Intelligent self-adaptive injection flow regulation and control system for geological storage

Abstract

The invention discloses an intelligent self-adaptive injection flow regulation and control system for geological storage. The system comprises a sensor module, a data processing module, an execution mechanism module and a self-adaptive algorithm module, the sensor modules are distributed in an injection well and a surrounding monitoring well and are used for acquiring pressure, temperature and flow data in real time; the data processing module performs normalization processing on the monitoring data and extracts dynamic characteristics; the adaptive algorithm module integrates a multi-parameter coupling regulation and control model constructed by fuzzy PID and an LSTM neural network, predicts the trend of a seepage field and generates a regulation and control instruction; the execution mechanism module adjusts an electric control valve group and a variable-frequency injection pump according to instructions, so that the error between an actual injection curve and a target parameter curve is smaller than 5%. All the modules form a closed-loop control system through an industrial Internet of Things platform, the LSTM neural network updates weight parameters every 30 minutes, and a fuzzy PID controls an integral time constant to be automatically adjusted. The system is suitable for geological storage of CO2, accurate regulation and control of the injection flow can be achieved, and the leakage risk and the injection energy consumption are reduced.

Description

Technical Field

[0001] This invention relates to the field of intelligent adaptive injection flow control for geological storage, specifically to an intelligent adaptive injection flow control system for geological storage. Background Art

[0002] In carbon sequestration technology, precise control of the injection flow rate is crucial for ensuring both efficiency and safety. Current technologies primarily utilize sensors for single-point monitoring of the injection well, making it difficult to capture the dynamic evolution of the formation seepage field and resulting in incomplete control methods. The lack of standardized processing and anomaly filtering of sensor data directly impacts the accuracy of the control model input. Furthermore, inconsistent industrial IoT communication protocols and a lack of edge computing capabilities lead to significant data transmission delays and noise interference. This makes it particularly challenging to balance injection efficiency and sequestration safety under complex geological conditions. Summary of the Invention

[0003] The purpose of this invention is to provide an intelligent adaptive injection flow control system for geological storage, so as to solve the problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: an intelligent adaptive injection flow control system for geological storage, comprising: a sensor module, a data processing module, an actuator module, and an adaptive algorithm module;

[0005] The sensor module consists of formation pressure sensors, temperature sensors, and flow sensors distributed in the injection well and surrounding monitoring wells, and is used to collect pressure gradient, temperature distribution, and fluid velocity data of the injection layer in real time.

[0006] The data processing module includes a central controller and a data storage unit connected in communication with it. The central controller is configured to normalize the monitoring data of the sensor module and extract the dynamic change characteristics of the injected parameters.

[0007] The adaptive algorithm module is integrated into the central controller and includes a multi-parameter coupled control model based on fuzzy PID control and LSTM neural network. This model predicts the evolution trend of the formation seepage field according to the dynamic change characteristics of the injection parameters and generates control commands that match the formation storage capacity in real time.

[0008] The actuator module includes an electric regulating valve group and a variable frequency injection pump installed at the injection wellhead. The electric regulating valve group is connected to the central controller through a PID closed-loop control circuit and is used to dynamically adjust the injection flow rate and pressure according to the control command, so that the error between the actual injection curve and the target parameter curve output by the model is less than 5%.

[0009] The sensor module, data processing module, and actuator module form a closed-loop control system through an industrial IoT platform. The LSTM neural network updates its weight parameters every 30 minutes based on the latest monitoring data, and the integral time constant of the fuzzy PID control is automatically adjusted according to the injection pressure fluctuation amplitude, with the adjustment amplitude satisfying the formula ΔT=K·(ΔP / P0). 2 Where K∈[0.5,1.2] is the formation permeability correction coefficient, ΔT is the adjustment amount of the integral time constant, ΔP is the pressure fluctuation amplitude in the current monitoring period, and P0 is the initial formation pressure of the target injection layer.

[0010] Furthermore, in the sensor module, the formation pressure sensors in the injection well are arranged at intervals along the wellbore axis above, below, and in the middle of the target injection layer, and the monitoring wells are arranged at 30°, 90°, 150°, 210°, 270°, and 330° circumferential positions around the injection well, within a range of 50-200 meters from the injection well.

[0011] Furthermore, the normalization process includes Z-score standardization of the pressure gradient data, calculated as follows:

[0012]

[0013] Where μ p σ represents the average pressure during the current monitoring period. p The standard deviation is used; the temperature data are normalized using the maximum-minimum method, calculated as follows: in T min This refers to the historical extreme temperatures of the monitoring well.

[0014] Furthermore, the fuzzy rule base of the fuzzy PID control contains 25 control rules, with the input variables being the pressure deviation e and the deviation change rate ec, and the output variable being the proportional coefficient K. p Integration time T i The adjustment amount, the membership function adopts a triangular-trapezoidal hybrid distribution, and the centroid method is used for defuzzification.

[0015] The fuzzy set of the pressure deviation e is defined as {negative large (NB), negative small (NS), zero (ZO), positive small (PS), positive large (PB)}, and the universe of discourse range is [-0.2P0, 0.2P0];

[0016] The fuzzy set of the deviation change rate ec is defined as {rapid decrease (ND), slow decrease (NS), stable (ZE), slow increase (PS), rapid increase (PD)}, and the universe of discourse ranges from [-0.15P0 / min, 0.15P0 / min];

[0017] The triangular distribution parameters of the membership function satisfy the following: the overlap of adjacent fuzzy sets is 15%-25% of the universe width, and the slope of the trapezoidal distribution at the boundary of the universe is 45°±5°.

[0018] The logical expression for the control rule is:

[0019] IFe = NB AND ec = ND THENΔK p =PB ANDΔT i =NB

[0020] IFe=ZO AND ec=ZE THENΔK p =ZO ANDΔT i =ZE

[0021] (The remaining 23 rules are arranged according to the same logical structure);

[0022] The centroid method of the defuzzification method is specifically calculated as follows:

[0023]

[0024] Among them, u * To ensure accurate output values, μ i (u i ) represents the membership value when the i-th rule is triggered.

[0025] Furthermore, the LSTM neural network contains three hidden layers with 128, 64, and 32 neurons in each layer, respectively. During training, monitoring data with a time window length of 6 hours is used as input features, and the loss function is mean squared error (MSE). The training dataset contains injection condition data for at least 10 formations with different permeability.

[0026] Furthermore, the electric regulating valve assembly includes a coarse regulating valve and a fine regulating valve installed in series. The coarse regulating valve adopts a V-type ball valve with a response time ≤3 seconds and an adjustment accuracy of ±2%FS; the fine regulating valve adopts a cage-type regulating valve with a response time ≤1 second and an adjustment accuracy of ±0.5%FS; the frequency adjustment range of the variable frequency injection pump is 5-80Hz, and the pressure control accuracy is ≤0.1MPa.

[0027] Furthermore, the industrial IoT platform uses the OPCUA protocol to transmit sensor data, sets up edge computing nodes at the injection wellhead, and is equipped with a data cache and an abnormal data filtering algorithm. The filtering algorithm removes abrupt data exceeding three times the standard deviation.

[0028] Furthermore, the weight parameters of the LSTM neural network are updated using stochastic gradient descent, and the following conditions are met:

[0029] The Adam optimizer is used for parameter updates, with a momentum coefficient μ = 0.9–0.99 and a learning rate η = 0.001–0.01, which decays exponentially with each epoch. The decay formula is as follows:

[0030] η t =η0·γ t / T

[0031] Where γ = 0.95-0.99 is the decay coefficient, t is the current epoch number, and T = 200-500 is the total number of training epochs;

[0032] Each time the parameters are updated, the gradient direction vectors of the previous 10 iterations are accumulated, and the momentum optimization term is calculated to satisfy:

[0033]

[0034] Where v t The current momentum vector, The current gradient;

[0035] During training, the batch size is 64-256, and a Dropout ratio of 0.2-0.5 is used in the first 20 epochs to prevent overfitting.

[0036] The training termination condition is that the decrease in the validation set loss function is less than 10 for 10 consecutive epochs. -5 Alternatively, the maximum number of epochs T can be reached, and training can be accelerated using TensorRT on the NVIDIA CUDA computing platform for graph optimization.

[0037] Furthermore, the method for determining the formation permeability correction coefficient K includes obtaining the static permeability K through core experiments. core By combining initial pressure test data, the dynamic permeability K can be inverted. dynamic ,Pick

[0038] Furthermore, the system is applicable to CO2 geological storage scenarios, and the real-time matching of storage capacity includes: when the caprock pressure is detected to exceed 85% of the fracture pressure threshold, an injection flow rate reduction program is automatically triggered, with a reduction gradient of 20% / 5min of the current flow rate, and the tracer detection module of the backup monitoring well is activated at the same time.

[0039] Compared with the prior art, the present invention has the following beneficial effects:

[0040] This invention uses pressure sensors arranged at intervals along the wellbore axis, combined with a monitoring well group covering 6 circumferential directions and a range of 50-200 meters, to achieve three-dimensional monitoring of formation pressure, temperature, and flow velocity. The sensor-data processing-actuator form a closed loop through the Industrial Internet of Things. The coordinated control of LSTM and fuzzy PID keeps the error between the actual injection curve and the target curve within 5%, which is 55% more efficient than traditional open-loop control and significantly reduces the risk of CO2 leakage and injection energy consumption. Attached Figure Description

[0041] Figure 1 This is a system flowchart of the present invention. Detailed Implementation

[0042] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0043] Please see Figure 1 The present invention provides a technical solution: an intelligent adaptive injection flow control system for geological sealing, comprising: a sensor module, a data processing module, an actuator module and an adaptive algorithm module;

[0044] The sensor module consists of formation pressure sensors, temperature sensors, and flow sensors distributed in the injection well and surrounding monitoring wells, used to collect real-time data on pressure gradient, temperature distribution, and fluid velocity at the injection site.

[0045] In the sensor module, formation pressure sensors in the injection well are arranged at intervals along the wellbore axis above, below, and in the middle of the target injection layer, and monitoring wells are arranged at 30°, 90°, 150°, 210°, 270°, and 330° circumferential positions around the injection well, within a range of 50-200 meters from the injection well.

[0046] Injection well sensor arrangement: Pressure sensors are installed every 10-15 meters along the wellbore axis, specifically located at:

[0047] 5 meters above the packer (monitoring annular pressure);

[0048] 3 meters below the packer (top pressure of the target injection layer);

[0049] The target injection layer is located in the middle (1 / 2 layer thickness from the top of the layer), where the average pressure within the layer is obtained.

[0050] A high-precision vibrating wire pressure sensor (accuracy ±0.05%FS, temperature compensation range -20℃~80℃) is selected, with a sampling frequency of 10Hz, which meets the requirements for stable operation in long-term downhole high-pressure environments (≤50MPa).

[0051] Monitoring well layout: Centered on the injection well, monitoring wells are arranged in six directions at 30°, 90°, 150°, 210°, 270° and 330°, with distances of 50m, 100m and 200m respectively (at least one well in each direction, with a total of ≥6 wells).

[0052] Temperature sensors (accuracy ±0.1℃) and electromagnetic flow sensors (accuracy ±0.3%FS) are installed simultaneously inside the monitoring well to capture radial seepage characteristics.

[0053] The data processing module includes a central controller and a data storage unit connected in communication with it. The central controller is configured to normalize the monitoring data from the sensor module and extract the dynamic change characteristics of the injected parameters.

[0054] Normalization processing includes Z-score standardization of the pressure gradient data, calculated as follows:

[0055]

[0056] Where μ p σ represents the average pressure during the current monitoring period. p The standard deviation is used; the temperature data are normalized using the maximum-minimum method, calculated as follows: in T min This refers to the historical extreme temperatures of the monitoring well.

[0057] The adaptive algorithm module is integrated into the central controller and includes a multi-parameter coupled control model based on fuzzy PID control and LSTM neural network. This model predicts the evolution trend of the formation seepage field according to the dynamic change characteristics of the injection parameters and generates control commands that match the formation storage capacity in real time.

[0058] The actuator module includes an electric regulating valve group and a variable frequency injection pump installed at the injection wellhead. The electric regulating valve group is connected to the central controller through a PID closed-loop control circuit, which is used to dynamically adjust the injection flow and pressure according to the control command, so that the error between the actual injection curve and the target parameter curve output by the model is less than 5%.

[0059] The electric regulating valve assembly includes a coarse adjustment valve and a fine adjustment valve installed in series. The coarse adjustment valve adopts a V-type ball valve with a response time of ≤3 seconds and an adjustment accuracy of ±2%FS. The fine adjustment valve adopts a cage-type regulating valve with a response time of ≤1 second and an adjustment accuracy of ±0.5%FS. The frequency adjustment range of the variable frequency injection pump is 5-80Hz, and the pressure control accuracy is ≤0.1MPa.

[0060] The fuzzy rule base for fuzzy PID control contains 25 control rules. The input variables are pressure deviation e and the rate of change of deviation ec, and the output variable is the proportional coefficient K. p Integration time T i The adjustment amount, the membership function adopts a triangular-trapezoidal hybrid distribution, and the centroid method is used for defuzzification;

[0061] The fuzzy set of pressure deviation e is defined as {negative large (NB), negative small (NS), zero (ZO), positive small (PS), positive large (PB)}, and the universe of discourse ranges [-0.2P0, 0.2P0];

[0062] The fuzzy set of the deviation change rate ec is defined as {rapid decrease (ND), slow decrease (NS), stable (ZE), slow increase (PS), rapid increase (PD)}, and the universe of discourse ranges from [-0.15P0 / min, 0.15P0 / min];

[0063] The triangular distribution parameters of the membership function satisfy the following: the overlap between adjacent fuzzy sets is 15%-25% of the universe width, and the slope of the trapezoidal distribution at the boundary of the universe is 45°±5°.

[0064] The logical expression for the control rule is:

[0065] IFe=NBANDec=NDTHENΔK p =PBANDΔT i =NB

[0066] IFe=ZOANDec=ZETHENΔK p =ZOANDΔT i =ZE

[0067] (The remaining 23 rules are arranged according to the same logical structure);

[0068] The centroid method for fuzzy resolution is specifically calculated as follows:

[0069]

[0070] Among them, u * To ensure accurate output values, μ i (u i ) represents the membership value when the i-th rule is triggered.

[0071] The sensor module, data processing module, and actuator module form a closed-loop control system through an industrial IoT platform. The LSTM neural network updates the weight parameters every 30 minutes based on the latest monitoring data, and the integral time constant of the fuzzy PID control is automatically adjusted according to the injection pressure fluctuation amplitude. The adjustment amplitude satisfies the formula ΔT=K·(ΔP / P0)2, where K∈[0.5,1.2] is the formation permeability correction coefficient, ΔT is the adjustment amount of the integral time constant, ΔP is the pressure fluctuation amplitude in the current monitoring period, and P0 is the initial formation pressure of the target injection layer.

[0072] The methods for determining the formation permeability correction factor K include obtaining the static permeability K through core experiments. core By combining initial pressure test data, the dynamic permeability K can be inverted. dynamic ,Pick

[0073] The LSTM neural network contains three hidden layers with 128, 64, and 32 neurons in each layer, respectively. During training, monitoring data with a time window of 6 hours is used as input features, and the loss function is mean squared error (MSE). The training dataset contains injection conditions data for at least 10 formations with different permeability.

[0074] Input layer: Time window length 6 hours (360 time steps, sampled once every 1 minute), features include pressure, temperature, flow rate, and permeability correction coefficient K, a total of 4-dimensional input.

[0075] Hidden layers: 3 layers (128 / 64 / 32 neurons), each layer has a tanh activation function, and Dropout is added between layers (0.3 ratio for the first 20 epochs).

[0076] Output layer: 1-dimensional (target injected flow Q_pred), with a linear activation function.

[0077] The weight parameters of the LSTM neural network are updated using stochastic gradient descent and satisfy the following conditions:

[0078] The Adam optimizer is used for parameter updates, with a momentum coefficient μ = 0.9–0.99 and a learning rate η = 0.001–0.01, which decays exponentially with each epoch. The decay formula is as follows:

[0079] η t =η0·γ t / T

[0080] Where γ = 0.95-0.99 is the decay coefficient, t is the current epoch number, and T = 200-500 is the total number of training epochs;

[0081] Each time the parameters are updated, the gradient direction vectors of the previous 10 iterations are accumulated, and the momentum optimization term is calculated to satisfy:

[0082]

[0083] Where v t The current momentum vector, The current gradient;

[0084] During training, the batch size is 64-256, and a Dropout ratio of 0.2-0.5 is used in the first 20 epochs to prevent overfitting.

[0085] The training termination condition is that the decrease in the validation set loss function is less than 10 for 10 consecutive epochs. -5 Alternatively, the maximum number of epochs T can be reached, and training can be accelerated using TensorRT on the NVIDIA CUDA computing platform for graph optimization.

[0086] The industrial IoT platform uses the OPCUA protocol to transmit sensor data. An edge computing node is set up at the injection wellhead, and it is equipped with a data cache and an abnormal data filtering algorithm. The filtering algorithm removes abrupt data that exceeds three times the standard deviation.

[0087] The system is suitable for CO2 geological storage scenarios, and the storage capacity is matched in real time. When the caprock pressure is detected to exceed 85% of the fracture pressure threshold, the injection flow rate is automatically triggered in a graded reduction program, with the reduction gradient being 20% ​​of the current flow rate every 5 minutes. At the same time, the tracer detection module of the backup monitoring well is started.

Claims

1. An intelligent adaptive injection flow control system for geological storage, characterized in that, include: Sensor module, data processing module, actuator module, and adaptive algorithm module; The sensor module consists of formation pressure sensors, temperature sensors, and flow sensors distributed in the injection well and surrounding monitoring wells, and is used to collect pressure gradient, temperature distribution, and fluid velocity data of the injection layer in real time. The data processing module includes a central controller and a data storage unit connected in communication with it. The central controller is configured to normalize the monitoring data of the sensor module and extract the dynamic change characteristics of the injected parameters. The adaptive algorithm module is integrated into the central controller and includes a multi-parameter coupled control model based on fuzzy PID control and LSTM neural network. This model predicts the evolution trend of the formation seepage field according to the dynamic change characteristics of the injection parameters and generates control commands that match the formation storage capacity in real time. The actuator module includes an electric regulating valve group and a variable frequency injection pump installed at the injection wellhead. The electric regulating valve group is connected to the central controller through a PID closed-loop control circuit and is used to dynamically adjust the injection flow rate and pressure according to the control command, so that the error between the actual injection curve and the target parameter curve output by the model is less than 5%. The sensor module, data processing module, and actuator module form a closed-loop control system through an industrial IoT platform. The LSTM neural network updates its weight parameters every 30 minutes based on the latest monitoring data, and the integral time constant of the fuzzy PID control is automatically adjusted according to the injection pressure fluctuation amplitude, with the adjustment amplitude satisfying the formula ΔT=K·(ΔP / P0). 2 Where K∈[0.5,1.2] is the formation permeability correction coefficient, ΔT is the adjustment amount of the integral time constant, ΔP is the pressure fluctuation amplitude in the current monitoring period, and P0 is the initial formation pressure of the target injection layer.

2. The intelligent adaptive injection flow control system for geological storage according to claim 1, characterized in that, In the sensor module, the formation pressure sensors in the injection well are arranged at intervals along the wellbore axis above, below and in the middle of the target injection layer, and the monitoring wells are arranged at 30°, 90°, 150°, 210°, 270° and 330° circumferential positions around the injection well, within a range of 50-200 meters from the injection well.

3. The intelligent adaptive injection flow control system for geological storage according to claim 1, characterized in that, The normalization process includes Z-score standardization of the pressure gradient data, calculated as follows: Where μ p σ represents the average pressure during the current monitoring period. p The standard deviation is used; the temperature data are normalized using the maximum-minimum method, calculated as follows: in T min This refers to the historical extreme temperatures of the monitoring well.

4. The intelligent adaptive injection flow control system for geological storage according to claim 1, characterized in that, The fuzzy rule base of the fuzzy PID control contains 25 control rules. The input variables are the pressure deviation e and the rate of change of deviation ec, and the output variable is the proportional coefficient K. p Integration time T i The adjustment amount, the membership function adopts a triangular-trapezoidal hybrid distribution, and the centroid method is used for defuzzification. The fuzzy set of the pressure deviation e is defined as {negative large (NB), negative small (NS), zero (ZO), positive small (PS), positive large (PB)}, and the universe of discourse range is [-0.2P0, 0.2P0]; The fuzzy set of the deviation change rate ec is defined as {rapid decrease (ND), slow decrease (NS), stable (ZE), slow increase (PS), rapid increase (PD)}, and the universe of discourse ranges from [-0.15P0 / min, 0.15P0 / min]; The triangular distribution parameters of the membership function satisfy the following: the overlap of adjacent fuzzy sets is 15%-25% of the universe width, and the slope of the trapezoidal distribution at the boundary of the universe is 45°±5°. The logical expression for the control rule is: IFe=NB AND ec=ND THEN ΔK p =PB AND ΔT i =NB IFe=ZO AND ec=ZE THEN ΔK p =ZO AND ΔT i =ZE (The remaining 23 rules are arranged according to the same logical structure); The centroid method of the defuzzification method is specifically calculated as follows: Among them, u * To ensure accurate output values, μ i (u i ) represents the membership value when the i-th rule is triggered.

5. The intelligent adaptive injection flow control system for geological storage according to claim 1, characterized in that, The LSTM neural network contains three hidden layers with 128, 64, and 32 neurons in each layer, respectively. During training, monitoring data with a time window of 6 hours is used as input features, and the loss function is mean squared error (MSE). The training dataset contains injection condition data for at least 10 formations with different permeability.

6. The intelligent adaptive injection flow control system for geological storage according to claim 1, characterized in that, The electric regulating valve assembly includes a coarse adjustment valve and a fine adjustment valve installed in series. The coarse adjustment valve is a V-type ball valve with a response time of ≤3 seconds and an adjustment accuracy of ±2%FS. The fine adjustment valve is a cage-type regulating valve with a response time of ≤1 second and an adjustment accuracy of ±0.5%FS. The frequency adjustment range of the variable frequency injection pump is 5-80Hz, and the pressure control accuracy is ≤0.1MPa.

7. The intelligent adaptive injection flow control system for geological storage according to claim 1, characterized in that, The industrial IoT platform uses the OPCUA protocol to transmit sensor data, sets up an edge computing node at the injection wellhead, and is equipped with a data cache and an abnormal data filtering algorithm. The filtering algorithm removes abrupt data exceeding three times the standard deviation.

8. The intelligent adaptive injection flow control system for geological storage according to claim 1, characterized in that, The weight parameters of the LSTM neural network are updated using stochastic gradient descent and satisfy the following conditions: The Adam optimizer is used for parameter updates, with a momentum coefficient μ = 0.9–0.99 and a learning rate η = 0.001–0.01, which decays exponentially with each epoch. The decay formula is as follows: or t =η0·γ t / T Where γ = 0.95-0.99 is the decay coefficient, t is the current epoch number, and T = 200-500 is the total number of training epochs; Each time the parameters are updated, the gradient direction vectors of the previous 10 iterations are accumulated, and the momentum optimization term is calculated to satisfy: Where v t The current momentum vector, The current gradient; During training, the batch size is 64-256, and a Dropout ratio of 0.2-0.5 is used in the first 20 epochs to prevent overfitting. The training termination condition is that the decrease in the validation set loss function is less than 10 for 10 consecutive epochs. -5 Alternatively, the maximum number of epochs T can be reached, and training can be accelerated using TensorRT on the NVIDIA CUDA computing platform for graph optimization.

9. The intelligent adaptive injection flow control system for geological storage according to claim 1, characterized in that, The method for determining the formation permeability correction factor K includes obtaining the static permeability K through core experiments. core By combining initial pressure test data, the dynamic permeability K can be inverted. dynamic ,Pick 10. The intelligent adaptive injection flow control system for geological storage according to claim 1, characterized in that, The system is applicable to CO2 geological storage scenarios. The real-time matching of storage capacity includes: when the caprock pressure is detected to exceed 85% of the fracture pressure threshold, an injection flow rate reduction program is automatically triggered, with a reduction gradient of 20% of the current flow rate per 5 minutes, and the tracer detection module of the backup monitoring well is activated at the same time.

Images