example, when the filter pipe becomes slightly clogged (increasing the formation disturbance index), it will lead to an increase in the pump's back pressure, which in turn accelerates the wear of the pump end bearings (and the equipment degradation index increases simultaneously). The product of these two factors, mapped by an exponential function, can more sensitively capture the deterioration trend of the system, allowing the comprehensive state index to produce significant numerical abrupt changes before a failure occurs.

[0060] like Figure 3 The figure presents a three-dimensional response surface plot of the nonlinear amplification effect of the system's comprehensive state index. The plot includes three coordinate axes: the two horizontal axes represent the equipment degradation index and the formation disturbance index, both in dimensionless units; the vertical axis represents the system's comprehensive state index, also in dimensionless units.

[0061] As can be seen from the data and spatial morphology of the three-dimensional surface in the figure, when the equipment degradation index and the formation disturbance index are in the low range below 0.2, the blue surface shows a relatively gentle trend of change. This corresponds to the stable operation stage of the injection system under normal conditions. At this time, the slight clogging of the filter pipe or the initial wear of the pump end bearing does not have a significant correlation with the whole.

[0062] When these two indicators rise simultaneously and exceed 0.5, the surface color gradually transitions from light blue to yellow and then to dark red. At the same time, the spatial climbing slope of the three-dimensional surface increases sharply, exhibiting a non-linear, steep upward shape.

[0063] When both the equipment degradation index and the formation disturbance index approach 1.0, the system comprehensive state index rapidly amplifies and exceeds 7.0, resulting in a significant numerical surge. This spatial surface change effect demonstrates that the logic mechanism can sensitively capture the nonlinear mutual amplification effect between variable frequency pump impeller wear and formation micropore blockage, thereby exceeding the preset judgment threshold and triggering emergency protection strategies in advance before a real danger occurs in the pipeline.

[0064] Meanwhile, observing the high-value areas in red in the figure reveals that the surface of the three-dimensional curved surface is accompanied by slight jagged undulations and data spikes. This reflects that under extremely high-risk and extreme fluid pressure conditions, the multi-source data fusion node is affected by the quantization error of the underlying sensor and the noise of the unsteady fluid, which will cause certain numerical calculation oscillations. This objectively indicates that the anti-interference smoothing ability of the index evaluation logic under extreme boundary conditions still has room for further upgrading and improvement.

[0065] Optionally, when the system's overall status index is less than a first preset threshold, it is determined to be in a normal state, and standard dynamic compensation control is executed; when the system's overall status index is greater than or equal to the first preset threshold and less than a second preset threshold, it is determined to be in a warning state, and a preventive control strategy of reducing the injection flow rate and increasing the sampling frequency is executed; when the system's overall status index is greater than or equal to the second preset threshold, it is determined to be in a dangerous state, and an emergency protection strategy of shutting down the injection pump is executed; wherein, the first preset threshold is less than the second preset threshold.

[0066] In actual logical execution, under normal conditions, the edge single-well controller executes according to the optimization target issued by the central platform. Once it enters the warning state, it means that there is a risk of deterioration in the current well condition. At this time, it is necessary to limit the output limit of the variable frequency speed control injection pump and activate the high-frequency sampling mode of the sensor to collect more data for disease diagnosis. This is a typical degraded operation approach, which aims to curb the risk from escalating without completely interrupting the repair process. If the system comprehensive status index continues to climb and exceeds the second preset threshold to enter a dangerous state, in order to avoid irreversible fracturing damage to the formation or pipeline rupture, the system will directly trigger a hardware protection interruption, directly cut off the power supply to the injection pump and output an alarm.

[0067] On a more macroscopic time scale, the generative geological model dynamic update module incorporates a conditional probability diffusion model, which receives pressure gradient data and formation disturbance index extracted and uploaded by the edge single-well controller based on the pressure data in the physical parameters. The conditional probability diffusion model adds Gaussian noise through a forward diffusion process and performs a reverse noise reduction process by inputting the real-time pressure gradient data to generate an updated three-dimensional formation permeability field. Based on the updated three-dimensional formation permeability field, the groundwater seepage equation and convection-diffusion response equation are solved to generate a predicted pollutant concentration distribution field.

[0068] Due to the high degree of invisibility of geological formations, traditional deterministic hydrogeological models often struggle to accurately reflect actual conditions. The conditional probability diffusion model initially adds Gaussian random noise in a Markov chain to the existing basic geological model, transforming it into a pure Gaussian noise field. Subsequently, in the reverse denoising generation stage, the model embeds real-time pressure gradient data uploaded from each injection well as a conditional guiding signal into the network. Since Darcy's law dictates a strict physical causal relationship between fluid pressure gradient and formation permeability, using the real pressure gradient as a constraint to guide the denoising process forces the model to generate a three-dimensional formation permeability field that is both hydrodynamically reasonable and consistent with actual field observations.

[0069] Next, based on this updated high-precision permeability coefficient field, the system uses the finite element method or finite volume method to solve partial differential equations that include terms of convection diffusion, hydrodynamic dispersion, and chemical reaction kinetics, and then extrapolates to future time points to output the predicted pollutant concentration distribution field in three-dimensional space.

[0070] Specifically, since the predicted values ​​of the generated model inevitably contain residuals, the central control platform also has a built-in module for calculating the remediation effect deviation. This module incorporates an anisotropic spherical semivariogram Kriging interpolation algorithm, which is used to input the measured in-situ pollutant concentration data, generate a global interpolated concentration field, and compare the global interpolated concentration field with the predicted pollutant concentration distribution field to calculate the global concentration deviation field and the global average concentration deviation. Because the sparse in-situ online remediation effect monitoring network can only provide concentration data at discrete points, the Kriging interpolation algorithm utilizes the spatial autocorrelation characteristics of formation attributes, particularly employing anisotropic spherical semivariograms, to address the spatial variability in concentration caused by groundwater flow direction (i.e., concentration correlation decays more slowly along the flow direction and more rapidly perpendicular to the flow direction), thereby generating a continuous global interpolated concentration field throughout the entire remediation grid space.

[0071] like Figure 4 This diagram presents a slice map of the anisotropic three-dimensional concentration distribution field based on Kriging interpolation. The map contains three spatial coordinate axes: the horizontal axis represents the distance in the direction of water flow and the distance perpendicular to the direction of water flow, both in meters, ranging from 0 to 100 meters; the vertical axis represents the formation depth, in meters, ranging from -50 to 0 meters; and the color bars in the legend represent the pollutant concentration in milligrams per liter.

[0072] The spatial color distribution of the bitmap shows that the high concentration of red and dark orange areas is mainly concentrated in the core area at a distance of 20 to 60 meters downstream and at a depth of approximately -20 meters, with the highest concentration reaching about 100 milligrams per liter.

[0073] As the spatial location expands outwards, the color gradually transitions to yellow, green, and blue, representing low concentrations, with the concentration value dropping below 20 mg / L. This slice-like image morphology, characterized by slower color decay in the downstream direction and a sharp contraction to blue in the vertical and depth directions, objectively reflects the physical interpolation characteristics of the anisotropic spherical semivariogram built into the remediation effect deviation calculation module when processing groundwater fluid dynamics. This demonstrates that the system can generate a global interpolated concentration field that conforms to the objective laws of fluid transport using limited monitoring data.

[0074] Meanwhile, observing the image edge regions, especially the low-concentration blue areas at depths of -40 to -50 meters and along the lateral edges, reveals obvious patchy color abrupt changes and irregular, jagged numerical fluctuations on the slice surface. This local color discontinuity indicates that the sparse in-situ online remediation monitoring network has a relatively limited number of sampling points in the background region far from the core of the pollution plume, leading to increased prediction variance of the Kriging interpolation algorithm in the distant boundary region, and consequently, some numerical interpolation distortion. This objectively demonstrates that the system's 3D spatial reconstruction accuracy under sparse sensing data conditions still has room for further optimization and improvement by adding in-situ online monitoring wells.

[0075] Subsequently, by calculating the difference between the predicted field and the interpolated field, the prediction error of the current geological model is quantitatively assessed. The system uses the following formula to calculate the global average concentration deviation:

[0076] ;

[0077] in, This represents the global average concentration deviation. The total calculated volume of the repair area. The total number of discretized grids. For the first Interpolation concentration of each grid cell, For the first The predicted concentration for each grid cell, For the first The volume of each grid.

[0078] The formula is essentially a volume-weighted error integral process that integrates the effects of local concentration overestimation or underestimation into a single scalar index, objectively reflecting the global deviation of the overall site remediation expectation.

[0079] For example, the central control platform also has a built-in generative geological model Bayesian self-calibration module; the generative geological model Bayesian self-calibration module has a built-in Markov chain Monte Carlo sampler, which is used to trigger parameter self-calibration when the global average concentration deviation exceeds the preset concentration deviation threshold or reaches the third control cycle; the Markov chain Monte Carlo sampler takes the global concentration deviation field and the predicted pollutant concentration distribution field as inputs, and outputs the optimal calibration parameter vector to update the generative geological model; the optimal calibration parameter vector includes a three-dimensional effective porosity field, a pollutant reagent second-order reaction rate constant field, and a pollutant solid phase adsorption and distribution coefficient field.

[0080] In in-situ injection engineering, porosity determines the storage space for the reagent, the second-order reaction rate constant determines the physicochemical rate of reagent consumption and pollutant degradation, and the solid-phase adsorption partition coefficient characterizes the slow-release process of pollutants desorbed from the soil framework into groundwater. These three core parameters often evolve non-stationarily over time. The Markov chain Monte Carlo sampler constructs a Markov chain with a stationary distribution in a multidimensional parameter space and continuously generates new parameter proposal distributions using the Metropolis-Hastings acceptance and rejection criterion.

[0081] The probability of accepting a parameter combination increases when the generated parameter combination makes the simulated concentration field more consistent with the measured interpolated concentration field (i.e., reduces the negative logarithm of the system likelihood function). After sufficient programming and iteration, the parameter combination that maximizes the posterior probability distribution is extracted as the optimal calibration parameter vector, thereby completing the high-precision correction of the underlying physicochemical boundary conditions.

[0082] It should also be noted that the repair effect-oriented multi-well collaborative optimization module has a built-in adaptive particle swarm optimization algorithm, which takes the minimum value of the multi-objective optimization function as the optimization objective and outputs the optimal target flow rate and the optimal target concentration of each injection well.

[0083] To drive the adaptive particle swarm optimization algorithm, the system first calculates the global average removal rate of pollutants, which represents the distance between the current remediation progress and the final target. The calculation formula is as follows:

[0084] ;

[0085] in, This represents the global average removal rate of pollutants. To restore the initial total mass of contaminants at startup, This represents the total remaining mass of pollutants during the current control period.

[0086] Next, the calculation formula for the multi-objective optimization function is defined as follows:

[0087] ;

[0088] in, The function value of the multi-objective optimization function. , and All are weighted coefficients and satisfy , This represents the global average removal rate of the pollutant. To improve overall drug utilization, This is the system's overall security index.

[0089] The adaptive particle swarm optimization algorithm treats the entire injection site as a search space, encoding the flow rate and concentration settings of each injection well into particle position vectors. During the iterative optimization process, the algorithm comprehensively weighs three mutually constraining engineering objectives:

[0090] The requirement is to achieve the highest possible pollutant removal rate, i.e., item Approaching 0; requiring the highest possible utilization rate of the reagent to avoid chemical waste and secondary pollution, i.e., item The index tends to 0; at the same time, the overall system safety index is required to be maintained at a high level to avoid large-scale geological damage, i.e., the item The algorithm adaptively adjusts the particle inertia weights and learning factors based on the diversity of the current population, eventually converging to the Pareto optimal solution set. The calculated optimal target flow rate and optimal target concentration are then sent to the corresponding edge single-well controllers for execution.

[0091] like Figure 5 The diagram presents a scatter plot of Pareto front data for adaptive particle swarm optimization (APSO). A three-dimensional coordinate system is constructed in the plot, with the two horizontal axes at the bottom representing pollutant removal rate (percentage) and reagent utilization rate (percentage), respectively. The vertical axis represents the system's global safety index (dimensionless).

[0092] The light blue dots in the figure represent historical search particles generated by the repair effect-oriented multi-well collaborative optimization module during the iterative optimization process, objectively demonstrating the algorithm's extensive exploration of the solution space for various combinations of injection flow and concentration. The red dots constitute the Pareto optimal solution set that finally converges. Observing the spatial distribution trend of the red dots reveals a significant mutual constraint relationship among the three engineering optimization objectives.

[0093] As the pollutant removal rate approaches a high level of 95%, the reagent utilization rate rapidly declines to around 65%, while the system's overall safety index also decreases to approximately 0.65. This three-dimensional data distribution pattern accurately reflects the objective physical boundaries in actual underground chemical injection projects. Specifically, forcibly increasing the output of the variable frequency pump in the single-well flow control command to pursue a high decontamination rate inevitably leads to excessive reagent loss in the dominant channel and increases the risk of formation pore fracturing. The formation of this frontier curve effectively demonstrates that this system possesses the comprehensive capability to calculate the optimal target flow rate and concentration under complex constraints.

[0094] Furthermore, observing the distribution curve of the red optimal solution set reveals that when the pollutant removal rate is within a specific range of 88% to 92%, the arrangement of the red scatter points exhibits obvious discontinuities and sparsity. This distribution characteristic reflects that when the current adaptive particle swarm optimization algorithm deals with strongly coupled nonlinear constraint boundaries, the particle population is prone to clustering in the compromise region, resulting in a loss of some distribution uniformity. This indicates that there is still room for further improvement in maintaining population diversity and adaptively adjusting inertial weights.

[0095] Under this control architecture, the first control cycle in the multi-timescale nested full closed-loop control architecture is shorter than the second control cycle, and the second control cycle is shorter than the third control cycle. The first control cycle is used to drive the real-time control closed loop of the edge single well, the second control cycle is used to drive the dynamic update closed loop of the generative geological model, and the third control cycle is used to drive the self-calibration and global optimization closed loop of the repair effect. The three control closed loops are nested to achieve dynamic compensation and optimization control of reagent injection.

[0096] For example, the first control cycle can be set to the millisecond to second level to ensure that mechanical or geological anomalies such as sudden pressure surges on site can be instantly captured and intercepted by the long short-term memory network; the second control cycle can be set to the minute to hour level to accumulate enough pressure gradient data to invert the dynamic evolution of underground seepage channels; the third control cycle can be set to the day or week level to use the chemical and concentration data enriched over a long period of time for in-depth calibration of global macroscopic parameters and overall scheduling of future operation plans.

[0097] Based on existing technologies, traditional chemical injection systems often employ open-loop control or a single pressure negative feedback mode. There is a lack of data interaction and overall coordination between injection wells, and the injection strategies implemented are fixed and rigid. This not only frequently leads to excessive loss of chemicals in high-permeability areas and failure of chemicals to reach low-permeability areas, but also easily causes engineering accidents such as well pipe rupture or dense fouling of filter layers.

[0098] The control system and method proposed in this embodiment achieve millisecond-level hardware security defense and near-wellbore micro-state monitoring through edge LSTM state identification in the first control cycle. The generative diffusion model in the second control cycle addresses the technical challenge of directly observing heterogeneous underground media. Finally, Bayesian self-calibration and multi-objective particle swarm optimization in the third control cycle achieve optimal spatial allocation of the reagents. This multimodal perception-driven, edge-cloud collaborative, and multi-timescale nested overall scheme significantly enhances the system's dynamic adaptability to complex and unknown underground conditions, and substantially improves the effective utilization and conversion rate of reagents, as well as the overall safety and economic benefits of site remediation projects.

[0099] Example 2:

[0100] In in-situ chemical remediation of groundwater and soil, the geological bodies are often highly heterogeneous, containing highly permeable gravel lenses, fracture networks, and low-permeability silty clay matrices. Traditional constant-flow continuous injection control methods are prone to chemical cross-flow when dealing with such complex formations. This means that under pressure, the chemical preferentially flows away rapidly along the high-permeability channels with the least fluid resistance, failing to effectively reach and penetrate the low-permeability matrix areas rich in contaminants. This not only leads to significant waste of expensive chemicals but also creates large blind spots in the remediation process.

[0101] In order to actively identify and dynamically suppress such non-ideal behavior in the microfluidic transport process, this embodiment further presents a specific scheme for dynamic suppression of the dominant channel based on the multi-timescale nested full closed-loop control architecture constructed in the above embodiments.

[0102] Specifically, the edge single-well controller also has a built-in dominant channel dynamic suppression module; the dominant channel dynamic suppression module is used to extract pressure data from the physical parameters and calculate the pressure gradient abrupt change rate along the friction during the execution of the single-well flow control command, and combine it with the high-frequency distortion value of the fluid acoustic signature in the acoustic parameters to calculate the spatial flow divergence rate; the formula for calculating the spatial flow divergence rate is:

[0103] ;

[0104] in, For spatial flow divergence rate, The friction gradient abrupt change rate along the pressure gradient is... This refers to the high-frequency distortion value of the fluid acoustic signature. This is the stress mutation weighting coefficient. This is the audio distortion weighting coefficient.

[0105] When the spatial flow divergence rate is greater than the set crossflow judgment threshold, the dominant channel dynamic suppression module intercepts and modifies the single-well flow control command, switching the continuous injection mode to the variable frequency pulse injection mode. In the variable frequency pulse injection mode, the dominant channel dynamic suppression module calculates the pulse compensation frequency in real time based on the spatial flow divergence rate, and dynamically adjusts the start-stop duty cycle of the distributed execution unit according to the pulse compensation frequency.

[0106] Specifically, the process by which the dominant channel dynamic suppression module acquires and calculates the friction gradient abrupt change rate relies on the arrayed fiber optic distributed pressure sensors deployed along the injection well tubing in the multimodal data sensing unit. Under normal radial uniform diffusion conditions, the fluid friction pressure distribution inside the injection well tubing should exhibit a smooth attenuation curve consistent with the fluid dynamics tubing friction loss.

[0107] However, when a high-permeability dominant channel develops in the formation at a certain depth outside the well casing, a large amount of reagent will rapidly escape laterally from the filter pipe at that depth towards the formation. This large local fluid loss causes the kinetic energy of the fluid inside the pipe in that specific depth range to be rapidly converted into pressure differential, thus forming a distinct "pressure drop funnel" on the spatial distribution curve of the pressure sensor. The dominant channel dynamic suppression module extracts the local extrema in the spatial derivative sequence by performing spatial differentiation on the real-time pressure data at each depth node, and defines the rate of change of the extrema as the friction gradient abrupt change rate. The magnitude of this abrupt change rate directly characterizes the severity of the non-uniform dispersion of the reagent into a single channel at the macroscopic fluid dynamics level.

[0108] For example, the flow state of fluids in different porous media exhibits fundamental acoustic differences. In homogeneous low- or medium-permeability formations, fluid flow typically follows Darcy's law and is in a laminar flow state. The acoustic emission signals generated by the friction between the fluid and the formation skeleton are mostly concentrated in the low-frequency range and have stable energy.

[0109] However, when the reagent rushes in at high speed and enters high-permeability fractures or coarse-grained dominant channels, the local surge in flow velocity can cause the flow pattern to change from laminar to non-Darcy flow or even turbulent flow. This microscopic turbulent vortex rupture and the shearing impact of the fluid on formation solid particles can excite high-frequency acoustic waves with specific frequency bands. The edge single-well controller continuously collects downhole acoustic data through a built-in high-precision hydrophone and uses a fast Fourier transform algorithm to convert the time-domain acoustic signals into frequency-domain signals.

[0110] Subsequently, the dominant channel dynamic suppression module extracts the high-frequency spectral energy characterizing turbulence features and compares it with the baseline spectral energy under normal laminar flow conditions to calculate the high-frequency distortion value of the fluid acoustic signature. This distortion value provides independent physical evidence from a micro-fluid dynamics perspective for determining whether abnormal scouring of the dominant channel has occurred within the formation.

[0111] like Figure 6 This diagram illustrates the coupling characteristics of high-frequency acoustic spectral distortion and pressure drop funnel under crossflow conditions. The diagram comprises two subplots. The upper subplot represents wellbore depth (meters) on the x-axis and fluid pressure along the flow path (megapascals) on the y-axis. The legend includes both normal radial diffusion pressure and funnel pressure under crossflow conditions. The lower subplot represents acoustic frequency (Hertz) on the x-axis and spectral energy density (decibels) on the y-axis. The legend includes both normal laminar acoustic spectrum and distorted spectrum under crossflow conditions.

[0112] Observing the upper subplot, it can be seen that the blue normal radial diffusion pressure curve shows a smooth linear decay trend with increasing depth, steadily decreasing from 2.0 MPa to around 1.5 MPa. This reflects that the agent is in a uniform radial diffusion state in the underground porous medium.

[0113] However, the red crossflow state funnel pressure curve shows a sharp dip near a depth of 30 meters, forming a distinct pressure drop funnel. The lowest point of the curve touches around 1.2 MPa. This physically demonstrates the local kinetic energy conversion and pressure differential change caused by the rapid lateral loss of a large amount of fluid from the dominant channel at this depth.

[0114] Looking at the frequency domain data of the lower half of the subplot, the blue normal laminar acoustic spectrum shows that its energy is mainly concentrated in the low-frequency range below 200 Hz and the waveform is stable, which is consistent with the physical laws of Darcy laminar flow.

[0115] Significantly coupled with this, when the pressure funnel appears, the distorted spectrum of the red crossflow state exhibits a huge energy jump in the high-frequency band of 400 to 800 Hz. This nonlinear distortion of the waveform confirms that the fluid velocity surges when entering the highly permeable fracture, triggering turbulent vortex rupture and generating a strong stripping impact on formation solid particles. This simultaneous coupling of spatial pressure characteristics and frequency domain acoustic characteristics provides a mutually corroborating physical basis for the dynamic suppression module of the dominant channel.

[0116] Meanwhile, observing the red crossflow acoustic curve reveals dense and sharp spikes in its high-frequency uplift region. This indicates that the high-precision hydrophone built into the edge single-well controller, while picking up turbulent acoustic signatures, still incorporates some high-frequency harmonics from the mechanical vibrations transmitted through the tubing by the variable-frequency speed-regulating injection pump. These signal spikes caused by strong on-site interference objectively demonstrate that the system still has room for further improvement in its underlying acoustic blind source separation and hardware vibration-damping filtering algorithms under extreme crossflow conditions.

[0117] Optionally, when calculating the selected spatial flow divergence rate, the system employs a linear weighted model to fuse macroscopic pressure characteristics with microscopic acoustic characteristics. In the above equation, the pressure mutation weighting coefficient... And audio distortion weighting coefficient The settings are not arbitrary but need to be pre-calibrated based on the specific hydrogeological conditions of the injection well site. Specifically, in sites rich in large-pore structures or karst fissures, the acoustic excitation mechanism during fluid flow is more pronounced, and the system will be adjusted accordingly. Higher weighting is applied to sites where the medium is predominantly fine sand, relatively homogeneous but with varying permeability, the spatial gradient of pressure is more sensitive, and the system will adjust accordingly. The weighting of multi-source heterogeneous data is effectively avoided by the misjudgment caused by local drift or environmental electromagnetic interference from a single sensor, ensuring the physical reliability of the core diagnostic indicator, spatial flow divergence rate.

[0118] It is also important to note that the control logic architecture of the edge well controller incorporates a strict interruption and interception mechanism. Under normal operating conditions, the edge well controller follows the single-well flow control commands issued by the central control platform based on the global geological model, maintaining continuous and stable reagent injection (i.e., continuous injection mode). However, once the spatial conductivity divergence rate calculated in real time by the dominant channel dynamic suppression module exceeds the pre-calibrated crossflow judgment threshold, it means that the local reagent loss in the formation has reached a level that cannot be ignored. Continuing to execute global commands would result in serious wasted engineering investment.

[0119] At this point, the dominant channel dynamic suppression module will trigger a local high-priority interrupt, directly intercepting and freezing the regular continuous traffic commands issued by the central control platform at the edge, forcing the execution logic to switch to variable frequency pulse injection mode. This demonstrates the architectural advantage of a distributed control system, which provides macro-level coordination at the central level and rapid response at the edge.

[0120] Specifically, after entering the variable frequency pulse injection mode, simple blind pulses cannot achieve the best anti-crossflow effect; adaptive matching must be performed based on the severity of the current crossflow. Therefore, the dominant channel dynamic suppression module calculates the pulse compensation frequency in real time based on the currently calculated spatial current divergence rate. The formula for calculating the pulse compensation frequency is as follows:

[0121] ;

[0122] in, The pulse compensation frequency is calculated in real time. The preset reference pulse frequency for the system, The frequency adjustment coefficient reflects the mechanical response sensitivity. It is the natural logarithm function. The calculated spatial divergence rate is given below. The threshold for determining cross-current is set.

[0123] As can be seen from the logarithmic adjustment function, when the spatial flow divergence rate slightly exceeds the threshold, the system pulses at a low frequency close to the reference frequency; while when the divergence rate increases sharply, i.e., when the crossflow phenomenon is extremely violent, the system will rapidly increase the compensation frequency. High-frequency pressure pulses can generate more concentrated transient shock waves in the fluid. Utilizing the viscosity changes of non-Newtonian fluids under high-frequency shear and the sharp increase in fluid inertia in porous media, a strong dynamic hydrodynamic blockage is formed on the high-permeability dominant channels, i.e., the "fluid blocking" effect.

[0124] For example, in addition to the pulse compensation frequency, dynamically adjusting the start-stop duty cycle of the distributed execution unit is also a core control parameter of the variable frequency pulse injection mode. The start-stop duty cycle is defined as the ratio of the pressurization time of the pump valve opening within one pulse cycle to the entire cycle time.

[0125] In terms of physical mechanism, when the pulse is in the "on" stage, the transient high-pressure water hammer effect can form a shock wave that radiates outward around the well. Due to the high fluid velocity in the high-permeability channel, the transient inertial resistance it experiences is also amplified many times over.

[0126] When the pulse is in the "stop" phase of the pressure relief interval, the hydrodynamic pressure inside the wellbore dissipates rapidly, and the capillary force inside the formation matrix becomes dominant. Because the pores of the low-permeability matrix are much smaller, its capillary adsorption capacity is much stronger than that of the large pores or fractures of the high-permeability matrix.

[0127] By reasonably adjusting the start-stop duty cycle and increasing the residence time window during the "stop" phase, the system can make full use of the capillary self-absorption effect, forcing the reagent solution accumulated near the well wall under the action of shock waves to slowly infiltrate into the low-permeability contamination blind zone that was originally difficult to reach through permeation.

[0128] like Figure 7 The figure displays the timing waveforms of the duty cycle and compensation frequency dynamically adjusted in variable frequency pulse injection mode. The horizontal axis represents the running time in seconds, the left vertical axis represents the actuator start / stop control signal in percentage, and the right vertical axis represents the transient pressure of the downhole tubing in megapascals. The legend indicates the blue start / stop control signal curve and the red transient pressure response curve.

[0129] In the initial phase of 0 to 10 seconds, the blue control signal stabilizes at 60, and the red pressure curve remains stable around 2.0 MPa, corresponding to the system being in the normal continuous injection mode. When the running time reaches 10 seconds and formation crossflow is detected, the system triggers an interception mechanism to switch to the variable frequency pulse injection mode.

[0130] At this point, the blue control signal changes from a stable linear signal to a high-frequency square wave alternating between 80 and 0. As time progresses, the pulse compensation frequency of this square wave gradually increases, while the pressurization time window in the on state gradually shortens, meaning the start-stop duty cycle dynamically decreases from 50 to around 30. This adjustment aims to increase the pressure relief interval time during the stop phase of the square wave.

[0131] Correspondingly, the red transient pressure response curve exhibits a violently jagged alternating waveform driven by the blue signal. During the pulse initiation phase, the red pressure instantly surges to approximately 3.5 MPa, utilizing the transient high-pressure shock wave to create strong hydrodynamic resistance to the fluid within the high-permeability channel; during the pulse termination phase, the pressure rapidly drops back to around 0.8 MPa, allowing the capillary self-absorption effect within the formation matrix to dominate, prompting the reagent to slowly infiltrate the low-permeability contaminated blind zone.

[0132] Meanwhile, observing the red transient pressure response curve reveals that at the rising edge of each pulse, significant oscillations, spikes, and overshoot phenomena appear at the peak position, accompanying the rapid pressure increase, with the highest instantaneous pressure occasionally approaching 4.0 MPa. This overshoot oscillation in the physical waveform accurately reflects that even under extremely high-frequency pulse switching, the mechanical coordination between the variable frequency speed-regulating injection pump and the proportional regulating solenoid valve is still affected by the inertial impact of water hammer reflection waves from the downhole fluid. This objectively indicates that there is still room for further improvement in the smooth suppression control of transient flow waveform shaping by the distributed actuator.

[0133] Optionally, at the actuator coordination level, this high-frequency, high-precision variable-frequency pulse injection mode places high demands on the hardware response of the distributed actuators. To achieve millisecond-level start / stop duty cycle adjustment and flow waveform shaping, the system abandons traditional mechanical throttle valves and instead coordinates the control of the variable-frequency speed-regulating injection pump and the proportional regulating solenoid valve. The edge single-well controller sends a torque command curve with a specific slope to the industrial frequency converter of the variable-frequency speed-regulating injection pump to control the acceleration and deceleration of the motor rotor.

[0134] Simultaneously, a high-frequency pulse width modulation (PWM) signal is sent to the proportional control solenoid valve. Through microsecond-level synchronization between the pump and valve, square wave or sawtooth wave pressure pulses with steep rise and fall edges can be generated in the downhole tubing. Meanwhile, to prevent fatigue damage to the surface pipeline flanges and downhole casing caused by repeated water hammer, the edge single-well controller also monitors the peak absolute pressure within the tubing in real time, ensuring that all pulse peaks are strictly limited within the safe envelope of formation fracture pressure and tubing yield strength.

[0135] It is important to note that the intervention of the dominant channel dynamic suppression module is a dynamic closed-loop correction process. During continuous operation in the variable frequency pulse injection mode, the module continuously recalculates the spatial conductivity divergence rate. With the continuous intervention of pulse dynamics, the fluid resistance within the high-permeability dominant channel gradually accumulates, forcing the reagent to achieve uniform dispersion into the surrounding formation. When the subsequently calculated spatial conductivity divergence rate gradually decreases and stabilizes below the set crossflow judgment threshold, it indicates that the crossflow state in the near-wellbore zone has been successfully broken and reshaped. At this point, the dominant channel dynamic suppression module will release the interception state, release the interruption, and allow the edge single-well controller to smoothly transition back to the initial continuous injection mode, continuing to execute the macroscopic optimization commands of the central control platform. To avoid control oscillations caused by frequent switching near the threshold critical point, a hysteresis loop interval design is also introduced into the control logic.

[0136] Based on the analysis of existing underground in-situ chemical injection technology, traditional equipment is often in an injection state without real-time feedback from the underground state. Since it is impossible to perceive the changes in the movement of micro-fluids deep underground in real time, the construction party can only judge the occurrence of crossflow by macroscopic lag phenomena such as surface bubbling and abnormal outflow of chemicals from surrounding monitoring wells. This results in the ineffective consumption of expensive chemical agents and increases the time cost of the project.

[0137] The dominant channel dynamic suppression technology solution provided in this embodiment cleverly utilizes the physical and acoustic sensing hardware integrated in the existing injection well. Through rigorous hydrodynamic feature extraction and multimodal data fusion algorithms, it achieves millisecond-level prediction and active interception of the dominant channel.

[0138] More importantly, this solution, through an innovative combination of variable frequency pulses and duty cycle adjustment, transforms mechanical kinetic energy into a hydrodynamic barrier deep within the formation. This achieves adaptive dynamic compensation for formation heterogeneity without adding any chemical water-blocking agents or physical flow-blocking tools in the well. The overall solution not only significantly improves the effective spatial sweep efficiency of expensive remediation agents and reduces the risk of secondary contamination, but also enables the distributed injection system to autonomously respond to complex geological disturbances at the lower levels.

[0139] Example 3:

[0140] In in-situ chemical remediation and porous media fluid injection projects for groundwater, the physicochemical compatibility between the geological body and the injected fluid is a key factor determining the long-term stable operation of the system. In conventional continuous injection operations, due to the violent redox reaction between high-concentration chemical agents (such as persulfate, permanganate, or Fenton's reagent) and the native groundwater, insoluble metal oxide precipitates (such as ferric hydroxide and manganese dioxide precipitates) are easily formed in the gaps of the filter pipes in the injection well and in the pores of the formation near the well.

[0141] Furthermore, changes in fluid pressure and ambient temperature can also trigger physical scaling of carbonates or sulfates in groundwater. The gradual accumulation of these precipitates and scaling substances significantly reduces formation porosity and permeability, leading to severe "interfacial fouling." Fouling not only prevents subsequent reagents from reaching the target contaminated area, causing abnormal increases in injection resistance and system pipeline pressure, but also frequently triggers system shutdown protection, resulting in the interruption of the entire remediation project.

[0142] To achieve intelligent prediction and online unblocking of deposits and fouling at the interface between equipment and formation, this embodiment, based on the multi-timescale nested closed-loop control architecture constructed in the above embodiments, further presents a specific scheme for near-wellbore fouling reversal.

[0143] Specifically, the edge single-well controller also has a built-in near-wellbore blockage reversal module; the near-wellbore blockage reversal module is used to extract pressure data from the physical parameters during the reagent injection process to generate injection pressure time series coefficients, and simultaneously collect the operating parameters of the injection pump motor in the distributed execution unit to calculate the interface fouling accumulation degree; the operating parameters include the real-time current of the injection pump motor and the reference steady-state current under non-fouling steady-state conditions; the formula for calculating the interface fouling accumulation degree is:

[0144] ;

[0145] in, The degree of dirt accumulation on the interface. The injection pressure time series coefficients are... The real-time current of the injection pump motor. This refers to the reference steady-state current of the injection pump motor under steady-state conditions without fouling. This is the pressure fluctuation weighting coefficient. This is the motor load weighting coefficient.

[0146] When the degree of fouling accumulation at the interface exceeds a preset fouling critical threshold, the near-wellbore fouling reversal module intercepts the single-well flow control command and triggers the distributed execution unit to perform a chemical-mechanical joint unblocking process. The chemical-mechanical joint unblocking process includes: suspending conventional agent injection and controlling the high-pressure clean water flushing pump to alternately inject clean water and unblocking agent until the re-extracted injection pressure time series coefficient and the real-time current both return to the preset safety threshold range.

[0147] Specifically, the primary task of the near-wellbore plugging reversal module is to perform early physical identification of the formation permeability decay trend. Traditional pressure monitoring often focuses only on the absolute pressure value at a single moment, which is easily affected by short-term water hammer impacts or local fluid oscillations, leading to false alarms. Therefore, this embodiment uses the injection pressure time series coefficient as a dynamic indicator to measure the evolution of plugging. The edge single-well controller, through its subordinate distributed physical parameter monitoring subunits, continuously acquires the dynamic pressure values ​​within the injection well string at a fixed sampling frequency and constructs a sliding time window on a one-dimensional time axis. Within the sliding time window, the complete algorithm for the near-wellbore plugging reversal module to calculate the injection pressure time series coefficient is as follows:

[0148] ;

[0149] in, The time series coefficients of the injected pressure are... This refers to the total number of pressure sampling points included within the set sliding time window. This is the time index number of the sampling point within the current sliding time window. For the first Dynamic pressure values ​​inside the tubing collected at each time index point. This is the baseline steady-state injection pressure value for the injection well when it was initially constructed and no fouling occurred. This is a time decay weighting factor that increases with the time index. A time decay weighting factor is introduced. The aim is to assign higher weight to pressure fluctuations that are closer to the present moment in time, thereby making the coefficient... It can more sensitively reflect the nonlinear trend of recent accelerated fouling. As near-wellbore pores are gradually filled with chemical deposits, the hydrodynamic pressure required to maintain the same injection flow rate will inevitably show an upward trend. The value also deviates significantly from zero.

[0150] For example, in a distributed execution unit, since injection wells are typically located in remote or complex field environments, relying solely on downhole hardware sensors carries the engineering risk of reading drift due to sensor probes being corroded by high-concentration chemicals or covered by suspended particles. Therefore, this embodiment introduces the electrical operating parameters of the injection pump motor as a dimension for redundancy and cross-validation.

[0151] Because there is a strict mechanical coupling between the output shaft torque of a positive displacement or centrifugal variable frequency speed-regulating injection pump and the back pressure of the pipeline fluid, when the formation becomes clogged and the back pressure increases, the motor must output a larger electromagnetic torque on its stator side in order to maintain the target flow rate and speed set by the edge single-well controller. This leads to a change in the electrical characteristics of the inverter output side. The near-well clog reversal module reads the stator current after low-pass filtering directly from the driver inside the variable frequency speed-regulating injection pump via an industrial fieldbus, and defines it as the real-time current. .

[0152] Simultaneously, the system uses clean water for pipeline calibration during the initial injection operation, records the baseline current characteristics when the motor is running smoothly under the predetermined flow target, and saves it in the controller's non-volatile memory, defining it as the reference steady-state current. This method of strictly distinguishing between real-time fluctuation data and steady-state baseline data provides an accurate basis for comparing electrical parameters to quantify the extent of extra work done by the motor.

[0153] It should be noted that, in order to scientifically quantify the overall fouling status at the wellbore interface and eliminate the local uncertainty of a single sensor signal, the near-wellbore fouling reversal module incorporates a multi-source data coupling calculation engine, which performs calculations strictly in accordance with the aforementioned interface fouling aggregation degree calculation formula.

[0154] In this coupled computational architecture, the first term on the right side of the equals sign The second term characterizes the evolution of system resistance at the fluid dynamics level. This characterizes the energy loss evolution at the electromechanical level. Pressure fluctuation weighting coefficient. With motor load weighting coefficient The value needs to be determined offline at the factory or in the initial field based on the specific pump model parameters and geological environment.

[0155] Specifically, in application scenarios where the pump's rated power has a large margin and the formation is a high-pressure, dense environment, Conversely, if the pipeline network is long and the fluid viscosity is significantly affected by the reagent concentration, the load fluctuation of the motor will be more indicative, and the system will then be configured as follows: .

[0156] Through this coupling formula, the system mathematically maps and dimensionlessly integrates the physical fluid resistance at the front end of the well with the electromechanical load pressure at the back end, resulting in the interfacial fouling aggregation degree. It has good engineering anti-interference capabilities and can accurately reflect the severity of physical and chemical densification of the filter pipe and the near-well aquifer.

[0157] Optionally, in the software architecture of the distributed reagent injection control system, the near-wellbore blockage reversal module is given high-priority system control interruption permissions. Under normal conditions, The value is running at a low level, and the edge single-well controller executes normal reagent injection according to the instructions issued by the LSTM state identification module and the central control platform. Once the near-wellbore blockage reversal module calculates that the interface fouling accumulation degree is greater than the pre-configured fouling critical threshold, it indicates that the interface fouling has entered a dangerous stage that may lead to equipment overload and burnout or irreversible formation damage. At this time, the module will immediately trigger a hard interrupt, intercepting and overriding the currently executing single-well flow control command, cutting off the conventional reagent injection loop. This interception mechanism ensures that the system will not blindly continue to push reagents into the already blocked pores, thereby avoiding the system deteriorating the downhole conditions through ineffective work.

[0158] Specifically, after successfully intercepting the conventional command, the system seamlessly connects and automatically triggers the chemical-mechanical combined unblocking process. This process breaks through the inefficient mode of manual well cleaning using drilling rigs after shutdown, transforming it into an automated physical-chemical combined unblocking sequence preset in the underlying control logic. The execution of the process consists of three alternating action phases:

[0159] The first stage is the "mechanical squeezing and hydraulic oscillation stage." The edge single-well controller starts the high-pressure clean water flushing pump in the distributed execution unit to pump pure groundwater into the wellbore in the form of variable frequency pulses. The high-pressure pulsed water flow uses transient kinetic energy to impact the loose sediments attached to the filter pipe wall and the near-well gravel filling layer, destroying their physical bridging structure and playing a preliminary hydraulic stripping role.

[0160] The second stage is the "chemical complexation and dissolution stage." The controller regulates the proportional control solenoid valve and the automatic dosing device to pump specific unblocking agents (such as weakly acidic solutions rich in chelating agents like citric acid and edema acid, or specific biological enzymes) into the wellbore at a low flow rate. After entering the near-wellbore pores, these unblocking agents undergo targeted complexation reactions or acid-base neutralization reactions with the metal oxides or carbonate scale generated in the early stage due to redox reactions, transforming the insoluble solid precipitates back into soluble ionic complexes.

[0161] The third stage is the "static soaking and backflow stage," in which the system shuts off all injection pump valves, maintains a certain pressure within the wellbore, and allows the unblocking agent sufficient reaction time in the porous medium to ensure the full chemical dissolution reaction. These three stages operate alternately according to a preset number of cycles under the controller's control, achieving a complementary effect between mechanical hydraulic disruption and chemical dissolution.

[0162] For example, the termination of the chemical-mechanical combined unblocking process is not based on a rigid, fixed time setting, but rather on a strict feedback measurement and state identification closed loop. At the end of each alternating well-washing cycle, the system briefly initiates a steady-state test injection of clean water and calls the multimodal data sensing unit to re-acquire pressure and current data. The near-wellbore blockage reversal module tracks and recalculates the injection pressure time series coefficient and acquires the real-time current in real time. Only when the re-extracted injection pressure time series coefficient shows a downward trend and stabilizes within the preset safety threshold range, and the real-time current also falls back and approaches the reasonable fluctuation range of the benchmark steady-state current under steady-state conditions, does the module determine that the permeability of the near-wellbore interface has been substantially restored. Subsequently, the module sends a reset command to end the unblocking process, release the interception of the conventional control loop, and allow the edge single-well controller to reconnect to the collaborative optimization network of the central control platform, resuming normal distributed chemical injection.

[0163] like Figure 8 The figure displays multi-parameter dynamic response curves of interface fouling evolution and combined chemical-mechanical declogging. The horizontal axis represents operating time in hours, the left vertical axis represents the injection pressure time series coefficient in dimensionless units, and the right vertical axis represents the real-time current of the injection pump motor in amperes. The legend indicates the blue pressure time series coefficient curve and the red motor real-time current curve.

[0164] In the initial stage of operation, from 0 to 30 hours, as the chemical agents reacted with the native groundwater and continuously generated precipitates in the near-wellbore formation pores, the blue pressure time series coefficient showed a non-linear, accelerated increase from around 0.05, reaching a peak of 0.4 at approximately 30 hours. Simultaneously, the red motor real-time current also climbed from the baseline steady-state current of 15 amperes to around 22 amperes. This simultaneous increase in both parameters objectively reflects the physical evolution process of increased system fluid resistance and intensified electromechanical energy loss caused by interfacial fouling.

[0165] When the running time reaches approximately 30 hours, the near-wellbore blockage reversal module automatically triggers a combined chemical-mechanical unblocking process because the interface fouling accumulation exceeds the preset fouling critical threshold. At this time, a violent alternating downward oscillation and a cliff-like drop waveform can be observed in the two curves. This corresponds to the dynamic process of the pulsed water flow impacting and destroying the physical bridging structure of the near-wellbore sediment, and the chemical agents dissolving and ablating the sparingly soluble solid phase precipitate, when the system controls the high-pressure clean water flushing pump and the unblocking agent to inject alternately.

[0166] After the 35-hour unblocking process was completed, the system resumed regular chemical injection, and both curves returned to the safe range and remained stable.

[0167] Observing the recovery operation phase from 35 to 50 hours, it can be seen that the blue pressure time series coefficient stabilized at around 0.08, and the red motor real-time current stabilized at around 15.5 amperes. Neither of them completely dropped to the absolute initial reference value at 0 hours, and the red current curve was accompanied by high-frequency slight spikes caused by the operation of the bottom-level variable frequency drive. This waveform characteristic indicates that although the current combined unblocking procedure has effectively cleared the near-well interface and restored the system's continuous injection capability, some minor irreversible pore blockages still remain deep in the formation. At the same time, the bottom-level electrical sampling still has data fluctuations when faced with high-frequency interference from the variable frequency drive, indicating that there is still room for improvement in the optimization of unblocking agent ratios for specific geological conditions and the bottom-level electromechanical signal denoising algorithms.

[0168] In terms of existing technology, traditional chemical injection systems generally suffer from technical deficiencies such as delayed detection and limited response methods when dealing with near-wellbore fouling. Operators often only discover severe wellbore blockage after the flow meter reading has plummeted or the injection pump has tripped due to severe overload and triggered a hardware thermal relay. By this time, the blockage surface is already highly dense and solidified, often requiring expensive and time-consuming large-scale well workover equipment for physical cleaning or even the complete abandonment of the original well.

[0169] The technical solution proposed in this embodiment rigorously couples the characteristics of deep well hydrodynamic changes with the electromechanical load characteristics of the surface, establishing an early detection and proactive defense mechanism for near-wellbore blockage. More importantly, through a chemical-mechanical combined unblocking process automatically triggered by underlying logic, the system can destroy and dissolve the deposits in the early stages of forming a dense, hard crust, achieving a technological improvement from manual shutdown maintenance to online automatic unblocking. This greatly ensures the continuous operation capability of the distributed injection network throughout its entire lifecycle in complex underground hydrogeochemical environments, significantly reduces the implicit maintenance costs of in-situ remediation projects, and shortens the overall remediation period, demonstrating extremely high industrial practical value.

[0170] Example 4:

[0171] like Figure 9 As shown in the background section, traditional in-situ remediation agent injection control for underground contamination mainly relies on periodic offline manual sampling and analysis, and the control mechanism for multiple injection devices on site is relatively simple, usually using a uniformly set fixed flow rate for continuous or open-circuit injection. This invention provides a distributed agent injection control method based on LSTM dynamic compensation.

[0172] Specifically, the method in this embodiment is applied to the aforementioned distributed drug injection control system based on LSTM dynamic compensation. The method is executed based on a multi-timescale nested fully closed-loop control architecture and includes the following steps:

[0173] Step S1: In-situ synchronous sensing of multi-source heterogeneous data.

[0174] The physical, chemical, and acoustic parameters at the injection well site are collected by the multimodal data sensing unit, and the in-situ pollutant concentration data are collected by the sparse in-situ online remediation effect monitoring network in the multimodal data sensing unit.

[0175] Specifically, in practical site remediation engineering applications, the physical environment and the chemical reaction environment are dynamically coupled. The multimodal data sensing unit relies on hardware sensing equipment deployed in each injection well string and wellhead. Among them, the acquisition of physical parameters depends on fiber optic distributed pressure sensors and high-precision electromagnetic flowmeters to obtain the fluid pressure gradient at different depths downhole and the actual volumetric flow rate of the injected agent. These data directly reflect the current permeability resistance state of the formation. The acquisition of chemical parameters depends on online pH sensors and redox potential sensors to monitor the physicochemical activity of the injected agent before it enters the formation in real time. Acoustic parameters are acquired through a high-precision hydrophone array deployed inside the well casing to capture the weak high-frequency acoustic signals generated when fluids move at high speed in porous media or fractures.

[0176] In addition, to obtain accurate remediation feedback, the system does not rely on manual, periodic offline sampling. Instead, it uses a sparse, in-situ online remediation effect monitoring network deployed across different areas within the contaminated site. The network is equipped with online gas chromatographs or spectral sensors to automatically extract and analyze groundwater samples, thereby outputting in-situ pollutant concentration data that has engineering guidance significance.

[0177] Step S2: High-frequency state identification and real-time control on the edge side.

[0178] Multiple edge single-well controllers receive the corresponding physical, chemical, and acoustic parameters, respectively. Using the built-in structured long short-term memory network state identification module and hierarchical risk response module, they extract the formation disturbance index according to the first control cycle and output single-well flow control commands.

[0179] For example, edge single-well controllers typically consist of industrial control computers (IPCs) or programmable logic controllers (PLCs) with high computing power and interference resistance, and are directly deployed in field control cabinets close to each injection well. To cope with abrupt changes in the subsurface fluid dynamics environment (such as instantaneous physical blockage or localized formation fracturing in the near-wellbore zone), the edge-end method execution is based on an extremely short timescale, namely the first control cycle (typically set to milliseconds to seconds).

[0180] Within each first control cycle, the edge single-well controller converts the received sensor electrical signals into a standardized time-series matrix. The structured long short-term memory network state identification module leverages its nonlinear fitting advantage in processing time-series data to extract a formation disturbance index that quantifies the degree of damage to the formation microstructure from the combined characteristics of pressure fluctuations, flow oscillations, and acoustic distortion.

[0181] Subsequently, the hierarchical risk response module assesses the current local operational risk based on this index. If it is determined that the current injection pressure has caused excessive formation disturbance, the controller will immediately generate corresponding single-well flow control instructions, aiming to implement local physical defenses as soon as possible through pressure reduction or flow restriction.

[0182] Step S3: Central side model update and global multi-objective optimization.

[0183] The central control platform receives the physical parameters and formation disturbance index uploaded by the edge single-well controller, as well as the in-situ pollutant concentration data. It uses the built-in generative geological model dynamic update module to update the geological model based on the physical parameters and formation disturbance index according to the second control cycle. It also uses the built-in repair effect-oriented multi-well collaborative optimization module to optimize global parameters based on the in-situ pollutant concentration data according to the third control cycle, and outputs the optimal target flow rate and optimal target concentration to the corresponding edge single-well controller.

[0184] It's also important to note that, unlike edge single-well controllers which focus on local and transient safety, the central control platform, deployed on a remote cloud server or control center, prioritizes macro-level and long-term trend coordination in its execution logic. Data from each edge node is continuously aggregated to the central control center via industrial IoT protocols (such as MQTT or OPCUA). During the second control cycle (typically set to minutes to hours), the generative geological model dynamic update module utilizes the aggregated pressure gradient and formation disturbance index to perform reverse calibration and update of the permeability field of the internal basic three-dimensional hydrogeological model, thereby eliminating model errors caused by poor visibility of the subsurface environment.

[0185] On a longer timescale, namely the third control period (usually set to daily or weekly), the remediation effect-oriented multi-well collaborative optimization module uses macroscopic pollutant concentration data transmitted from the sparse in-situ online remediation effect monitoring network as feedback input. Addressing the problem of independent operation and lack of coordination among wells in traditional methods, this module, based on an updated geological model, calculates the parameter combination that maximizes the overall plume degradation efficiency while ensuring global safety and total reagent constraints. This combination is then decoupled into optimal target flow rates and optimal target concentrations, which are then distributed to each specific injection well.

[0186] Step S4: Instruction execution and dynamic compensation of distributed physical units.

[0187] The distributed execution unit receives the single-well flow control command issued by the edge single-well controller based on the optimal target flow rate and the optimal target concentration, and executes the reagent injection action of the corresponding injection well.

[0188] Specifically, the final implementation of this method relies on the physical actions of the on-site electromechanical equipment. The distributed execution unit receives underlying electrical control signals from the edge single-well controller. When the edge single-well controller does not determine the existence of a local high risk within the first control cycle, the single-well flow control command it issues will directly track the optimal target flow and optimal target concentration issued by the central control platform.

[0189] At the execution level, the industrial frequency converter of the variable frequency speed control injection pump adjusts the stator output frequency according to the flow command, changes the pump impeller speed to regulate the main pipeline flow; the proportional control solenoid valve changes the valve opening according to the command to fine-tune the fluid resistance; at the same time, the metering pump in the automatic dosing device is linked to adjust to ensure that the mixing ratio of the reagent mother liquor and dilution water strictly meets the requirements of the optimal target concentration, thereby completing the physical injection action in the entire closed-loop control cycle.

[0190] Based on actual industrial operations, the control method provided in this embodiment effectively overcomes the challenge of existing fixed injection control mechanisms being unable to adapt to complex formation evolution by constructing a multi-timescale nested closed-loop architecture that combines edge control and central remote coordination. The high-frequency edge identification in the first control cycle enables the system to react promptly in the early stages of local blockage or crossflow, reducing the risk of reagent deviation or formation fracturing caused by formation heterogeneity. The central model update and global optimization in the second and third control cycles change the traditional single "open-loop" or "blind injection" mode, ensuring that the reagent flow rate and concentration of injection wells in different areas correspond to the dynamic degradation requirements of the underground contaminant plume. This method effectively coordinates the collaborative operations of distributed injection wells while achieving a dynamic balance between precise spatial allocation of reagents and formation structure safety protection, improving the overall accuracy and operational stability of in-situ remediation projects.

[0191] Example 5:

[0192] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0193] like Figure 10The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0194] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0195] Bus 102 may include a pathway for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 10 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0196] The memory 103 stores a computer program corresponding to the distributed drug injection control method based on LSTM dynamic compensation according to the above embodiments of the present invention. This computer program is executed by the processor 101. The processor 101 executes the computer program stored in the memory 103 to implement the content shown in the aforementioned method embodiments.

[0197] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 10 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0198] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A distributed drug injection control system based on LSTM dynamic compensation, characterized in that, The system adopts a multi-timescale nested fully closed-loop control architecture, including: The multimodal data sensing unit includes a sparse in-situ online remediation effect monitoring network, which is used to collect physical, chemical and acoustic parameters at the injection well site, and to collect in-situ pollutant concentration data through the sparse in-situ online remediation effect monitoring network. Multiple edge single-well controllers are deployed at the corresponding injection well sites, receiving the physical, chemical, and acoustic parameters. They incorporate a structured long short-term memory (LSTM) network state identification module and a hierarchical risk response module. These modules output single-well flow control commands according to the first control cycle and extract the formation disturbance index. The structured LTM network state identification module employs a three-layer stacked LTM network structure. The input layer receives a standardized multimodal feature vector matrix composed of the physical, chemical, and acoustic parameters from multiple past time points. Temporal features are extracted through the LTM units of the hidden layer. The output layer outputs the equipment degradation index and the formation disturbance index. The equipment degradation index reflects the decline in mechanical efficiency. The formation disturbance index characterizes the degree of pathological evolution of the subsurface microstructure. The hierarchical risk response module is used to classify the system's comprehensive state index into three levels of operational risk and execute corresponding control strategies. The formula for calculating the system's comprehensive state index is: ;in, This is the system's overall state index. The degradation index of the equipment is [value missing]. The term refers to the formation disturbance index. The system is defined as an exponential function with a base of the natural constant. When the overall system state index is less than a first preset threshold, it is considered to be in a normal state, and standard dynamic compensation control is executed. When the overall system state index is greater than or equal to the first preset threshold and less than a second preset threshold, it is considered to be in a warning state, and a preventative control strategy of reducing injection flow and increasing sampling frequency is executed. When the overall system state index is greater than or equal to the second preset threshold, it is considered to be in a dangerous state, and an emergency protection strategy of shutting down the injection pump is executed. Wherein, the first preset threshold is less than the second preset threshold. The central control platform, deployed in the remote monitoring center, receives the physical parameters and formation disturbance index uploaded by the edge single-well controllers, as well as the in-situ pollutant concentration data. It has a built-in generative geological model dynamic update module and a remediation effect-oriented multi-well collaborative optimization module. It is used to update the geological model based on the physical parameters and formation disturbance index according to the second control cycle, and to optimize the global parameters based on the in-situ pollutant concentration data according to the third control cycle, and output the optimal target flow rate and optimal target concentration to the edge single-well controllers. The distributed execution unit is used to receive the single-well flow control command issued by the edge single-well controller based on the optimal target flow rate and the optimal target concentration, and execute the reagent injection action of the corresponding injection well.

2. The system according to claim 1, characterized in that, The multimodal data sensing unit also includes: The distributed physical parameter monitoring subunit includes an array of fiber optic distributed pressure sensors deployed along the injection well string, as well as a high-precision electromagnetic flowmeter, temperature sensor, and density sensor integrated into the wellhead. The chemical parameter monitoring subunit includes an online pH sensor, a redox potential sensor, and an ion-selective electrode integrated into the wellhead reagent mixing pipeline; The acoustic parameter monitoring subunit includes high-precision hydrophones installed at the bottom and middle of the well casing; The sparse in-situ online remediation effect monitoring network includes in-situ online monitoring wells respectively deployed in the core area, boundary area and background area of ​​the pollution plume, and is equipped with an online gas chromatograph and a fiber optic Raman spectroscopy sensor.

3. The system according to claim 1, characterized in that, The generative geological model dynamic update module has a built-in conditional probability diffusion model, which is used to receive the pressure gradient data and the formation disturbance index extracted and uploaded by the edge single well controller based on the pressure data in the physical parameters. The conditional probability diffusion model adds Gaussian noise through a forward diffusion process and performs a reverse noise reduction process by inputting the real-time pressure gradient data to generate an updated three-dimensional formation permeability coefficient field. Based on the updated three-dimensional formation permeability field, the groundwater seepage equation and convection-diffusion reaction equation are solved to generate the predicted pollutant concentration distribution field.

4. The system according to claim 3, characterized in that, The central control platform also has a built-in module for calculating the deviation of the repair effect; The repair effect deviation calculation module incorporates an anisotropic spherical semi-variogram function Kriging interpolation algorithm, which is used to input the measured in-situ pollutant concentration data, generate a global interpolated concentration field, and compare the global interpolated concentration field with the predicted pollutant concentration distribution field to calculate the global concentration deviation field and the global average concentration deviation.

5. The system according to claim 4, characterized in that, The central control platform also has a built-in generative geological model Bayesian self-calibration module. The generative geological model Bayesian self-calibration module has a built-in Markov chain Monte Carlo sampler, which is used to trigger parameter self-calibration when the global average concentration deviation exceeds the preset concentration deviation threshold or reaches the third control cycle. The Markov chain Monte Carlo sampler takes the global concentration bias field and the predicted pollutant concentration distribution field as inputs, and outputs the optimal calibration parameter vector to update the generative geological model. The optimal calibration parameter vector includes a three-dimensional effective porosity field, a pollutant-reagent second-order reaction rate constant field, and a pollutant solid-phase adsorption and partition coefficient field.

6. The system according to claim 1, characterized in that, The repair effect-oriented multi-well collaborative optimization module incorporates an adaptive particle swarm optimization algorithm, which uses the minimum value of a multi-objective optimization function as the optimization objective, and outputs the optimal target flow rate and the optimal target concentration for each injection well. The formula for calculating the multi-objective optimization function is as follows: ； in, The function value of the multi-objective optimization function. , and All are weighted coefficients and satisfy , This represents the global average removal rate of pollutants. To improve overall drug utilization, This is the system's overall security index.

7. The system according to any one of claims 1-6, characterized in that, In the multi-timescale nested full closed-loop control architecture, the first control cycle is shorter than the second control cycle, and the second control cycle is shorter than the third control cycle; The first control cycle is used to drive the real-time control closed loop of the edge single well, the second control cycle is used to drive the dynamic update closed loop of the generative geological model, and the third control cycle is used to drive the self-calibration and global optimization closed loop of the repair effect. The three control closed loops are nested to achieve dynamic compensation and optimization control of the reagent injection.

8. A distributed drug injection control method based on LSTM dynamic compensation, characterized in that, Applied to the system of any one of claims 1 to 7, the method is executed based on a multi-timescale nested fully closed-loop control architecture and includes the following steps: The physical, chemical, and acoustic parameters at the injection well site are collected by the multimodal data sensing unit, and the in-situ pollutant concentration data are collected by the sparse in-situ online remediation effect monitoring network in the multimodal data sensing unit. Multiple edge single-well controllers receive the corresponding physical, chemical, and acoustic parameters, respectively. Utilizing a built-in structured long short-term memory (LSTM) network state identification module and a hierarchical risk response module, they extract the formation disturbance index according to the first control cycle and output single-well flow control commands. The structured LTM network state identification module employs a three-layer stacked LTM network structure. The input layer receives a standardized multimodal feature vector matrix composed of the physical, chemical, and acoustic parameters from multiple past time points. Temporal features are extracted through the LTM units of the hidden layer. The output layer outputs the equipment degradation index and the formation disturbance index. The equipment degradation index reflects the decline in mechanical efficiency; the formation disturbance index characterizes the degree of pathological evolution of the subsurface microstructure. The hierarchical risk response module is used to classify three levels of operational risk based on the system comprehensive state index and execute corresponding control strategies. The formula for calculating the system comprehensive state index is: ;in, This is the system's overall state index. The degradation index of the equipment is [value missing]. The term refers to the formation disturbance index. The system is defined as an exponential function with a base of the natural constant. When the overall system state index is less than a first preset threshold, it is considered to be in a normal state, and standard dynamic compensation control is executed. When the overall system state index is greater than or equal to the first preset threshold and less than a second preset threshold, it is considered to be in a warning state, and a preventative control strategy of reducing injection flow and increasing sampling frequency is executed. When the overall system state index is greater than or equal to the second preset threshold, it is considered to be in a dangerous state, and an emergency protection strategy of shutting down the injection pump is executed. Wherein, the first preset threshold is less than the second preset threshold. The central control platform receives the physical parameters and formation disturbance index uploaded by the edge single-well controller, as well as the in-situ pollutant concentration data. It uses the built-in generative geological model dynamic update module to update the geological model based on the physical parameters and formation disturbance index according to the second control cycle. It also uses the built-in repair effect-oriented multi-well collaborative optimization module to optimize global parameters based on the in-situ pollutant concentration data according to the third control cycle, and outputs the optimal target flow rate and optimal target concentration to the corresponding edge single-well controller. The distributed execution unit receives the single-well flow control command issued by the edge single-well controller based on the optimal target flow rate and the optimal target concentration, and executes the reagent injection action of the corresponding injection well.

Images