Method for intelligently predicting and optimally controlling operation state of oil field sewage settling tank

Abstract

The invention relates to a method for intelligently predicting and optimally controlling the running state of an oilfield sewage settling tank, which comprises the following steps of: acquiring time sequence data by using a modular design reconfigurable settling system fused with multi-modal perception; combining a physical information neural network (PINN) algorithm to construct a mapping relation of working characteristics between an actual settling tank and an experimental settling tank: predicting the performance of the actual settling tank through data of the experimental settling tank; identifying and predicting the operation state of the experimental settling tank, and establishing a multi-parameter fuzzy comprehensive evaluation model; intelligent decision making and automatic control are carried out on the settlement system; a Markov decision process MDP is adopted as an optimization framework of a control decision, and the control decision is optimized in combination with a migration enhanced PPO algorithm, so that automatic adjustment of system performance is realized; and continuous self-optimization for the MDP decision process of the Markov decision process is realized. The problem that the operation state of the settling tank is difficult to monitor in real time and dynamically regulate and control in the traditional oilfield sewage treatment process is solved.

Description

Technical fields:

[0001] This invention relates to oilfield wastewater treatment technology, specifically a method for intelligent prediction and optimized control of the operating status of oilfield wastewater settling tanks. Background technology:

[0002] As most oilfields in my country enter the later stages of development, oil recovery technologies are gradually shifting towards tertiary oil recovery (EOR) technologies (such as polymer flooding and chemical flooding). This has led to significant changes in the composition of produced water, increasing its complexity. Produced water contains polyacrylamide (HPAM), residual surfactants, and other chemical agents, forming stable emulsified oil and colloidal suspensions. Traditional gravity sedimentation separation efficiency has decreased by 30% to 50%. At the same time, water quality standards are becoming more stringent. Reinjected water must meet the requirements of "SY / T 5329-2022 Water Quality Indicators for Injection Water in Clastic Rock Reservoirs," including suspended solids (≤10mg / L), oil content (≤15mg / L), and median particle size (≤2μm). Existing technologies struggle to consistently meet these standards. High dosage of chemicals (such as demulsifiers and flocculants) increases treatment costs by 20% to 40%, and substandard reinjection water can cause formation blockage (permeability decreases by 50% to 70%) and double the injection pressure (from 5 MPa to over 12 MPa), significantly increasing equipment maintenance and oil production energy costs, resulting in both economic and environmental pressures.

[0003] Current oilfield wastewater settling and separation processes mainly rely on static gravity settling tanks, which have very limited efficiency. Traditional settling tanks suffer from dead zones and short-circuiting, uneven heat transfer, and an effective separation volume utilization rate of less than 60%. When influent water quality fluctuates (such as changes in polymer concentration and temperature), there is a lack of real-time control methods, resulting in dynamic response lag, prolonged settling time, and excessive reliance on manual experience to adjust process parameters, making it difficult to match optimal operating conditions under complex water quality. Summary of the Invention:

[0004] The purpose of this invention is to provide a method for intelligent prediction and optimized control of the operating status of settling tanks for oilfield wastewater. This method is used to solve the technical problem that the operating status of settling tanks is difficult to monitor and dynamically control in real time during traditional oilfield wastewater treatment.

[0005] The technical solution adopted by this invention to solve its technical problem is as follows: This intelligent prediction and optimization control method for the operating status of oilfield wastewater settling tanks includes the following steps:

[0006] Step 1: A modular design for reconfigurable settling system integrating multimodal sensing is used to collect time-series data, including flow rate, particle concentration, and flow field vorticity. The settling system integrates a PIV particle image velocimeter, a distributed fiber optic sensor, a conductivity-turbidity multi-parameter sensor array, and temperature staining technology to construct a multi-dimensional physical field dynamic monitoring network, thereby realizing three-dimensional visualization reconstruction of the flow characteristics inside the experimental settling tank and quantitative analysis of particle transport trajectories.

[0007] Step 2: Combining the Physical Information Neural Network (PINN) algorithm, construct a mapping relationship between the working characteristics of the actual settling tank and the experimental settling tank: predict the performance of the actual settling tank using data from the experimental settling tank.

[0008] Step 3: Identification and prediction of the operating status of the experimental settling tank; Establish a multi-parameter fuzzy comprehensive evaluation model, and use the multi-parameter fuzzy comprehensive evaluation model to evaluate the operating status of the experimental settling tank in real time based on the current operating data of the experimental settling tank, and establish evaluation indicators; then combine machine learning algorithms to predict the possible situations that may occur when the settling system continues to operate under the current state, as well as the changes in the future operating status over time.

[0009] Step four, the settling system operates with intelligent decision-making and automatic control; by accurately monitoring dynamic variables directly related to the settling process, a Markov decision process (MDP) is used as the optimization framework for control decisions, and a migration-enhanced PPO algorithm is combined to optimize control decisions, thereby achieving automatic adjustment of system performance. The dynamic variables include oil layer thickness, turbidity, temperature, pH value, and suspended solids concentration.

[0010] Step 5: Decision verification and dynamic correction closed-loop optimization control of the Markov Decision Process (MDP). Through deep coupling of GBDT and reinforcement learning, continuous self-optimization of the MDP decision process is achieved.

[0011] The multi-parameter fuzzy comprehensive evaluation model in step three of the above scheme is as follows:

[0012]

[0013] In the formula: ρ w ρ is the density of water. o ρ is the density of the oil; g is the acceleration due to gravity; α is the kinetic coefficient; Q is the liquid flow rate; k DAF A is the efficiency coefficient of the air flotation unit. b v is the total surface area flux of the bubbles; b d represents the rising velocity of the bubble. 32 C represents the average bubble diameter of Sauter; oil Oil content; C oil,crit E is the critical emulsification concentration of the oil phase. aR is the apparent activation energy; T is the ideal gas constant; T is the heating coil temperature; T ref Reference temperature; f coil (x c ,y c ,z c ) is the spatial location (x c ,y c ,z c The influence function of the spatial position of the heating coil at point P; i Let P be the spatial coordinates of the water inlet i=1, the heating coil i=2, and the flotation unit i=3; opt For the optimal layout of each component; L char The characteristic length is taken as 20% of the tank diameter; w i For weighting coefficients; μ(T) is the fluid viscosity; A settlle For effective settlement area; Q max C is the critical flow rate; TSS β represents the turbidity of the wastewater; β is the oil phase correction coefficient.

[0014] The modular design of the above scheme that integrates multimodal perception and can reconfigure the settlement system includes an experimental module, a data acquisition module, and a dataset parsing and control module. The modules work together to form a complete experimental feedback loop.

[0015] The experimental module includes an experimental settling tank, a power unit, a heating unit, and a flotation unit. The experimental settling tank is equipped with a central column tube and a central reaction cylinder. The flotation unit is a dissolved air pump, which uses a multiphase dissolved air pump to draw in air and wastewater together. The impeller of the dissolved air pump cuts the wastewater and air into fine foam, making them fully mixed. At the same time, the high pressure generated by the high-speed rotation of the impeller fully dissolves the drawn-in gas into the water. Together, they form stable dissolved air water. When this dissolved air water is released at normal pressure through a pressure reducing valve, it produces tiny bubbles with a diameter between 20 and 50 μm.

[0016] The data acquisition module integrates a PIV particle image velocimetry device, a CCD camera array, a multispectral turbidimeter, and a distributed sensor group. The PIV particle image velocimetry device and the CCD camera are synchronously triggered by laser pulses to achieve submicron-level dynamic analysis of the oil droplet particle size distribution. The multispectral turbidimeter uses multi-wavelength scattered light intensity detection technology combined with an embedded temperature and pressure compensation algorithm to achieve online calibration of turbidity changes during the settling process. The distributed sensor group includes a micro differential pressure transmitter and an array of temperature probes, forming a multi-point physical parameter monitoring network for the flow field inside the settling tank.

[0017] The data analysis and control module is equipped with a motion feature extraction unit, a multiphysics coupling analysis unit, and a sedimentation efficiency evaluation unit. The motion feature extraction unit uses an improved PTV particle tracking algorithm to perform Gaussian fitting and noise reduction on continuous frame images, and calculates the oil droplet velocity field, acceleration field, and fractal dimension of the motion trajectory. The multiphysics coupling analysis unit analyzes the coalescence and breakup dynamics of oil droplets by establishing a correlation matrix model of turbidity-particle size-flow velocity. The sedimentation efficiency evaluation unit constructs a particle swarm sedimentation prediction model based on the Fokker-Planck equation and outputs key indicators such as sedimentation rate and interface clarity.

[0018] Step two in the above scheme is specifically as follows:

[0019] Step 1: Construct a physical information neural network architecture. Input layer: Parameters obtained from the experimental settling tank operation; Output layer: Actual settling tank performance parameters, including particle settling efficiency and turbidity removal rate.

[0020] Define the loss function:

[0021]

[0022] In the formula: For data matching, a small amount of actual settling tank calibration data Y is used in the calculation. real ;

[0023]

[0024] For physical constraints;

[0025] —Similarity criterion constraints;

[0026]

[0027] The allocation of each item is dynamically adjusted through the validation set;

[0028] Step 2: Time series data preprocessing and training of the PINN neural network model for physical information;

[0029] The time series data is augmented by making it dimensionless and interpolating it to match the actual tank time axis. The experimental tank settling time series data is used to train the basic PINN to establish a high-precision mapping relationship between the experimental settling tank and the actual settling tank.

[0030] Step 3: Dynamically update and continuously fine-tune the model with new data to optimize the PINN physical information neural network model.

[0031] The specific method for step three in the above scheme is as follows:

[0032] Step 1: Obtain data, including water flow rate, flotation intensity, heating temperature, wastewater turbidity, and oil content;

[0033] Step 2: Establish a multi-parameter fuzzy comprehensive evaluation model;

[0034] A sedimentation efficiency η model is constructed based on the mass conservation and kinetic equations:

[0035]

[0036] Subfunction expansion:

[0037] (1) Air flotation enhancement factor Γ:

[0038]

[0039] In the formula: k DAF The efficiency coefficient of the non-air flotation unit is related to the spatial location of the flotation unit; A b The total surface area flux of the bubbles is m. 2 / s;v b d represents the rising velocity of the bubble. 32 C represents the average bubble diameter of Sauter; oil,crit This is the critical emulsification concentration of the oil phase;

[0040] (2) Thermal effect correction term T eff :

[0041]

[0042] In the formula: E a R is the apparent activation energy; T is the ideal gas constant; ref Reference temperature; f coil The function representing the influence of the spatial position of the heating coil is positively correlated with the uniformity of the thermal field.

[0043] (3) Spatial Coordination Factor ξ space :

[0044]

[0045] In the formula: P i P represents the spatial coordinates of the water inlet (i=1), heating coil (i=2), and flotation unit (i=3); opt For the optimal layout of each component; L char The characteristic length is taken as 20% of the tank diameter; w i These are the weighting coefficients;

[0046] (4) Foundation settlement rate η0

[0047]

[0048] In the formula: μ(T) is the fluid viscosity; A settle For effective settlement area; Q max The critical flow rate;

[0049] Summarized as follows:

[0050]

[0051] Using the data obtained in step 1, a multi-parameter fuzzy comprehensive evaluation model is trained. The multi-parameter fuzzy comprehensive evaluation model is used to input the above-mentioned water distribution flow rate, air flotation intensity, heating temperature, sewage turbidity, oil content and dynamic gradient characteristics. The target variables are the oil content of the dewatering water, the water content of the dewatering water and the settling time. After continuous iteration, the optimal experimental parameters are comprehensively evaluated.

[0052] Step 4: Data dimensionality reduction, determine the correlation between each time series data and the efficiency of the settling tank, and determine the influence weight of each time series data.

[0053] Step 5, State Identification and Prediction.

[0054] Beneficial effects:

[0055] (I) This invention innovatively combines similarity theory principles with physical information neural network technology to construct a scaled-down experimental system for wastewater settling tanks. First, similarity theory guides the design of experimental models, significantly reducing experimental costs and implementation difficulties while ensuring flow field similarity. Second, physical information neural networks are used to fuse measured data with fluid dynamics equations, breaking through the limitations of traditional empirical formulas and achieving high-precision mapping of multi-physics characteristics between the experimental tank and the prototype tank. Third, a hybrid modeling method combining data-driven and mechanism model co-optimization is adopted, which can effectively overcome the poor generalization of pure data models and significantly improve the adaptability to complex working conditions. By combining small-scale experiments with intelligent algorithm extrapolation, the high cost of full-scale experiments can be avoided, and key parameters such as flow distribution and settling efficiency can be obtained simultaneously, providing a reliable basis for optimizing actual tank structure parameters and operation control strategies.

[0056] (II) This invention innovatively integrates mechanistic models and machine learning algorithms to construct an optimized control system driven by physical laws and coordinated with intelligent decision-making: A mechanistic model based on multiphase flow dynamics equations provides interpretable physical constraints for the system, ensuring that the core settling process conforms to fluid mechanics principles; a multi-parameter fuzzy comprehensive evaluation model uses membership functions and fuzzy rule bases to handle the uncertainty and nonlinear correlation between process parameters and settling conditions, identifying the state of experimental equipment and determining the optimal operating parameters; a Markov decision process model generates dynamic decisions through state transition probability matrices to optimize equipment operating parameters, maximizing global benefits under time-varying conditions; a transfer-enhanced deep learning algorithm optimizes the generated decisions; and finally, a gradient-boosting decision tree algorithm is used to correct the decision model based on the deviation between theoretical efficiency improvement and actual conditions. The synergy between these algorithms can improve settling efficiency in oilfield wastewater settling tanks and promote unmanned control. Relying on virtual sensors and online parameter correction mechanisms, it effectively overcomes the technical bottlenecks of model mismatch and decision lag in traditional methods, possessing both physical interpretability and dynamic optimization capabilities.

[0057] (III) This invention focuses on the coordinated optimization design of the structural parameters and process parameters of settling tanks. It integrates intelligent sensor networks and automatic actuators by constructing a modular experimental device, establishes a working state prediction model by combining machine learning algorithms, and develops an adaptive control strategy to achieve dynamic optimization of the settling process, including computational fluid dynamics simulation, intelligent sensor data fusion, predictive control model construction, and adaptive regulation method based on real-time monitoring.

[0058] (IV) This invention focuses on two core directions: improving the efficiency of sedimentation and separation of polymer-containing wastewater and intelligent control. It innovates traditional process modes by constructing an intelligent control technology that combines machine learning algorithms through multi-physics field coupling experiments. The development of an adaptive control algorithm based on machine learning to optimize sedimentation tank operating parameters in real time will significantly reduce the overall cost of oilfield wastewater treatment, while ensuring the stability of reinjected water quality and extending the formation water injection development cycle.

[0059] (V) This invention solves the technical challenge of real-time monitoring and dynamic control of settling tank operation status (normal / abnormal, high-efficiency / low-efficiency) by intelligently identifying the operating status of settling tanks in real time, performing multi-parameter diagnostic analysis on abnormal and low-efficiency conditions, and automatically generating optimized decision-making schemes. This provides an innovative solution for achieving intelligent and unmanned operation and maintenance management of oilfield water treatment equipment. Through the deep integration of deep learning algorithms and IoT technology, the system achieves accurate judgment of operating status, early warning of faults, and autonomous optimization of operating parameters, significantly improving the operating efficiency and management level of oilfield wastewater treatment systems.

[0060] (vi) This invention constructs an experimental device that can simulate the oil-water separation process in an oilfield wastewater settling tank and visualize its flow field and temperature field. This provides research conditions for further studying the heat transfer and flow characteristics of the settling separation process, forming a control method for the flotation process, and further optimizing its heater structure, water inlet structure, and process parameters.

[0061] (vii) This invention can visualize the flow and temperature of the settling tank during uninterrupted and intermittent liquid preparation processes, obtain detailed three-dimensional temperature and velocity field data, provide rich and accurate test data, and precisely control experimental conditions. It is conducive to in-depth research on the temperature and flow coupling characteristics of the oilfield wastewater settling process, and can further calculate parameters such as the coordination angle and uniformity, providing experimental conditions for studying the flow and heat transfer coupling characteristics of the oilfield wastewater settling tank under different processes.

[0062] (viii) The present invention has a particle image velocimetry (PIV) device, which can obtain flow and temperature distribution data of the medium in the experimental settling tank by photographing tracer particles and temperature dyes premixed in the experimental medium. The acquired data is complete, detailed and synchronized with the temperature field and velocity field. By combining the temperature data of the temperature sensor with the temperature data of the temperature dye, more accurate three-dimensional temperature field data can be obtained.

[0063] (ix) The experimental settling tank and its corresponding pipelines of this invention are equipped with multiple valves, allowing for free switching of processes and facilitating the simulation of various workflows, such as closing the dispensing port and opening the collecting port to lower the liquid level, closing the collecting port and opening the dispensing port to raise the liquid level, and adjusting the opening degree of the dispensing and collecting ports. Furthermore, the pipeline walls are covered with insulation material, and the temperature of the fluid entering the experimental settling tank can be more accurately and effectively controlled by adjusting the heating device. The centrifugal pump speed control in the pipeline system, in conjunction with the flow sensor, allows for precise flow control.

[0064] (x) The present invention simultaneously measures the three-dimensional temperature field and velocity field. The measurement data can be used to analyze the coupling characteristics of the three-dimensional temperature and velocity fields, such as, but not limited to, calculating the coordination angle between the temperature field and the velocity field, the uniformity of the temperature field, the energy utilization rate of the device, the average flow velocity, and the average temperature.

[0065] (XI) This invention achieves precise reconfiguration of multiple modules in three-dimensional space through a high-precision quick-assembly interface and an intelligent positioning system, restoring the actual equipment structure. By analyzing the operating parameters of the experimental device, the actual working characteristics of the settling tank are mapped, forming a closed-loop R&D system that supports orthogonal experimental design and real-time feedback optimization. Simultaneously, other emerging technologies can be introduced to assess their efficiency improvement capabilities in settling and utilize machine learning algorithms to optimize the structural and operating parameters of these emerging technologies. This creates an economical research platform with extremely high module reusability, while reserving multidisciplinary expansion interfaces, providing an innovative solution for emerging experimental technologies for wastewater settling tanks that combines precision, flexibility, and scalability. Attached image description:

[0066] Figure 1 This is a schematic diagram of the experimental settling tank;

[0067] Figure 2 A schematic diagram of a modular design for a reconfigurable settlement system that integrates multimodal sensing;

[0068] Figure 3 A schematic diagram illustrating the implementation process of a neural network algorithm for mapping physical information from experimental settling tanks to actual settling tanks;

[0069] Figure 4 A schematic diagram of the principal component analysis algorithm for dimensionality reduction of experimental data from a settling tank.

[0070] Figure 5 A schematic diagram illustrating the implementation process of a multi-parameter fuzzy comprehensive evaluation method for identifying and predicting equipment operating status.

[0071] Figure 6 A schematic diagram illustrating the implementation process of the Markov decision process algorithm;

[0072] Figure 7 A schematic diagram illustrating the implementation process of the migration-enhanced PPO algorithm for selecting the best decision-making solution;

[0073] Figure 8 A schematic diagram illustrating the implementation process of the gradient boosting decision tree algorithm for backtracking and validating decision schemes;

[0074] Figure 9 This is a technical roadmap for the present invention.

[0075] In the diagram: 1-Liquid mixing pipe; 2-Oil collecting tank; 3-Liquid collecting pipe; 4-Liquid outlet pipe; 5-Liquid inlet pipe; 6-Heating coil; 7-Gas distribution pipe; 8-Tank wall; 9-Tank top; 10-Tank bottom; 11-Central reaction cylinder; 12-Central column tube; 13-Experimental settling tank; 14-Data acquisition unit; 15-Water bath heating box; 16-Centrifugal pump; 17-Power supply; 18-Cylindrical lens; 19-Pulsed laser; 20-Synchronizer; 21-High-speed CCD camera; 22-Square cavity; 23-Sensor. Detailed implementation method:

[0076] The present invention will be further described below with reference to the accompanying drawings:

[0077] Example 1:

[0078] See Figures 1-9 This intelligent prediction and optimization control method for the operating status of oilfield wastewater settling tanks includes the following steps:

[0079] Step 1: Utilize modular design to reconfigure the settling system by integrating multimodal sensing to collect time-series data, including flow rate, particle concentration, and flow field eddy current.

[0080] Step two: Combine the physical information neural network PINN algorithm to construct a mapping relationship between the working characteristics of the actual settling tank and the experimental settling tank: predict the performance of the actual settling tank using the experimental settling tank data.

[0081] Step 3: Identification and prediction of the operating status of the experimental settling tank; Establish a multi-parameter fuzzy comprehensive evaluation model, and use the multi-parameter fuzzy comprehensive evaluation model to evaluate the operating status of the experimental settling tank in real time based on the current operating data of the experimental settling tank, and establish evaluation indicators; then combine machine learning algorithms to predict the possible situations that may occur when the settling system continues to operate under the current state, as well as the changes in the future operating status over time.

[0082] Step four: The settling system operates with intelligent decision-making and automatic control. By accurately monitoring dynamic variables directly related to the settling process, a Markov decision process (MDP) is used as the optimization framework for control decisions. Combined with a migration-enhanced procedural optimization (PPO) algorithm, the system performance is automatically adjusted. The dynamic variables include oil layer thickness, turbidity, temperature, pH value, and suspended solids concentration.

[0083] Step 5: Decision verification and dynamic correction of the decision model through closed-loop optimization control. Through deep coupling of GBDT and reinforcement learning, continuous self-optimization of the decision-making process of the MDP model is achieved.

[0084] The modular design of the reconfigurable settlement system that integrates multimodal perception in this invention includes an experimental module, a data acquisition module, and a dataset parsing and control module. These modules work together to form a complete experimental feedback loop.

[0085] (1) The experimental module includes a basic settling tank body, a power unit, a heating unit, and a flotation unit. The basic settling tank (i.e., the experimental settling tank) body includes a settling tank wall 8, a tank top 9, a tank bottom 10, an inlet pipe 5, an outlet pipe 4, an oil outlet pipe, a cloth liner, a central column pipe 12, and a central reaction cylinder 11. The central column pipe 12 is connected to a collecting pipe 3, and the central reaction cylinder 11 is connected to a dispensing pipe 1. The oil-water mixture enters the central reaction cylinder 11 from the inlet pipe 5. As the liquid level in the central reaction cylinder 11 rises, it reaches the dispensing pipe 1, and the oil-water mixture enters the experimental settling tank from the dispensing pipe 1. The liquid in the experimental settling tank flows out from the collecting pipe 3 and enters the central column pipe 12. The liquid level in the central column pipe 12 rises until it reaches the height of the outlet pipe 4, and the liquid is discharged from the outlet pipe 4. The main equipment of the power unit is a centrifugal pump, which is used to provide power for dispensing, collecting, and collecting oil. The heating unit is a water bath device connected to the heating coil inside the experimental settling tank 13, making it a constant-temperature heat source. The flotation unit's mechanical equipment is a dissolved air pump, which uses a multiphase dissolved air pump to draw in air and wastewater together. The impeller of the dissolved air pump cuts the wastewater and air into fine foam, ensuring thorough mixing. At the same time, the high pressure generated by the high-speed rotation of the impeller can also fully dissolve the drawn-in gas into the water, working together to form stable dissolved air water. When this dissolved air water is released at atmospheric pressure through a pressure reducing valve, it produces tiny bubbles with a diameter between 20 and 50 μm. The dissolved air pump is located outside the experimental settling tank 1, and the gas distribution pipe 7 is installed inside the experimental settling tank 1.

[0086] (2) The data acquisition module integrates a PIV particle image velocimetry device, a high-speed CCD camera array, a multispectral turbidimeter, and a distributed sensor group. The PIV particle image velocimetry device and the high-speed CCD camera 21 are synchronously triggered by laser pulses to achieve sub-micron level dynamic analysis of the oil droplet particle size distribution (range 0.1-500μm). The turbidimeter uses multi-wavelength scattered light intensity detection technology (working wavelength range 400-900nm), combined with an embedded temperature and pressure compensation algorithm, to achieve online calibration of turbidity changes during sedimentation. The distributed sensor group includes a micro-differential pressure transmitter (range 0-10kPa, accuracy ±0.1%FS) and an array of temperature probes, forming a multi-point physical parameter monitoring network for the flow field inside the sedimentation tank. The PIV particle image velocimetry device includes a high-speed CCD camera 21, a square cavity 22, a cylindrical lens 18, a pulsed laser 19, and a synchronizer 20.

[0087] (3) The data analysis and control module is built on a high-performance computing workstation and is equipped with: 1) a motion feature extraction unit, which uses an improved PTV particle tracking algorithm to perform Gaussian fitting and noise reduction on continuous frame images and calculates the oil droplet velocity field, acceleration field and fractal dimension of motion trajectory; 2) a multi-physics coupling analysis unit, which analyzes the dynamic characteristics of oil droplet aggregation and breakup by establishing a correlation matrix model of turbidity-particle size-flow velocity; 3) a sedimentation efficiency evaluation unit, which constructs a particle group sedimentation prediction model based on the Fokker-Planck equation and outputs key indicators such as sedimentation rate and interface clarity.

[0088] This invention is based on similarity theory and physical information neural networks, and designs a scaled-down experimental model according to an actual sewage settling tank:

[0089] (1) The experimental setup was designed strictly with reference to the structural parameters and operating conditions of the actual settling tank. By establishing precise geometric and dynamic similarity criteria, the experimental model was ensured to maintain a high degree of consistency with the actual settling tank in terms of fluid dynamic characteristics. The experimental model was established by scaling down the actual settling tank according to its geometric characteristics, with a scaling factor of 1:12 and a geometric similarity error of less than 1%, to ensure that the experimental tank could accurately map the working characteristics of the actual equipment.

[0090] (2) By analyzing the physical processes under actual working conditions, the Froude number (Fr), Reynolds number (Re), or Stokes number (Stk) are selected as the dominant similarity criteria, and the fluid viscosity, flow velocity, or particle size of the experimental model are adjusted to meet the similarity conditions. Sensor arrays, turbidimeters, and PIV particle image velocimetry devices are arranged in the experimental model to collect time-series data including flow rate, particle concentration, temperature, and flow field vorticity;

[0091] (3) By combining the Physical Information Neural Network (PINN) algorithm, a mapping relationship between the working characteristics of actual and experimental settling tanks is constructed. This integrates physical laws with machine learning, incorporating fundamental physical laws such as fluid mechanics and particle settling during training. This ensures that the network not only fits experimental data but also guarantees that the model output conforms to known physical constraints. This reduces the need for large amounts of labeled data and improves the model's prediction accuracy and generalization ability.

[0092] This invention relates to a method for exploring the optimal equipment structure based on a modular, detachable settling tank experimental device.

[0093] First, a basic settling tank body integrating multimodal monitoring functions is constructed, with an optical sensor observation window at the top and a detachable mud and sand collection bin at the bottom. Second, adjustable structural parameter modules are installed in key parts of the tank body, covering adjustable collection / dispensing systems (number of interfaces, height distribution, geometry) to control inlet flow field distribution and fluid dynamics characteristics; adjustable heating units (number of coils, pipe diameter, spatial layout) to optimize the internal temperature field gradient; and adjustable air flotation devices (gas release head density, gas distribution pipe diameter parameters, installation height) to optimize microbubble distribution. The system includes adjustable components for oil droplet capture efficiency and an adjustable inlet / outlet piping system. A data acquisition sensor network is deployed concurrently during the experiment. Comparative experimental groups with different combinations of structural parameters are set up. Standard wastewater samples (with controllable oil content, temperature, and flow rate) are used for dynamic testing. CFD numerical simulation is used for offline verification. High-speed camera technology is used to capture oil droplet aggregation behavior. The system records core performance indicators such as oil phase flotation rate, effluent oil phase concentration, and sludge deposition morphology. Finally, multi-dimensional parameter correlation analysis determines the optimal collaborative configuration scheme for each module.

[0094] The modular design of the reconfigurable settling system, which integrates multimodal sensing, incorporates a PIV particle image velocimetry device, a distributed fiber optic sensor, a conductivity-turbidity multi-parameter sensor array, and temperature staining technology to construct a multi-dimensional physical field dynamic monitoring network. This enables three-dimensional visualization reconstruction of the flow characteristics inside the settling tank and quantitative analysis of particulate matter transport trajectories.

[0095] Under the action of the power unit, oily wastewater is pumped into the experimental settling tank by a centrifugal pump. The tank body is made of highly transparent plexiglass, maintaining the size proportions of settling tanks in oilfields. The experimental settling tank has an arched top with multiple test holes arranged on it. Test tubes extend into the experimental settling tank 1 through the test holes, and the part extending out of the arched top is connected to the data acquisition and control system via wires. Multiple temperature sensors 23 are installed on the test tubes. The upper and lower parts of the tank body are respectively equipped with a settling tank dispensing port and a settling tank collecting port. The settling tank inlet, settling tank outlet, and heating coil 6 inlet are all connected to corresponding branch pipelines. Each branch pipeline is equipped with valves, pressure sensors, and flow sensors, forming an experimental pipeline. The other end of the experimental pipeline is connected to a set of water bath heating boxes 15. A set of centrifugal pumps 16 is installed between the water bath heating boxes 15 and the experimental settling tank 13. By switching valves, different branch pipelines are interconnected, the experimental process is switched, and different experimental conditions are constructed. The experimental settling tank chamber is designed to simulate the actual flow environment of oily wastewater, providing a venue for studying the behavior of oil droplets under different conditions.

[0096] This invention focuses on key similarity parameters such as Reynolds number and Froude number to achieve accurate simulation of flow field characteristics. The sedimentation system is equipped with advanced measurement devices and data acquisition systems, enabling precise measurement of key parameters such as velocity field and concentration distribution.

[0097] (1) Multi-dimensional parameter system construction and correlation modeling: The process parameter set (inlet and outlet flow rate, air flotation intensity, heating temperature) and the environmental parameter set (oil content, temperature, pH value, suspended solids concentration) are used as input variables, and the quality of the stripping medium (oil content / oil content of stripped water) and sedimentation efficiency (treatment time) are used as decision variables to establish a mathematical model of parameter interaction.

[0098] (2) Data feature space reconstruction: Principal component analysis is used to decouple the features of multidimensional monitoring data. Principal component load spectrum is extracted by covariance matrix decomposition. Key sensitive parameters are screened and weight coefficients are quantified based on variance contribution rate to reveal the nonlinear mapping relationship between process-environment parameters and sedimentation efficiency.

[0099] (3) Intelligent identification of operating status, construction of fuzzy comprehensive evaluation model, integration of real-time monitoring data stream and historical database, use of membership function to quantify the influence intensity of parameters, realize dynamic diagnosis and risk prediction of operating conditions, conduct multi-step trend inference through recurrent neural network, and establish a comprehensive evaluation index system including stability index, energy efficiency index and other dimensions.

[0100] (4) Adaptive optimization control: Based on reinforcement learning algorithm, a parameter optimization mechanism is constructed, process control parameters and environmental compensation coefficients are iteratively updated, multi-objective collaborative optimization is analyzed and realized, forming a closed-loop control architecture of "monitoring-evaluation-optimization", and finally outputting the optimal parameter combination scheme that meets decision constraints.

[0101] (5) Mapping the working characteristics of actual production equipment: By analyzing the experimental parameters of the experimental settling tank and combining them with the aforementioned Physical Information Neural Network (PINN) algorithm, the working characteristics of actual oilfield equipment are inverted. By combining experimental data with the constraints of the physical model, the deep-seated relationship between experimental parameters and the working characteristics of actual equipment is uncovered. Inverting the working characteristics of actual oilfield equipment can provide theoretical support for equipment optimization design, operation strategy adjustment, and performance improvement.

[0102] This invention achieves precise control and autonomous optimization of the settling process by constructing a three-in-one intelligent control architecture of "physical device-sensing system-intelligent algorithm".

[0103] (1) Structural parameter optimization unit: The topology of the settling tank component layout is optimized by combining a multi-parameter comprehensive fuzzy evaluation algorithm with orthogonal experimental design, and the simulation results of three-dimensional flow field generated by CFD are compared.

[0104] (2) Process parameter adjustment unit: Control components are installed at each component of the settling tank to dynamically calculate the inlet and outlet flow rates (adjustment accuracy ±0.5m) based on the MDP model and migration-enhanced PPO algorithm.3 The optimal decision for the output is determined by factors such as the hourly rate, temperature gradient (control resolution ±0.1℃), and flocculant dosage (metering error <0.2%).

[0105] (3) Human-computer interaction unit: supports three-dimensional visualization reconstruction of the settling process and early warning of abnormal working conditions, and provides self-tuning PID control curves for process parameters;

[0106] (4) Backtracking verification unit: used to verify the effectiveness of the implemented decision. By dynamically monitoring and analyzing the key parameters of the settling tank, the deviation between the theoretical settling effect and the actual effect is evaluated, and the machine learning model is iteratively optimized accordingly to improve the accuracy and practicality of the system decision.

[0107] Example 2:

[0108] In this embodiment, step two specifically involves:

[0109] Data from the experimental tank (input X) exp Predict actual tank performance (output Y) real And ensure that the prediction conforms to the laws of physics.

[0110] Step 1: Construct a physical information neural network architecture

[0111] Input layer: Parameters obtained from the experimental settling tank operation; Output layer: Actual settling tank performance parameters, such as particle settling efficiency and turbidity removal rate. Define loss function:

[0112]

[0113] In the formula: —Data matching item, a small amount of actual tank calibration data Y is used in the calculation. real ;

[0114]

[0115] —Physical constraints;

[0116] —Similarity criterion constraints;

[0117]

[0118] The allocation of each item is dynamically adjusted through the validation set.

[0119] Step 2, Data Preprocessing and Model Training

[0120] The experimental data was augmented by making the experimental parameters dimensionless and interpolating the experimental data to match the actual tank time axis. The basic PINN was trained using only the experimental tank data to learn Φ:X.exp →Y exp Add 5% to 10% actual tank monitoring data, freeze the first 3 layers of the network, and only update the weights of the last 2 layers to fine-tune the model and adjust the network parameters.

[0121] After training, model validation and uncertainty quantification are performed using spatiotemporal cross-validation to ensure the training / test sets include different traffic phases (startup, steady state, shutdown) and to check the physical feasibility of the experimental equipment. Monte Carlo Dropout is employed (20% Dropout is retained during training); confidence intervals are added to the predicted output.

[0122]

[0123] Step 3, Dynamically update the model

[0124] The model is fine-tuned with new data every 24 hours (incremental learning, learning rate = 1e-6). If the KL divergence detects a data drift of >5%, retraining is triggered.

[0125] Through the above steps, the PINN algorithm not only establishes a high-precision mapping relationship between the experimental tank and the actual tank, but also ensures engineering interpretability through physical constraints, significantly reducing the dependence of traditional pure data-driven methods on a large amount of actual data.

[0126] Example 3:

[0127] In this embodiment, step three specifically involves:

[0128] This method optimizes the control of a settlement system based on multi-factor orthogonal experiments and dynamic disturbance coupling analysis. It intelligently analyzes and identifies the operating characteristics of experimental equipment based on real-time operating parameters and predicts subsequent operating characteristics. Specifically, it involves the technical implementation process of experimental design and data acquisition combined with a multi-parameter fuzzy comprehensive evaluation model. The method achieves accurate optimization and robustness verification of the system's operating parameters by constructing a multi-parameter coupling matrix and an extreme condition simulation test system. The specific implementation process is as follows:

[0129] Step 1, Parameter Definition and Mechanism of Action

[0130] Clarify the physical meaning of the parameters and define the mechanism by which each parameter affects the settlement process:

[0131] Water flow rate (Q): determines the residence time of wastewater in the tank, affecting the efficiency of oil droplet aggregation and separation;

[0132]

[0133] Air flotation intensity (G): Oil droplets are adsorbed by microbubbles, accelerating their rise. The equation is modified according to Stokes' law.

[0134]

[0135] Heating temperature (T): Reduces wastewater viscosity and promotes oil-water separation;

[0136] μ=μ0e -bT

[0137] Wastewater turbidity (C) TSS ): Reflects the concentration of suspended solids, affecting light scattering and sedimentation resistance;

[0138] Oil content (C) oil ): Determines the collision probability and coalescence efficiency of oil droplets (Smoluchowski equation).

[0139] Step 2, Experimental Design and Data Acquisition

[0140] After the experimental settling tank equipment was set up and all data acquisition devices were connected to the experimental instruments, the orthogonal experimental design was carried out, using L25(5 6 Orthogonal arrays, for example, selecting water distribution flow rates (1 / 2 / 3 / 4 / 5m³). 3 Five levels of operating parameters, including flow rate (L / min), flotation intensity (0.5 / 1.0 / 1.5 / 2.0 / 2.5 L / min), and heating temperature (60 / 70 / 80 / 90 / 100℃), were tested using an L25 orthogonal array. Dynamic disturbance experiments were then conducted, including extreme conditions such as a step change in oil content (from 5% to 10%) and fluctuations in water flow rate (±30%), to verify the system's robustness. Using data acquisition systems such as PIV, the optimal parameters among the five levels were compared and analyzed.

[0141] Step 3, Build the model

[0142] A sedimentation efficiency (η) model is constructed based on mass conservation and kinetic equations:

[0143]

[0144] Subfunction expansion:

[0145] (1) Air flotation enhancement factor Γ:

[0146]

[0147] In the formula: k DAF —Efficiency coefficient of the air flotation unit (related to the spatial location of the flotation unit); A b —Total surface area flux of bubbles, m 2 / s;v b — Bubble rising velocity, m / s; d 32 —Sauter's average bubble diameter, mm;

[0148] C oil,crit —Critical emulsification concentration of the oil phase, mg / L.

[0149] (2) Thermal effect correction term T eff :

[0150]

[0151] In the formula: E a — Apparent activation energy, J / mol; R — Ideal gas constant; T ref —Reference temperature, °C; f coil —Influence function of heating coil spatial location (positively correlated with thermal field uniformity)

[0152] (3) Spatial Coordination Factor ξ space :

[0153]

[0154] In the formula: P i — Spatial coordinates of the water inlet (i=1), heating coil (i=2), and flotation unit (i=3); P opt —Optimal layout positions for each component; L char —Characteristic length, taken as 20% of the tank diameter; w i —Weighting coefficient.

[0155] (4) Foundation settlement rate η0

[0156]

[0157] Where: μ(T) — fluid viscosity, Pa·s; A settle —Effective settlement area, m 2 Q max — Critical flow rate; exceeding this value will result in a short flow.

[0158] Summarized as follows:

[0159]

[0160] In the formula: ρ w —Density of water, kg / m³ 3 ;ρ o —Density of oil, kg / m³ 3 g—acceleration due to gravity, m / s² 2 α—kinetic coefficient, related to particle-fluid interaction and collision efficiency; Q—liquid flow rate, m³ 3 / s;k DAF —Efficiency coefficient of the air flotation unit (related to the spatial location of the flotation unit); Ab —Total surface area flux of bubbles, m 2 / s;v b — Bubble rising velocity, m / s; d 32 —Sauter's average bubble diameter, mm; C oil —Oil content; C oil,crit —Critical emulsification concentration of oil phase, mg / L; E a — Apparent activation energy, J / mol; R — Ideal gas constant; T — Heating coil temperature, °C; T ref —Reference temperature, °C; f coil (x c ,y c ,z c )——Spatial location is (x c ,y c ,z c The influence function of the spatial position of the heating coil at point P (positively correlated with thermal field uniformity); i — Spatial coordinates of the water inlet (i=1), heating coil (i=2), and flotation unit (i=3); P opt —Optimal layout positions for each component; L char —Characteristic length, taken as 20% of the tank diameter, in meters (m); w i —Weighting coefficient; μ(T) —Fluid viscosity, Pa·s; A settlle —Effective settlement area, m 2 Q max —Critical flow rate; exceeding this value will cause a short circuit, m 3 / s;C TSS — Wastewater turbidity; β — Oil phase correction coefficient, reflecting the influence of oil phase on sedimentation efficiency.

[0161] A multi-parameter fuzzy comprehensive evaluation model was used to train the model with experimental data to quantify nonlinear coupling effects (such as the nonlinear improvement in air flotation efficiency at high temperatures). 100 sets of experimental data were collected (75 sets under normal operating conditions and 25 sets under disturbed operating conditions), and divided into training, validation, and test sets in a 7:2:1 ratio. Using the multi-parameter fuzzy comprehensive evaluation model, the five characteristic parameters and dynamic gradient features were input, with the target variables being the oil content in the dewatering water, the water content in the dewatering oil, and the settling time. After continuous iteration, the optimal experimental parameters were comprehensively evaluated.

[0162] Step 4, Data Dimensionality Reduction

[0163] Principal component analysis (PCA) was chosen as an unsupervised machine learning method to reduce the dimensionality of the experimental data while maximizing the variance. The main variables that best reflect the equipment's operating status were extracted from the original high-dimensional features and ranked according to their contribution to the principal components. This allowed for the selection of parameters with a significant impact on the equipment's operating status. Based on the aforementioned operational characteristic analysis and model building results, the correlation between various operating parameters, control parameters, and settling tank efficiency could be determined, and the influence weight of each variable could be assessed.

[0164] (1) Data standardization

[0165] Each feature is standardized so that its mean is 0 and its standard deviation is 1.

[0166]

[0167] Where: μ—characteristic mean; σ—standard deviation;

[0168] (2) Calculate the covariance matrix

[0169] Reflecting the correlation between features, off-diagonal elements represent covariance. Standardized data matrix X scaled The covariance matrix is:

[0170]

[0171] (3) Characteristic decomposition

[0172] Perform eigenvalue decomposition on the covariance matrix to obtain the eigenvalues ​​λ. i and the corresponding feature vector v i The magnitude of the eigenvalues ​​represents the variance contribution of the corresponding principal components.

[0173] (4) Select the number of principal components

[0174] Based on the variance contribution rate, select the top k largest eigenvalues ​​so that the cumulative contribution rate is >95%.

[0175]

[0176] (5) Data projection

[0177] The first k eigenvectors are used to construct a projection matrix W, which projects the original data onto a new space, resulting in the dimensionality-reduced new data X. pca :

[0178] X pca =X scaled ·W

[0179] Step 5, State Identification and Prediction

[0180] A multi-parameter fuzzy comprehensive evaluation model is employed to intelligently predict and identify operational conditions, clarifying the specific impact of each variable on the experimental settling tank's operating status. Based on current experimental settling tank operating data, the operating status is assessed in real time, and evaluation indicators are established. Combining the aforementioned relationship model and machine learning algorithms, the model predicts potential scenarios for continued operation under the current condition, as well as future changes in operating status over time. By integrating current and future development trends, the model identifies the equipment status and provides risk warnings, delivering an evaluation result. New data is then incorporated into the database after the prediction is completed.

[0181] (1) Determine the set of evaluation factors

[0182] Input parameters: Select key parameters affecting the operating status of the settling tank (such as temperature, pressure, flow rate, flotation intensity, etc.). Constituent factor set:

[0183] U = {u1, u2, ..., u} n}

[0184] Target output: Define a set of comments for the running status: "Normal", "Good settlement effect", "Average settlement effect", "Poor settlement effect", "Very poor settlement effect", which constitute the comment set.

[0185] V = {v1, v2, v3, v4, v5}

[0186] (2) Based on the weights of each parameter determined by the previous principal component analysis method, establish a fuzzy relation matrix:

[0187]

[0188] (3) Fuzzy synthesis operation

[0189] Fuzzy synthesis was performed using the maximum membership method, resulting in comprehensive evaluation result B:

[0190]

[0191] By using fuzzy logic to handle nonlinear and uncertain problems, the influence of multiple parameters on the state of the settling tank is clarified, making it suitable for intelligent monitoring and prediction of complex industrial systems. Combined with the aforementioned PINN algorithm, the actual operating characteristics of wastewater settling tanks are mapped, enabling the analysis of the actual equipment's operating characteristics on the experimental platform described in this invention.

[0192] Example 4:

[0193] In this embodiment, step four specifically involves:

[0194] Based on a smart control architecture applicable to wastewater settling tanks, this system utilizes dynamic variable monitoring and Markov Decision Process (MDP) to improve settling efficiency and system stability through high-precision real-time monitoring and fine-tuning of process parameters. The core idea of ​​the smart control architecture is to achieve automated adjustment of system performance by accurately monitoring dynamic variables directly related to the settling process and combining this with a migration-enhanced Proximal Policy Optimization (PPO) algorithm for optimal control decisions.

[0195] Step 1, MDP Algorithm Preparation

[0196] 1. State Space Definition and Dynamic Variable Acquisition: To achieve intelligent control of the settling process, this invention first defines dynamic variables closely related to settling efficiency as the state space of a Markov Decision Process (MDP). These dynamic variables include, but are not limited to, oil layer thickness, turbidity, temperature, pH value, and suspended solids concentration. These variables comprehensively reflect the changing state of the settling process and have a significant impact on settling efficiency. Therefore, accurately monitoring the real-time changes of these variables is key to the intelligent control architecture of this invention.

[0197] To ensure high-precision real-time acquisition of state variables, this invention employs a fusion technology based on particle image velocimetry (PIV), an optical turbidimeter, and multi-source sensors. Data captured through a multimodal sensor network enables comprehensive monitoring of fluid states, especially in complex sedimentation environments, ensuring high-precision real-time acquisition of state variables and providing a reliable basis for subsequent decision-making.

[0198] 2. Action Space Mapping and Fine Control: In the intelligent control architecture, for different process requirements, this invention maps adjustable parameters, such as the power of the water bath electric heating rod, the opening degree of the sludge discharge valve, and the power of the centrifugal pump, into the action space of the Markov decision process. Each adjustable parameter corresponds to a control action. The intelligent control system calculates the optimal control strategy based on the changes in the current state variables, thereby adjusting these parameters to optimize the efficiency and stability of the settling process.

[0199] This invention particularly emphasizes the design of a discrete-continuous hybrid action space, which solves the problem that traditional discrete action control methods cannot meet the requirements of fine control. In practical applications, many control parameters are not only discrete values, but can be continuously adjusted within a certain range. For example, the power of the water bath electric heater and the opening degree of the sludge discharge valve can be finely adjusted according to the real-time status of the system. Through the design of the discrete-continuous hybrid action space, the control system can adapt more flexibly to various operating conditions while ensuring control accuracy, thereby improving the response speed and control accuracy of the entire system.

[0200] 3. Markov Decision Process Optimization Strategy: This invention employs a Markov Decision Process (MDP) as the optimization framework for control decisions. By defining the state space, action space, and reward function, the MDP can select the optimal control strategy based on the current system state. Through learning from historical data and real-time feedback, the MDP continuously optimizes the control strategy, thereby maximizing system performance. The introduction of the MDP endows the intelligent control architecture of this invention with adaptability and learning capabilities, enabling it to automatically adjust the control strategy under changing process conditions, thus maintaining the high efficiency and stability of the settling process.

[0201] Step 2, Implementation of Markov Decision Algorithm

[0202] Construct a multi-objective composite reward function, for example:

[0203] R t =α·(T) oil / T target )+β·(C chemical -1 )+γ·(E settling / E baseline )

[0204] α, β, and γ are learnable dynamic weight coefficients used to achieve multi-objective optimization of sedimentation efficiency, reagent consumption, and energy consumption.

[0205] Step 3, Decision Optimization

[0206] The transfer-enhanced PPO algorithm, which combines reinforcement learning framework and transfer learning technology, enables intelligent optimization of control decisions for oilfield wastewater settling tanks.

[0207] First, the multi-source sensing experimental data are fused and the features are encoded.

[0208] (1) Basic PPO Algorithm

[0209]

[0210] δ t =γV(s) t+1 )-V(s t )

[0211] In the formula: A t —The advantage function at time t, i.e., the degree of advantage or deviation of action selection; δ—represents the TD error at that time, i.e., the difference between the current estimate and the actual observation; γ—discount factor, used to weigh the importance of future rewards. γ∈[0,1]; λ—decay factor, determining the degree of influence of the TD error on past time steps. A larger λ value indicates that rewards over a longer time span are considered; a smaller λ value indicates that the TD error is mainly affected by the current time step; V(st )——State s at time t t The value function, i.e., from state s t Initially, the expected total reward in the future.

[0212] (2) Strategy objective function

[0213]

[0214] In the formula: L CLIP (θ) — The objective function of the PPO, representing the objective of updating the policy using the shearing method under the current policy. The objective function optimizes the policy parameter θ by maximizing this expression; E t —The expected value over all time steps t. Represents the expectation of the trajectory (state-action sequence) altered by the execution policy; ρ t (θ) — Importance sampling ratio, representing the ratio between the current strategy and the old strategy, specifically:

[0215]

[0216] Wherein, πθ(a t |s t ) is the current policy in state s t Choose action a t The probability, πθ old (a t |s t ) is the probability of choosing the old strategy. This ratio measures the degree of change in the current strategy relative to the old strategy.

[0217] clip(ρ t (θ), 1-∈, 1+∈) — shearing operation, restricting ρ t The value of (θ) is in the range [1-∈, 1+∈]. The purpose of pruning is to prevent excessive policy updates and reduce the risk of over-optimization. Specifically:

[0218] If ρ t If (θ) exceeds the range of [1-∈, 1+∈], it is clipped (i.e. set as a boundary value);

[0219] If ρ t (θ) will not be modified within this interval.

[0220] This operation is key to PPO (Program Point Optimization), as it helps control the magnitude of policy updates, preventing drastic changes to the policy and thus improving training stability.

[0221] ∈ -- a hyperparameter that represents the size of the cut range.

[0222] (3) Implementation of migration enhancement

[0223] The basic policy network is trained in the source MDP decision scheme and digital twin system, covering simulation data under multiple operating conditions. Then, transfer optimization is implemented on the PPO by adding a transfer-related regularization term to the original loss function of the PPO. The transfer loss term is calculated using KL divergence, resulting in the adaptive loss function as follows:

[0224] L total =L PPO +λ·D KL (π old ||π new )

[0225] In the formula: L total —The overall objective function, which is the target to be optimized during PPO training. It includes the PPO objective function itself (i.e., the part that limits the policy update magnitude through shearing operations) and an additional KL divergence penalty term. PPO —PPO strategy objective function; D KL (π old ||π new — KL divergence, which measures the new strategy (π) new ) and the old strategy (π) old Differences between:

[0226]

[0227] After model training, multi-MDP policies are fused, and for each candidate policy a... i Calculate its value assessment under each MDP:

[0228]

[0229] Simultaneously, process constraint filtering is applied to the integrated results to ensure the feasibility of the optimized strategy, and to select and execute the optimal decision.

[0230] Example 5:

[0231] In this embodiment, step five specifically involves:

[0232] After implementing the theoretically optimal decision, the real-time operating parameters of the experimental settling tank are continuously monitored to evaluate the improvement effect of the decision on operational efficiency. Simultaneously, the deviation between theoretical settling results and actual observation data is analyzed, and the model algorithm is iteratively optimized based on this, prompting the model to continuously learn in practice. Through this process, the prediction accuracy of the model is gradually improved, further enhancing the overall operating efficiency of the experimental equipment after the decision is implemented.

[0233] Step 1: Quantitatively analyze the deviation between the theoretical improvement in settlement efficiency and the actual situation.

[0234] After executing the decision proposed by the Markov model, the MDP algorithm makes a function prediction of the decision value:

[0235]

[0236] In the formula: γ - discount factor, taken as 0.95; r - instant reward function (weighted terms such as oil content in effluent, energy consumption, etc.).

[0237] After the decision was implemented, the actual efficiency improvement of the experimental settling tank was calculated using actual operating data:

[0238]

[0239] In the formula: η exp (t) — Composite efficiency index after decision-making; η base (t) — Efficiency index before decision-making;

[0240] T – Settlement test time.

[0241] Efficiency Difference Quantitative Analysis:

[0242] Δ=E theory -E real

[0243] Step 2: Establish an algorithm for the dynamic compensation mechanism to address the discrepancy between theoretical and practical efficiency improvements.

[0244] Gradient Boosting Decision Tree (GBDT) achieves continuous optimization of the decision model through closed-loop feedback between theoretical prediction and actual performance. Specifically, it innovatively designs a three-stage dynamic correction process to address the prediction-actual deviations caused by idealized assumptions in existing decision algorithms. Its key features are:

[0245] (1) Establish a multidimensional feature space and fuse the parameter set of the theoretical execution strategy with the experimental environment variables using tensors;

[0246] (2) Using the Boosting integration strategy of GBDT, a nonlinear mapping model of theoretical-actual deviation is constructed by weighted residual approximation method;

[0247] (3) An incremental learning mechanism is constructed based on the Bayesian optimization framework to dynamically update the node splitting rules of the decision tree using real-time experimental data. This architecture forms a complete iterative optimization loop of "prediction-execution-feedback-correction" by constructing an interpretable feature importance matrix. The specific implementation process is as follows:

[0248] Step 3, GBDT Deviation Modeling Framework

[0249] (1) Input feature set

[0250] Theoretical parameter: The theoretical settlement effect value E output by the fuzzy evaluation algorithm. theory ;

[0251] Real-time monitoring parameters: liquid level, wastewater turbidity, oil-water separation rate, etc.

[0252] Historical deviation sequence: mean, variance, and trend term of deviation within a sliding window (default length 30 periods); Operating parameters: control variables such as power of external heating rod of coil and speed of centrifugal pump.

[0253] (2) Target variable: Actual deviation value

[0254] Δ=E theory -E real

[0255] (3) Model building and optimization

[0256] Construct a regression-based GBDT model with the objective of minimizing the absolute error of the bias:

[0257]

[0258] Where: GBDT(X) i ;θ)——represents the GBDT model for sample X i The prediction; θ—the set of parameters in GBDT; λ—the regularization intensity hyperparameter.

[0259] The optimal hyperparameter combination was determined using a grid search strategy: to effectively suppress overfitting, the maximum depth of the decision tree was limited to 7 layers; a learning rate of 0.15 was used to control the iteration step size of the gradient descent process, balancing convergence speed and optimization accuracy; and a subsample sampling ratio of 0.8 was set to enhance model robustness through random feature subset selection. This parameter combination, after cross-validation, significantly improved the model's generalization ability while maintaining high prediction accuracy.

[0260] Step 4, Deviation-driven iterative optimization of the model

[0261] (1) Calculate the contribution of each feature to the bias prediction, select the top 20% of features by contribution, and dynamically adjust the multi-parameter weights w. i :

[0262]

[0263] In the formula: w i —Weight of the i-th element; β —Learning rate, default value 0.01; P i —This is the i-th standardized parameter; —The gradient of the prediction result indicates how the error changes with P. i It changes with the changes.

[0264] (2) Online incremental learning

[0265] When the deviation analysis module triggers the optimization command, the following steps are executed:

[0266] a. Extract newly added abnormal data samples from the time series database;

[0267] b. Update the GBDT model using a partial fitting algorithm:

[0268] c. If model performance degrades, restart global training.

[0269] (3) Collaborative reinforcement learning for model optimization

[0270] The key features output by GBDT are mapped to the state space of a Markov MDP to reconstruct the decision model:

[0271] s′=[s origin GBDT(X) t )]

[0272] Reward function correction:

[0273] r′=r base -η·|Δ t |

[0274] Where η is an adaptively adjustable deviation penalty coefficient.

[0275] By deeply coupling GBDT with reinforcement learning, continuous self-optimization of the decision-making process of the MDP model is achieved, providing an innovative solution for intelligent control of wastewater treatment.

Claims

1. A method for intelligent prediction and optimized control of the operating status of oilfield wastewater settling tanks, characterized in that... The steps include: Step 1: A modular design for reconfigurable settling system integrating multimodal sensing is used to collect time-series data, including flow rate, particle concentration, and flow field vorticity. The settling system integrates a PIV particle image velocimeter, a distributed fiber optic sensor, a conductivity-turbidity multi-parameter sensor array, and temperature staining technology to construct a multi-dimensional physical field dynamic monitoring network, thereby realizing three-dimensional visualization reconstruction of the flow characteristics inside the experimental settling tank and quantitative analysis of particle transport trajectories. Step 2: Combining the Physical Information Neural Network (PINN) algorithm, construct a mapping relationship between the working characteristics of the actual settling tank and the experimental settling tank: predict the performance of the actual settling tank using data from the experimental settling tank. Step 3: Identification and prediction of the operating status of the experimental settling tank; Establish a multi-parameter fuzzy comprehensive evaluation model, and use the multi-parameter fuzzy comprehensive evaluation model to evaluate the operating status of the experimental settling tank in real time based on the current operating data of the experimental settling tank, and establish evaluation indicators; then combine machine learning algorithms to predict the possible situations that may occur when the settling system continues to operate under the current state, as well as the changes in the future operating status over time. Step four, the settling system operates with intelligent decision-making and automatic control; by accurately monitoring dynamic variables directly related to the settling process, a Markov decision process (MDP) is used as the optimization framework for control decisions, and a migration-enhanced PPO algorithm is combined to optimize control decisions, thereby achieving automatic adjustment of system performance. The dynamic variables include oil layer thickness, turbidity, temperature, pH value, and suspended solids concentration. Step 5: Decision verification and dynamic correction closed-loop optimization control of the Markov Decision Process (MDP). Through deep coupling of GBDT and reinforcement learning, continuous self-optimization of the MDP decision process is achieved.

2. The intelligent prediction and optimized control method for the operating status of oilfield wastewater settling tanks according to claim 1, characterized in that: The multi-parameter fuzzy comprehensive evaluation model in step three is as follows: In the formula: ρ w ρ is the density of water. o ρ is the density of the oil; g is the acceleration due to gravity; α is the kinetic coefficient; Q is the liquid flow rate; k DAF A is the efficiency coefficient of the air flotation unit. b v is the total surface area flux of the bubbles; b d represents the rising velocity of the bubble. 32 C is the average bubble diameter of Sauter; oil Oil content; C oil,crit E is the critical emulsification concentration of the oil phase. a R is the apparent activation energy; T is the ideal gas constant; T is the heating coil temperature; T ref Reference temperature; f coil (x c ,y c ,z c ) is the spatial location (x c ,y c ,z c The influence function of the spatial position of the heating coil at point P; i Let P be the spatial coordinates of the water inlet i=1, the heating coil i=2, and the flotation unit i=3; opt The optimal layout positions for each component; L char The characteristic length is taken as 20% of the tank diameter; w i For weighting coefficients; μ(T) is the fluid viscosity; A settlle For effective settlement area; Q max C is the critical flow rate; TSS β represents the turbidity of the wastewater; β is the oil phase correction coefficient.

3. The intelligent prediction and optimized control method for the operating status of oilfield wastewater settling tanks according to claim 2, characterized in that: The modular design reconfigurable settlement system integrating multimodal perception includes an experimental module, a data acquisition module, and a dataset parsing and control module. These modules work together to form a complete experimental feedback loop. The experimental module includes an experimental settling tank, a power unit, a heating unit, and a flotation unit. The experimental settling tank is equipped with a central column tube and a central reaction cylinder. The flotation unit is a dissolved air pump, which uses a multiphase dissolved air pump to draw in air and wastewater together. The impeller of the dissolved air pump cuts the wastewater and air into fine foam, making them fully mixed. At the same time, the high pressure generated by the high-speed rotation of the impeller fully dissolves the drawn-in gas into the water. Together, they form stable dissolved air water. When this dissolved air water is released at normal pressure through a pressure reducing valve, it produces tiny bubbles with a diameter between 20 and 50 μm. The data acquisition module integrates a PIV particle image velocimetry device, a CCD camera array, a multispectral turbidimeter, and a distributed sensor group. The PIV particle image velocimetry device and the CCD camera are synchronously triggered by laser pulses to achieve submicron-level dynamic analysis of the oil droplet particle size distribution. The multispectral turbidimeter uses multi-wavelength scattered light intensity detection technology combined with an embedded temperature and pressure compensation algorithm to achieve online calibration of turbidity changes during the settling process. The distributed sensor group includes a micro differential pressure transmitter and an array of temperature probes, forming a multi-point physical parameter monitoring network for the flow field inside the settling tank. The data parsing and control module is equipped with a motion feature extraction unit, a multi-physics coupling analysis unit, and a settlement efficiency evaluation unit; The motion feature extraction unit uses an improved PTV particle tracking algorithm to perform Gaussian fitting and noise reduction on continuous frame images, and calculates the oil droplet velocity field, acceleration field and fractal dimension of motion trajectory. The multiphysics coupling analysis unit analyzes the dynamic characteristics of oil droplet coalescence and breakup by establishing a correlation matrix model of turbidity-particle size-flow velocity; the sedimentation efficiency evaluation unit constructs a particle group sedimentation prediction model based on the Fokker-Planck equation and outputs key indicators such as sedimentation rate and interface clarity.

4. The intelligent prediction and optimized control method for the operating status of oilfield wastewater settling tanks according to claim 3, characterized in that: Step two specifically involves: Step 1: Construct a physical information neural network architecture. Input layer: Parameters obtained from the experimental settling tank operation; Output layer: Actual settling tank performance parameters, including particle settling efficiency and turbidity removal rate. Define the loss function: In the formula: For data matching, a small amount of actual settling tank calibration data Y is used in the calculation. real ; For physical constraints; —Similarity criterion constraints; The allocation of each item is dynamically adjusted through the validation set; Step 2: Time series data preprocessing and training of the PINN neural network model for physical information; The time series data is augmented by making it dimensionless and interpolating it to match the actual tank time axis. The experimental tank settling time series data is used to train the basic PINN to establish a high-precision mapping relationship between the experimental settling tank and the actual settling tank. Step 3: Dynamically update and continuously fine-tune the model with new data to optimize the PINN physical information neural network model.

5. The intelligent prediction and optimized control method for the operating status of oilfield wastewater settling tanks according to claim 4, characterized in that: The specific method for step three is as follows: Step 1: Obtain data, including water flow rate, flotation intensity, heating temperature, wastewater turbidity, and oil content; Step 2: Establish a multi-parameter fuzzy comprehensive evaluation model; A sedimentation efficiency η model is constructed based on the mass conservation and kinetic equations: Subfunction expansion: (1) Air flotation enhancement factor Γ: In the formula: k DAF The efficiency coefficient of the non-air flotation unit is related to the spatial location of the flotation unit; A b The total surface area flux of the bubbles is m. 2 / s;v b d represents the rising velocity of the bubble. 32 C is the average bubble diameter of Sauter; oil,crit This is the critical emulsification concentration of the oil phase; (2) Thermal effect correction term T eff : In the formula: E a R is the apparent activation energy; T is the ideal gas constant; ref Reference temperature; f coil The function representing the influence of the spatial position of the heating coil is positively correlated with the uniformity of the thermal field. (3) Spatial Coordination Factor ξ space : In the formula: P i P represents the spatial coordinates of the water inlet (i=1), heating coil (i=2), and flotation unit (i=3); opt The optimal layout positions for each component; L char The characteristic length is taken as 20% of the tank diameter; w i These are the weighting coefficients; (4) Foundation settlement rate η0 In the formula: μ(T) is the fluid viscosity; A settle For effective settlement area; Q max The critical flow rate; Summarized as follows: Using the data obtained in step 1, a multi-parameter fuzzy comprehensive evaluation model is trained. The multi-parameter fuzzy comprehensive evaluation model is used to input the above-mentioned water distribution flow rate, air flotation intensity, heating temperature, sewage turbidity, oil content and dynamic gradient characteristics. The target variables are the oil content of the dewatering water, the water content of the dewatering water and the settling time. After continuous iteration, the optimal experimental parameters are comprehensively evaluated. Step 4: Data dimensionality reduction, determine the correlation between each time series data and the efficiency of the settling tank, and determine the influence weight of each time series data. Step 5, State Identification and Prediction.

Images