Constructed wetland sewage treatment system and method based on multi-Agent collaborative optimization

Abstract

The invention discloses a constructed wetland sewage treatment system and method based on multi-Agent collaborative optimization. The system comprises a physical wetland unit and a digital twin Agent module, the environment sensing Agent collects multi-source data of the physical wetland unit and sends the multi-source data to the predictive analysis Agent, the predictive analysis Agent predicts water quality through a water quality predictive model, and parameters are updated by measured values regulated and controlled by the execution control Agent; if the water quality is abnormal, the alarm is triggered and sent to the optimization decision Agent, decision logic is generated by the optimization decision Agent, and then the decision logic is converted into a real-time regulation and control instruction through the execution control Agent. According to the method, dynamic optimization and self-adaptive regulation and control of the system are realized through multi-Agent collaboration and digital twinning interaction.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of environmental engineering and intelligent control, and in particular to an artificial wetland wastewater treatment system and method based on multi-agent collaborative optimization. Background Technology

[0002] Constructed wetland wastewater treatment technology has attracted much attention in the treatment of urban sewage and agricultural non-point source pollution due to its advantages such as eco-friendliness and low cost. However, fluctuations in water quality and quantity, complex biochemical processes, and the need for dynamic regulation urgently require the introduction of intelligent control methods.

[0003] Current intelligent control technologies for constructed wetlands are exploring multiple paths. At the control algorithm level, fuzzy PID control (adjusting aeration through error feedback) and BP neural networks (predicting denitrification efficiency based on historical data) have achieved basic automation, but are mostly limited to single-parameter closed-loop control, lacking multi-variable coupled optimization capabilities. In terms of modeling methods, mechanistic models (such as the Monod equation describing microbial degradation kinetics) and data-driven models coexist, but the former is insufficient in characterizing plant-microbe interactions, while the latter suffers from poor generalization due to small sample data. Regarding system architecture, while centralized SCADA systems simplify operation and maintenance, they pose a single point of failure risk and have high expansion costs. While emerging edge computing solutions improve real-time performance, inter-unit coordination mechanisms still rely on preset rules, making it difficult to cope with sudden pollution loads. Furthermore, some studies have attempted to introduce reinforcement learning, but excessive discretization of the state space leads to coarse decision granularity and low global optimization efficiency.

[0004] While current technological advancements have partially achieved automation and local optimization, the following key bottlenecks still exist: (1) Model-algorithm separation: Mechanistic models (such as the Monod equation) are difficult to accurately characterize the multidimensional coupling of plant-microbe-matrix, while data-driven models (such as XGBoost and BP neural networks) are limited by small sample data and sudden changes in working conditions, resulting in weak generalization ability and a disconnect between prediction and control. (2) Defects of multivariate coordination: Fuzzy PID, reinforcement learning and other algorithms focus on the closed-loop regulation of a single parameter (such as aeration volume, hydraulic load), lack a dynamic game optimization mechanism with multiple objectives (energy efficiency, removal rate, shock resistance), and the overall coordination efficiency is less than 20%; (3) Rigidity of architecture: Centralized SCADA systems rely on the central node for decision-making, which poses a single point of failure risk and has high expansion costs; although edge computing solutions improve real-time performance, inter-unit collaboration relies on preset rules and cannot adapt to sudden pollution (such as a sudden increase of 50% in concentration). (4) Coarse decision granularity: Existing reinforcement learning strategies (such as Q-learning) have low precision of control commands (such as aeration intensity adjustment) and significant response lag due to excessive state space discretization, making it difficult to match the dynamic characteristics of wetland biochemical reactions.

[0005] These problems collectively make it difficult for existing technologies to break through the predicament of "local optimization, static response, and high operation and maintenance costs," thus restricting their engineering applicability in complex scenarios. Summary of the Invention

[0006] To address the shortcomings of existing technologies, such as model-algorithm disconnect, insufficient multivariate collaboration, and response lag, this invention proposes an artificial wetland wastewater treatment system and method based on multi-agent collaborative optimization. This system adopts a collaborative design of modular physical wetland units and a digital twin agent architecture.

[0007] The specific technical solution is as follows: An artificial wetland wastewater treatment system based on multi-agent collaborative optimization includes a physical wetland unit and a digital twin agent module. The physical wetland unit comprises several wetland sub-units connected in series or parallel. Each wetland sub-unit includes: a backup adsorption matrix module, a plant community layer, a matrix filling layer, and a water distribution system arranged sequentially from top to bottom. The backup adsorption matrix module is isolated from the other components of the wetland sub-unit. A sensor network is set up to monitor multi-source data of the physical wetland unit in real time and send it to the digital twin agent module. The digital twin agent module includes: an environmental perception agent, a predictive analysis agent, an optimization decision agent, and an execution control agent; the environmental perception agent is used to receive the multi-source data and periodically send it to the predictive analysis agent; if the concentration of a certain pollutant exceeds the standard or the water quality changes abruptly, an alarm signal is triggered and sent to the optimization decision agent; The predictive analysis agent uses a water quality prediction model to predict water quality based on multi-source data and obtain water quality prediction results. The parameters of the water quality prediction model are dynamically updated by the predictive analysis agent using the measured water quality values ​​obtained after being regulated by the execution control agent. The optimization decision agent uses Nash equilibrium to coordinate energy efficiency, removal rate and shock resistance based on water quality prediction results or early warning levels in alarm signals to achieve multi-objective optimization, obtain the optimization results of the parameters of each component of the physical wetland unit, and generate decision logic. The execution control agent is used to convert the decision logic into real-time control instructions for physical wetland units.

[0008] Furthermore, the matrix filling layer comprises, from top to bottom: a coarse sand layer, a zeolite-volcanic rock mixed layer, and a gravel layer; The water distribution system includes: a water supply component, an aeration component, and a water collection component; the water supply component's inlet is located above the standby adsorption substrate module and is evenly distributed in the plant community layer, and its inlet flow rate is controlled by a water flow control pump and valves; the aeration component is located between the zeolite-volcanic rock mixed layer and the gravel layer, and its aeration rate is controlled by an aerator; the water collection component is located at the bottom of the gravel layer and is used to collect the treated wastewater and output it to the outlet of the wetland subunit.

[0009] Furthermore, the real-time control commands include: adjusting the water distribution pulse frequency of the water distribution system, adjusting the inlet flow rate of the water supply component and the outlet flow rate of the water collection component, adjusting the power of the aerator in the aeration component, and adjusting the plant harvesting cycle.

[0010] Furthermore, the multi-source data includes: organic matter concentration, total nitrogen, total phosphorus, pH, dissolved oxygen, water level, and redox potential.

[0011] Furthermore, the backup adsorption materials pre-embedded in the backup adsorption matrix module include: modified zeolite, activated alumina, and composite materials supported on Fe and Mn; The backup adsorption matrix module is provided with a drug injection port and an adsorption material replenishment port at the upper end, which are used to inject specified reagents and add new adsorption materials into the backup adsorption matrix module, respectively.

[0012] Furthermore, in the optimization decision agent, the objective function for multi-objective optimization is: In the formula, η j w represents the removal rate of the j-th pollutant to be removed. j λ is the dynamic weighting coefficient for the j-th pollutant; λ is the energy consumption penalty coefficient, and Q... i Let A be the hydraulic load of the i-th wetland sub-unit. i Let be the aeration intensity of the i-th wetland sub-unit.

[0013] Furthermore, the water quality prediction model is built based on an LSTM network, with historical concentration sequences and real-time multi-source data as inputs and pollutant concentration change trends for the next 24 hours as outputs.

[0014] A constructed wetland wastewater treatment method based on multi-agent collaborative optimization, implemented using the aforementioned constructed wetland wastewater treatment system based on multi-agent collaborative optimization, includes the following steps: S1: The wastewater to be treated is input into the physical wetland unit for treatment and then output. Multi-source data of the physical wetland unit is collected through the sensor network and sent to the digital twin agent module. S2: The environment perception agent receives multi-source data and sends the multi-source dataset within the period to the prediction analysis agent according to the periodic prediction task; When multi-source data shows that pollutant concentration exceeds the standard or water quality changes abruptly, an alarm signal is immediately triggered, and the alarm signal and real-time multi-source dataset are sent to the optimization decision agent, and the process jumps to execute S4. S3: The predictive analysis agent receives the multi-source dataset sent by the environmental perception agent, uses the water quality prediction model to output the predicted future pollutant concentration change trend, and sends the predicted future pollutant concentration to the optimization decision agent. S4: In the normal path, the optimization decision agent receives the prediction results from the prediction analysis agent, obtains the decision logic, and sends it to the execution control agent; In the emergency response scenario, the optimization decision agent receives the alarm signal and the corresponding real-time multi-source dataset. Based on the warning level in the alarm signal, and with the goal of maximizing pollutant removal rate while minimizing energy consumption, it dynamically allocates the weight coefficients in the objective function of multi-objective optimization to obtain the decision logic, which is then sent to the execution control agent. The warning level includes mild pollution risk and severe pollution risk. S5: The execution control agent receives the decision logic from the optimization decision agent and transforms it into real-time control instructions for the physical wetland units; S6: The physical wetland unit achieves dynamic correction through bidirectional interaction between the digital twin model and the physical layer; based on the feedback of the control agent's regulation effect on the physical wetland unit, the parameters of the water quality prediction model are dynamically corrected to improve the prediction accuracy.

[0015] Furthermore, in S5, if the warning level is a slight pollution risk, the control instructions include: ① fine-tuning the water distribution system: extending the residence time of water flowing through the matrix filling layer rich in iron and aluminum oxides, i.e., the hydraulic residence time, by 10%-20%; ② fine-tuning the aeration: increasing the aeration intensity by 10%-25%, changing the oxidation-reduction point, and promoting the conversion of soluble phosphates into solid precipitates. If the warning level is a severe pollution risk, the control instructions include: ① Opening the inlet to the backup adsorption matrix module to activate the backup adsorption matrix module, allowing the wastewater to be treated to flow through the backup adsorption material pre-embedded in the backup adsorption matrix module; ② Starting the dosing pump to add chemical agents to the backup adsorption matrix module, quickly fixing phosphorus through chemical precipitation; ③ Coordinating hydraulic control: adjusting the water distribution pulse frequency to extend the hydraulic retention time to more than 20%; ④ Adjusting and increasing aeration to increase the aeration intensity to more than 25%.

[0016] Furthermore, in multi-agent collaborative optimization, based on the preliminary strategy generated by each wetland sub-unit based on local performance data, the marginal contribution of each participant in different collaborative alliances is calculated through distributed negotiation to achieve global optimality. The participants are the wetland sub-unit or the agent of that unit. The distributed negotiation uses the Shapley value algorithm, and the marginal contribution calculation expression is as follows: In the formula, φ i Let $\frac{i}{i}$ be the contribution of participant $i$ to the overall optimization objective of the system, $N$ be the set of all participants in the system, $S$ be the alliance, representing a subset of participants, $|S|$ be the size of the alliance, representing the number of participants in the alliance $S$, ${i}$ be the specific wetland sub-unit whose contribution is being calculated, $v(S)$ be the value function of the alliance $S$, representing the system performance index that can be achieved when the units in the alliance $S$ cooperate with each other, and $v(S∪{i})$ be the system performance index that can be achieved without participant $i$.

[0017] The beneficial effects of this invention are: (1) Multi-Agent collaborative architecture and digital twin interaction: Through the collaboration of four types of intelligent agents—environmental perception, predictive analysis, optimization decision-making, and execution control—and combined with real-time data and digital twin model bidirectional correction, dynamic optimization and adaptive regulation are achieved.

[0018] (2) Emergency mechanism for backup adsorption matrix: Set up backup adsorption matrix module and pre-embed encapsulated adsorption material. In the event of sudden pollution, activate the isolation chamber and inject chemical agents to temporarily increase the adsorption capacity by 20%-30% and quickly respond to sudden concentration changes.

[0019] (3) Game theory-driven multi-objective optimization: Based on Nash equilibrium, multiple objectives such as energy efficiency, removal rate and shock resistance are coordinated, and optimization weights are dynamically allocated to achieve dual optimization of water quality stability and energy consumption reduction. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of an artificial wetland wastewater treatment system based on multi-agent collaborative optimization in an embodiment of the present invention.

[0021] Figure 2 This is a schematic diagram of the physical wetland unit in an embodiment of the present invention. Detailed Implementation

[0022] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The objectives and effects of the present invention will become clearer as a result. The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0023] like Figure 1 As shown, an artificial wetland wastewater treatment system based on multi-agent collaborative optimization includes a physical wetland unit (physical layer) and a digital twin agent module (intelligent layer); the digital twin agent module includes: an environmental perception agent, a predictive analysis agent, an optimization decision agent, and an execution control agent.

[0024] The physical wetland unit collects multi-source data through a high-precision sensor network. After receiving the multi-source data, the environmental perception agent uses an LSTM model to predict water quality trends, the predictive analysis agent generates control strategies based on game theory, and the execution control agent executes commands (adjusting equipment, activating backup adsorption matrix modules, etc.) within 10 seconds. The execution results are fed back to the LSTM model of the predictive analysis agent, dynamically correcting the model parameters. In the event of a sudden increase in pollution, an emergency response involving backup adsorption matrix modules and reagent injection is triggered, achieving multi-objective collaborative optimization and improving treatment efficiency and stability.

[0025] The physical wetland unit is the core physical carrier of this system, and it adopts a modular design to improve pollutant removal efficiency. The modular structure includes several wetland sub-units connected in series or parallel.

[0026] like Figure 2 As shown, each wetland subunit includes: a plant community layer, a substrate filling layer, and a water distribution system.

[0027] The plant community layer is planted with pollution-tolerant plants such as reeds and cattails, at a density of 3-5 plants / m². 2 It promotes microbial metabolism through root oxygen secretion.

[0028] The substrate filling layer is arranged below the plant community layer and, from top to bottom, includes: a coarse sand layer (surface layer, porosity 40%-45%), a zeolite-volcanic rock mixed layer (middle layer, adsorption specific surface area > 200m²). 2 / g) and gravel layer (bottom layer, porosity 30%-35%), to optimize hydraulic permeability and pollutant adsorption capacity through stratification.

[0029] The water distribution system includes: a water supply component, an aeration component, and a water collection component. The water supply component uses perforated pipes, distributed throughout the vegetation layer, with the outlet holes arranged in an array to ensure uniform water flow distribution within the wetland sub-unit. The inlet pipe of the water supply component is equipped with a water flow control pump and valves to control the flow rate of the incoming water (i.e., the water to be treated). Furthermore, the single-hole flow rate of the water supply component's pipes... Where γ is the flow coefficient, A is the orifice area, and h is the head height; adjust the parameters affecting the flow rate of a single orifice to ensure that the hydraulic load range is stable within 0.1-0.5m. 3 / (m2 •d). The aeration components' water pipes are arranged between the zeolite-volcanic rock mixed layer and the gravel layer. These pipes are also perforated to provide sufficient oxygen and enhance biodegradation. The aeration rate of the aeration components is controlled by a pump. The water collection component includes porous collection pipes located at the bottom of the gravel layer to collect treated wastewater, which is then discharged through the outlet of the wetland subunit for subsequent monitoring of the wastewater treatment results.

[0030] The wetland subunit also includes a backup adsorption matrix module, which is separated from the rest of the wetland subunit by an impermeable partition, forming an isolated chamber. Preferably, the backup adsorption matrix module is located above the gravel layer, separated from the plant community layer, coarse sand layer, and zeolite-volcanic rock mixed layer. This design is based on functional zoning considerations, namely, the main function of the backup adsorption matrix module, coarse sand layer, and zeolite-volcanic rock mixed layer is to purify water, while the main function of the gravel layer is to filter water. During normal operation when the pollutant concentration is low, the backup adsorption matrix module is in a closed state by default and does not come into direct contact with the water flow. When the pollutant concentration exceeds the threshold, the backup adsorption matrix module is activated to increase the effective adsorption area and temporarily increase the adsorption capacity of the physical wetland unit by 20%-30%. This is used when the matrix filling layer cannot meet the purification needs, and the inlet to the matrix filling layer is closed. The volume of the backup adsorption matrix module accounts for 8% to 12% of the total effective matrix volume of the physical wetland unit, preferably 10%. The backup adsorption matrix module is pre-embedded with highly efficient backup adsorption materials, including modified zeolite, activated alumina, and composite materials loaded with Fe and Mn. The backup adsorption matrix module is equipped with a dosing port and an adsorption material replenishment port at the top. An external pipeline is connected to the dosing port through a dosing pump to inject a specified reagent solution into the backup adsorption matrix module. New adsorption material is introduced into the backup adsorption matrix module through a screw conveyor from the adsorption material replenishment port.

[0031] Each wetland subunit can treat wastewater based on a pollutant removal mechanism, which is achieved through the synergistic action of multiple physicochemical processes. Organic matter (COD) degradation depends on microbial metabolism in the aerobic-anaerobic alternating zone, following a first-order kinetic model. In the formula, C is the COD concentration, and k is the reaction rate constant (0.15-0.25dB). -1 ).

[0032] Total nitrogen (TN) removal efficiency η TN Related to the dissolved oxygen (DO) gradient, the calculation formula is: In the formula, η TN Indicates total nitrogen removal efficiency, DO 表层 DO represents the surface dissolved oxygen in the water body of the wetland subunit.底层 This represents the dissolved oxygen at the bottom layer of the water body in the wetland subunit.

[0033] Phosphorus adsorption is primarily mediated by the chemical precipitation of Fe oxides or Al oxides in the matrix-filled layer, and its adsorption capacity is described by the Langmuir isotherm. In the formula, q e q represents the phosphorus adsorption capacity. max To determine the maximum adsorption capacity, q is used in this embodiment. max The concentration is 2.5-3.5 mg / g; K represents the Langmuir adsorption equilibrium constant, reflecting the affinity of the adsorbent for phosphorus; C e This indicates the concentration of phosphorus in the solution at adsorption equilibrium.

[0034] The system operates based on a high-precision sensor network for real-time monitoring and control, collecting multi-source data from the physical layer. The sensor network includes: water quality sensors, water level sensors, and root detection electrodes. Water quality sensors are deployed at the inlet and outlet of each wetland subunit and in the zeolite-volcanic rock mixture layer of the substrate filling layer. The monitored multi-source data includes: COD concentration, total nitrogen (TN), total phosphorus (TP), pH, and dissolved oxygen (DO), with a sampling interval of 5 minutes. Water level sensors (accuracy ±0.5 cm, located in the plant community layer above the coarse sand layer) and root monitoring electrodes (monitoring oxidation-reduction potential -300 to +300 mV, located in the coarse sand layer) further provide feedback on hydraulic loading and microbial activity, monitoring multi-source data including water level and oxidation-reduction potential.

[0035] Hydraulic characteristics are optimized by dynamically adjusting the hydraulic retention time (HRT). The formula for calculating HRT is: In the formula, V 有效 Q is the effective volume of the wetland subunit. 进水 This refers to the inflow rate.

[0036] In addition, by configuring the matrix porosity gradient in the matrix filling layer and using intermittent water distribution (pulse frequency 2-4 times / hour), the risk of clogging is significantly reduced, ensuring long-term stable operation of the system.

[0037] In summary, the physical wetland unit, through its modular structure, multi-physical and chemical process-based pollutant removal mechanism, and precise monitoring and control via sensor networks, constructs an efficient and scalable physical foundation for wastewater treatment, providing reliable data support and an execution environment for the multi-agent collaborative optimization of the digital twin agent module.

[0038] The digital twin agent module is the core of the intelligent layer of the constructed wetland wastewater treatment system. Through real-time interaction between four types of functional agents (environmental perception agent, predictive analysis agent, optimization decision-making agent, and execution control agent) and the physical wetland unit, it achieves dynamic optimization and control of the wastewater treatment process. The specific roles of the four types of functional agents are as follows: (1) Environmental perception agent. It is responsible for receiving multi-source data from the physical layer collected by the sensor network, forming a multi-source dataset. According to the set periodic prediction task, the environmental perception agent sends the multi-source dataset to the prediction analysis agent. If the multi-source data shows that the concentration of a certain pollutant in the outlet of the physical wetland unit exceeds the standard or the water quality changes abruptly, an alarm signal is triggered, and the alarm signal and the real-time multi-source dataset are sent to the optimization decision agent.

[0039] Furthermore, the multi-source data includes COD concentration, TN, TP, pH, DO, water level, and redox potential, with a sampling frequency of 5 minutes per sampling to ensure that the virtual model is synchronized with the physical system.

[0040] (2) Predictive Analysis Agent. A water quality prediction model is constructed based on an LSTM network, with historical concentration sequences (such as C) as input. t-24 C t-12 ,…,C t The system uses real-time multi-source data to output the trend of pollutant concentration changes over the next 24 hours. The prediction formula is abstracted as follows: In the formula, These are predicted pollutant concentrations. For the historical concentration sequence within the time window (tn,t), f LSTM This represents the nonlinear mapping relationship of the LSTM network. The predictive analysis agent sends the prediction results of the water quality prediction model to the optimization decision agent.

[0041] (3) Optimize the decision-making agent. A game theory framework is used to coordinate the operating parameters of each unit, and a multi-objective optimization problem is defined. The goal is to maximize the pollutant removal rate and minimize energy consumption, and to optimize the hydraulic load Q of the i-th wetland sub-unit. i With aeration intensity A i The optimization results are transformed into decision logic and sent to the execution control agent.

[0042] The objective function for multi-objective optimization is expressed as follows. The global optimal strategy is solved through Nash equilibrium to synergistically optimize pollutant removal rate, system energy consumption, and resistance to shock loads: In the formula, η jw represents the removal rate of the j-th pollutant (including COD, TN, TP, etc.) to be removed. j Let be the dynamic weighting coefficient for the j-th pollutant, whose value is calculated in real time by the optimization decision agent based on the system status (e.g., normal operation, mild warning, severe warning); λ is the energy consumption penalty coefficient, Q i Let A be the hydraulic load of the i-th wetland sub-unit. i Let be the aeration intensity of the i-th wetland sub-unit.

[0043] In this embodiment, the pollutants to be removed from the wastewater are COD and TN. Therefore, the objective function expression for multi-objective optimization is: In the formula, η COD η represents the COD removal rate. TN w represents the removal rate of TN. COD w is the dynamic weighting coefficient for COD. TN λ is the dynamic weighting coefficient of TN; λ is the energy consumption penalty coefficient. The global optimal strategy is solved through Nash equilibrium.

[0044] (4) Execution Control Agent. The decision logic provided by the optimization decision agent is converted into real-time control instructions for the physical equipment and sent to the physical wetland unit. The real-time control instructions include: adjusting the water flow of the water supply components in the water distribution system by controlling the pump and valve opening, adjusting the power of the aerator in the aeration components, and adjusting the plant harvesting cycle. Preferably, the response time of the execution control agent is <10 seconds.

[0045] A method for treating wastewater from constructed wetlands based on multi-agent collaborative optimization includes the following steps: S1: The wastewater to be treated is input into the physical wetland unit for treatment and then output. Multi-source data of the physical wetland unit is collected through the sensor network and sent to the digital twin agent module.

[0046] S2: The environment perception agent of the digital twin agent module receives multi-source data and sends the multi-source dataset within the period to the prediction analysis agent according to the periodic prediction task (set to once per hour in this embodiment).

[0047] Specifically, when the concentration of pollutants at the outlet of the physical wetland unit exceeds the standard (e.g., TP>0.5mg / L) or there is a sudden change in water quality (e.g., the concentration of pollutants increases by 50%), an alarm signal is immediately triggered, and the alarm signal and real-time multi-source dataset are sent to the optimization decision agent, and the process jumps to execute S4.

[0048] S3: The predictive analysis agent receives the multi-source dataset sent by the environmental perception agent according to the periodic prediction task (once per hour in this embodiment), outputs the predicted future pollutant concentration change trend (the time limit obtained in this embodiment is within the next 24 hours), i.e. the water quality prediction result, and sends it to the optimization decision agent.

[0049] S4: In the normal path, the optimization decision agent receives the prediction results from the predictive analysis agent, obtains the decision logic, and sends it to the execution control agent. This is implemented through the following sub-steps: Step 1: Optimization Decision Agent. The agent first uses the water quality forecast results (such as the trend of pollutant concentration changes over the next 24 hours) from the predictive analysis agent as the core input to construct a specific mathematical optimization problem. All controllable physical parameters in the system are identified as decision variables. These decision variables collectively define the system's "operation space." The decision variables include: the hydraulic load Q of each wetland sub-unit. i Aeration intensity A of each wetland subunit i The pulse frequency T of the water distribution system i Activation status M of the standby adsorption matrix module i (0 means off, 1 means on).

[0050] To achieve the objectives of this invention, a comprehensive multi-objective function is established. This function aims to synergistically optimize pollutant removal rate, system energy consumption, and resistance to shock loads. Its typical mathematical expression is as follows: Use prediction results to define the boundaries of the problem. For example, if the prediction shows that a certain pollutant will exceed the standard in the next few hours, then impose a hard constraint that "the concentration of the pollutant in the effluent must be lower than the limit value". At the same time, the physical limits of the equipment itself (such as the maximum pump speed and the upper limit of aeration) also constitute constraints.

[0051] Step 2: Dynamic Weight Allocation. This is crucial for demonstrating the system's intelligence and adaptability. The optimization decision agent dynamically adjusts the dynamic weight coefficients w in the multi-objective optimization objective function based on the system state reflected in the prediction results. j To achieve different optimization strategies.

[0052] Step 3: Collaborative Solution – Seeking the Global Optimal Solution.

[0053] After defining the objective function with weights and constraints, the optimization decision agent uses a game-theoretic algorithm (such as Nash equilibrium) to solve it. Within this framework, each wetland sub-unit or its control agent is considered a "participant." To achieve fair and efficient collaboration, the optimization decision agent uses the Shapley value algorithm to calculate the marginal contribution (φ) of each participant in different cooperative alliances (S). i This calculation is based on a preliminary strategy formed from the local performance data (such as the pollutant gradient ∇C) reported by each wetland sub-unit. By integrating the contributions of all participants, a Nash equilibrium point for a set of decision variables (such as {Q1, A1, Q2, A2, ...}) is finally found. At this point, any change in any single parameter can no longer improve the overall system performance, which constitutes the globally optimal strategy.

[0054] Marginal contribution (φ) i The globally optimal expression, calculated through distributed negotiation (Shapley value algorithm), is as follows: In the formula, φ i Shapley value for participant i represents the contribution of a participant (i.e., a wetland sub-unit or its agent) to the overall optimization goal of the system. A higher value indicates a greater contribution of the wetland sub-unit to the collaborative processing. N represents the set of all participants in the system; for example, if the system has 5 connected units, N = {1, 2, 3, 4, 5}. S represents a coalition, a subset of participants; |S| is the size of the coalition, representing the number of participants in coalition S; for example, if S = {1, 2, 4}, it means that units 1, 2, and 4 form a temporary collaborative group, and |S| = 3. {i} represents the specific wetland sub-unit whose contribution is being calculated. v(S) is the value function of coalition S, representing the system performance index achievable when units within coalition S collaborate; in this embodiment, this can be a comprehensive score, such as (improved pollutant removal rate - energy consumption penalty). v(S∪{i}) is the coalition value excluding participant i (i.e., the achievable system performance index).

[0055] Step 4: Decision Logic Output. The optimal set of decision variables obtained from the solution is directly converted into structured decision logic. This decision logic is a clear "list of action instructions," for example: {Wetland Subunit 1: {Target hydraulic load: 0.25m} 3 / (m 2 ·d), Target aeration intensity: 75% (operating at 75% load under maximum aeration), Standby module: Off}, Wetland subunit 2: {Target hydraulic load: 0.30m 3 / (m2 ·d), Target aeration intensity: 85%, Backup module: On.

[0056] In the emergency response scenario, the optimization decision agent receives alarm signals from the environmental perception agent and corresponding real-time multi-source datasets. Based on the alarm signal warning level (mild or severe pollution risk), and with the goal of maximizing pollutant removal rate while minimizing energy consumption, it dynamically allocates optimization weights (such as prioritizing the adsorption capacity of which pollutant, how to adjust hydraulic load and aeration intensity, and determining the specific value of λ in the objective function of multi-objective optimization), obtains the decision logic, and sends it to the execution control agent.

[0057] In this embodiment, the alarm signal is generated by the total phosphorus (TP) concentration exceeding the standard. The specific decision logic is as follows: If the warning level is mild pollution risk, the decision logic is to increase the optimization weight of the "phosphorus removal" target by 5%-15%, while maintaining the overall system balance. If the warning level is severe pollution risk, the decision logic is to prioritize maximizing the "phosphorus adsorption capacity," at which point the weights of other targets such as energy consumption are temporarily reduced.

[0058] S5: The execution control agent receives the decision logic from the optimization decision agent and translates it into real-time control instructions for the physical equipment. If the warning level is a mild pollution risk, the control instructions include: ① Fine-tuning the water distribution system: extending the residence time of water flowing through the matrix filling layer rich in iron and aluminum oxides by 10%-20%. ② Fine-tuning aeration: increasing the aeration intensity by 10%-25% to change the redox potential and promote the conversion of soluble phosphates into solid precipitates.

[0059] If the warning level is severe pollution risk, the control instructions include: ① Activating the backup adsorption matrix module: Opening the inlet to the backup adsorption matrix module, allowing high-concentration phosphorus-containing wastewater to flow through the pre-embedded modified zeolite, activated alumina, and other highly efficient adsorption materials in the backup adsorption matrix module. ② Injecting chemical agents: Starting the dosing pump to add FeCl3 or PAC solution to the backup adsorption matrix module, rapidly fixing phosphorus through chemical precipitation. ③ Coordinated hydraulic control: Adjusting the water distribution pulse frequency, regulating the inlet flow rate of the water supply component and the outlet flow rate of the water collection component, further extending the hydraulic retention time (HRT) to ensure sufficient contact and reaction between the wastewater and the adsorbent and / or agents. ④ Adjusting aeration: Adjusting the aerator power to further increase aeration, enhance dissolved oxygen (DO), strengthen biodegradation capacity, and create conditions for subsequent biochemical recovery.

[0060] In this embodiment, the decision logic generated by the excessive total phosphorus (TP) concentration is transformed into the following control instruction: ① Activate the standby adsorption matrix module (increasing the capacity to treat wastewater by 20%-30%), start the dosing pump, and add FeCl3 or PAC solution to the standby adsorption matrix module; or start the screw conveyor to replenish the standby adsorption matrix module with fresh adsorption material.

[0061] ② Adjust the water distribution pulse frequency (i.e., the pulse frequency of the water supply component) (2-8 times / hour), adjust the water flow rate of the water collection component, and extend the HRT; adjust the power of the aerator in the aeration component (0-100%).

[0062] S6: The physical wetland unit achieves dynamic correction through bidirectional interaction between the digital twin model and the physical layer; specifically, based on the feedback of the control agent's regulation effect on the physical wetland unit (such as measured data after DO adjustment), the parameters of the water quality prediction model are dynamically corrected to improve prediction accuracy.

[0063] The physical wetland unit achieves dynamic correction through bidirectional interaction between the digital twin model and the physical layer, specifically including two processes: (1) Data-driven model parameter update: The residual between the measured and predicted water quality values ​​after the execution of the control agent (ε=C) 实测 -C 预测 With the goal of setting the target value below the prediction threshold, the network parameters of the water quality prediction model (LSTM) are dynamically adjusted using the backpropagation algorithm to improve its prediction accuracy.

[0064] (2) Simultaneously, the key state variables in the system are calibrated online based on measured data. For example, the actual adsorption capacity q of the matrix filling layer. e It will be updated based on runtime and measured performance; its empirical decay model can be expressed as follows: Where α is the matrix aging coefficient. After calibration The value can be used as a system state input, for optimizing decisions or as an input to assist the LSTM model.

[0065] To verify the effectiveness of the system and method of the present invention, the following embodiments are provided to compare the system of the present invention with ordinary artificial wetland systems.

[0066] The system of this invention was used as the experimental group, with key operations including real-time monitoring (COD, TN, TP, DO, etc., every 5 minutes), dynamic adjustment of aeration rate, hydraulic load, and activation of the backup adsorption substrate module. A conventional artificial wetland system (without intelligent control) was used as the control group, with its key parameters set as: fixed hydraulic load (0.3 m³ / s). 3 / (m 2 •d)), no aeration equipment or fixed aeration volume (if present), no emergency (backup) adsorption substrate module.

[0067] The three-phase test will be conducted as follows: (1) First stage: Steady-state operation test (7 days) ① Input conditions: Influent COD=200mg / L, TN=30mg / L, TP=5mg / L, and stable flow rate.

[0068] ② Monitoring indicators include: Pollutant removal rate: Record the concentrations of COD, TN, and TP in the effluent every hour (sensor accuracy ±2%).

[0069] Energy consumption: Electricity consumption of aerator and water distribution system in the experimental group (kW•h / d); if the control group includes aeration equipment, the fixed energy consumption is recorded.

[0070] The test results are shown in Table 1 below.

[0071] Table 1. Steady-state operating data (average values) (2) Second stage: Dynamic load test (simulating sudden pollution) ① Input conditions: On the 8th day, a sudden high concentration of wastewater was injected (COD suddenly increased to 400mg / L, lasting for 2 hours).

[0072] ② Monitoring indicators include: peak inhibition rate (the decrease in pollutant concentration within 2 hours) and recovery time (the time required for the system to recover to steady-state removal rate). The test results are shown in Table 2 below.

[0073] Table 2 Dynamic load test results (3) Third stage: Long-term stability test (30 days) ① Input conditions: Periodic water quality fluctuations (±20% COD, TN, or TP).

[0074] ② Monitoring indicators include: Matrix clogging rate: the rate of decrease in matrix layer permeability (m / d•month).

[0075] Operation and maintenance costs: The experimental group was calculated based on the amount of chemical reagents replenished, while the control group was calculated based on the frequency of manual cleaning.

[0076] The test results are shown in Table 3 below.

[0077] Table 3 Long-term stability data Although the experimental group consumed a certain amount of energy (approximately 120 kW•h / d), compared to the control group, the pollutant removal rate was increased by 15%-21.3%. Furthermore, by dynamically adjusting the system, the use of chemical reagents was reduced, resulting in a 67% reduction in long-term operation and maintenance costs. Through the setup of a backup adsorption matrix module and real-time intelligent control, the experimental group shortened the recovery time to 3 hours under sudden pollution events, compared to 12 hours for the control group relying on natural degradation, representing a 75% reduction in recovery time and demonstrating significant effectiveness. This proves that the system of this invention has significant technical advantages over ordinary artificial wetlands in terms of pollutant removal efficiency, dynamic response capability, and long-term stability, resulting in improved energy efficiency, faster emergency response, and reduced operation and maintenance costs. It also possesses dynamic adaptability, high energy efficiency, and scalability, making it suitable for urban sewage, agricultural non-point source pollution, and industrial wastewater treatment scenarios.

[0078] This invention employs a modular physical wetland unit and a digital twin agent module architecture for collaborative design. The physical wetland unit optimizes pollutant removal through a substrate filling layer, a plant community layer, and a dynamic water distribution system. The digital twin agent module integrates four types of intelligent agents: environmental perception, LSTM prediction, game theory multi-objective optimization, and rapid execution, adjusting parameters such as hydraulic load and aeration intensity in real time. Through bidirectional interaction between the multi-agent collaborative architecture and the digital twin model, the mechanistic model and the data-driven model are integrated to solve the model-algorithm disconnect problem, improving prediction accuracy and operational adaptability. Based on a game theory framework, dynamic coordination of multi-objective optimization (energy efficiency, removal rate, and shock resistance) globally improves pollutant removal efficiency and can cope with sudden pollution events. The combination of modular physical wetlands and a distributed intelligent system supports plug-and-play expansion, avoiding the single-point failure risk of centralized architectures. LSTM network prediction and real-time feedback closed-loop optimization, combined with high-precision execution control, achieve improved pollutant removal rate and reduced overall energy consumption. The backup adsorption substrate emergency mechanism and adaptive control strategy significantly enhance system stability and long-term operational reliability, reducing operation and maintenance costs.

[0079] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A constructed wetland wastewater treatment system based on multi-agent collaborative optimization, characterized in that, It includes a physical wetland unit and a digital twin agent module; the physical wetland unit includes several wetland sub-units connected in series or parallel; the wetland sub-unit includes: a backup adsorption matrix module, a plant community layer, a matrix filling layer, and a water distribution system arranged from top to bottom; the backup adsorption matrix module is arranged in isolation from the other components of the wetland sub-unit; a sensor network is set up to monitor multi-source data of the physical wetland unit in real time and send it to the digital twin agent module; The digital twin agent module includes: an environmental perception agent, a predictive analysis agent, an optimization decision agent, and an execution control agent; the environmental perception agent is used to receive the multi-source data and periodically send it to the predictive analysis agent; if the concentration of a certain pollutant exceeds the standard or the water quality changes abruptly, an alarm signal is triggered and sent to the optimization decision agent; The predictive analysis agent uses a water quality prediction model to predict water quality based on multi-source data and obtain water quality prediction results. The parameters of the water quality prediction model are dynamically updated by the predictive analysis agent using the measured water quality values ​​obtained after being regulated by the execution control agent. The optimization decision agent uses Nash equilibrium to coordinate energy efficiency, removal rate and shock resistance based on water quality prediction results or early warning levels in alarm signals to achieve multi-objective optimization, obtain the optimization results of the parameters of each component of the physical wetland unit, and generate decision logic. The execution control agent is used to convert the decision logic into real-time control instructions for physical wetland units.

2. The constructed wetland wastewater treatment system based on multi-agent collaborative optimization according to claim 1, characterized in that, The matrix filling layer, from top to bottom, includes: a coarse sand layer, a zeolite-volcanic rock mixed layer, and a gravel layer; The water distribution system includes: a water supply component, an aeration component, and a water collection component; the water supply component's inlet is located above the standby adsorption substrate module and is evenly distributed in the plant community layer, and its inlet flow rate is controlled by a water flow control pump and valves; the aeration component is located between the zeolite-volcanic rock mixed layer and the gravel layer, and its aeration rate is controlled by an aerator; the water collection component is located at the bottom of the gravel layer and is used to collect the treated wastewater and output it to the outlet of the wetland subunit.

3. The constructed wetland wastewater treatment system based on multi-agent collaborative optimization according to claim 2, characterized in that, The real-time control commands include: adjusting the water distribution pulse frequency of the water distribution system, adjusting the inlet flow rate of the water supply component and the outlet flow rate of the water collection component, adjusting the power of the aerator in the aeration component, and adjusting the plant harvesting cycle.

4. The constructed wetland wastewater treatment system based on multi-agent collaborative optimization according to claim 1, characterized in that, The multi-source data includes: organic matter concentration, total nitrogen, total phosphorus, pH, dissolved oxygen, water level, and redox potential.

5. The constructed wetland wastewater treatment system based on multi-agent collaborative optimization according to claim 1, characterized in that, The backup adsorption materials pre-embedded in the backup adsorption matrix module include: modified zeolite, activated alumina, and composite materials loaded with Fe and Mn. The backup adsorption matrix module is provided with a drug injection port and an adsorption material replenishment port at the upper end, which are used to inject specified reagents and add new adsorption materials into the backup adsorption matrix module, respectively.

6. The constructed wetland wastewater treatment system based on multi-agent collaborative optimization according to claim 1, characterized in that, In the optimization decision agent, the objective function for multi-objective optimization is: In the formula, η j w represents the removal rate of the j-th pollutant to be removed. j λ is the dynamic weighting coefficient for the j-th pollutant; λ is the energy consumption penalty coefficient, and Q... i Let A be the hydraulic load of the i-th wetland sub-unit. i Let be the aeration intensity of the i-th wetland sub-unit.

7. The constructed wetland wastewater treatment system based on multi-agent collaborative optimization according to claim 1, characterized in that, The water quality prediction model is built on an LSTM network. Its inputs are historical concentration sequences and real-time multi-source data, and its output is the trend of pollutant concentration changes over the next 24 hours.

8. A constructed wetland wastewater treatment method based on multi-agent collaborative optimization, implemented based on the constructed wetland wastewater treatment system based on multi-agent collaborative optimization as described in any one of claims 1-7, characterized in that, Includes the following steps: S1: The wastewater to be treated is input into the physical wetland unit for treatment and then output. Multi-source data of the physical wetland unit is collected through the sensor network and sent to the digital twin agent module. S2: The environment perception agent receives multi-source data and sends the multi-source dataset within the period to the prediction analysis agent according to the periodic prediction task; When multi-source data shows that pollutant concentration exceeds the standard or water quality changes abruptly, an alarm signal is immediately triggered, and the alarm signal and real-time multi-source dataset are sent to the optimization decision agent, and the process jumps to execute S4. S3: The predictive analysis agent receives the multi-source dataset sent by the environmental perception agent, uses the water quality prediction model to output the predicted future pollutant concentration change trend, and sends the predicted future pollutant concentration to the optimization decision agent. S4: In the normal path, the optimization decision agent receives the prediction results from the prediction analysis agent, obtains the decision logic, and sends it to the execution control agent; In the emergency response scenario, the optimization decision agent receives the alarm signal and the corresponding real-time multi-source dataset. Based on the warning level in the alarm signal, and with the goal of maximizing pollutant removal rate while minimizing energy consumption, it dynamically allocates the weight coefficients in the objective function of multi-objective optimization to obtain the decision logic, which is then sent to the execution control agent. The warning level includes mild pollution risk and severe pollution risk. S5: The execution control agent receives the decision logic from the optimization decision agent and transforms it into real-time control instructions for the physical wetland units; S6: The physical wetland unit achieves dynamic correction through bidirectional interaction between the digital twin model and the physical layer; based on the feedback of the control agent's regulation effect on the physical wetland unit, the parameters of the water quality prediction model are dynamically corrected to improve the prediction accuracy.

9. The constructed wetland wastewater treatment method based on multi-agent collaborative optimization according to claim 8, characterized in that, In S5, if the warning level is a light pollution risk, the control instructions include: ① fine-tuning the water distribution system: extending the residence time of water flowing through the matrix filling layer rich in iron and aluminum oxides, i.e., the hydraulic residence time, by 10%-20%; ② fine-tuning the aeration: increasing the aeration intensity by 10%-25%, changing the oxidation-reduction point, and promoting the conversion of soluble phosphates into solid precipitates. If the warning level is a severe pollution risk, the control instructions include: ① Opening the inlet to the backup adsorption matrix module to activate the backup adsorption matrix module, allowing the wastewater to be treated to flow through the backup adsorption material pre-embedded in the backup adsorption matrix module; ② Starting the dosing pump to add chemical agents to the backup adsorption matrix module, quickly fixing phosphorus through chemical precipitation; ③ Coordinating hydraulic control: adjusting the water distribution pulse frequency to extend the hydraulic retention time to more than 20%; ④ Adjusting and increasing aeration to increase the aeration intensity to more than 25%.

10. The constructed wetland wastewater treatment method based on multi-agent collaborative optimization according to claim 8, characterized in that, In multi-agent collaborative optimization, a preliminary strategy is generated based on local performance data of each wetland subunit, and then the marginal contribution of each participant in different cooperative alliances is calculated through distributed negotiation to achieve global optimality. The participants are wetland subunits or agents of that unit. The distributed negotiation uses the Shapley value algorithm, and the marginal contribution calculation expression is as follows: In the formula, φ i Let $\frac{i}{i}$ be the contribution of participant $i$ to the overall optimization objective of the system, $N$ be the set of all participants in the system, $S$ be the alliance, representing a subset of participants, $|S|$ be the size of the alliance, representing the number of participants in the alliance $S$, ${i}$ be the specific wetland sub-unit whose contribution is being calculated, $v(S)$ be the value function of the alliance $S$, representing the system performance index that can be achieved when the units in the alliance $S$ cooperate with each other, and $v(S∪{i})$ be the system performance index that can be achieved without participant $i$.

Images