An intelligent control method for a three-stage artificial wetland recirculating aquaculture system

Abstract

This invention provides an intelligent control method for a three-stage artificial wetland recirculating aquaculture system, belonging to the field of machine learning technology. First, water quality data and operational control data from each control unit are continuously collected to construct a multi-source heterogeneous dataset on a unified time scale. Subsequently, a pollution evolution trend prediction model is constructed to capture the dynamic evolution trend of water quality over time and changes in control behavior. Then, guided by the prediction results, by analyzing the similarity of historical states and the sensitivity of control responses, the key control parameters currently affecting the system's water quality changes are automatically identified and their dynamically adjustable boundaries are inferred. Finally, a reinforcement learning strategy network is constructed that integrates state prediction, parameter boundaries, and control objectives. This network achieves a multi-objective trade-off between pollutant removal efficiency, water quality compliance rate, and operational energy consumption, and outputs an efficient and robust control strategy through continuous interactive training. This invention can achieve efficient, accurate, and robust operation of the wetland system.

Description

Technical Field

[0001] The present invention belongs to the field of machine learning technology, and in particular relates to an intelligent control method for a three-level artificial wetland recirculating aquaculture system. Background Art

[0002] With the increasing demand for ecological and environmental protection, artificial wetlands, as a cost-effective means of water purification and ecological restoration, have been widely used in aquaculture, rural sewage treatment, and agricultural non-point source pollution control. In particular, in recirculating aquaculture systems, three-level artificial wetlands, through the construction of different functional zones in series and parallel, achieve efficient removal of pollutants such as nitrogen, phosphorus, and organic matter, ensuring the ecological safety of water recycling. However, due to factors such as complex environmental disturbances, frequent fluctuations in water quality indicators, and nonlinear coupling of control parameters, current wetland systems generally suffer from problems such as reliance on experience in the regulation process, delayed regulation response, and rigid control strategies, making it difficult to accurately perceive the dynamic environment and achieve real-time optimization and scheduling. In this context, developing an intelligent management method that can integrate multi-source environmental information, dynamically identify key system states, and possess adaptive regulation capabilities has become a core requirement for improving the operational efficiency and intelligence level of wetland systems.

[0003] At present, the regulation of artificial wetland recirculating aquaculture systems mainly includes the following methods:

[0004] (1) Artificial experience control method: Artificial experience control method is still widely used in many small and medium-sized farms. It mainly relies on on-site operators to make adjustments based on experience judgment and historical water quality monitoring data. For example, when the water becomes turbid or the odor becomes worse, the operator will subjectively judge whether it is necessary to increase the return ratio, extend the hydraulic retention time, or adjust the pump operating frequency. This method has the advantages of simple operation and low investment cost, but the control is highly dependent on individual experience and lacks data support. It is prone to problems such as response lag, inaccurate control, and non-replicable operation. Especially when facing complex and dynamically changing water quality conditions, the control effect fluctuates greatly;

[0005] (2) Single parameter feedback control method: The single parameter feedback control method is based on real-time monitoring of a key water quality indicator (such as ammonia nitrogen concentration or dissolved oxygen concentration). When the monitoring value exceeds the set threshold, the system automatically triggers the corresponding control instruction, such as adjusting the water intake or starting the aeration device, thereby forming a simple closed-loop control. This method has a certain degree of real-time and automation and can achieve rapid response, but its control logic usually revolves around a single variable, ignoring the multi-factor coupling between water quality indicators and the overall operating status of the system, which can easily lead to control imbalance, insufficient regulation or system oscillation. In particular, it is difficult to ensure system stability in the scenario of the coordinated evolution of multiple pollutants;

[0006] (3) Rule base + PLC automation control method: This method is a relatively mature engineering control method that manages the operating parameters of the wetland system through preset "condition-action" logic (such as IF-THEN rules). A typical practice is: "If the ammonia nitrogen is greater than 1.5 mg / L and the water temperature is higher than 25°C, then increase the pump operating frequency to 80%." This type of method uses a programmable logic controller to automatically execute control commands and has strong engineering adaptability and execution efficiency. However, this method relies on pre-set fixed rules and lacks the ability to adapt to new scenarios. Rule updates rely on manual adjustments, making it difficult to dynamically optimize or explain the complex logic in the control process. Once the system state deviates from the scope of application of the rules, it may fail or be miscontrolled. In particular, its performance is limited in highly nonlinear and multivariable interference scenarios. Summary of the Invention

[0007] In response to the above problems, the present invention proposes an intelligent control method for a three-stage artificial wetland recirculating aquaculture system, comprising the following steps:

[0008] S1, based on the key structural units of the three-level artificial wetland system and the multi-type environmental sensors deployed in the aquaculture pond, continuously collects water quality time series data and control time series data, and constructs multi-source heterogeneous data under a unified time scale;

[0009] S2 inputs the water quality time series data and control time series data collected by S1 into the constructed and trained pollution evolution trend prediction model, and outputs multi-dimensional water quality prediction results;

[0010] S3 averages and aggregates the multi-dimensional water quality prediction results of S2 in the time dimension to obtain the state vector to be evaluated; performs unsupervised clustering on the historical water quality time series data set to obtain the cluster center and the corresponding cluster; performs state similarity matching and response sensitivity analysis on the state vector to be evaluated and the cluster center to obtain the key control parameters with the most regulatory value at the current moment, and infers their safe and adjustable dynamic response range;

[0011] S4, based on the constructed reinforcement learning control strategy network, takes water quality time series data, control time series data and multi-dimensional water quality prediction results as environmental state input, key control parameters as adjustable actions, and dynamic response intervals as action constraints. It designs a comprehensive reward function under the multi-objective optimization goals of maximizing pollutant removal rate, stably meeting water quality indicators, and minimizing system energy consumption, and generates the optimal control parameter adjustment action through the reinforcement learning strategy network.

[0012] Furthermore, the parameters of the water quality time series data include pH value , dissolved oxygen concentration , ammonia nitrogen concentration , nitrite concentration , water temperature and turbidity ;

[0013] The parameters of the control time series data include water flow , water flow , hydraulic retention time , wetland module enabled status , valve opening , Pump operating frequency Reflux ratio ;

[0014] Multi-dimensional water quality prediction label parameters including pollutant indicators ; Reaction environment indicators ; System disturbance index ; together constitute the water quality prediction label combination .

[0015] Furthermore, the data set collection methods for model training and obtaining key control parameters and their dynamically adjustable response ranges include:

[0016] Input water quality time series data collection: for each control unit, at the time point Collect the surrounding water quality data of all control units to obtain the water quality data at that moment ,in Indicates at the time point , No. Water quality data collected at each control unit, and Based on this principle, we collected Complete input water quality time series data at each time point ,and ;

[0017] Input control timing data collection: at time point Collect control data corresponding to all control units ;in, Indicates at the time point , No. The control data collected from each control unit, and Similarly, collection Complete input control timing data at each moment ,and .

[0018] Furthermore, the pollution evolution trend prediction model includes a water quality dynamic perception channel, a control parameter encoding channel, and a fusion decoding prediction channel;

[0019] The water quality dynamic perception channel introduces a three-branch parallel structure, and different branches are designed with different receptive fields to enhance the model's sensitivity to the changing trends of various water quality characteristics over time; the input is the collected water quality time series data, capturing short-term and medium-term trends to generate water quality dynamic coding features ;

[0020] The control parameter encoding channel input is the collected control time series data, and the gated fusion mechanism is introduced to perform weighted integration on multiple control dimensions, dynamically adjust the representation contribution of different control variables in a specific state, and output the gated fusion feature. ;

[0021] Then, the fusion features are controlled The first enhanced GRU unit is fed into the initial time series modeling; after that, the output of the first GRU unit is combined with the gated fusion feature The cascade is fed into the second enhanced GRU unit to characterize the deep interaction relationship of the control variables under the state combination conditions; the output of the second GRU unit is processed by layer normalization to obtain the final feature of the first stacked unit ;

[0022] Based on the above structure, five stackings are performed to extract multi-layer temporal features from local response to global regulation layer by layer, and finally the characterization features of the control parameter encoding channel are obtained through the Sigmoid activation function. ;

[0023] The fusion decoding prediction channel dynamically encodes the water quality characteristics Characterization characteristics of the control parameter encoding channel Perform splicing and fusion in the channel dimension, and output multi-dimensional water quality prediction results based on the dual Transformer decoding layer-gating fusion structure .

[0024] Furthermore, the specific structure of the water quality dynamic sensing channel includes:

[0025] Short-term perception branch: Water quality time series data first passes through The dilated convolution block is used to extract local fluctuation features, and then The dilated convolution block further expands the short-term context perception range and obtains short-term dynamic features ;

[0026] Long-term perception branch: Water quality time series data passes through Dilated convolutional blocks and The dilated convolution block uses a larger dilation rate to capture the trend evolution law across time slices, which is used to model the cumulative impact of changes in operating conditions on pollutants and output medium-term evolution characteristics. ;

[0027] Original retention branch: retain the original water quality time series data as the residual connection path, retain low-level detail information, and form the basic retention feature ;

[0028] Afterwards, the output features of the three branches are concatenated in the channel dimension to form a fusion feature , and sent to the global average pooling layer to enhance the global representation ability, and the fusion result is input into the next round of three-branch parallel structure for deep feature extraction; the above structure is repeatedly stacked for a total of 5 layers to construct a progressive multi-scale time series modeling path; after the output of the last layer, the Sigmoid activation function is connected to enhance the nonlinear expression ability, and finally the dynamic coding feature of water quality is generated. .

[0029] Furthermore, the specific structure of the fusion decoding prediction channel includes:

[0030] First, the dynamic coding characteristics of water quality Characterization characteristics of the control parameter encoding channel Splicing is performed in the channel dimension; then, the spliced ​​features are sent to the gated fusion mechanism to obtain the interactive features ;

[0031] Interaction Features The model is fed into two consecutive Transformer decoding layers to capture long-range dependencies and potential sequence structures between multi-dimensional features. Each Transformer decoding layer includes a multi-head attention mechanism, a feedforward neural network, and a residual connection structure. The Transformer decoding results are then fed into a gated fusion mechanism.

[0032] The dual Transformer decoding layer-gating fusion constitutes a decoding module unit. This decoding module unit is repeatedly stacked in 5 layers to form a multi-layer information abstraction and trend reasoning path; it is finally sent to the Dropout layer to enhance the generalization ability, and then processed by the Linear mapping layer and activation function to output the multi-dimensional water quality prediction results. .

[0033] Furthermore, the specific process of S3 includes:

[0034] Construction of the state vector to be evaluated: multi-dimensional water quality prediction results for the output , average aggregation is performed on each control unit in the time dimension to obtain the state vector to be evaluated at the current control moment ;

[0035] Historical status data aggregation analysis: Based on the constructed water quality time series dataset, the water quality status data at each single moment in the past is extracted from it , and the scale is The historical water quality state sample set is obtained by performing unsupervised clustering on all historical water quality state vectors. Cluster centers ,in Indicates the cluster centers, and ;

[0036] Next, calculate the current state vector to be evaluated The Euclidean distance from all cluster centers and a set of distance vectors are obtained ;in Represents the state vector to be evaluated With the The Euclidean distance between cluster centers;

[0037] After that, select the cluster center closest to , and extract all historical water quality status data in the corresponding cluster as the candidate sample set; further calculate all samples in the cluster and The cosine similarity between them; finally the top with the highest similarity is selected The control time series data corresponding to the historical sample water quality time series data , which constitutes a representative control sample set for response analysis ;

[0038] Response sensitivity assessment and key control parameter screening: for control sample sets For each set of control parameter data in the , a disturbance experiment is performed to calculate each control parameter The corresponding sensitivity coefficient ;in, and Respectively represent control parameters and their sensitivity coefficients;

[0039] Finally, the sensitivity coefficient will be satisfied Greater than the preset threshold The parameters are used as the key control parameter set at the current control moment ;

[0040] Based on a set of key control parameters, the safe and adjustable dynamic response range is inferred.

[0041] Furthermore, the process of inferring a safe and adjustable dynamic response range based on a set of key control parameters is as follows:

[0042] For each key parameter , from the control sample set Extract the maximum and minimum values ​​from the historical values ​​to form key parameters Initial response range: ,in Represents parameters respectively The minimum and maximum values ​​in the control sample set;

[0043] Afterwards, combined with the system preset operation safety constraints Perform double boundary clipping to obtain the dynamically adjustable response range of this parameter:

[0044] ;

[0045] in, Parameters for the system Preset minimum and maximum values, and To obtain the maximum and minimum value operations; Key parameters The corresponding dynamically adjustable response range and the structured control parameter-response range pair ;

[0046] Based on the The corresponding dynamically adjustable response intervals of all key parameters in the set are obtained to obtain the response interval set ; The complete set of control parameter-response interval pairs is ,in ,and .

[0047] Furthermore, the reinforcement learning control strategy network constructed in S4 adopts an Actor-Critic dual-head architecture;

[0048] First, the current water quality time series data , control timing data and predicted status The data are concatenated into a unified state vector and sent to the network input layer for normalization and feature embedding to extract basic information.

[0049] The feature extraction layer consists of a three-layer fully connected network and a ReLU activation unit, which performs high-dimensional feature conversion on the state vector and extracts deep control signal features;

[0050] Strategy Head-to-Head Action Each key control parameter in the Gaussian distribution is constructed to output the value, and finally the expected value and corresponding variance of the control action are output, and the corresponding value range is obtained; that is, the value range of each potential key variable is: the value range of the water inlet flow , the value range of the water flow rate , the range of hydraulic retention time , wetland module state discrete value interval , 1 is enabled, 0 is disabled; the range of valve opening ; Pump frequency range , the range of reflux ratio ; Afterwards, through the dynamic response interval Further define the action constraints of each key variable and perform Sub-sampling, to obtain a set of potential actions ;

[0051] Value header calculations are performed separately for potential actions The long-term cumulative expected rewards that can be obtained , in order to guide network optimization; the value head consists of two layers of fully connected networks, which serves as the Critic end in the Actor-Critic framework to provide a gradient estimation basis for policy optimization and guide the policy network to converge to a better solution.

[0052] Furthermore, the environmental state is composed of the water quality time series data of the control system at the current moment, the control time series data and the multi-dimensional water quality prediction results; including the current water quality data ,for Time series data of water quality of all control units at all times , real-time control data ,for The executed control timing data corresponding to all control units at the moment ; Prediction status , for the future Multi-dimensional water quality prediction results at all control units ;

[0053] Among them, action Key control parameters for each control unit The adjustment amount; the action value range is determined by the dynamic response interval constraint;

[0054] Among them, the reward function It is a weighted multi-objective comprehensive reward mechanism that encourages efficient pollutant removal, stable index compliance, and minimization of energy consumption, that is, the multi-objective comprehensive reward function:

[0055] ;

[0056] in, is the pollutant removal rate, which is defined as the normalized average of the changes in ammonia nitrogen and nitrite concentrations per unit time; is the water quality compliance rate, which represents the weighted accumulation of the number of times all water quality parameters meet the standards in each control unit; is the total energy consumption of the system, which is represented by the normalized average value of pump frequency and valve opening.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] (1) Pollution evolution trend prediction mechanism based on dual-channel fusion modeling: A joint prediction structure integrating the “dynamic water quality perception channel” and the “control parameter encoding channel” is designed to achieve forward-looking modeling of pollutant concentration change trends; the model’s ability to express sudden changes and delayed responses is improved through the parallel multi-scale dilated convolution and gated GRU structure, effectively enhancing the system’s perception and prediction performance of complex dynamic water quality evolution laws;

[0059] (2) Key control parameter screening method based on state similarity and response sensitivity: A parameter screening strategy combining clustering-matching mechanism and disturbance sensitivity analysis is proposed. It can automatically identify the key control parameters that are most sensitive to the current pollution evolution without introducing manual experience, and infer their dynamic response range based on historical control boundaries, significantly improving the strategy search efficiency and the accuracy of control parameter selection.

[0060] (3) A reinforcement learning control framework integrating prediction guidance and strategy optimization: A control strategy network based on the Actor-Critic structure is constructed, and the water quality prediction results are introduced into the state definition process to achieve dynamic enhancement of the strategy input; at the same time, a multi-objective weighted reward function and PPO optimization mechanism are adopted to jointly optimize the pollutant removal efficiency, water quality compliance and energy consumption control, so as to achieve efficient convergence of the control strategy and multi-objective coordination. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is the overall flow chart of the intelligent control method of the present invention.

[0062] Figure 2 This is the network structure diagram of the pollution evolution trend prediction model.

[0063] Figure 3 Generate flow charts for key control parameter screening and response intervals.

[0064] Figure 4 This is a graph comparing the RMSE results of different models in the pollutant prediction task according to an embodiment of the present invention.

[0065] Figure 5 This is a comparison chart of the MAE results of different models in the pollutant prediction task according to the embodiment of the present invention.

[0066] Figure 6This is a heat map of the sensitivity of key control parameters in an embodiment of the present invention. DETAILED DESCRIPTION

[0067] This invention proposes a three-level artificial wetland recirculating aquaculture system intelligent control method, the overall technical path is as follows Figure 1 As shown:

[0068] S1, Construction of a Multi-source Environmental Time Series Dataset: Multiple types of environmental sensors are deployed in the key structural units and aquaculture ponds of the three-level constructed wetland system to continuously collect water quality time series data and control time series data, constructing multi-source heterogeneous data at a unified time scale; this provides comprehensive data support for subsequent model training and inference;

[0069] S2, pollution evolution trend prediction model construction: The water quality time series data and control time series data collected in S1 are input into the constructed and trained pollution evolution trend prediction model, and multi-dimensional water quality prediction results are output, including pollutant indicators, reaction environment indicators, and system disturbance indicators. These are used for key control parameter screening in S3 and for prospective guidance of regulation strategies in S4.

[0070] S3, key control parameter screening and response range generation: Guided by the prediction results of S2, combined with the historical operation database, state similarity matching and response sensitivity analysis are performed to identify the key control parameters with the most control value at the current moment, and their safe and adjustable dynamic response ranges are inferred, providing parameter boundaries and search space constraints for subsequent strategy optimization;

[0071] S4, reinforcement learning-driven optimal control strategy generation: Based on the multi-dimensional water quality prediction results output by the S2 module and the key control parameters and their dynamically adjustable response ranges determined by the S3 module, a reinforcement learning control strategy network is constructed; the current system state (including real-time water quality, control data and prediction information) is used as the environmental state input, and the key control parameters are used as adjustable actions. A comprehensive reward function is designed under the multi-objective optimization goals of maximizing pollutant removal rate, stably meeting water quality indicators, and minimizing system energy consumption. The optimal control parameter adjustment action is generated through the reinforcement learning strategy network to achieve intelligent collaborative scheduling and dynamic control strategy output for multiple control units.

[0072] The invention will be further described below with reference to specific embodiments.

[0073] 1. Construction of Multi-source Environment Time Series Dataset

[0074] This paper aims to achieve dynamic prediction and intelligent control of key water quality indicators in wetland systems. Therefore, we first construct a standardized multi-source time series dataset that integrates water quality parameters and operational status parameters. Specific steps include:

[0075] 1. Selection of water quality parameters and operation control parameters: Based on the operating characteristics of the wetland system and the laws of water quality evolution, select parameter types that have a significant impact on pollutant changes and water purification processes to form a water quality-operation joint parameter combination;

[0076] (1) Water quality parameter selection: Based on the "Surface Water Environmental Quality Standard" and the pollutant transformation law in the typical artificial wetland water purification process, six most representative water quality monitoring indicators are selected, including: pH value , dissolved oxygen concentration , ammonia nitrogen concentration , nitrite concentration , water temperature and turbidity ;Together constitute the water quality parameter combination ;

[0077] (2) Control parameter selection: Based on the common adjustment methods and energy efficiency management mechanisms in the operation and control of wetland systems, seven parameters closely related to hydraulic process control and facility start-up and shutdown status are selected, including: water inflow , water flow , hydraulic retention time , wetland module enabled status , valve opening , Pump operating frequency Reflux ratio ;Together constitute the operating parameter combination ;

[0078] 2. Multi-unit input data collection: Wetland systems are usually composed of multiple functional modules connected in series and parallel. The processing capacity and water quality response of each module vary in space. Therefore, at each main control unit (total (a) deploy corresponding environmental sensors to collect water quality parameter data and monitor the corresponding operating parameters of each control unit; specifically,

[0079] (1) Input water quality time series data collection: For each control unit, at the time point Collect the surrounding water quality data of all control units to obtain the water quality data at that moment ,in Indicates at the time point , No. Water quality data collected at each control unit, and Based on this principle, we collected Complete input water quality data at each time point ,and ;

[0080] (2) Input control timing data collection: at time point Collect control data corresponding to all control units ;in, Indicates at the time point , No. The control data collected from each control unit, and Similarly, collection Complete input control data at each time point ， and ;

[0081] Therefore, the input data of the multi-source environmental time series dataset include: water quality data at all control units , and the control data corresponding to all control units ;

[0082] 3. Water quality prediction label: The purpose of model prediction includes: (1) to determine key control parameters; (2) to enhance the state of reinforcement learning; therefore, the water quality prediction label needs to be sensitive to the control parameters and cover the potential risk trend of the system. Therefore, the prediction label parameters include pollutant indicators. ; Reaction environment indicators ; System disturbance index ; together constitute the water quality prediction label combination ;

[0083] After that, set the prediction time step to , collection time point Water quality data at all control units ,in Indicates at the time point , No. Water quality prediction label data collected from each control unit, and , ;

[0084] Finally, collect The water quality prediction label data corresponding to all time points is used to obtain the complete water quality prediction label , and use it as the output data of the multi-source environment time series dataset;

[0085] 4. Construction of multi-source environment time series data set: collect the input data and , and the constructed water quality prediction label output data Perform one-to-one alignment and combine to form a complete multi-source environment input and output data pair. Based on the above method, a total of The input and output samples with a unified time base are grouped together, and finally a multi-source environment time series dataset is constructed.

[0086] 2. Design of a pollution evolution trend prediction model

[0087] In order to achieve the prediction and intelligent control of key water quality indicators in wetland systems, this paper designs a pollution evolution trend prediction model based on the construction of a multi-source environmental time series data set, which is oriented to multi-source input and supports multi-step prediction. The model includes a water quality dynamic perception channel, a control parameter encoding channel, and a fusion decoding prediction channel. The model structure is as follows: Figure 2 As shown;

[0088] 1. Design of dynamic water quality sensing channel

[0089] In order to effectively capture the dynamic evolution of water quality indicators in wetland systems at multiple time scales, this paper designs a water quality dynamic perception channel to enhance the model's sensitivity to the temporal trends of various water quality characteristics. The input of this channel is the collected water quality time series data. ;

[0090] First, to fully explore the temporal dependencies at different time scales, the channel introduces a three-branch parallel structure. Different branches have different receptive field designs to capture short-term and medium- to long-term trends. The structure is as follows:

[0091] (1) Short-term perception branch: First pass The dilated convolution block is used to extract local fluctuation features, and then The dilated convolution block further expands the short-term context perception range, thereby characterizing the response pattern of water quality parameters under short-term disturbances, and finally obtaining the short-term dynamic features ;

[0092] (2) Long-term perception branch: Pass by in sequence Dilated convolutional blocks and The dilated convolution block uses a larger dilation rate to capture the trend evolution law across time slices, which is used to model the cumulative impact of changes in operating conditions on pollutants and output medium-term evolution characteristics. ;

[0093] (3) Original retention branch: retain the original water quality time series data as the residual connection path, retain low-level detail information, and form the basic retention feature ;

[0094] Afterwards, the output features of the three branches are concatenated in the channel dimension to form a fusion feature , and sent to the Global Average Pooling (GAP) layer to enhance the global representation ability. The fusion result is then input into the next round of three-branch parallel structure for deep feature extraction. The above structure is repeatedly stacked for a total of 5 layers to construct a progressive multi-scale time series modeling path. After the output of the last layer, the Sigmoid activation function is connected to enhance the nonlinear expression ability, and finally the dynamic coding feature of water quality is generated. ;

[0095] in, The dilated convolution block consists of the following three parts: First, the kernel is , the expansion rate is of Dilated The TCN layer is used to expand the receptive field while maintaining computational efficiency, enabling modeling of non-uniform temporal dependencies. Feature processing is then achieved through layer normalization and ReLU activation. This structure supports multi-scale parallel perception and is suitable for multi-layer extraction and dynamic modeling of temporal features in complex environments.

[0096] 2. Control parameter encoding channel design

[0097] In view of the characteristics of the operating parameters in the wetland system, such as diverse dimensions and obvious control delay effects, a control parameter response modeling channel was invented to enhance the model's dynamic expression ability and control sensitivity to changes in key operating states. The input of this channel is the collected operation control time series data. ;

[0098] First, a gated fusion module is introduced to make full use of the joint effects of various operating parameters in different time periods. This module performs weighted integration on multiple control dimensions, thereby dynamically adjusting the representation contribution of different control variables in a specific state and outputting gated fusion features. ;

[0099] Then, the fusion features are controlled The first enhanced GRU unit is sent to perform preliminary time series modeling, so as to use the time series network with memory mechanism to deal with the nonlinear dependence characteristics of the wetland system operation process; then, in order to further enhance the joint modeling ability between control behavior and system response, the output of the first GRU unit is combined with the gated fusion feature The cascade is fed into the second enhanced GRU unit to characterize the deep interaction relationship of the control variables under the state combination conditions; the output of the second GRU unit is processed by layer normalization to obtain the final feature of the first stacked unit ;

[0100] In order to achieve a progressive understanding and deep abstract expression of the evolution law of control parameters, this module is stacked five times based on this structure, extracting multi-layer temporal features from local response to global regulation layer by layer, and finally obtaining the representation features of the control parameter encoding channel through the Sigmoid activation function. ;

[0101] The enhanced GRU unit consists of the following components: first, it uses the standard GRU network to implement time series dependency modeling, then embeds a time series pooling module to improve the representation of sudden and slowly changing control modes, and finally, combines the ReLU activation function to enhance the nonlinear modeling of complex control responses. This structure can improve the compressed expression of control inputs while maintaining temporal structure coherence.

[0102] 3. Fusion decoding prediction channel design

[0103] To achieve high-precision prediction of the future water quality status of the wetland system, this module designs a fusion decoding prediction channel to jointly model the deep correlation between water quality evolution characteristics and operation control dynamics;

[0104] First, the dynamic coding characteristics of water quality Characterization characteristics of the control parameter encoding channel The splicing is performed in the channel dimension; then, the splicing features are sent to the gate control fusion mechanism to adaptively adjust the contribution weights of water quality and control information and strengthen their key interaction paths to obtain the interaction features ;

[0105] On this basis, the interactive features The data is fed into two consecutive Transformer decoding layers to capture long-range dependencies and potential sequential structures between multi-dimensional features. Each Transformer decoding layer incorporates a multi-head attention mechanism, a feedforward neural network, and a residual connection structure. The Transformer decoding results are then fed into a gated fusion mechanism to further enhance response consistency and prediction reliability across key feature dimensions.

[0106] The above structure (dual Transformer decoding layer-gated fusion) constitutes a decoding module unit with the ability to collaboratively process features. The decoding module unit is repeatedly stacked in 5 layers to form a multi-layer information abstraction and trend inference path, which is finally sent to the Dropout layer to enhance generalization ability. It is then processed by the Linear mapping layer and activation function to output multi-dimensional water quality prediction results. ;

[0107] 4. Model training: Based on the constructed multi-source environmental time series dataset, this paper uses Mean Squared Error (MSE) as the loss function to measure the water quality prediction results output by the model. with actual water quality labels The model parameters are iteratively updated through the SGD stochastic gradient descent algorithm, and the training is terminated after reaching the preset maximum number of training rounds, and finally a prediction model with the ability to model the pollution evolution trend is obtained.

[0108] 3. Key Control Parameter Screening and Response Range Generation

[0109] In order to achieve a balance between the regulation efficiency and execution accuracy of the three-stage artificial wetland circulation system in dynamic control, this paper proposes a key control parameter screening and response interval generation method based on historical state similarity and response sensitivity analysis; this process is guided by the multi-dimensional water quality prediction results output by the model. The overall process is as follows Figure 3 As shown;

[0110] 1) Construction of the state vector to be evaluated: Multi-dimensional water quality prediction results output by the model , average aggregation is performed on each control unit in the time dimension to obtain the state vector to be evaluated at the current control moment ;

[0111] 2) Historical state data aggregation analysis: Based on the water quality data in the constructed multi-source environmental time series data set, the data at each single moment in the past are extracted ( ) water quality status data , and the scale is The historical water quality state sample set is obtained; then, in order to improve the similarity matching efficiency, all historical water quality state vectors are clustered unsupervisedly to obtain Cluster centers ,in Indicates the cluster centers, and ;

[0112] Next, calculate the current state vector to be evaluated The Euclidean distance from all cluster centers and a set of distance vectors are obtained ;in Represents the state vector to be evaluated With the The Euclidean distance between cluster centers;

[0113] After that, select the cluster center closest to , and extract all historical water quality status data in the corresponding cluster as the candidate sample set; further calculate all samples in the cluster and The cosine similarity between them; finally the top with the highest similarity is selected The control parameter data corresponding to the historical water quality sample data , which constitutes a representative control sample set for response analysis ;

[0114] 3) Response sensitivity assessment and key control parameter screening: for control sample sets Each set of control parameter data in , which corresponds to the specific values ​​of 7 control parameters (including water flow , water flow , hydraulic retention time , wetland module enabled status , valve opening , Pump operating frequency Reflux ratio );

[0115] Afterwards, disturbance experiments were performed to calculate each control parameter The corresponding sensitivity coefficient Specifically, the gradient approximation is used to calculate the cumulative response intensity of each control parameter as the sensitivity coefficient ;in, and Respectively represent control parameters and their sensitivity coefficients, and , corresponding to the above 7 control parameters respectively;

[0116] Finally, the sensitivity coefficient will be satisfied Greater than the preset threshold The parameters are used as the key control parameter set at the current control moment The parameters in this set are considered to have the most significant impact on the current water quality evolution trend and are the core variables in the design of subsequent control strategies;

[0117] 4) Dynamically adjustable response interval reasoning: For each selected key parameter , from the control sample set Extract the maximum and minimum values ​​from the historical values ​​to form key parameters Initial response range: ,in Represents parameters respectively The minimum and maximum values ​​in the control sample set;

[0118] Afterwards, combined with the system preset operation safety constraints Perform double boundary clipping to obtain the dynamically adjustable response range of this parameter:

[0119] ;

[0120] in, Parameters for the system The preset minimum and maximum values, and To obtain the maximum and minimum value operations; Key parameters The corresponding dynamically adjustable response range and the structured control parameter-response range pair ;

[0121] Based on this method, The corresponding dynamically adjustable response intervals of all key parameters in the set, and the response interval set is obtained ; Therefore, the complete set of control parameter-response interval pairs is ,in ,and .

[0122] 4. Generation of Optimal Control Strategies Driven by Reinforcement Learning

[0123] Future water quality prediction results based on output and the selected key control parameters and their response ranges , build a reinforcement learning (RL)-driven optimal strategy generation network, dynamically output the optimal control action sequence for multiple control units, and achieve multi-objective collaborative optimization of maximizing pollutant removal efficiency, stably meeting water quality indicators, and minimizing system operation energy consumption;

[0124] 1. Definition of the reinforcement learning framework: The core is a reinforcement learning controller based on policy optimization. Its key elements are designed as follows:

[0125] (1) Define the state: The environmental state is composed of the water quality time series data of the control system at the current moment, the control time series data and the multi-dimensional water quality prediction results. Specifically including: current water quality data ,for Water quality data of all control units at all times . Real-time control data ,for The executed control data corresponding to all control units at the moment ; Prediction status , for the future of output Water quality prediction results at all control units , used to assist in forward-looking regulation;

[0126] (2) Define actions :Key control parameters for each control unit The action value range is determined by the dynamic response range provided by the S3 module. constraint;

[0127] (3) Reward Function :Design a weighted multi-objective comprehensive reward mechanism to encourage efficient pollutant removal, stable performance of indicators, and minimization of energy consumption. That is, the multi-objective comprehensive reward function:

[0128] ;

[0129] in, is the pollutant removal rate, which is defined as the standardized average value of ammonia nitrogen and nitrite per unit time; the standardized average value of the change in acid salt concentration; is the water quality compliance rate, which represents the weighted accumulation of the number of times all water quality parameters meet the standards in each control unit; is the total energy consumption of the system, which is represented by the normalized average value of pump frequency and valve opening;

[0130] 2. Policy Architecture Network Design: To enhance the expressiveness and generalization performance of policy generation, the policy network structure adopts an Actor-Critic dual-head architecture. Specifically:

[0131] (1) First, the current water quality data , real-time control data , and predicted status The data are concatenated into a unified state vector and sent to the network input layer for normalization and feature embedding to extract basic information.

[0132] (2) The feature extraction layer consists of a three-layer fully connected network and a ReLU activation unit, which performs high-dimensional feature conversion on the state vector and extracts deep control signal features;

[0133] (3) Strategy head-to-head action Each key control parameter in the Gaussian distribution is constructed to output the value, and finally the expected value and corresponding variance of the control action are output, and the corresponding value range is obtained; that is, the value range of each potential key variable is: the value range of the water inlet flow , the value range of the water flow rate , the range of hydraulic retention time , discrete value interval of wetland module state , 1 is enabled, 0 is disabled; the range of valve opening ; Pump frequency range , the range of reflux ratio ; Afterwards, through the dynamic response interval Further define the action constraints of each key variable and perform Sub-sampling, thus obtaining a set of potential actions ;

[0134] (4) Value head calculation executes potential actions separately The long-term cumulative expected rewards that can be obtained , in order to guide network optimization; specifically, the value head consists of two layers of fully connected networks, which serve as the Critic side in the Actor-Critic framework to provide gradient estimation basis for policy optimization and guide the policy network to converge to a better solution;

[0135] 3. Policy Network Training Mechanism Design: To achieve robust training and efficient convergence of the policy network in the complex wetland system operating environment, an end-to-end reinforcement learning training mechanism based on the Actor-Critic architecture was constructed. An improved Proximal Policy Optimization (PPO) algorithm was used to optimize the policy network parameters. The specific training and implementation technical routes are as follows:

[0136] (1) The system constructs the environment interaction process through the simulator, using the current state (including real-time water quality data , control data and future water quality forecast information ) Generate control actions , and build training trajectories based on rewards and next states obtained based on environmental feedback; the policy network continuously optimizes action outputs based on the principle of maximizing cumulative rewards;

[0137] (2) The training objective consists of two parts: the strategy head uses the PPO loss function in the form of a clipped probability ratio to improve the strategy stability; while the value head estimates the expected return of the current state by minimizing the time difference error; the two together drive the strategy to converge towards high pollutant removal rate, good water quality stability, and low energy consumption;

[0138] (3) Introducing policy entropy regularization during training to maintain policy exploratory nature, and setting early stopping criteria to prevent overfitting, further improving the policy’s convergence speed and generalization ability;

[0139] (4) After the strategy network training is completed, the system can output the optimal adjustment action combination of subsequent key control parameters according to the environmental state in each control cycle. Finally, the mean of the continuous sampling action set output by the strategy is used as the optimal control action of the system , realizing coordinated scheduling and dynamic optimization of multiple control units.

[0140] 5. Analysis of experimental results

[0141] To systematically verify the effectiveness of the proposed "intelligent control method for a three-stage artificial wetland recirculating aquaculture system," two comparative experiments were designed, corresponding to: (1) the evaluation of the accuracy of pollution evolution trend prediction; and (2) the sensitivity analysis of key control parameters. The comparative models included the traditional time series modeling methods LSTM and GRU to highlight the predictive advantages and control mechanism explanatory power of the proposed multimodal modeling structure.

[0142] 1. Pollutant concentration prediction accuracy assessment

[0143] Based on the constructed multi-source environmental time series dataset, LSTM, GRU and the pollution evolution trend prediction model of the present invention are used to make multi-step predictions on the changes in water quality data in the future period. All prediction and control experiments are modeled and evaluated based on the monitoring data of 8 typical control units. Each unit covers different wetland structure modules, has strong heterogeneity and differentiated operation responses, and can effectively verify the adaptability and stability of this method under multiple working conditions and multiple structure scenarios. The error comparison between the model output results and the actual observation values ​​is carried out. The evaluation indicators include root mean square error (RMSE) and mean absolute error (MAE). The results are as follows Figure 4 and Figure 5 As shown;

[0144] Statistical results show that the proposed method achieves the lowest error across all pollutant indicators, with an overall RMSE range of 0.098–0.114, with an average of approximately 0.106; the MAE remains stable between 0.078–0.088, with an average of approximately 0.083, significantly outperforming the comparison models. In comparison, the LSTM model achieves an RMSE range of 0.132–0.150 and a MAE of 0.101–0.113; the GRU model achieves even higher errors, with an RMSE of 0.157–0.176 and a MAE of 0.122–0.134. In medium- and high-complexity scenarios with multiple control units, the proposed method outperforms traditional sequence models in both error volatility and convergence stability. This demonstrates the effectiveness and adaptability of the proposed model's multi-scale perception structure and fusion decoding mechanism in processing multi-source environmental time series data, demonstrating its enhanced generalization capabilities and practical engineering value.

[0145] 2. Key control parameter sensitivity heat map analysis

[0146] To further verify the scientific nature of the key control parameter screening mechanism, a control parameter perturbation experiment was designed to quantitatively calculate the response intensity of seven operating parameters under multiple control states. The selected parameters include: (1) inlet flow rate; (2) outlet flow rate; (3) hydraulic retention time; (4) wetland module activation status; (5) valve opening; (6) pump operating frequency; (7) return flow ratio; the sensitivity coefficient of each parameter is measured by the average impact of its perturbation on the change of pollutant prediction value, and the results are summarized as a sensitivity heat map, as shown in the figure below. Figure 6 As shown;

[0147] Figure 6 The sensitivity heat map shown is statistically calculated based on the historical operating data of 8 control units during a complete control cycle. The sensitivity coefficient of each control parameter is quantified by the average impact of its disturbance on the change of pollutant concentration and is displayed in the form of a heat map in the figure.

[0148] From the results in the figure, it can be observed that the sensitivity performance of different control parameters in different units has certain fluctuations, indicating that the system's response to various parameters has local dependence and state coupling. Overall, the "reflux ratio" and "pump operating frequency" show relatively higher sensitivity levels in most units, and the sensitivity coefficients of some units even exceed 0.8, verifying their core position in system stability control. In addition, in specific control scenarios, parameters such as "water inlet flow" and "valve opening" also show local high response characteristics, suggesting that they have a phased regulatory effect on the dynamic changes of the system. Therefore, the control parameter sensitivity analysis method proposed in the present invention can combine the system's historical state distribution and pollutant response trends to dynamically identify the most valuable parameter combinations for regulation, avoid redundant control inputs, and improve the efficiency and operability of strategy deployment.

[0149] Therefore, through systematically designed experiments on pollutant prediction accuracy assessment and control parameter sensitivity analysis, the modeling capabilities and intelligent control performance of the present invention in a multi-source heterogeneous environment were fully verified. In the prediction phase, the present method demonstrated error control capabilities and stability superior to traditional models under multiple control units and complex operating conditions, fully demonstrating the advantages of its multi-scale perception structure and fusion decoding mechanism in handling the dynamic evolution of pollutants. In the control phase, through parameter perturbation analysis and heat map visualization, it accurately identified operating parameter combinations with key control value. In particular, the "recirculation ratio" and "pump operating frequency" showed significant influence in most system states, verifying the interpretability and adaptability of the method for obtaining key control parameters in complex control scenarios.

[0150] The above are merely preferred embodiments of the present application and are not intended to limit the present application. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application shall be included within the scope of protection of the present application.

[0151] Although the above describes the specific implementation methods of the present invention, it does not limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A three-stage artificial wetland circulating aquaculture system intelligent control method, characterized in that: The following steps are involved: S1, based on the key structural units of the three-level artificial wetland system and the multi-type environmental sensors deployed in the aquaculture pond, continuously collects water quality time series data and control time series data, and constructs multi-source heterogeneous data under a unified time scale; S2 inputs the water quality time series data and control time series data collected by S1 into the constructed and trained pollution evolution trend prediction model, and outputs multi-dimensional water quality prediction results; The pollution evolution trend prediction model includes a water quality dynamic perception channel, a control parameter encoding channel, and a fusion decoding prediction channel; The water quality dynamic perception channel introduces a three-branch parallel structure, with different branches designed with different receptive fields to enhance the model's sensitivity to the temporal trends of various water quality characteristics; The input is the collected water quality time series data, which captures short-term and medium-term trends and generates dynamic coding features of water quality. ; The control parameter encoding channel input is the collected control time series data, and the gated fusion mechanism is introduced to perform weighted integration on multiple control dimensions, dynamically adjust the representation contribution of different control variables in a specific state, and output the gated fusion feature. ; Then, the fusion features are controlled The first enhanced GRU unit is fed into the initial time series modeling; after that, the output of the first GRU unit is combined with the gated fusion feature The cascade is fed into the second enhanced GRU unit to characterize the deep interactive relationship between the control variables under the state combination conditions; The output of the second GRU unit is normalized by the layer to obtain the final features of the first stacked unit. ; Based on the above structure, five stackings are performed to extract multi-layer temporal features from local response to global regulation layer by layer, and finally the characterization features of the control parameter encoding channel are obtained through the Sigmoid activation function. ; The fusion decoding prediction channel dynamically encodes the water quality characteristics Characterization characteristics of the control parameter encoding channel Perform splicing and fusion in the channel dimension, and output multi-dimensional water quality prediction results based on the dual Transformer decoding layer-gating fusion structure ; S3 averages and aggregates the multi-dimensional water quality prediction results of S2 in the time dimension to obtain the state vector to be evaluated; performs unsupervised clustering on the historical water quality time series data set to obtain the cluster center and the corresponding cluster; performs state similarity matching and response sensitivity analysis on the state vector to be evaluated and the cluster center to obtain the key control parameters with the most regulatory value at the current moment, and infers their safe and adjustable dynamic response range; S4, based on the constructed reinforcement learning control strategy network, takes water quality time series data, control time series data and multi-dimensional water quality prediction results as environmental state input, key control parameters as adjustable actions, and dynamic response intervals as action constraints. It designs a comprehensive reward function under the multi-objective optimization goals of maximizing pollutant removal rate, stably meeting water quality indicators, and minimizing system energy consumption, and generates the optimal control parameter adjustment action through the reinforcement learning strategy network.

2. The intelligent control method for a three-stage artificial wetland recirculating aquaculture system according to claim 1, characterized in that: The parameters of the water quality time series data include pH value , dissolved oxygen concentration , ammonia nitrogen concentration , nitrite concentration , water temperature and turbidity ; The parameters of the control time series data include water flow , water flow , hydraulic retention time , wetland module enabled status , valve opening , Pump operating frequency Reflux ratio ; Multi-dimensional water quality prediction label parameters including pollutant indicators ; Reaction environment indicators ; System disturbance index ; Together they form a water quality prediction label combination .

3. The intelligent control method for a three-stage artificial wetland recirculating aquaculture system according to claim 2, characterized in that: Methods for collecting data sets used for model training and obtaining key control parameters and their dynamically adjustable response ranges include: Input water quality time series data collection: for each control unit, at the time point Collect the surrounding water quality data of all control units to obtain the water quality data at that moment ,in Indicates at the time point , No. Water quality data collected at each control unit, and Based on the above collection method, a total of Complete input water quality time series data at each time point ,and ; Input control timing data collection: at time point Collect control data corresponding to all control units ;in, Indicates at the time point , No. The control data collected from each control unit, and Similarly, collection Complete input control timing data at each moment ， and .

4. The intelligent control method for a three-stage artificial wetland recirculating aquaculture system according to claim 1, characterized in that: The specific structure of the water quality dynamic sensing channel includes: Short-term perception branch: Water quality time series data first passes through The dilated convolution block is used to extract local fluctuation features, and then The dilated convolution block further expands the short-term context perception range and obtains short-term dynamic features ; Long-term perception branch: Water quality time series data passes through Dilated convolutional blocks and The dilated convolution block uses a larger dilation rate to capture the trend evolution law across time slices, which is used to model the cumulative impact of changes in operating conditions on pollutants and output medium-term evolution characteristics. ; Original retention branch: retain the original water quality time series data as the residual connection path, retain low-level detail information, and form the basic retention feature ; Afterwards, the output features of the three branches are concatenated in the channel dimension to form a fusion feature , and sent to the global average pooling layer to enhance the global representation ability, and the fusion result is input into the next round of three-branch parallel structure for deep feature extraction; the above structure is repeatedly stacked for a total of 5 layers to construct a progressive multi-scale time series modeling path; after the output of the last layer, the Sigmoid activation function is connected to enhance the nonlinear expression ability, and finally the dynamic coding feature of water quality is generated. .

5. The intelligent control method for a three-stage artificial wetland recirculating aquaculture system according to claim 1, characterized in that: The specific structure of the fusion decoding prediction channel includes: First, the dynamic coding characteristics of water quality Characterization characteristics of the control parameter encoding channel Splicing is performed in the channel dimension; then, the spliced ​​features are sent to the gated fusion mechanism to obtain the interactive features ; Interaction Features The model is fed into two consecutive Transformer decoding layers to capture long-range dependencies and potential sequence structures between multi-dimensional features. Each Transformer decoding layer includes a multi-head attention mechanism, a feedforward neural network, and a residual connection structure. The Transformer decoding results are then fed into a gated fusion mechanism. The dual Transformer decoding layer-gating fusion constitutes a decoding module unit. This decoding module unit is repeatedly stacked in 5 layers to form a multi-layer information abstraction and trend reasoning path; it is finally sent to the Dropout layer to enhance the generalization ability, and then processed by the Linear mapping layer and activation function to output the multi-dimensional water quality prediction results. .

6. The intelligent control method for a three-stage artificial wetland recirculating aquaculture system according to claim 1, characterized in that: The specific process of S3 includes: Construction of the state vector to be evaluated: multi-dimensional water quality prediction results for the output , average aggregation is performed on each control unit in the time dimension to obtain the state vector to be evaluated at the current control moment ; Historical status data aggregation analysis: Based on the constructed water quality time series dataset, the water quality status data at each single moment in the past is extracted from it , and the scale is The historical water quality state sample set is obtained by performing unsupervised clustering on all historical water quality state vectors. Cluster centers ,in Indicates the cluster centers, and ; Next, calculate the current state vector to be evaluated The Euclidean distance from all cluster centers and a set of distance vectors are obtained ;in Represents the state vector to be evaluated With the The Euclidean distance between cluster centers; After that, select the cluster center closest to , and extract all historical water quality status data in the corresponding cluster as the candidate sample set; further calculate all samples in the cluster and The cosine similarity between them; finally the top with the highest similarity is selected The control time series data corresponding to the historical sample water quality time series data , which constitutes a representative control sample set for response analysis ; Response sensitivity assessment and key control parameter screening: for control sample sets For each set of control parameter data in the , a disturbance experiment is performed to calculate each control parameter The corresponding sensitivity coefficient ;in, and Respectively represent control parameters and their sensitivity coefficients; Finally, the sensitivity coefficient will be satisfied Greater than the preset threshold The parameters are used as the key control parameter set at the current control moment .

7. The intelligent control method for a three-stage artificial wetland recirculating aquaculture system according to claim 6, characterized in that: Based on the set of key control parameters, the safe and adjustable dynamic response range is inferred. The specific process is as follows: For each key parameter , from the control sample set Extract the maximum and minimum values ​​from the historical values ​​to form key parameters Initial response range: ,in Represents parameters respectively The minimum and maximum values ​​in the control sample set; Afterwards, combined with the system preset operation safety constraints Perform double boundary clipping to obtain the dynamically adjustable response range of this parameter: in, Parameters for the system The preset minimum and maximum values, and To obtain the maximum and minimum value operations; Key parameters The corresponding dynamically adjustable response range and the structured control parameter-response range pair .

8. The intelligent control method for a three-stage artificial wetland recirculating aquaculture system according to claim 1, characterized in that: The reinforcement learning control strategy network constructed in S4 adopts an Actor-Critic dual-head architecture; First, the current water quality time series data , control timing data and predicted status The data are concatenated into a unified state vector and sent to the network input layer for normalization and feature embedding to extract basic information. The feature extraction layer consists of a three-layer fully connected network and a ReLU activation unit, which performs high-dimensional feature conversion on the state vector and extracts deep control signal features; Strategy Head-to-Head Action Each key control parameter in the Gaussian distribution is constructed to output the value, and finally the expected value and corresponding variance of the control action are output, and the corresponding value range is obtained; then, the dynamic response interval is used to calculate the expected value of the control action. Define the action constraints for each key variable and perform Sub-sampling, to obtain a set of potential actions ; Value header calculations are performed separately for potential actions The long-term cumulative expected rewards that can be obtained , in order to guide network optimization; the value head consists of two layers of fully connected networks, which serves as the Critic end in the Actor-Critic framework to provide a gradient estimation basis for policy optimization and guide the policy network to converge to a better solution.

9. The intelligent control method for a three-stage artificial wetland recirculating aquaculture system according to claim 8, characterized in that: in, The environmental state is composed of the water quality time series data of the control system at the current moment, the control time series data and the multi-dimensional water quality prediction results; including the current water quality data ,for Time series data of water quality of all control units at all times , real-time control data ,for The executed control timing data corresponding to all control units at the moment ; Prediction status , for the future Multi-dimensional water quality prediction results at all control units ; Among them, action Key control parameters for each control unit The adjustment amount; the action value range is determined by the dynamic response interval constraint; Among them, the reward function It is a weighted multi-objective comprehensive reward mechanism that rewards efficient removal of pollutants, stable achievement of indicators and minimization of energy consumption.

Images