Control method and system for mid-universe wetland simulation system

Abstract

The invention discloses a control method and system for a midspace wetland simulation system, and belongs to the technical field of wetland simulation and control, and the method comprises the steps: determining preset parameters according to experimental requirements, starting the system, and collecting and preprocessing environmental parameters in real time through a sensor network; establishing a wetland ecosystem dynamic model by fusing machine learning algorithms such as LSTM, Transform and a graph neural network, and predicting an environment change trend; an optimal control strategy is generated in combination with a reinforcement learning algorithm, and parameters of an environment control module are adjusted in real time; the big data analysis technology is used for mining data rules, and system preset parameters are automatically optimized. Meanwhile, a physical model is introduced to calculate a theoretical value to correct a prediction model error, and through space-time fusion modeling, reinforcement learning decision and big data driven optimization, the control problems of multi-factor coupling and strong dynamic nature in wetland simulation are solved, and the environmental change prediction precision and the control strategy optimality are remarkably improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wetland simulation and control, and more particularly to a control method and system for a mesocosm wetland simulation system. BACKGROUND

[0002] Wetland ecosystems have important ecological functions, such as water purification, flood regulation, and biodiversity protection. However, due to changes in the natural environment and human activities, wetland ecosystems are facing many threats. Currently, indoor devices are used for experiments in wetland ecosystem simulation, which is a relatively mature and widely used technical means. By providing specific wetland simulation conditions to the samples in the device, the effects of these conditions on the samples are observed. The mesocosm wetland simulation system can simulate the actual environmental conditions of wetlands and provide a relatively controllable experimental platform for wetland ecological research.

[0003] Currently, in the control of mesocosm wetland simulation systems, traditional control methods are used, which cannot accurately simulate the complex environmental conditions of wetland ecosystems, resulting in large differences between experimental results and actual wetland ecosystems. At the same time, the collection and processing of wetland environmental data are not efficient enough to accurately reflect the dynamic changes of the wetland ecosystem in a timely manner. The control strategy is relatively fixed, lacks intelligence and adaptability, and it is difficult to adjust the control parameters in real time according to environmental changes, so as to effectively maintain the stable state of the wetland ecosystem.

[0004] Therefore, how to provide a control method and system for a mesocosm wetland simulation system that can accurately simulate a wetland ecosystem and adaptively adjust control strategies is a problem that needs to be solved by those skilled in the art. SUMMARY

[0005] Therefore, the present application provides an intelligent control method for a mesocosm wetland simulation system, which realizes accurate simulation and regulation of the wetland ecosystem by integrating sensor technology, data analysis, and artificial intelligence algorithms.

[0006] To achieve the above purpose, the present application adopts the following technical solutions: On the one hand, the present application provides a control method for a mesocosm wetland simulation system, comprising: determining the preset device parameters and preset environmental conditions of the mesocosm wetland simulation system according to experimental requirements; starting the mesocosm wetland simulation system, collecting environmental parameters of the simulation system in real time through a sensor network, and performing preprocessing; using a machine learning algorithm to analyze the collected and preprocessed environmental parameters, establishing a dynamic model of the wetland ecosystem, and predicting the trend of environmental changes; According to the preset device parameters, the preset environmental conditions and the predicted environmental change trend, an optimal control strategy is generated by using a reinforcement learning algorithm to adjust the parameters of the environmental control module in real time; Real-time acquisition of environmental data, real-time analysis of data using big data analysis technology, and mining of potential relationships and rules between data; According to the data analysis results, the preset device parameters and the preset environmental conditions of the mesocosm wetland simulation system are automatically adjusted to realize dynamic optimization control of the mesocosm wetland simulation system.

[0007] Preferably, the collected and preprocessed environmental parameters are analyzed using a machine learning algorithm to establish a dynamic model of the wetland ecosystem and predict the environmental change trend, specifically including: Feature extraction is performed on the preprocessed environmental parameters to obtain time series features and spatial sequence features; The long short-term memory network and the graph convolution network are fused to construct the dynamic model of the wetland ecosystem.

[0008] Preferably, the long short-term memory network and the graph convolution network are fused to construct the dynamic model of the wetland ecosystem, including: The LSTM neural network and the Transformer model are combined to construct a time series prediction model of the wetland ecosystem to obtain time series prediction results of the environmental change trend; Based on the graph neural network, a spatial correlation model of the time series prediction model of the wetland ecosystem is constructed to obtain spatial change prediction results of the environmental change trend; The time series prediction results and the spatial change prediction results are fused in chronological order to obtain environmental change trend prediction results; The theoretical value calculated based on the physical model is used as prior knowledge, and the system error of the time series prediction model and the spatial correlation model is corrected through residual learning.

[0009] Preferably, the LSTM neural network and the Transformer model are combined to construct the time series prediction model of the wetland ecosystem, specifically including: A data preprocessing layer is used to construct an environmental change sequence with a length based on the collected environmental data; An LSTM feature extraction module uses a bidirectional LSTM network to capture the forward / backward dependency relationship of the environmental change sequence to obtain local time series features; A Transformer global dependency modeling module uses a multi-head self-attention mechanism to model the spatial interaction relationship between multiple nodes to obtain global dependency features; A feature fusion layer concatenates the local time series features output by the LSTM and the global dependency features output by the Transformer. a prediction output layer, which maps the fused features to an output dimension through a fully connected network to obtain a time series prediction result of the environmental change trend.

[0010] Preferably, the spatial correlation model of the wetland ecosystem is constructed based on a graph neural network, comprising: The wetland simulation system is divided into multiple spatial nodes, each node representing a micro-ecological unit, and a node feature vector is extracted from the collected environmental parameters; An adjacency matrix is constructed based on the physical distance and ecological correlation between nodes, and the edge weight is calculated using a Gaussian kernel function; A hybrid graph neural network architecture is constructed, which includes a feature extraction layer, a hybrid layer of a multi-layer graph convolutional network and a graph attention network, and an output layer; A time embedding layer and a spatio-temporal attention weight matrix are determined based on the environmental parameters to adaptively adjust the importance of different time and space nodes, and a spatial correlation model is obtained; The spatial correlation model is optimized using a composite loss function combining node prediction loss and graph structure reconstruction loss; Based on the optimized spatial correlation model, wetland ecological analysis is performed to obtain a spatial change prediction result of the environmental change trend.

[0011] Preferably, the theoretical value calculated based on the physical model is used as prior knowledge to correct the system error of the time series prediction model and the spatial correlation model, comprising: Based on the physical and chemical principles of the wetland ecosystem, a physical model describing the relationship between environmental parameters is established to calculate the theoretical value of the change in environmental parameters; The theoretical value is spatio-temporally aligned with the prediction results of the time series prediction model and the spatial correlation model to extract residual features; The weights of the time series prediction model and the spatial correlation model are dynamically adjusted according to the residual features to realize adaptive updating of the models.

[0012] Preferably, according to the preset device parameters, the preset environmental conditions and the predicted environmental change trend, an optimal control strategy is generated using a reinforcement learning algorithm to adjust the parameters of the environmental control module in real time, comprising: The preset device parameters, the preset environmental conditions and the environmental change trend are input into a policy network sub-model in the reinforcement learning model to generate a control strategy set containing multiple control strategies; The multiple control strategies in the control strategy set are evaluated through a value network sub-model in the reinforcement learning model to select the optimal control strategy; The device parameters of the extraterrestrial wetland simulation system are adjusted based on the optimal control strategy.

[0013] In another aspect, the present application provides a control system for a mesocosm wetland simulation system, comprising: a setting module configured to determine preset equipment parameters and preset environmental conditions of the mesocosm wetland simulation system according to experimental requirements; a collection module configured to start the mesocosm wetland simulation system, collect environmental parameters of the simulation system in real time through a sensor network, and perform preprocessing; a prediction module configured to analyze the collected and preprocessed environmental parameters using a machine learning algorithm, establish a dynamic model of the wetland ecosystem, and predict environmental change trends; a decision module configured to generate an optimal control strategy using a reinforcement learning algorithm based on the preset equipment parameters, the preset environmental conditions, and the predicted environmental change trends, and adjust parameters of an environmental control module in real time; an analysis module configured to collect environmental data in real time, analyze the data in real time using big data analysis techniques, and mine potential relationships and rules between the data; a setting optimization module configured to automatically adjust the preset equipment parameters and the preset environmental conditions of the mesocosm wetland simulation system based on the data analysis results, and realize dynamic optimization control of the mesocosm wetland simulation system.

[0014] According to the above technical solution, compared with the prior art, the present application provides a control method and system for a mesocosm wetland simulation system, which fuses LSTM, Transformer, and graph neural networks to construct a dynamic model combining temporal and spatial fusion with physical constraints, captures time series dependence of environmental parameters and spatial node ecological correlation, corrects errors based on physical model theoretical values, greatly improves environmental change trend prediction accuracy, and truly reflects wetland ecological dynamic rules. Then, a reinforcement learning algorithm is combined to generate and evaluate control strategies with preset parameters and prediction trends as inputs, dynamically optimize environmental control module parameters, autonomously explore optimal strategies in complex scenarios, balance simulation reality and experimental goals, and avoid artificial intervention lag or deviation. Finally, big data analysis is used to mine potential data correlations, automatically adjust preset equipment parameters and environmental conditions, form a collection-analysis-optimization-feedback-closed loop, adapt to nonlinear changes in ecological systems, and improve system flexibility and robustness. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only embodiments of the present application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0016] Figure 1 A flowchart is provided for the present application.

[0017] Figure 2 A flowchart of data transmission and signal feedback for the present application.

[0018] Figure 3 A structural diagram is provided for the present application. DETAILED DESCRIPTION

[0019] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0020] The embodiments of the present application disclose a control method for a middle-universe wetland simulation system, as shown in Figures 1-2 The control method comprises the following steps: Step 1: Determine the preset equipment parameters and preset environmental conditions of the middle-universe wetland simulation system according to experimental requirements.

[0021] When determining the equipment parameters and environmental conditions of the middle-universe wetland simulation system, the specific target of the experiment should be fully combined, such as studying the material cycle and energy flow of the wetland ecosystem, or exploring the influence of different environmental factors on the wetland biological community. In terms of equipment parameters, the key performance indicators of various types of hardware equipment are covered, for example: the flow rate, head and start-stop frequency of the water pump, the wind speed, wind direction and running time of the fan, the light intensity, spectral distribution (such as the proportion of red light and blue light) and light period of the lighting equipment, the power and temperature control accuracy of the heating and refrigeration equipment, the amplitude, period and number of cycles of the pressure wave body, etc.

[0022] The setting of environmental conditions includes multi-dimensional ecological factors, specifically as follows: temperature parameters, the temperature range and variation gradient of different regions of the wetland (such as the surface water body and the bottom mud) need to be determined, considering the diurnal temperature difference and seasonal variation mode; humidity parameters, for the atmospheric humidity of the wetland, a reasonable humidity interval and fluctuation rule are set; water level parameters, the average water level, water level fluctuation amplitude and variation rate of the wetland are determined to simulate the hydrological process of the natural wetland; water quality parameters, including dissolved oxygen concentration, nutrient salt (such as nitrogen and phosphorus) content, heavy metal ion concentration, organic pollutant concentration, etc.; soil parameters, such as the water content, organic matter content, pH value and porosity of the bottom mud, which affect the growth of wetland plants and the activity of microorganisms; biological parameters, if the experiment involves biological communities, the types of introduced plants and animals and their population density need to be determined.

[0023] Step 2: Start the middle universe wetland simulation system, collect environmental parameters of the simulation system in real time through the sensor network, and preprocess.

[0024] The sensor network is composed of multiple types of sensors distributed at key positions in the wetland simulation system, including: water body sensors installed at different water depths to collect water temperature, dissolved oxygen, conductivity, pH value, turbidity, and nutrient salt concentration; atmospheric sensors arranged at different heights above the wetland to monitor air temperature, humidity, wind speed, wind direction, carbon dioxide concentration, and oxygen concentration; soil sensors inserted into the sediment to obtain sediment temperature, moisture content, conductivity, and organic matter content; biological sensors, if biological monitoring is required, sensors can be deployed to monitor plant growth status (such as stem diameter, leaf humidity), and animal activity (such as fish swimming speed, position).

[0025] The preprocessing steps include: data cleaning, removing obvious outliers (such as sudden data caused by sensor failure), which can use statistical methods (such as Z-score method, IQR method) or machine learning-based anomaly detection algorithms; data calibration, correcting the original data according to the calibration parameters of the sensor to ensure the accuracy of the data; data normalization, converting the output data of different sensors to a unified numerical range to facilitate subsequent analysis and modeling, common methods include min-max normalization, standardization; data interpolation, for missing data caused by sensor failure or communication problems, linear interpolation, polynomial interpolation, model-based interpolation, etc. are used to fill in; data filtering, using moving average filtering, Gaussian filtering, etc. to remove high-frequency noise and smooth the data curve.

[0026] Step 3: Use machine learning algorithms to analyze the collected and preprocessed environmental parameters, establish a dynamic model of the wetland ecosystem, and predict environmental change trends.

[0027] Step 4: According to the preset device parameters, preset environmental conditions, and predicted environmental change trends, use reinforcement learning algorithms to generate optimal control strategies, and adjust the parameters of the environmental control module in real time.

[0028] Step 5: Collect environmental data in real time, use big data analysis technology to analyze the data in real time, and mine the potential relationships and rules between the data.

[0029] The real-time collected environmental data is huge, and contains rich potential relationships and rules, which need to be deeply mined with the help of big data analysis technology. The data mining methods used include: association rule analysis, mining the association relationship between different environmental parameters, such as finding that when the dissolved oxygen concentration is lower than a certain threshold, the ammonia nitrogen concentration will tend to rise in the next period of time, thereby establishing the association rule between the two; clustering analysis, clustering the running state of the wetland ecological system, identifying different typical working conditions (such as normal running working condition, abnormal fluctuation working condition), providing basis for fault diagnosis and optimization control of the system; time series analysis, analyzing the time series data of environmental parameters, extracting periodic characteristics (such as diurnal cycle, seasonal cycle) and trend characteristics, and predicting the long-term trend in the future; feature engineering in machine learning, reducing the data dimension through principal component analysis (PCA), factor analysis and other methods, extracting the key factors affecting the wetland ecological system, and simplifying the model structure.

[0030] Step 6: According to the data analysis results, automatically adjusting the preset device parameters and preset environmental conditions of the middle universe wetland simulation system, realizing dynamic optimization control of the middle universe wetland simulation system.

[0031] For example, through analysis of long-term experimental data, it is determined that the root stability of rigid plants changes the sedimentation pattern of the gully edge, and then affects the drainage efficiency and geometric characteristics of the gully, so by adjusting the number of plants, the best plant number for drainage efficiency is found.

[0032] Specifically, the machine learning algorithm is used to analyze the preprocessed environmental parameters collected, and a dynamic model of the wetland ecological system is established to predict the trend of environmental change, which specifically includes: Feature extraction is performed on the preprocessed environmental parameters to obtain time series features and spatial sequence features; The long short-term memory network and the graph convolution network are fused to construct the dynamic model of the wetland ecological system.

[0033] Further, the long short-term memory network and the graph convolution network are fused to construct the dynamic model of the wetland ecological system, which includes: The LSTM neural network and the Transformer model are combined to construct the time series prediction model of the wetland ecological system, and the time series prediction result of the trend of environmental change is obtained; Based on the graph neural network, a spatial correlation model of the time series prediction model of the wetland ecological system is constructed, and a spatial change prediction result of the trend of environmental change is obtained; The time series prediction result and the spatial change prediction result are fused in time sequence to obtain the environmental change trend prediction result; The theoretical value calculated based on the physical model is used as prior knowledge to correct the system error of the time series prediction model and the spatial correlation model through residual learning.

[0034] Further, the LSTM neural network and the Transformer model are combined to construct the time series prediction model of the wetland ecosystem, and specifically include: A data preprocessing layer is configured to construct an environment change sequence with a length according to the collected environment data; An LSTM feature extraction module is configured to capture a forward / backward dependency relationship of the environment change sequence by using a bidirectional LSTM network to obtain local time sequence features; A Transformer global dependency modeling module is configured to model a spatial interaction relationship among multiple nodes by using a multi-head self-attention mechanism to obtain global dependency features; A feature fusion layer is configured to splice the local time sequence features output by the LSTM and the global dependency features output by the Transformer; A prediction output layer is configured to map the fused features to an output dimension by using a fully connected network to obtain a time series prediction result of an environment change trend.

[0035] Further, the spatial correlation model of the wetland ecosystem is constructed based on a graph neural network, including: The wetland simulation system is divided into multiple spatial nodes, and a node feature vector is extracted from the collected environment parameters.

[0036] According to the terrain features, hydrological connectivity and ecological function zoning of the wetland simulation system, the system is divided into a plurality of spatial nodes. The node scale needs to consider both accuracy and computational efficiency. Each node represents a micro-ecological unit, i.e., the smallest spatial unit with relatively independent ecological functions. The boundary of each node needs to match the sensor layout position to ensure that the data can be directly associated.

[0037] From the environment parameters collected by the sensor network, a high-dimensional feature vector is extracted for each node, including physical parameters, chemical parameters, biological parameters and space-time identifiers. The feature vector needs to be standardized to eliminate the influence of dimensional differences on the model.

[0038] An adjacency matrix is constructed based on the physical distance and ecological correlation between nodes, and the edge weight is calculated by using a Gaussian kernel function; A hybrid graph neural network architecture is constructed, which includes a feature extraction layer, a hybrid layer of a multi-layer graph convolutional network and a graph attention network, and an output layer.

[0039] The feature extraction layer adopts a 1D convolution layer or a fully connected layer to pre-process the node feature vector, map the original features to a high-dimensional hidden space (e.g., convert a 64-dimensional feature vector to 128-dimensional), enhance the non-linear expression ability of the features, and highlight the signals of key ecological factors (e.g., dissolved oxygen, plant biomass).

[0040] The bottom GCN captures the local dependence of short-distance, strongly correlated nodes (e.g., nutrient exchange between a certain emergent plant zone and adjacent water area) through local aggregation operations (e.g., weighted sum of features of node neighbors); the middle GAT introduces an attention mechanism to assign dynamic weights to each node's neighbors (e.g., in drought, nodes close to water sources have higher weights for drought nodes), addressing the limitation of GCN's "equal treatment" of all neighbors; the high-level hybrid layer fuses the local structure perception of GCN and the dynamic attention of GAT, alleviating the gradient vanishing problem of deep networks through residual connections, and capturing long-distance correlations across regions (e.g., water quality influence of upstream nodes on downstream nodes).

[0041] The output layer uses a fully connected layer and an activation function to map high-level features into a predicted feature vector of nodes (e.g., dissolved oxygen concentration, plant growth rate in the next 24 hours), or output an association strength matrix between nodes (for subsequent spatial trend analysis).

[0042] Based on the environmental parameters, a time embedding layer and a spatio-temporal attention weight matrix are determined to adaptively adjust the importance of different time and space nodes, obtaining a spatial correlation model.

[0043] To capture the dynamics of spatial correlation, the sampling timestamp is converted into a low-dimensional vector through time embedding, and concatenated with the node feature vector, enabling the model to perceive the spatial correlation at a certain time. The spatio-temporal attention weight is dynamically adjusted based on real-time environmental parameters, with the formula: ; where q i and k j are the feature query / key vectors of nodes i and j, d k is the vector dimension, and s t is the time decay coefficient. Matrix α i,j,t is used to highlight key nodes at different time steps t, achieving adaptive modeling of spatial correlation over time.

[0044] The spatial correlation model is optimized using a composite loss function that combines node prediction loss and graph structure reconstruction loss.

[0045] The model is constrained by dual loss, balancing node feature prediction accuracy and spatial structure authenticity. Specifically: The mean squared error (MSE) is used to calculate the difference between the predicted node features and the true observed values, with the formula: ; wherein N is the number of nodes, is the predicted feature, y i is the true feature, ensuring that the model can accurately predict the ecological state of individual nodes.

[0046] The graph structure reconstruction loss is constrained by the cross-entropy loss by comparing the difference between the correlation matrix output by the model and the real adjacency matrix, and the formula is: ; wherein, is the correlation strength predicted by the model, forcing the model to learn the real spatial connection pattern.

[0047] The composite loss function is: ; wherein λ1, λ2 are weight coefficients.

[0048] Based on the optimized spatial correlation model, wetland ecological analysis is performed to obtain the spatial change prediction result of the environmental change trend. The final output spatial change prediction result can provide fine spatial positioning basis for subsequent control strategy generation. The optimized model can output the spatial change trend in the following ways: based on the correlation strength matrix output by the model, the key ecological correlation path is extracted to reveal the spatial conduction chain of pollution diffusion or ecological disturbance; a spatial distribution heat map of environmental parameters is generated to predict high-risk areas; the interaction strength change of different functional areas is quantified to predict the influence range and degree of disturbance in a certain area on the surrounding nodes.

[0049] In another embodiment, the theoretical value calculated based on the physical model is used as prior knowledge to correct the system error of the time series prediction model and the spatial correlation model, including: Based on the physical and chemical principles of the spatial correlation model of the wetland ecosystem, a physical model describing the relationship between environmental parameters is established to calculate the theoretical value of the change in environmental parameters; The theoretical value is spatiotemporally aligned with the prediction results of the time series prediction model and the spatial correlation model, and residual features are extracted; According to the residual features, the weights of the time series prediction model and the spatial correlation model are dynamically adjusted to realize adaptive updating of the model.

[0050] In another embodiment, according to the preset device parameters, the preset environmental conditions, and the predicted environmental change trend, an optimal control strategy is generated using a reinforcement learning algorithm to real-time adjust the parameters of the environmental control module, including: The preset device parameters, the preset environmental conditions, and the environmental change trend are input into a policy network sub-model in a reinforcement learning model to generate a control strategy set containing multiple control strategies; The plurality of control strategies in the control strategy set are evaluated by a value network submodel in the reinforcement learning model, and an optimal control strategy is selected; The device parameters of the mesocosm wetland simulation system are adjusted based on the optimal control strategy.

[0051] The specific steps of generating the optimal control strategy using the reinforcement learning algorithm are as follows: The parameters of the input strategy network are converted into a computable vector form; A deep neural network (multi-layer perceptron + residual connection) is used, the input layer receives the above-mentioned quantized parameters, the hidden layer captures the coupling relationship between the parameters through nonlinear transformation (ReLU activation function), and the output layer generates a strategy set containing multiple control strategies through Gaussian distribution sampling or discrete action space enumeration.

[0052] The specific form of the control strategy is a device parameter adjustment scheme, specifically: Strategy 1: Increase the light intensity by 10%, increase the aeration frequency by 5 times / hour, and keep the water pump flow unchanged; Strategy 2: Reduce the heating power by 15%, raise the water level by 0.2m, and shorten the light period by 1 hour; Strategy 3: Stop aeration, increase the water pump flow by 20%, and lower the target water temperature by 2℃; The size of the strategy set is set according to the system complexity (usually containing 10-50 strategies), which needs to cover the main adjustment direction and amplitude of the device parameters.

[0053] The optimal control strategy is evaluated and selected by the value network submodel, and the value network submodel takes the comprehensive benefit of the control strategy as the core evaluation index, and the benefit function integrates multiple dimensions: Target achievement: After the implementation of the strategy, the deviation of the environmental parameters from the preset target (such as dissolved oxygen deviation, water temperature deviation), the smaller the deviation, the higher the benefit; System stability: After the implementation of the strategy, the fluctuation amplitude of the environmental parameters (such as water temperature change within 1 hour not more than ±1℃), the smaller the fluctuation, the higher the benefit; Energy cost: The incremental energy consumption of device adjustment (such as the increase in power consumption caused by the increase in water pump flow), the lower the energy consumption, the higher the benefit; Ecological safety: Whether the strategy avoids extreme environments (such as dissolved oxygen below 2mg / L leading to biological death), the benefit is positive within the safety threshold, and negative beyond the safety threshold.

[0054] The value network inputs each control strategy in the strategy set into an environment response simulator (a simplified model trained based on historical data, quickly predicting the environmental change after the strategy is implemented), and outputs the benefit value of the strategy. By comparing the benefit values of all strategies, the strategy with the highest benefit value is selected as the optimal control strategy; if there are multiple high-benefit strategies, the strategy with lower energy consumption or smaller adjustment range is preferred.

[0055] The present embodiment takes the verification of the ecological restoration efficiency of urban river wetlands as an application scenario, and uses the control method and system of the mesocosm wetland simulation system to simulate the "submerged plant + microbial collaborative restoration". The following steps are included: S1: Determine the preset equipment parameters and preset environmental conditions of the mesocosm wetland simulation system according to the experimental requirements: Referring to the actual ecological characteristics of a certain city's ring city river (1:10 scale), a mesocosm wetland simulation system with a size of 15m x 6m x 1.2m (length x width x height) is built, and the following preset parameters are determined by setting the modules: The preset equipment parameters include: Hydrological simulation unit: centrifugal pump flow 0.8m 3 / h, dry season 4h per day (maintain water flow rate 0.1m / s), wet season 8h per day (flow rate 0.2m / s); Water quality regulation unit: automatic dosing pump accuracy ±0.1mL / min, aeration device (nano aeration disc) power 80W; Biological unit: set 3 submerged plant density groups (black algae, bitter grass): 5 plants / m 2 , 10 plants / m 2 , 15 plants / m 2 ; 2 microbial concentration groups (nitrifying bacteria): 10 6 CFU / mL, 10 7 CFU / mL; The preset environmental conditions include: Water temperature: 22±2℃, pH: 7.2±0.3, initial total nitrogen (TN): 8mg / L, dissolved oxygen (DO): 1.8mg / L, hydrological period: one cycle every 30 days, alternating simulation of dry season (water level 0.6m) and wet season (water level 1.0m).

[0056] S2: Start the mesocosm wetland simulation system, and collect the environmental parameters of the simulation system in real time through the sensor network and perform preprocessing; After starting the system, the environmental parameter collection and preprocessing are implemented through the collection module: (1) Sensor network deployment: divide the system into 9 monitoring nodes (3x3 grid), and deploy: Water quality sensors: DO (precision ± 0.05 mg / L), TN (precision ± 0.02 mg / L), pH (precision ± 0.01), sampling frequency 1 time / 15 minutes.

[0057] Biological sensors: plant growth monitor (measuring black algae height and fresh weight, 1 time / 7 days), microbial activity sensor (detecting bacterial quantity, 1 time / 3 days).

[0058] Hydrological sensors: water level gauge (1 time / 5 minutes), flow meter (1 time / 30 minutes).

[0059] (2) Data preprocessing: Outlier rejection: "3σ rule" is used to filter sensor fault data, a total of 186 invalid data (0.2% of the total data) are removed.

[0060] Calibration correction: The sensor is calibrated with standard solution, and the black algae height measurement value is corrected (measurement deviation caused by light shielding).

[0061] Missing value completion: For missing data caused by communication interruption, the "adjacent node interpolation method" is used to complete.

[0062] S3: Use machine learning algorithms to analyze the collected and preprocessed environmental parameters, establish a dynamic model of the wetland ecosystem, and predict the trend of environmental changes.

[0063] Construct a spatio-temporal fusion dynamic model through the prediction module: (1) Time series prediction model construction (LSTM+Transformer): Take "7-day historical data" as the input window (a total of 672 sampling points), use bidirectional LSTM to capture the short-term dependence of DO and TN (such as the 12-hour lag effect of DO on TN degradation after aeration), and use the Transformer multi-head self-attention mechanism to mine the long-term correlation between microbial concentration and TN removal, output "14-day TN prediction value".

[0064] (2) Spatial correlation model construction (GNN): Take 9 monitoring nodes as "nodes" of the graph, and construct an adjacency matrix based on physical distance and ecological correlation.

[0065] Output "TN spatial distribution heat map" through graph neural network, and find the diffusion law of "high plant density node→low TN area".

[0066] (3) Physical model correction: Based on the physical and chemical principles of wetland nitrogen cycle, a physical model is established to calculate the theoretical removal rate of TN: TN theoretical removal rate=k×DO×t×B; where k is the reaction rate constant (0.03 / d), t is the reaction time (days), and B is the microbial concentration coefficient; The TN prediction error is reduced from 6% to 4.2% by adjusting the model weights through residual learning. S4: According to the preset device parameters, the preset environmental conditions and the predicted environmental change trend, the optimal control strategy is generated by using reinforcement learning algorithm, and the parameters of the environmental control module are adjusted in real time.

[0067] The optimal control strategy is generated by the decision module: (1) Reinforcement learning model design: State space: current TN value, DO value, plant density, microbial concentration, device operating parameters; Action space: water pump running time ±1h, aeration start threshold ±0.2mg / L, microbial dosage ±1 CFU / mL; Reward function: R=0.7×(TN initial-TN current)-0.3×aeration energy consumption.

[0068] (2) Optimal strategy generation and execution: TN=6.2mg / L and DO=2.1mg / L were monitored on the 30th day of the experiment; Three groups of candidate strategies were generated and evaluated by the value network, and the optimal strategy was selected: water pump +1h, microbial +10 6 CFU / mL, aeration threshold 2.0mg / L; After execution, on the 7th day, the system TN dropped to 3.0mg / L, DO rose to 4.3mg / L, and aeration energy consumption was 4.9kW h / day; S5: Real-time collection of environmental data, real-time analysis of data using big data analysis technology, and mining of potential relationships and rules between data.

[0069] Through the analysis module, the 120-day experimental data was analyzed, and the core rules were mined: When the submerged plant density is 10 plants / m 2 + microbial concentration is 10 7 CFU / mL, the TN removal rate is the highest (up to 68%), and there is a significant "synergistic effect". During the wet season, the water pump runs for 8h and the aeration threshold is 2.2mg / L, the energy consumption is the lowest (4.5kW h / day), and the DO is maintained at 3.5-4.0mg / L (optimal plant growth); S6: According to the data analysis results, automatically adjust the preset device parameters and preset environmental conditions of the Zhongyuzhongxing wetland simulation system to realize dynamic optimization control; The parameters are automatically adjusted by setting the optimization module: Submerged plant density is optimized to 10 plants / m 2 Microbial concentration is optimized to 10 7 CFU / mL, water pump operation time is adjusted to 8h in the wet season, and the aeration starting threshold is adjusted to 2.2mg / L.

[0070] In another aspect, the present application provides a control system for a mesocosm wetland simulation system, as shown, comprising: Figure 3 a setting module for determining preset device parameters and preset environmental conditions of the mesocosm wetland simulation system according to experimental requirements; a collection module for starting the mesocosm wetland simulation system, collecting environmental parameters of the simulation system in real time through a sensor network, and performing preprocessing; a prediction module for analyzing the collected and preprocessed environmental parameters using a machine learning algorithm, establishing a dynamic model of the wetland ecosystem, and predicting environmental change trends; a decision module for generating an optimal control strategy using a reinforcement learning algorithm according to the preset device parameters, the preset environmental conditions, and the predicted environmental change trends, and adjusting parameters of the environmental control module in real time; an analysis module for collecting environmental data in real time, performing real-time analysis of the data using big data analysis techniques, and mining potential relationships and rules between the data; a setting optimization module for automatically adjusting the preset device parameters and the preset environmental conditions of the mesocosm wetland simulation system according to the data analysis results, and realizing dynamic optimization control of the mesocosm wetland simulation system.

[0071] The various embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the various embodiments can be mutually referred to. For the device disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple, and the related parts can be referred to the method part.

[0072] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.​

Claims

1. A control method for a medium-scale cosmic wetland simulation system, characterized in that, include: Determine the preset equipment parameters and preset environmental conditions for the cosmic wetland simulation system based on experimental requirements; The Zhongyuzhou Wetland Simulation System was launched, and environmental parameters of the simulation system were collected in real time through a sensor network and preprocessed. Machine learning algorithms are used to analyze the preprocessed environmental parameters collected, establish a dynamic model of the wetland ecosystem, and predict environmental change trends. Based on preset equipment parameters, preset environmental conditions, and predicted environmental change trends, a reinforcement learning algorithm is used to generate the optimal control strategy and adjust the parameters of the environmental control module in real time. Real-time collection of environmental data, and real-time analysis of the data using big data analytics to uncover potential relationships and patterns between the data. Based on the data analysis results, the preset equipment parameters and preset environmental conditions of the Zhongyuzhou Wetland Simulation System are automatically adjusted to achieve dynamic optimization control of the Zhongyuzhou Wetland Simulation System.

2. The control method for a medium-scale cosmic wetland simulation system according to claim 1, characterized in that, Machine learning algorithms are used to analyze preprocessed environmental parameters to establish a dynamic model of the wetland ecosystem and predict environmental change trends. Specifically, this includes: Feature extraction is performed on the preprocessed environmental parameters to obtain time series features and spatial series features; A dynamic model of the wetland ecosystem is constructed by integrating a long short-term memory network and a graph convolutional network.

3. The control method for a medium-scale cosmic wetland simulation system according to claim 2, characterized in that, The dynamic model of the wetland ecosystem is constructed by integrating long short-term memory networks and graph convolutional networks, including: A time series prediction model for the wetland ecosystem was constructed by combining LSTM neural network and Transformer model to obtain time series prediction results of environmental change trends. Based on a graph neural network, a spatial correlation model of the wetland ecosystem is constructed to obtain spatial variation prediction results of environmental change trends. By fusing the time series prediction results and the spatial change prediction results in chronological order, an environmental change trend prediction result is obtained. The theoretical values ​​calculated based on the physical model are used as prior knowledge, and the systematic errors of the time series prediction model and the spatial correlation model are corrected through residual learning.

4. The control method for a cosmic wetland simulation system according to claim 3, characterized in that, The time series prediction model for the wetland ecosystem is constructed by combining LSTM neural networks and Transformer models, specifically including: The data preprocessing layer is used to construct an environmental change sequence of length based on the collected environmental data. The LSTM feature extraction module uses a bidirectional LSTM network to capture the forward / backward dependencies of the environmental change sequence to obtain local temporal features. The Transformer global dependency modeling module uses a multi-head self-attention mechanism to model the spatial interaction relationships between multiple nodes and obtain global dependency features. The feature fusion layer concatenates the local temporal features output by the LSTM with the global dependency features output by the Transformer. The prediction output layer uses a fully connected network to map the fused features to the output dimension, thus obtaining time-series prediction results of environmental change trends.

5. The control method for a medium-scale cosmic wetland simulation system according to claim 3, characterized in that, A spatial association model of the wetland ecosystem is constructed based on a graph neural network, including: The wetland simulation system is divided into multiple spatial nodes, each node representing a micro-ecological unit, and node feature vectors are extracted from the collected environmental parameters. An adjacency matrix is ​​constructed based on the physical distance and ecological association between nodes, and the edge weights are calculated using a Gaussian kernel function. A hybrid graph neural network architecture is constructed, which includes a feature extraction layer, a hybrid layer of multi-layer graph convolutional network and graph attention network, and an output layer; Based on the environmental parameters, the temporal embedding layer and spatiotemporal attention weight matrix are determined, and the importance of different time and space nodes is adaptively adjusted to obtain a spatial association model. The spatial association model is optimized using a composite loss function that combines node prediction loss and graph structure reconstruction loss. Wetland ecological analysis was conducted based on the optimized spatial correlation model to obtain spatial variation prediction results of environmental change trends.

6. The control method for a medium-scale cosmic wetland simulation system according to claim 3, characterized in that, Theoretical values ​​calculated based on physical models are used as prior knowledge to correct the systematic errors of the time series prediction model and the spatial correlation model, including: Based on the physicochemical principles of the spatial correlation model of wetland ecosystems, a physical model describing the relationship of environmental parameters is established, and the theoretical values ​​of environmental parameter changes are calculated. The theoretical values ​​are spatiotemporally aligned with the prediction results of the time series prediction model and the spatial correlation model to extract residual features; The weights of the time series prediction model and the spatial correlation model are dynamically adjusted based on the residual characteristics to achieve adaptive model updates.

7. The control method for a medium-scale cosmic wetland simulation system according to claim 1, characterized in that, Based on preset equipment parameters, preset environmental conditions, and predicted environmental change trends, a reinforcement learning algorithm is used to generate the optimal control strategy and adjust the parameters of the environmental control module in real time, including: The preset device parameters, the preset environmental conditions, and the environmental change trend are input into the policy network sub-model in the reinforcement learning model to generate a control policy set containing multiple control policies. The optimal control strategy is selected by evaluating multiple control strategies in the control strategy set through the value network sub-model in the reinforcement learning model. The equipment parameters of the aforementioned cosmic wetland simulation system are adjusted based on the optimal control strategy.

8. A control system for a medium-scale cosmic wetland simulation system, characterized in that, include: The settings module is used to determine the preset equipment parameters and preset environmental conditions of the cosmic wetland simulation system according to experimental requirements. The data acquisition module is used to start the Zhongyuzhou Wetland Simulation System, collect environmental parameters of the simulation system in real time through a sensor network, and perform preprocessing. The prediction module is used to analyze the preprocessed environmental parameters collected using machine learning algorithms, establish a dynamic model of the wetland ecosystem, and predict environmental change trends. The decision-making module is used to generate the optimal control strategy based on preset equipment parameters, preset environmental conditions and predicted environmental change trends, and adjust the parameters of the environmental control module in real time. The analysis module is used to collect environmental data in real time, and use big data analysis technology to analyze the data in real time to uncover potential relationships and patterns between the data. An optimization module is set up to automatically adjust the preset equipment parameters and preset environmental conditions of the Zhongyuzhou Wetland Simulation System based on data analysis results, so as to realize dynamic optimization control of the Zhongyuzhou Wetland Simulation System.

Images