Reservoir circulating water industrial aquaculture system control method combining Internet of Things and AI

Abstract

The invention provides a reservoir circulating water industrial aquaculture system control method combining the Internet of Things and AI, and relates to the technical field of control, and the method comprises the steps: collecting water body parameters through a distributed sensing unit to construct space-time state representation data, recognizing a stress factor distribution position and a metabolic load accumulation characteristic, calculating the environment regulation and control demand quantity of each region, and calculating the environmental regulation and control demand quantity of each region. The spatial response priority and the time sequence matching relation of the functional units are determined, cooperative work parameters are optimized by applying a reinforcement learning model, and the accuracy and the intelligent level of aquaculture water environment regulation and control are improved with the growth rate, the health level and the water stability of aquaculture objects as targets.

Description

Technical Field

[0001] This invention relates to the field of control technology, and in particular to a control method for a reservoir recirculating water industrial aquaculture system that combines the Internet of Things and AI. Background Technology

[0002] With the rapid development of aquaculture, recirculating aquaculture systems (RAS), as a water-saving, environmentally friendly, and efficient modern aquaculture model, have become an important way to solve pollution problems and improve aquaculture efficiency in traditional aquaculture. Large-scale recirculating aquaculture systems are a new type of aquaculture model developed on the basis of traditional RAS, and can better adapt to the needs of large-scale aquaculture production.

[0003] In recent years, the application of IoT and AI technologies in agriculture has become increasingly widespread, providing new technical means for the intelligent control of large-scale recirculating aquaculture systems. Currently, environmental monitoring mainly relies on discrete sensors to collect basic parameters such as water temperature, dissolved oxygen, and pH. Control systems adjust the operation of equipment such as water pumps and aerators based on simple threshold judgments. This monitoring and control method is no longer sufficient to meet the refined management needs of modern large-scale recirculating aquaculture. Summary of the Invention

[0004] This invention provides a control method for a reservoir recirculating water industrialized aquaculture system that combines the Internet of Things and AI, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a control method for a reservoir recirculating aquaculture system combining the Internet of Things (IoT) and AI, comprising: Environmental parameters and biological behavior parameters of the aquaculture water body are collected by distributed sensing units. The environmental parameters and biological behavior parameters are correlated and mapped according to the collection time and spatial location to establish spatiotemporal state characterization data of the aquaculture water body. Based on the spatiotemporal state characterization data, the spatial distribution of stress factors deviating from the normal range in the aquaculture water body and the concentration gradient of local metabolic load accumulation characteristics are identified. Combined with the dissolved oxygen demand threshold and ammonia nitrogen tolerance threshold of the cultured organisms at different growth stages, the environmental regulation requirements of each spatial region of the aquaculture water body are calculated. Based on the environmental regulation requirements, the spatial response priority and temporal coordination relationship of each functional unit in the circulating water system are determined. The spatial response priority is determined based on the spatial distribution location of the stress factors, and the temporal coordination relationship is determined based on the concentration gradient of the metabolic load accumulation characteristics, thereby generating the collaborative working parameters of each functional unit. The collaborative working parameters are iteratively optimized using a reinforcement learning model. The reinforcement learning model takes the growth rate of the aquaculture object and the stability of the water body as optimization objectives, and energy consumption constraints and equipment load constraints as boundary conditions, and outputs a cyclical control command sequence.

[0006] Based on the spatiotemporal state characterization data, the spatial distribution locations of stress factors deviating from the normal range in aquaculture water bodies and the concentration gradients of localized metabolic load accumulation characteristics include: The spatiotemporal state representation data includes spatial dimension data and temporal dimension data; Based on the spatial dimension data, a multi-level spatial grid of the aquaculture water body is constructed. The environmental parameters within each spatial grid are judged to be within the normal range. The stress factor types corresponding to the environmental parameters that exceed the normal range are identified. The spatial grid positions corresponding to the stress factor types are recorded as the spatial distribution positions of the stress factors. Among them, the stress factor types include hypoxia stress factors where dissolved oxygen is lower than the oxygen demand threshold and nitrogen metabolism stress factors where ammonia nitrogen concentration is higher than the tolerance threshold. Extract the feeding and activity characteristic change curves of the cultured objects from the time dimension data, calculate the generation and accumulation rates of metabolic products in the culture water at adjacent time points, and determine the transmission path and accumulation region of metabolic load between different spatial grids by combining the water flow connectivity relationship of the spatial grid, and calculate the concentration gradient of metabolic load in the accumulation region.

[0007] Calculate the generation and accumulation rates of metabolic products in the aquaculture water at adjacent time points, and, based on the water flow connectivity of the spatial grid, determine the transmission paths and accumulation regions of metabolic load between different spatial grids, including: Based on the difference in feed intake at adjacent moments in the feeding characteristic change curve and the metabolic conversion coefficient of the cultured organisms, the generation rate of ammonia nitrogen metabolites in the culture water at adjacent moments is calculated. Based on the difference in activity intensity at adjacent moments in the activity characteristic change curve and the excretion pattern parameters of the cultured organisms, the generation rate of organic waste in the culture water at adjacent moments is calculated. An input-output balance relationship for metabolic products is established for each spatial grid, where the input includes the generation rate of organic waste and the output includes the amount of metabolic products transferred to adjacent spatial grids. The accumulation rate of metabolic products within each spatial grid cell is calculated based on the input-output balance relationship. The water flow connectivity includes the flow direction and flow exchange intensity between adjacent spatial grids; a directed transport network of metabolites between different spatial grids is constructed based on the flow direction; and the transport flux of metabolites along the directed transport network from the source spatial grid to the downstream spatial grid is calculated based on the flow exchange intensity and the accumulation rate of the metabolites. The transport flux characterizes the amount of metabolites migrating between adjacent spatial grids per unit time. Spatial grids in the directed transport network with transport flux less than the outflow threshold are identified as regions where metabolic load accumulates.

[0008] Based on the environmental control requirements, the spatial response priority and temporal coordination relationships of each functional unit in the circulating water system are determined as follows: The spatial domain information of each functional unit in the circulating water system is obtained. The spatial domain information includes the installation location and adjustable spatial range of each functional unit. The spatial intersection area between the adjustable spatial range of each functional unit and the influence range of stress factors is calculated. The ratio of the spatial intersection area to the total area of ​​the influence range of stress factors is used as the spatial coverage matching degree. Extract the spatial distance between each functional unit and the center of the influence range of the stress factor, take the inverse value of the spatial distance as the response time coefficient, and then perform weighted fusion of the spatial coverage matching degree and the response time coefficient after normalization to obtain the spatial response priority value. The spatial overlap ratio between the spatial domain of each functional unit and the high-concentration aggregation area is calculated as the positional correlation degree. The processing efficiency parameters of each functional unit for metabolic load substances are obtained. The positional correlation degree and the processing efficiency parameters are weighted to obtain the start-up priority index of each functional unit. The start-up order is determined by arranging each functional unit in descending order according to the start-up priority index. The spatial overlap ratio between the spatial domain of each functional unit and the aggregation area of ​​the concentration gradient greater than a preset concentration threshold is calculated as the positional correlation degree. The processing efficiency parameter of each functional unit for metabolic load substances is obtained. The positional correlation degree and the processing efficiency parameter are weighted to obtain the activation priority index of each functional unit. The functional units are arranged in descending order according to the activation priority index to determine the temporal coordination relationship.

[0009] The collaborative working parameters are iteratively optimized using a reinforcement learning model. This model uses the growth rate of the aquaculture organisms and the stability of the water body as optimization objectives, and energy consumption constraints and equipment load constraints as boundary conditions. The output cyclic control command sequence includes: Obtain the total energy consumption value of the circulating water system and the equipment load rate of each functional unit at the current moment, calculate the extent by which the total energy consumption value exceeds the preset energy consumption limit, and calculate the extent by which the equipment load rate of each functional unit exceeds the preset load limit. When the total energy consumption or the equipment load rate of any functional unit exceeds the limit, a continuous penalty coefficient is calculated based on the extent of the exceedance. The continuous penalty coefficient is multiplied by the comprehensive reward value to achieve gradient penalty. When the total energy consumption and the equipment load rate of all functional units do not exceed the limit, the continuous penalty coefficient is set to a unit value to keep the comprehensive reward value unchanged. The comprehensive reward value modulated by the continuous penalty coefficient is output as the boundary constraint reward value, wherein the comprehensive reward value is the weighted sum of the growth rate reward component and the water body stability reward component. The policy parameters of the reinforcement learning model are updated based on the boundary constraint reward value, and the policy parameters describe the mapping relationship from the execution state of the collaborative work parameters to the adjustment action of the collaborative work parameters; Based on the updated strategy parameters, collaborative working parameter adjustment actions are generated, including adjustments to the operating power, operating time, and spatial range of each functional unit.

[0010] Calculating a continuous penalty coefficient based on the aforementioned over-limit range, and then multiplying the continuous penalty coefficient by the comprehensive reward value to achieve gradient penalty includes: The method further includes constructing a bi-objective reward function, which includes a growth rate reward component and a water body stability reward component. The growth rate reward component is calculated based on the weight growth rate and feed conversion rate of the cultured organism under the current collaborative working parameters. The water body stability reward component is calculated based on the fluctuation range of dissolved oxygen concentration and the fluctuation range of ammonia nitrogen concentration. Calculate a first ratio of the energy consumption exceedance to a preset energy consumption limit and a second ratio of the load exceedance to a preset load limit, and select the maximum value between the first ratio and the second ratio as the comprehensive exceedance ratio; Determine whether the overall over-limit ratio exceeds a preset segmentation threshold. If the overall over-limit ratio does not exceed the preset segmentation threshold, multiply the overall over-limit ratio with the first attenuation coefficient to obtain a linear penalty amount, and subtract the linear penalty amount from the unit value to obtain a continuous penalty coefficient. When the overall over-limit ratio exceeds a preset segmented threshold, the difference between the overall over-limit ratio and the preset segmented threshold is calculated as the over-threshold deviation. The preset segmented threshold is multiplied by a first attenuation coefficient to obtain a linear penalty at the threshold. The linear penalty at the threshold is subtracted from the unit value to obtain a penalty reference value at the threshold. The over-threshold deviation is multiplied by a second attenuation coefficient to obtain an exponential attenuation index. The negative value of the exponential attenuation index raised to the power of the natural constant is calculated as an exponential attenuation factor. The penalty reference value at the threshold is multiplied by the exponential attenuation factor to obtain a continuous penalty coefficient.

[0011] A second aspect of the present invention provides a control system for a reservoir recirculating aquaculture system that combines the Internet of Things (IoT) and AI, comprising: The first unit is used to collect environmental parameters and biological behavior parameters of the aquaculture water body through a distributed sensing unit, and to associate and map the environmental parameters and biological behavior parameters according to the collection time and spatial location to establish spatiotemporal state characterization data of the aquaculture water body; The second unit is used to identify the spatial distribution of stress factors that deviate from the normal range in the aquaculture water body and the concentration gradient of the cumulative metabolic load characteristics in local areas based on the spatiotemporal state characterization data. Combined with the dissolved oxygen demand threshold and ammonia nitrogen tolerance threshold of the cultured organisms at different growth stages, it calculates the environmental regulation requirements of each spatial area of ​​the aquaculture water body. The third unit is used to determine the spatial response priority and temporal coordination relationship of each functional unit in the circulating water system according to the environmental regulation demand. The spatial response priority is determined based on the spatial distribution location of the stress factors, and the temporal coordination relationship is determined based on the concentration gradient of the metabolic load accumulation characteristics, thereby generating the collaborative working parameters of each functional unit. The fourth unit is used to iteratively optimize the collaborative working parameters through a reinforcement learning model. The reinforcement learning model takes the growth rate of the aquaculture object and the stability of the water body as optimization objectives, and energy consumption constraints and equipment load constraints as boundary conditions, and outputs a cyclic control command sequence.

[0012] A third aspect of the embodiments of the present invention, An electronic device is provided, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0013] Fourth aspect of the present invention, A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0014] The beneficial effects of this application are as follows: This invention uses distributed sensing units to accurately collect and correlate environmental parameters and biological behavior parameters, thereby realizing the spatiotemporal dynamic characterization of aquaculture water conditions. It can accurately identify the distribution of stress factors and the characteristics of metabolic load accumulation in the water, thus accurately calculating the environmental regulation requirements of each area and improving the accuracy and targeting of water environment management.

[0015] The control strategy based on spatial response priority and temporal coordination enables the functional units of the recirculating aquaculture system to work together and carry out differentiated control of the water environment in different areas. This avoids the energy waste and environmental fluctuations caused by the single control method of traditional systems, and realizes the intelligent and precise operation of the recirculating aquaculture system.

[0016] The reinforcement learning model is used to iteratively optimize the collaborative working parameters, with the growth rate of the cultured organisms and the stability of the water body as the optimization objectives. At the same time, energy consumption and equipment load constraints are considered, and the adaptive adjustment of the system operation is realized. While ensuring the healthy growth of the cultured organisms, energy consumption is reduced and the system operation efficiency is improved, which has good ecological and economic benefits. Attached Figure Description

[0017] Figure 1 This is a schematic flowchart of the control method for a reservoir recirculating water industrialized aquaculture system that combines the Internet of Things and AI, according to an embodiment of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0020] Figure 1 This is a schematic flowchart of the control method for a reservoir recirculating water aquaculture system combining IoT and AI, as described in an embodiment of the present invention. Figure 1 As shown, the method includes: Environmental parameters and biological behavior parameters of the aquaculture water body are collected by distributed sensing units. The environmental parameters and biological behavior parameters are correlated and mapped according to the collection time and spatial location to establish spatiotemporal state characterization data of the aquaculture water body. Based on the spatiotemporal state characterization data, the spatial distribution of stress factors deviating from the normal range in the aquaculture water body and the concentration gradient of local metabolic load accumulation characteristics are identified. Combined with the dissolved oxygen demand threshold and ammonia nitrogen tolerance threshold of the cultured organisms at different growth stages, the environmental regulation requirements of each spatial region of the aquaculture water body are calculated. Based on the environmental regulation requirements, the spatial response priority and temporal coordination relationship of each functional unit in the circulating water system are determined. The spatial response priority is determined based on the spatial distribution location of the stress factors, and the temporal coordination relationship is determined based on the concentration gradient of the metabolic load accumulation characteristics, thereby generating the collaborative working parameters of each functional unit. The collaborative working parameters are iteratively optimized using a reinforcement learning model. The reinforcement learning model takes the growth rate of the aquaculture object and the stability of the water body as optimization objectives, and energy consumption constraints and equipment load constraints as boundary conditions, and outputs a cyclical control command sequence.

[0021] In implementing this invention, distributed sensing units are deployed at different spatial locations and depths within the aquaculture water body. These sensing units include a dissolved oxygen sensor, a pH sensor, a temperature sensor, an ammonia nitrogen sensor, a water flow velocity sensor, and an underwater camera. The dissolved oxygen sensor uses an electrochemical or optical probe, with a measurement range of 0 to 20 mg / L, a resolution of 0.01 mg / L, and a response time of less than 30 seconds. The ammonia nitrogen sensor uses an ion-selective electrode method, with a measurement range of 0 to 5 mg / L, a resolution of 0.01 mg / L, and a calibration cycle of 7 days. The temperature sensor uses a platinum resistance thermometer or a thermistor, with a measurement range of 0 to 50°C and an accuracy of ±0.1°C. The underwater camera has a resolution of at least 1920×1080 pixels, a frame rate of 25 frames per second, and is equipped with an infrared supplementary lighting device to adapt to low-light environments. The sensing units are connected to edge computing nodes wirelessly or via wired connections. The sampling cycle is set according to the parameter type: dissolved oxygen and temperature are collected every 5 minutes, and ammonia nitrogen is collected every 30 minutes. Video data is continuously collected and stored in 30-minute segments.

[0022] The collected environmental and biological behavior parameters are mapped to the timestamps of the data collection time and the spatial coordinates of the sensor deployment. The timestamps are in UTC format, accurate to the second, and the spatial coordinates are represented in a three-dimensional Cartesian coordinate system, with the center bottom of the aquaculture tank as the origin, the horizontal X and Y axes, and the vertical Z axis. Each sensor node records its three-dimensional coordinates upon deployment; for example, a dissolved oxygen sensor at coordinates (1.5, 2.0, 0.6) indicates a horizontal distance of 1.5 meters and 2.0 meters from the origin, and a height of 0.6 meters from the bottom. The collected data is a structured record containing the timestamp, sensor number, spatial coordinates, parameter type, measured value, unit, and quality flag. The quality flag is used to mark data validity; sensors that malfunction or measured values ​​exceeding the physically possible range are marked as invalid. Video data is processed using image processing techniques to extract behavioral features such as swimming speed, swimming frequency, and population density of the aquaculture animals. The extraction algorithm uses background subtraction to identify moving targets, optical flow to calculate velocity vectors, and clustering algorithms to calculate population density distribution. The extracted behavioral features are also linked to the timestamps and the spatial coordinates of the area covered by the camera's field of view. All data is stored in a time-series database, supporting fast queries by time window and spatial range. Data retention period is no less than 90 days, and expired data is downsampled and archived.

[0023] The aquaculture water body was divided into a multi-level three-dimensional spatial grid. For a cylindrical aquaculture tank with a diameter of 6 meters and a water depth of 2 meters, it was divided into 8×8 grids horizontally and 5 layers vertically, with each grid measuring approximately 0.5 meters × 0.5 meters × 0.4 meters. Grid numbers were represented by a three-dimensional index (i, j, k), where i and j represent the horizontal position and k represents the vertical layer, with k=1 for the bottom layer and k=5 for the surface layer. Environmental parameters within each grid were calculated using spatial interpolation. If there were no sensors within a grid, inverse distance-weighted interpolation was used to estimate parameters based on measurements from adjacent grids, with the interpolation weight inversely proportional to the square of the distance. The metabolic load characteristics within the grid were calculated based on the density of the cultured organisms, feed intake, and a metabolic product generation model. The model inputs included the average weight of the cultured organisms, feed protein content, and feeding interval, with the output being the ammonia nitrogen generation rate in milligrams per liter per hour.

[0024] When identifying stress factors deviating from normal ranges in aquaculture water, normal ranges for environmental parameters are pre-defined. For tilapia farming, the normal range for dissolved oxygen is 4.5 to 8.0 mg / L, for ammonia nitrogen it is 0 to 0.5 mg / L, and for water temperature it is 25 to 32℃. When the measured parameter value of a certain grid exceeds the normal range, the grid is marked as containing a stress factor, and the type of stress factor, the degree of exceedance, and the duration are recorded. For example, the dissolved oxygen value of grid (3, 4, 2) is 3.8 mg / L, which is below the lower limit of 4.5 mg / L, and the degree of exceedance is 0.7 mg / L. The duration is counted from the first detection of the exceedance. The spatial distribution of stress factors is formed into a three-dimensional distribution map, which is visualized through isosurface rendering or heatmaps.

[0025] When predicting the decreasing trend of dissolved oxygen and the increasing trend of ammonia nitrogen, the rate of change of environmental parameters within a continuous time window is calculated. The time window length is set to 1 hour, and the rate of change is calculated every 5 minutes. The dissolved oxygen rate of change is obtained by subtracting the measurement from 1 hour ago from the current measurement and then dividing by the time interval, with the unit being milligrams per liter per hour. When the dissolved oxygen rate of change is below -0.3 mg / L per hour, or the ammonia nitrogen rate of change is above +0.05 mg / L per hour, it is determined that there is an adverse trend, triggering the early warning mechanism. The early warning level is divided into three levels: Level 1 indicates that the rate of change has just exceeded the threshold, Level 2 indicates that the rate of change has exceeded the threshold by 1.5 times, and Level 3 indicates that the parameter will deviate from the normal range within 30 minutes. The early warning information includes grid location, parameter type, current value, predicted value, and suggested intervention measures. The predicted value is calculated using a linear extrapolation method, assuming that the rate of change remains constant over the next 30 minutes, to estimate the future value of the parameter.

[0026] The cumulative metabolic load characteristic is obtained by calculating the concentration difference of metabolites in each grid at adjacent time points. Taking ammonia nitrogen as an example, the ammonia nitrogen concentration of each grid is collected every hour, and the difference between two adjacent measurements is the concentration change during that time period. The concentration gradient is calculated by dividing the concentration difference between adjacent grids by the grid spacing, with units of milligrams per liter per meter. The direction of the concentration gradient points in the direction of increasing concentration, indicating the area of ​​metabolic product accumulation. According to the fixed water flow pattern of the aquaculture tank, the aerator is located at the bottom center of the tank, generating a bottom-up mainstream with a vertical flow velocity of approximately 3.5 cm / s in the central area. A bottom-up backflow is formed at the edge of the tank with a velocity of approximately 1.2 cm / s, and a radial flow is generated horizontally from the center to the edge with a velocity of approximately 2.0 cm / s. During system initialization, a water flow connectivity model is established through a one-time water flow measurement. The model records the water flow direction and velocity between each grid, constructing a directed graph structure where nodes are grids, edges are water flow channels, and the edge weight is the ratio of flow velocity to grid spacing, representing the mass transfer time coefficient. The transport path of metabolites is determined by a graph search algorithm, which traces the metabolites from the high-concentration grid to the low-concentration grid along the water flow direction. The path length is the number of grids traversed, and the transport time is the sum of the weights of the edges along the path.

[0027] The dissolved oxygen (DO) requirement and ammonia nitrogen (AM) tolerance thresholds for cultured fish at different growth stages were determined based on their average body weight. For juvenile fish weighing less than 50 grams, the DO requirement was 5.0 mg / L and the AM tolerance threshold was 0.3 mg / L. For medium-sized fish weighing 50 to 150 grams, the DO requirement was 4.5 mg / L and the AM tolerance threshold was 0.5 mg / L. For adult fish weighing more than 150 grams, the DO requirement was 4.0 mg / L and the AM tolerance threshold was 0.6 mg / L. The average body weight was obtained through periodic sampling and weighing, with a sampling period of 7 days. At least 30 individuals were randomly caught each time, and the average value was calculated and the growth stage identification was updated.

[0028] When calculating the environmental regulation requirements for each spatial area of ​​the aquaculture water body, for dissolved oxygen, the requirement is the difference between the actual dissolved oxygen value of the grid and the required threshold. If the difference is negative, oxygenation is required, and the unit of oxygenation is milligrams per liter. For ammonia nitrogen, the requirement is the difference between the actual ammonia nitrogen concentration of the grid and the tolerance threshold. If the difference is positive, nitrogen reduction is required. Nitrogen reduction is achieved through water replacement or biological filtration, and the unit of replacement is cubic meters. The calculation method is to multiply the grid volume by the concentration difference and divide by the concentration difference between the influent and effluent. The regulation requirements are summarized by grid to form a three-dimensional demand distribution map.

[0029] The recirculating aquaculture system comprises an aeration unit, a water circulation pump, a biological filtration unit, an ultraviolet disinfection unit, and a feeding unit. The aeration unit offers bottom micro-pore aeration and pure oxygen supply, with bottom aeration ranging from 0 to 30 cubic meters per hour and surface spray flow ranging from 0 to 1.5 L / min. The water circulation pump has a flow rate ranging from 0 to 200 cubic meters per hour and a head of 3 meters. The biological filtration unit has a treatment capacity of 100 cubic meters per hour and an ammonia nitrogen removal rate of over 80%. The ultraviolet disinfection unit has a power of 120 watts, a flow rate of 50 cubic meters per hour, and a sterilization rate of 99.9%. The feeding unit provides a daily feed amount ranging from 0 to 3% of the fish's total body weight, divided into three even feedings.

[0030] Spatial response priority is determined based on the degree of exceedance and the range of influence of the stress factor. The degree of exceedance is calculated by dividing the absolute value of the parameter's deviation from the normal range by the width of the normal range, yielding a dimensionless degree of exceedance. The range of influence is calculated by the number of consecutive grids containing the same type of stress factor. The priority score is calculated by multiplying the degree of exceedance by the range of influence and then by the sensitivity coefficient for the growth stage: 1.5 for juvenile fish, 1.0 for intermediate fish, and 0.8 for adult fish. The region with the highest score receives priority in response and is allocated more functional unit resources.

[0031] The timing coordination is determined based on the concentration gradient and transport path of the accumulated metabolic load. When a sustained increase in ammonia nitrogen concentration is detected in a certain area and the concentration gradient points towards that area, the water circulation pump near that area is activated first to pump out the high-concentration water and send it to the biological filtration unit. The filtered, low-concentration water is returned to the adjacent low-concentration area to achieve concentration equilibrium. The start-up delay time of the circulation pump is calculated based on the metabolic product transport time to ensure that replacement is completed before the metabolic products diffuse to the sensitive area. The treatment load of the biological filtration unit is calculated based on the circulating water flow rate and ammonia nitrogen concentration. When the load exceeds 80% of the rated treatment capacity, the circulation pump flow rate is reduced or diverted to the standby filtration unit.

[0032] A feedback regulation mechanism was established by combining water quality indicators and growth data of the cultured organisms. Water quality indicators included average dissolved oxygen, average ammonia nitrogen, and pH stability, with stability expressed as standard deviation. Growth data included weight gain rate and feed conversion ratio. The weight gain rate was calculated by subtracting the last measured average weight from the current average weight and then dividing by the time interval, expressed in grams per day. The feed conversion ratio was the weight gain divided by the total amount of feed given. When the weight gain rate was less than 80% of the standard growth rate for the species, growth was considered restricted. Water quality indicators for the corresponding time period were analyzed. If the average dissolved oxygen was below the required threshold, the operating time of the aeration unit was increased; if the average ammonia nitrogen was above the tolerance threshold, the frequency of water circulation was increased. Feeding strategies were adjusted based on the feeding response time, extracted through video analysis. Under normal circumstances, feeding behavior occurred within 30 seconds after feeding; under stress, this was delayed to more than 2 minutes. When the feeding response time exceeded 1 minute for three consecutive times, the amount of feed given per feeding was reduced or the feeding interval was extended to avoid uneaten feed increasing the metabolic load. The oxygenation strategy is adjusted according to the diurnal variation curve of dissolved oxygen. Since dissolved oxygen decreases more rapidly at night when photosynthesis ceases, aeration intensity is increased in advance at 6 PM to maintain dissolved oxygen levels above the required threshold. The water circulation strategy is adjusted according to temperature stratification. In summer, when surface water temperature is higher than bottom water temperature, vertical circulation mode is activated to transport warm surface water downwards, promoting more uniform temperature distribution.

[0033] The collaborative working parameters include the start and stop times, workload, and duration of each functional unit; the air volume or flow rate setpoint of the aeration unit; the flow rate and direction of the water circulation pump; the opening of the inlet valve of the biological filtration unit; and the feeding amount and timing of the feeding unit. Parameter generation follows the principle of minimizing energy consumption. While meeting environmental control requirements, priority is given to using functional units with low energy consumption. For example, bottom aeration consumes less energy than surface spraying, so the air volume of bottom aeration is increased first. Equipment load constraints are monitored through cumulative operating time. The continuous operating time of a single unit cannot exceed 8 hours; after exceeding this time, a forced shutdown and 30-minute cooling period is implemented, and the load is taken over by backup equipment.

[0034] The reinforcement learning model employs a deep Q-network architecture. The state space includes environmental parameter values, stress factor distribution, metabolic load gradient, functional unit operating status, and the growth stage of the cultured organisms for each grid cell. The action space includes parameter adjustments for each functional unit, with adjustments increments of ±10% of the current value. The reward function consists of a growth rate reward, a water stability reward, and an energy consumption penalty. The growth rate reward is the ratio of the weight gain rate to the standard growth rate; the water stability reward is the reciprocal of the standard deviations of dissolved oxygen and ammonia nitrogen; and the energy consumption penalty is the ratio of total power to baseline power. Model training utilizes an experience replay mechanism with a 10,000-data experience pool, sampling 128 data points for gradient updates each time. The initial learning rate is 0.001, decaying by 10% every 1000 epochs. The discount factor is 0.95, and the exploration rate linearly decays from 1.0 to 0.1 over 5000 epochs. The model is trained offline using historical data, covering at least one complete 90-day culture cycle, with the validation set comprising 20% ​​of the dataset. After online deployment, the model's parameters are updated every 24 hours, fine-tuned using the most recent 7 days' worth of running data, with the learning rate reduced to 0.0001 to prevent overfitting. The model's output sequence of cyclic control commands represents the parameter settings for the functional units for the next 2 hours, with commands updated every 30 minutes. A safety check is performed before execution to ensure that parameter changes do not exceed 20% of the previous settings, avoiding drastic fluctuations that could trigger stress.

[0035] In one optional implementation, identifying the spatial distribution location of stress factors deviating from the normal range and the concentration gradient of local metabolic load accumulation characteristics in aquaculture water bodies based on the spatiotemporal state characterization data includes: The spatiotemporal state representation data includes spatial dimension data and temporal dimension data; Based on the spatial dimension data, a multi-level spatial grid of the aquaculture water body is constructed. The environmental parameters within each spatial grid are judged to be within the normal range. The stress factor types corresponding to the environmental parameters that exceed the normal range are identified. The spatial grid positions corresponding to the stress factor types are recorded as the spatial distribution positions of the stress factors. Among them, the stress factor types include hypoxia stress factors where dissolved oxygen is lower than the oxygen demand threshold and nitrogen metabolism stress factors where ammonia nitrogen concentration is higher than the tolerance threshold. Extract the feeding and activity characteristic change curves of the cultured objects from the time dimension data, calculate the generation and accumulation rates of metabolic products in the culture water at adjacent time points, and determine the transmission path and accumulation region of metabolic load between different spatial grids by combining the water flow connectivity relationship of the spatial grid, and calculate the concentration gradient of metabolic load in the accumulation region.

[0036] In implementing this invention, the spatiotemporal state characterization data is based on multi-dimensional information collected from the aquaculture water body, including spatial and temporal data. The spatial data mainly consists of environmental parameters collected by sensor nodes distributed at different locations in the aquaculture water body, such as dissolved oxygen, pH value, temperature, and ammonia nitrogen concentration; the temporal data includes information on the changes in the behavioral and metabolic characteristics of the cultured organisms over time.

[0037] Constructing a multi-level spatial grid for aquaculture water is based on spatial dimensional data. This step involves dividing the aquaculture water into multiple grid units to form a three-dimensional grid structure. For a standard aquaculture tank environment, it can be divided into an 8×8×5 three-dimensional grid, i.e., 8×8 grids horizontally and 5 layers vertically. The size of each grid can be set to 0.5m×0.5m×0.3m, so that the entire structure covers an aquaculture water space of 6m×6m×2m.

[0038] Determining the normal range for environmental parameters within each spatial grid is a crucial step in identifying stress factors. Normal ranges for environmental parameters are pre-defined; for example, for tilapia farming, the normal range for dissolved oxygen is 4.5-8.0 mg / L, for ammonia nitrogen concentration is 0-0.5 mg / L, and for water temperature is 25-32℃. When a parameter in a grid cell is detected to exceed its normal range, a corresponding stress factor is identified at that location.

[0039] The method for identifying stress factor types is as follows: when the dissolved oxygen value at grid position (3,4,2) is detected to be 3.8 mg / L, which is lower than the oxygen demand threshold of 4.5 mg / L, this position is marked as containing a hypoxic stress factor; when the ammonia nitrogen concentration at grid position (5,6,1) is detected to be 0.68 mg / L, which is higher than the tolerance threshold of 0.5 mg / L, this position is marked as containing a nitrogen metabolism stress factor. The coordinates of all grid positions where stress factors were identified are recorded to form a spatial distribution map of stress factors.

[0040] The characteristic change curves of farmed fish in the time dimension data were extracted through video analysis of farmed fish behavior and recording of feed consumption. Underwater cameras were used to collect 24-hour video footage of fish activity. Image processing techniques were used to extract activity characteristics such as swimming speed, swimming frequency, and group aggregation / dispersion, and activity characteristic change curves were plotted at 30-minute intervals. Simultaneously, parameters such as feeding response time, feeding intensity, and feeding duration after each feeding were recorded to construct feeding characteristic change curves. For example, under normal conditions, the fish showed a strong feeding response within 30 seconds of feeding, with a feeding duration of approximately 5 minutes; however, under stress, the feeding response was delayed to 2 minutes, and the feeding duration was only 2 minutes.

[0041] Calculating the generation and accumulation rates of metabolites in aquaculture water at adjacent time points is a step in quantifying metabolic load. A metabolite generation model is established based on factors such as stocking density, feed intake, and fish growth stage. Taking ammonia nitrogen as an example, ammonia nitrogen concentration data for each grid point is collected every hour. The difference between two adjacent measurements divided by the time interval yields the ammonia nitrogen generation rate within that grid. In a practical example, at grid (4,4,3), the ammonia nitrogen concentration measured at 9:00 was 0.32 mg / L, and at 10:00 it was 0.39 mg / L. Therefore, the ammonia nitrogen generation rate at that location is 0.07 mg / L·h.

[0042] Determining metabolic load transfer paths by combining spatial grid-based water flow connectivity is based on a fixed water flow pattern within the aquaculture tank. According to the aeration and circulation system configuration of the tank, a water flow connectivity model is established during system initialization through a one-time water flow measurement or fluid dynamics simulation. The aeration pipes are located around the bottom circumference of the tank, forming a top-down main flow at the edge, generating a bottom-up vertical flow velocity of approximately 3.5 cm / s in the edge region, a bottom-down backflow velocity of approximately 1.2 cm / s in the center region, and a radial flow from the edge to the center with a velocity of approximately 2.0 cm / s. Based on this fixed water flow pattern, a connectivity graph between the grids is constructed to determine the transfer path of metabolic products from high-concentration areas to low-concentration areas. When equipment configuration changes (such as adjusting aerator power or location), the water flow model can be recalibrated.

[0043] The accumulation zone of metabolic load was determined by analyzing the water flow connectivity graph and the distribution of metabolite concentrations. Regions with low water flow velocities and high metabolite concentrations in surrounding grids were identified as potential accumulation zones. In practical applications, it was found that the average water flow velocity in grid regions (5,5,1) to (6,6,1) was only 0.5 cm / s, while the ammonia nitrogen generation rate in the surrounding grids all exceeded 0.05 mg / L·h. Therefore, this region was marked as the accumulation zone of metabolic load.

[0044] The concentration gradient of metabolic load within the accumulation region was calculated by spatial interpolating the metabolic product concentrations at each grid point within the accumulation region. Dense sampling was performed on the marked accumulation region; for example, 25 sampling points were added within the region from (5,5,1) to (6,6,1), forming a refined 5×5 grid. The ammonia nitrogen concentration at the edge grid (5,5,1) was measured to be 0.52 mg / L, and the ammonia nitrogen concentration at the central grid (5.5,5.5,1) was 0.78 mg / L. Based on this, the ammonia nitrogen concentration gradient within the accumulation region was calculated to be approximately 0.05 mg / L per 0.1 meter. In this way, a metabolic load concentration gradient map of the accumulation region was generated, visually demonstrating the spatial distribution characteristics of the metabolic load.

[0045] In one optional implementation, the generation and accumulation rates of metabolic products in the aquaculture water at adjacent time points are calculated. Combined with the water flow connectivity of the spatial grid, the transmission paths and accumulation regions of metabolic load between different spatial grids are determined, including: Based on the difference in feed intake at adjacent moments in the feeding characteristic change curve and the metabolic conversion coefficient of the cultured organisms, the generation rate of ammonia nitrogen metabolites in the culture water at adjacent moments is calculated. Based on the difference in activity intensity at adjacent moments in the activity characteristic change curve and the excretion pattern parameters of the cultured organisms, the generation rate of organic waste in the culture water at adjacent moments is calculated. An input-output balance relationship for metabolic products is established for each spatial grid, where the input includes the generation rate of organic waste and the output includes the amount of metabolic products transferred to adjacent spatial grids. The accumulation rate of metabolic products within each spatial grid cell is calculated based on the input-output balance relationship. The water flow connectivity includes the flow direction and flow exchange intensity between adjacent spatial grids; a directed transport network of metabolites between different spatial grids is constructed based on the flow direction; and the transport flux of metabolites along the directed transport network from the source spatial grid to the downstream spatial grid is calculated based on the flow exchange intensity and the accumulation rate of the metabolites. The transport flux characterizes the amount of metabolites migrating between adjacent spatial grids per unit time. Spatial grids in the directed transport network with transport flux less than the outflow threshold are identified as regions where metabolic load accumulates.

[0046] For example, the cultured species was bass, and the culture water was divided into 10×10 grid units. Each grid unit was 5 meters × 5 meters in size, with a water depth of 2 meters. The stocking density was 80 bass per cubic meter, with an average weight of 250 grams per fish. Feeding characteristic variation curves obtained through monitoring showed that from 6:00 to 8:00, feed intake increased from 0 to a maximum of 25 grams per kilogram of body weight, while from 18:00 to 20:00, feed intake decreased from 25 grams per kilogram of body weight back to 0. Activity characteristic variation curves showed that the activity intensity of bass gradually increased after sunrise, reaching a peak of 100% at noon, and gradually decreased to 20% at night.

[0047] When calculating the ammonia nitrogen metabolite production rate in aquaculture water at adjacent time points based on feeding characteristic change curves, calculations are performed with a 1-hour time interval. For example, between 6:00 and 7:00, the difference in feed intake is 12.5 g / kg body weight. The ammonia nitrogen metabolism conversion coefficient for bass is 0.03, meaning that 1 gram of feed ingested produces 0.03 grams of ammonia nitrogen. Therefore, the ammonia nitrogen metabolite production rate during this period is 12.5 × 0.03 = 0.375 g / kg body weight / hour.

[0048] Considering the stocking density and the average weight of the fish, the total ammonia nitrogen generation in each grid unit (5×5×2 cubic meters) is 0.375×0.25×3×5×5×2=28.125 grams / hour.

[0049] The organic waste generation rate was calculated based on the difference in activity intensity between adjacent moments and the excretion pattern parameters in the activity characteristic change curve. The excretion pattern parameters for bass were set as follows: for every 10% increase in activity intensity, the amount of organic waste excreted increased by 0.05 g / kg body weight / hour. Between 8:00 and 9:00, the activity intensity increased from 60% to 80%, a difference of 20%. Therefore, the organic waste generation rate was 20% × 0.05 ÷ 10% = 0.1 g / kg body weight / hour. Considering the stocking density and fish weight, the total organic waste generated within this grid cell was 0.1 × 0.25 × 3 × 5 × 5 × 2 = 7.5 g / hour.

[0050] When establishing the input-output balance of metabolic products for each spatial grid, we take the grid cell located at coordinates (3,4) as an example. During the period from 9:00 to 10:00, the ammonia nitrogen generation rate of this grid cell is 25 g / h, and the organic waste generation rate is 8 g / h. The water flow direction shows that this grid cell exchanges water with four adjacent grid cells at coordinates (3,3), (3,5), (2,4), and (4,4). According to the water exchange intensity data, the outflow volume from this grid cell is as follows: 5 m³ / h towards (3,3), 8 m³ / h towards (3,5), 3 m³ / h towards (2,4), and 6 m³ / h towards (4,4). Assuming the ammonia nitrogen concentration in the grid cell during this period is 0.5 mg / L, the outflow volume of ammonia nitrogen is (5+8+3+6)×0.5=11 g / h. Similarly, the outflow of organic waste is (5+8+3+6)×0.2=4.4 grams / hour, of which 0.2 mg / L is the concentration of organic waste.

[0051] Therefore, the accumulation rate of ammonia nitrogen in this grid cell is 25-11=14 grams / hour, and the accumulation rate of organic waste is 8-4.4=3.6 grams / hour. This means that 14 grams of ammonia nitrogen and 3.6 grams of organic waste accumulate in this grid cell per hour.

[0052] When constructing a directed transport network for metabolites between different spatial grids, the transport path is determined based on the direction of water flow. Taking the grid cell at coordinate (3,4) as an example, the water flowing out of this cell points to four grid cells at (3,3), (3,5), (2,4), and (4,4). Therefore, in the directed transport network, there are directed edges from (3,4) to these four grid cells.

[0053] When calculating the transport flux of metabolites along a directed transport network, take the transport from (3,4) to (3,5) as an example. The water exchange intensity is 8 cubic meters per hour, and the ammonia nitrogen concentration is 0.5 mg / L. Therefore, the ammonia nitrogen transport flux is 8 × 0.5 = 4 g / hour. Similarly, the transport flux of organic waste is 8 × 0.2 = 1.6 g / hour.

[0054] When identifying areas of metabolic load accumulation, an outflow threshold of 2 g / h was assumed. In the simulation of the entire aquaculture area, the ammonia nitrogen transfer flux of grid cells located at coordinates (7,8), (7,9), (8,8), and (8,9) was found to be less than 2 g / h; therefore, these grid cells were identified as ammonia nitrogen accumulation areas. Similarly, the organic waste transfer flux of grid cells located at coordinates (9,1), (9,2), and (10,1) was less than 2 g / h, and these were identified as organic waste accumulation areas.

[0055] The above methods can be used to dynamically monitor the distribution of metabolic products in aquaculture water, promptly identify areas of metabolic load accumulation, provide a scientific basis for aquaculture water quality control and disease prevention, and improve aquaculture efficiency and environmental sustainability.

[0056] In one optional implementation, determining the spatial response priority and temporal coordination relationship of each functional unit in the circulating water system based on the environmental control demand includes: The spatial domain information of each functional unit in the circulating water system is obtained. The spatial domain information includes the installation location and adjustable spatial range of each functional unit. The spatial intersection area between the adjustable spatial range of each functional unit and the influence range of stress factors is calculated. The ratio of the spatial intersection area to the total area of ​​the influence range of stress factors is used as the spatial coverage matching degree. Extract the spatial distance between each functional unit and the center of the influence range of the stress factor, take the inverse value of the spatial distance as the response time coefficient, and then perform weighted fusion of the spatial coverage matching degree and the response time coefficient after normalization to obtain the spatial response priority value. The spatial overlap ratio between the spatial domain of each functional unit and the high-concentration aggregation area is calculated as the positional correlation degree. The processing efficiency parameters of each functional unit for metabolic load substances are obtained. The positional correlation degree and the processing efficiency parameters are weighted to obtain the start-up priority index of each functional unit. The start-up order is determined by arranging each functional unit in descending order according to the start-up priority index. The spatial overlap ratio between the spatial domain of each functional unit and the aggregation area of ​​the concentration gradient greater than a preset concentration threshold is calculated as the positional correlation degree. The processing efficiency parameter of each functional unit for metabolic load substances is obtained. The positional correlation degree and the processing efficiency parameter are weighted to obtain the activation priority index of each functional unit. The functional units are arranged in descending order according to the activation priority index to determine the temporal coordination relationship.

[0057] First, obtain the spatial domain information of each functional unit in the circulating water system. This information includes the installation location and adjustable spatial range of the functional unit. For example, in an aquaculture system with an area of ​​100 square meters, there are three functional units: filter A, aerator B, and pump C. Filter A is installed in the northeast corner of the system, and its adjustable range covers the eastern half of the system, approximately 50 square meters; aerator B is located in the center of the system, with a radiation range of a circular area with a radius of 3 meters, approximately 28 square meters; pump C is installed in the southwest corner, and its influence range covers the southwest area, approximately 35 square meters.

[0058] When an increase in ammonia nitrogen concentration is detected in the southeast region, forming a stress factor influence area of ​​approximately 20 square meters, the spatial overlap area between each functional unit and this stress area is calculated. For example, filter A overlaps with the stress area by 15 square meters, aerator B by 8 square meters, and pump C by 5 square meters. Dividing these overlap areas by the total stress factor influence area of ​​20 square meters yields the spatial coverage matching degree: 0.75 for filter A, 0.4 for aerator B, and 0.25 for pump C.

[0059] Measure the spatial distance between each functional unit and the center of the stress factor's influence range. Assume filter A is 4 meters from the stress center, aerator B is 2.5 meters away, and pump C is 6 meters away. Inverse values ​​of these distances are used as response time coefficients: -4, -2.5, and -6, respectively. After normalization, the response time coefficients become 0.6, 0.8, and 0.4, respectively.

[0060] The normalized spatial coverage matching degree and response time coefficient are weighted and fused, with weights of 0.6 and 0.4 respectively. The spatial response priority of filter A is calculated as 0.75×0.6+0.6×0.4=0.69, aerator B as 0.4×0.6+0.8×0.4=0.56, and pump C as 0.25×0.6+0.4×0.4=0.31. According to the calculation results, the spatial response priorities are ranked from high to low as follows: filter A, aerator B, and pump C.

[0061] To determine the activation sequence of functional units, the spatial overlap ratio between the spatial domain of each functional unit and the high-concentration accumulation area is calculated as the positional correlation. Assuming the high-concentration ammonia nitrogen area in the system is 15 square meters, the spatial domain of filter A overlaps with the high-concentration area by 12 square meters, a ratio of 0.8; aerator B overlaps by 7 square meters, a ratio of 0.47; and pump C overlaps by 3 square meters, a ratio of 0.2.

[0062] Simultaneously, the treatment efficiency parameters of each functional unit for metabolic load substances were obtained. Filter A's treatment efficiency for ammonia nitrogen is a reduction of 5 mg / L per hour, aerator B's is a reduction of 3 mg / L per hour, and pump C's is a reduction of 2 mg / L per hour. Assuming a weight ratio of location correlation to treatment efficiency of 0.7:0.3, after normalizing the treatment efficiency, the normalized treatment efficiencies of filter A, aerator B, and pump C are 1.0, 0.6, and 0.4, respectively.

[0063] The system calculates the startup priority index of each functional unit based on a weighted average of location correlation and processing efficiency parameters. The startup priority index for filter A is 0.8 × 0.7 + 1.0 × 0.3 = 0.86, for aerator B it is 0.47 × 0.7 + 0.6 × 0.3 = 0.509, and for pump C it is 0.2 × 0.7 + 0.4 × 0.3 = 0.26. Based on the calculation results, the system determines the startup sequence of the functional units as: filter A, aerator B, and pump C.

[0064] To determine the timing coordination relationship, the percentage of the spatial overlap between the spatial domain of each functional unit and the area where the concentration gradient exceeds a preset concentration threshold is calculated. Assuming the area in the system where the concentration gradient exceeds the preset threshold is 25 square meters, filter A overlaps with this area by 18 square meters, accounting for 0.72; aerator B overlaps by 15 square meters, accounting for 0.6; and pump C overlaps by 10 square meters, accounting for 0.4.

[0065] Applying the aforementioned processing efficiency parameters and setting the weight ratio of location correlation to processing efficiency to 0.6:0.4, the start-up priority index of each functional unit is calculated as follows: Filter A is 0.72×0.6+1.0×0.4=0.832, Aerator B is 0.6×0.6+0.6×0.4=0.6, and Pump C is 0.4×0.6+0.4×0.4=0.4.

[0066] Based on the above calculations, the timing sequence of each functional unit in the circulating water system is determined as follows: first, filter A is activated to treat the high-concentration area; after 45 seconds, aerator B is activated to increase dissolved oxygen and promote nitrification; and after 90 seconds, pump C is activated to accelerate water circulation and diffusion. In practical applications, the system will also dynamically adjust the activation interval and runtime based on real-time monitoring data to ensure optimal treatment results.

[0067] This implementation method achieves precise control of each functional unit of the circulating water system through accurate calculation and weighted fusion of spatial coverage matching degree, response time coefficient, location correlation degree and processing efficiency parameters. It can effectively cope with various stress factors in the water environment, improve system processing efficiency, reduce energy consumption and extend equipment service life.

[0068] In one optional implementation, the collaborative working parameters are iteratively optimized using a reinforcement learning model. The reinforcement learning model uses the growth rate of the aquaculture object and water stability as optimization objectives, and energy consumption constraints and equipment load constraints as boundary conditions. The output cyclical control command sequence includes: Obtain the total energy consumption value of the circulating water system and the equipment load rate of each functional unit at the current moment, calculate the extent by which the total energy consumption value exceeds the preset energy consumption limit, and calculate the extent by which the equipment load rate of each functional unit exceeds the preset load limit. When the total energy consumption or the equipment load rate of any functional unit exceeds the limit, a continuous penalty coefficient is calculated based on the extent of the exceedance. The continuous penalty coefficient is multiplied by the comprehensive reward value to achieve gradient penalty. When the total energy consumption and the equipment load rate of all functional units do not exceed the limit, the continuous penalty coefficient is set to a unit value to keep the comprehensive reward value unchanged. The comprehensive reward value modulated by the continuous penalty coefficient is output as the boundary constraint reward value, wherein the comprehensive reward value is the weighted sum of the growth rate reward component and the water body stability reward component. The policy parameters of the reinforcement learning model are updated based on the boundary constraint reward value, and the policy parameters describe the mapping relationship from the execution state of the collaborative work parameters to the adjustment action of the collaborative work parameters; Based on the updated strategy parameters, collaborative working parameter adjustment actions are generated, including adjustments to the operating power, operating time, and spatial range of each functional unit.

[0069] In recirculating aquaculture systems, iterative optimization of collaborative working parameters using reinforcement learning models is an efficient control method. This method uses the growth rate of the cultured organisms and water stability as optimization objectives, while considering energy consumption constraints and equipment load constraints as boundary conditions, thereby outputting a sequence of cyclic control commands. Implementing this process involves several technical steps, including acquiring system parameters, calculating the extent of exceeding limits, setting penalty mechanisms, updating the learning model, and generating parameter adjustment actions.

[0070] When acquiring the current system status parameters, the system collects the total energy consumption of the recirculating aquaculture system and the equipment load rate of each functional unit in real time. For example, assuming the aquaculture system includes three functional units: aeration, filtration, and disinfection, and at a certain moment the total system energy consumption is 15 kWh, the preset energy consumption limit is 20 kWh, and the load rates of each functional unit are 85% for the aeration unit, 70% for the filtration unit, and 50% for the disinfection unit, while the preset load limits for each are 90%. At this time, the system calculates that the total energy consumption exceeds the limit by 0 (because it does not exceed the limit), the load exceedance of the aeration unit is 0 (because it does not exceed the limit), the load exceedance of the filtration unit is 0 (because it does not exceed the limit), and the load exceedance of the disinfection unit is 0 (because it does not exceed the limit). Simultaneously, the system collects key water quality indicators, including dissolved oxygen concentration, ammonia nitrogen concentration, and suspended solids concentration. For example, if the detected dissolved oxygen concentration is 4.2 mg / L, the ammonia nitrogen concentration is 0.62 mg / L, and the suspended solids concentration is 35 mg / L, these water quality indicators will serve as important bases for parameter adjustment decisions.

[0071] In calculating the over-limit range, the system determines the over-limit range by comparing the current energy consumption value with the preset energy consumption upper limit. For example, if at another moment the system's total energy consumption rises to 22 kWh, exceeding the preset upper limit by 2 kWh, then the over-limit range is 10% (calculated by dividing the excess by the preset upper limit). Similarly, for equipment load rates, the system calculates the over-limit situation for each functional unit. For instance, if the oxygenation unit's load rate rises to 95%, exceeding the preset upper limit by 5%, then the unit's over-limit range is 5.56% (calculated by dividing the excess by the preset upper limit).

[0072] The calculation of the continuous penalty coefficient is a crucial step in implementing boundary constraints. When the total energy consumption or the load rate of any functional unit exceeds the preset upper limit, the system calculates the continuous penalty coefficient based on the extent of the exceedance. For example, a penalty coefficient of 0.8 corresponds to an energy consumption exceedance of 10%, and a penalty coefficient of 0.9 corresponds to a load rate exceedance of 5.56%. When multiple exceedances exist, the minimum penalty coefficient is taken as the final continuous penalty coefficient. When the total system energy consumption and the load rate of all functional units are within limits, the continuous penalty coefficient is set to 1.0, meaning no penalty adjustment is applied to the overall reward value.

[0073] The calculation of the comprehensive reward value considers two key indicators: the growth rate of the cultured organisms and the stability of the water body. The growth rate reward component is calculated based on the weight gain rate of the cultured organisms. For example, if the average weight of fish increases from 100 grams to 115 grams during the observation period, the weight gain rate is 15%, which can be converted into a growth rate reward component of 0.75 (assuming a linear transformation method is used). The water stability reward component is based on the fluctuation of key water quality parameters. For example, if the dissolved oxygen fluctuation range is controlled between 6.5-7.5 mg / L, with a fluctuation range of 15%, it can be converted into a water stability reward component of 0.85. Assuming that the weights of growth rate and water stability are 0.6 and 0.4 respectively, the comprehensive reward value is 0.75 × 0.6 + 0.85 × 0.4 = 0.79.

[0074] The boundary constraint reward value is calculated by multiplying the continuous penalty coefficient by the comprehensive reward value. In the example above, if the continuous penalty coefficient is 0.8 (due to a 10% energy consumption exceedance), the boundary constraint reward value is 0.79 × 0.8 = 0.632, reflecting the gradient penalty for exceeding the limit. If the system parameters do not exceed the limit, the continuous penalty coefficient is 1.0, and the boundary constraint reward value is equal to the comprehensive reward value of 0.79.

[0075] The policy parameters of a reinforcement learning model are updated based on the boundary constraint reward value. Assuming a policy gradient algorithm is used, when the boundary constraint reward value is 0.632, the system calculates the policy gradient based on this value and updates the model parameters according to a preset learning rate (e.g., 0.01). The updated policy parameters reflect the optimized mapping relationship from system state to operational actions. These parameters are stored in the policy network for subsequent decision-making processes.

[0076] The final output step is generating collaborative working parameter adjustment actions. Based on the updated strategy parameters, the system combines real-time water quality indicators to generate operational parameter adjustment suggestions for each functional unit. When the dissolved oxygen concentration is detected to be lower than the set threshold (e.g., 4.5 mg / L), the system suggests that the oxygenation unit increase its operating power by 15% (from 2 kW to 2.3 kW), extend its operating time by 1 hour (from 24 hours to 25 hours), and expand its spatial coverage area by 12% (from covering 100 square meters to 112 square meters) to quickly improve the dissolved oxygen level in the water. When the ammonia nitrogen concentration or suspended solids concentration exceeds the tolerance threshold (e.g., ammonia nitrogen exceeding 0.5 mg / L or suspended solids exceeding 30 mg / L), the system suggests increasing the time and frequency of sewage discharge operations, for example, extending the sewage discharge time from 5 minutes each time to 8 minutes, increasing the sewage discharge frequency from once every 4 hours to once every 3 hours, while increasing the operating power of the filtration unit by 10% (from 1.5 kW to 1.65 kW) to accelerate the water purification speed. For the filtration and disinfection units, the system will also output corresponding parameter adjustment values, forming a complete collaborative working parameter adjustment scheme.

[0077] Through the aforementioned technologies, reinforcement learning models can effectively control energy consumption and equipment load while ensuring the growth performance of aquaculture organisms and the stability of the aquatic environment, thereby achieving intelligent and refined management of recirculating aquaculture systems.

[0078] In one optional implementation, calculating a continuous penalty coefficient based on the over-limit magnitude and multiplying the continuous penalty coefficient by the comprehensive reward value to achieve gradient penalty includes: The method further includes constructing a bi-objective reward function, which includes a growth rate reward component and a water body stability reward component. The growth rate reward component is calculated based on the weight growth rate and feed conversion rate of the cultured organism under the current collaborative working parameters. The water body stability reward component is calculated based on the fluctuation range of dissolved oxygen concentration and the fluctuation range of ammonia nitrogen concentration. Calculate a first ratio of the energy consumption exceedance to a preset energy consumption limit and a second ratio of the load exceedance to a preset load limit, and select the maximum value between the first ratio and the second ratio as the comprehensive exceedance ratio; Determine whether the overall over-limit ratio exceeds a preset segmentation threshold. If the overall over-limit ratio does not exceed the preset segmentation threshold, multiply the overall over-limit ratio with the first attenuation coefficient to obtain a linear penalty amount, and subtract the linear penalty amount from the unit value to obtain a continuous penalty coefficient. When the overall over-limit ratio exceeds a preset segmented threshold, the difference between the overall over-limit ratio and the preset segmented threshold is calculated as the over-threshold deviation. The preset segmented threshold is multiplied by a first attenuation coefficient to obtain a linear penalty at the threshold. The linear penalty at the threshold is subtracted from the unit value to obtain a penalty reference value at the threshold. The over-threshold deviation is multiplied by a second attenuation coefficient to obtain an exponential attenuation index. The negative value of the exponential attenuation index raised to the power of the natural constant is calculated as an exponential attenuation factor. The penalty reference value at the threshold is multiplied by the exponential attenuation factor to obtain a continuous penalty coefficient.

[0079] For example, the construction of the bi-objective reward function is implemented in the reward evaluation module of the system control logic. This module receives growth monitoring data of the cultured organisms and aquatic environment monitoring data as input. The calculation of the growth rate reward component depends on the real-time acquisition of weight growth rate data and feed conversion rate data. The weight growth rate is obtained by non-contact weighing of the cultured organisms at fixed times every day using an underwater weighing device. The weighing device is deployed at a specific channel location in the culture water area. When the cultured organisms pass through this channel during their daily patrols, the weighing sensor is triggered. The system records the weighing data for seven consecutive days and calculates the average daily weight increase. The weight growth rate value is obtained by dividing the average daily weight increase by the initial weight.

[0080] The feed conversion rate is calculated as the ratio of the cumulative feeding amount recorded by the feeding system to the weight gain of the cultured organism. The feeding system records the feed mass at each feeding and accumulates it to form the total feeding amount for the period. Dividing the weight gain by the total feeding amount for the period yields the feed conversion rate, which reflects the efficiency of converting unit mass of feed into the weight of the cultured organism. The growth rate bonus component is calculated by weightedly combining the weight growth rate and the feed conversion rate. The weighting coefficient for the weight growth rate is set to 0.6, and the weighting coefficient for the feed conversion rate is set to 0.4. The weighted combination result is multiplied by a preset growth bonus amplification coefficient to obtain the final value of the growth rate bonus component. The typical range of the amplification coefficient is five to fifteen, and it is adjusted according to the culture stage and species characteristics.

[0081] The calculation of the water stability bonus component is based on real-time monitoring data of dissolved oxygen concentration fluctuations and ammonia nitrogen concentration fluctuations. Dissolved oxygen concentration monitoring is achieved through a dissolved oxygen sensor array distributed throughout the aquaculture water area. The sensor nodes are deployed according to the principle of spatial uniformity, and each node collects dissolved oxygen concentration values ​​at a fixed sampling period of ten minutes. The concentration data collected continuously for seventy-two hours constitutes a time series, and the difference between the maximum and minimum concentration values ​​in this time series is calculated as the dissolved oxygen concentration fluctuation range. The ammonia nitrogen concentration fluctuation range is obtained in the same way as the dissolved oxygen concentration. The ammonia nitrogen sensor nodes adopt the same spatial deployment strategy and sampling period, and the difference between the maximum and minimum concentration values ​​is calculated after continuously collecting data for seventy-two hours.

[0082] The water stability reward component is calculated by constructing penalty functions for the fluctuation amplitudes of dissolved oxygen concentration and ammonia nitrogen concentration. These penalty functions map the fluctuation amplitude to a negative reward value; the larger the fluctuation amplitude, the larger the absolute value of the negative reward. The weighting coefficients for both dissolved oxygen and ammonia nitrogen concentration fluctuation amplitudes are set to 0.5. After weighted combination, the water stability reward component is obtained by subtracting the weighted fluctuation penalty from the preset stability benchmark value. A typical setting for the stability benchmark value is 10, representing the maximum stability reward under ideal, fluctuation-free conditions. The comprehensive reward value of the bi-objective reward function is obtained by summing the growth rate reward component and the water stability reward component. This comprehensive reward value serves as the benchmark input for subsequent penalty corrections.

[0083] The first ratio between the energy consumption exceedance and the preset energy consumption limit is calculated in real time by the system power monitoring module. The power monitoring module collects the instantaneous power data of each functional unit and sums them to obtain the total system power. The instantaneous power sampling period is set to five seconds. After continuously collecting the total power data for one hour, the moving average is calculated. The sliding window length is set to twelve sampling points, and the sliding step size is set to one sampling point. The moving average total power is compared with the preset energy consumption limit. The preset energy consumption limit is determined according to the system design capacity. For a system with a total installed power of 100 kilowatts, the energy consumption limit is set to 85 kilowatts. When the moving average total power exceeds the energy consumption limit, the excess part is the energy consumption exceedance. The first ratio is obtained by dividing the energy consumption exceedance by the preset energy consumption limit.

[0084] The second ratio, calculated by monitoring the load rate of key actuators (including circulating water pumps and aerators), is obtained through load rate monitoring. Each actuator is equipped with a load sensor, with a sampling period of ten seconds. After continuously collecting data for one hour, the sliding average load rate of each actuator is calculated. The sliding window is consistent with the power monitoring. The maximum value of the sliding average load rate among all actuators is selected as the system's highest load rate. The preset load limit is set to 90% of the rated load. When the system's highest load rate exceeds the preset load limit, the excess portion is the load over-limit. Dividing the load over-limit by the preset load limit yields the second ratio. The comprehensive over-limit ratio is determined by comparing the first and second ratios. The system selects the larger of the two as the comprehensive over-limit ratio. This ratio quantifies the degree to which the current operating state encroaches on the system's safety margin.

[0085] The preset segmentation threshold is set to 0.2. This threshold divides the penalty calculation into two intervals. The comparison between the overall over-limit ratio and the preset segmentation threshold determines the choice of penalty calculation path. When the overall over-limit ratio does not exceed the preset segmentation threshold, the system performs linear penalty calculation, multiplying the overall over-limit ratio by a first attenuation coefficient set to 0.5. The result of the multiplication is the linear penalty amount, which represents the attenuation ratio of the overall reward value. The continuous penalty coefficient is obtained by subtracting the linear penalty amount from the unit value. The unit value is defined as 1.0, and the continuous penalty coefficient ranges from 0.9 to 1.0. This coefficient represents the retention ratio of the overall reward value.

[0086] When the overall over-limit ratio exceeds the preset segmented threshold, the system performs nonlinear penalty calculation. The difference between the overall over-limit ratio and the preset segmented threshold is calculated as the over-threshold deviation. The over-threshold deviation reflects the extent to which the over-limit exceeds the safety threshold. The preset segmented threshold is multiplied by the first attenuation coefficient to obtain the linear penalty at the threshold. The linear penalty at the threshold is subtracted from the unit value to obtain the penalty benchmark value at the threshold. The penalty benchmark value at the threshold is used as the starting benchmark for nonlinear penalty calculation.

[0087] The deviation from the threshold is multiplied by a second decay coefficient to obtain an exponential decay index. The second decay coefficient is set to 3.0. The exponential decay index controls the nonlinear decay rate of the penalty function. The exponential decay index is calculated by raising the natural constant to the power of the exponential decay index and taking its negative value as the exponential decay factor. The natural constant is set to 2.71828. The value of the exponential decay factor ranges from 0 to 1.0 and rapidly approaches zero as the deviation from the threshold increases. The penalty baseline value at the threshold is multiplied by the exponential decay factor to obtain a continuous penalty coefficient. This coefficient exhibits exponential decay characteristics as the overall over-limit ratio increases within the nonlinear penalty range. The continuous penalty coefficient is multiplied by the overall reward value to achieve gradient-based penalty. The multiplication operation is performed at the end of each control cycle, which is set to five minutes. The result of the multiplication operation is the final reward value after penalty correction. This value is input to the reinforcement learning policy update module for gradient calculation of policy parameters.

[0088] The present invention provides a control system for a reservoir recirculating aquaculture system that combines the Internet of Things (IoT) and AI, comprising: The first unit is used to collect environmental parameters and biological behavior parameters of the aquaculture water body through a distributed sensing unit, and to associate and map the environmental parameters and biological behavior parameters according to the collection time and spatial location to establish spatiotemporal state characterization data of the aquaculture water body; The second unit is used to identify the spatial distribution of stress factors that deviate from the normal range in the aquaculture water body and the concentration gradient of the cumulative metabolic load characteristics in local areas based on the spatiotemporal state characterization data. Combined with the dissolved oxygen demand threshold and ammonia nitrogen tolerance threshold of the cultured organisms at different growth stages, it calculates the environmental regulation requirements of each spatial area of ​​the aquaculture water body. The third unit is used to determine the spatial response priority and temporal coordination relationship of each functional unit in the circulating water system according to the environmental regulation demand. The spatial response priority is determined based on the spatial distribution location of the stress factors, and the temporal coordination relationship is determined based on the concentration gradient of the metabolic load accumulation characteristics, thereby generating the collaborative working parameters of each functional unit. The fourth unit is used to iteratively optimize the collaborative working parameters through a reinforcement learning model. The reinforcement learning model takes the growth rate of the aquaculture object and the stability of the water body as optimization objectives, and energy consumption constraints and equipment load constraints as boundary conditions, and outputs a cyclic control command sequence.

[0089] A third aspect of the present invention provides an electronic device, comprising: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0090] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0091] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for controlling a reservoir circulating water factory farming system combining the Internet of Things and AI, characterized by, The application relates to a method for realizing automatic environmental regulation of aquaculture water bodies. The method comprises the following steps: Collecting environmental parameters and biological behavior parameters of the aquaculture water bodies by a distributed sensing unit, and mapping the environmental parameters and the biological behavior parameters according to the collection time and the spatial position to establish time-space state characterization data of the aquaculture water bodies; Identifying the spatial distribution position of stress factors deviating from the normal range and the concentration gradient of the metabolic load accumulation characteristics of local areas in the aquaculture water bodies based on the time-space state characterization data, combining the dissolved oxygen demand threshold and the ammonia nitrogen tolerance threshold of the aquaculture objects in different growth stages to calculate the environmental regulation demand of each spatial area of the aquaculture water bodies; According to the environmental regulation demand, the spatial response priority and the time sequence coordination relationship of each functional unit in the circulating water system are determined, the spatial response priority is determined based on the spatial distribution position of the stress factors, the time sequence coordination relationship is determined based on the concentration gradient of the metabolic load accumulation characteristics, and the cooperative working parameters of each functional unit are generated; 2. The method of claim 1, wherein, The cooperative working parameters are iteratively optimized by a reinforcement learning model, the growth rate of the aquaculture objects and the water body stability are taken as the optimization objectives, the energy consumption constraint and the equipment load constraint are taken as the boundary conditions, and a circulating regulation instruction sequence is output. The method comprises the following steps: The time-space state characterization data comprises spatial dimension data and time dimension data; A multi-level spatial grid of the aquaculture water bodies is constructed based on the spatial dimension data, the environmental parameters in each spatial grid are subjected to normal interval determination, the stress factor types corresponding to the environmental parameters exceeding the normal interval are identified, and the spatial grid positions corresponding to the stress factor types are recorded as the spatial distribution positions of the stress factors, wherein the stress factor types include hypoxia stress factors with dissolved oxygen lower than the oxygen demand threshold and nitrogen metabolism stress factors with ammonia nitrogen concentration higher than the tolerance threshold; 3. The method of claim 2, wherein, The feeding characteristic change curve and the activity characteristic change curve of the aquaculture objects in the time dimension data are extracted, the generation rate and the accumulation rate of metabolic products in the aquaculture water bodies at adjacent time points are calculated, the transmission path and the accumulation area of the metabolic load between different spatial grids are determined based on the water flow connection relationship of the spatial grids, and the concentration gradient of the metabolic load in the accumulation area is calculated. The method comprises the following steps: According to the feeding amount difference of adjacent time points in the feeding characteristic change curve and the metabolic conversion coefficient of the aquaculture objects, the generation rate of ammonia nitrogen metabolic products in the aquaculture water bodies at adjacent time points is calculated, and according to the activity intensity difference of adjacent time points in the activity characteristic change curve and the excretion rule parameters of the aquaculture objects, the generation rate of organic waste in the aquaculture water bodies at adjacent time points is calculated. establishing an input-output balance relationship of metabolic products for each spatial grid, the input including a generation rate of organic waste, and the output including a metabolic product transfer amount flowing to an adjacent spatial grid, calculating a metabolic product accumulation rate in the spatial grid unit based on the input-output balance relationship; the water body flow communication relationship includes a water flow direction and a water flow exchange intensity between adjacent spatial grids; a directed transfer network of metabolic products between different spatial grids is constructed according to the water flow direction, and a metabolic product transfer flux along the directed transfer network from a source spatial grid to a downstream spatial grid is calculated according to the water flow exchange intensity and the metabolic product accumulation rate, the transfer flux representing the migration amount of metabolic products between adjacent spatial grids per unit time; identify the spatial grid with a transfer flux less than the outflow threshold in the directed transfer network as a metabolic load accumulation area.

4. The method of claim 1, wherein, determining the spatial response priority and timing coordination relationship of each functional unit in the circulating water system according to the environmental regulation demand amount includes: obtaining spatial scope information of each functional unit in the circulating water system, the spatial scope information including the installation position and the controllable space range of each functional unit, calculating the spatial intersection area of the controllable space range of each functional unit and the stress factor influence range, and taking the ratio of the spatial intersection area to the total area of the stress factor influence range as the spatial coverage matching degree; extracting the spatial distance between each functional unit and the center position of the stress factor influence range, taking the negated value of the spatial distance as the response time coefficient, and performing weighted fusion on the spatial coverage matching degree and the response time coefficient after normalization to obtain the spatial response priority value; calculating the spatial overlap area ratio of the spatial scope of each functional unit and the high concentration aggregation area as the location correlation degree, obtaining the processing efficiency parameter of each functional unit to the metabolic load material, and performing weighted calculation on the location correlation degree and the processing efficiency parameter to obtain the starting priority index of each functional unit, and arranging each functional unit in descending order according to the starting priority index to determine the starting order; calculating the spatial overlap area ratio of the spatial scope of each functional unit and the aggregation area with a concentration gradient greater than a preset concentration threshold as the location correlation degree, obtaining the processing efficiency parameter of each functional unit to the metabolic load material, and performing weighted calculation on the location correlation degree and the processing efficiency parameter to obtain the starting priority index of each functional unit, and arranging each functional unit in descending order according to the starting priority index to determine the timing coordination relationship.

5. The method of claim 1, wherein, iteratively optimizing the collaborative work parameters through a reinforcement learning model, the reinforcement learning model taking the growth rate of the cultured object and the water body stability as the optimization target, and taking the energy consumption constraint and the device load constraint as the boundary condition, and outputting a circulating control instruction sequence including: obtaining the total energy consumption value of the circulating water system and the device load rate of each functional unit at the current time, calculating the overrun amplitude of the total energy consumption value exceeding the preset energy consumption upper limit, and calculating the overrun amplitude of the device load rate of each functional unit exceeding the preset load upper limit; When the total energy consumption value or the equipment load rate of any functional unit exceeds the limit, a continuous penalty coefficient is calculated according to the exceeding limit amplitude, and the continuous penalty coefficient is multiplied by the comprehensive reward value to realize gradientized punishment; when the total energy consumption value and the equipment load rate of all functional units do not exceed the limit, the continuous penalty coefficient is set to a unit value to keep the comprehensive reward value unchanged, and the comprehensive reward value after the continuous penalty coefficient modulation is output as the boundary constraint reward value, wherein the comprehensive reward value is the weighted sum of the growth rate reward component and the water body stability reward component; The policy parameters of the reinforcement learning model are updated based on the boundary constraint reward value, and the policy parameters describe the mapping relationship from the cooperative working parameter execution state to the cooperative working parameter adjustment action; The cooperative working parameter adjustment action is generated according to the updated policy parameters, and the cooperative working parameter adjustment action includes the running power adjustment amount, the running time length adjustment amount and the space action range adjustment amount of each functional unit.

6. The method of claim 5, wherein, The continuous penalty coefficient is calculated according to the exceeding limit amplitude, and the continuous penalty coefficient is multiplied by the comprehensive reward value to realize gradientized punishment, including: The method further includes constructing a double-target reward function, the double-target reward function including a growth rate reward component and a water body stability reward component, the growth rate reward component being calculated according to the weight gain rate and the feed conversion rate of the cultured objects under the action of the current cooperative working parameters, and the water body stability reward component being calculated according to the fluctuation amplitudes of the dissolved oxygen concentration and the ammonia nitrogen concentration; A first ratio of the energy consumption exceeding limit amplitude to a preset energy consumption upper limit and a second ratio of the load exceeding limit amplitude to a preset load upper limit are calculated, and the maximum value of the first ratio and the second ratio is selected as a comprehensive exceeding limit ratio; It is judged whether the comprehensive exceeding limit ratio exceeds a preset segmentation threshold value, when the comprehensive exceeding limit ratio does not exceed the preset segmentation threshold value, a linear penalty amount is obtained by multiplying the comprehensive exceeding limit ratio by a first attenuation coefficient, and a continuous penalty coefficient is obtained by subtracting the linear penalty amount from a unit value; When the comprehensive exceeding limit ratio exceeds the preset segmentation threshold value, a difference between the comprehensive exceeding limit ratio and the preset segmentation threshold value is calculated as an over-threshold deviation, a threshold linear penalty amount is obtained by multiplying the preset segmentation threshold value by a first attenuation coefficient, a threshold penalty reference value is obtained by subtracting the threshold linear penalty amount from a unit value, an exponential attenuation index is obtained by multiplying the over-threshold deviation by a second attenuation coefficient, a negative value of the exponential attenuation index power of a natural constant is calculated as an exponential attenuation factor, and the continuous penalty coefficient is obtained by multiplying the threshold penalty reference value by the exponential attenuation factor.

7. A reservoir circulating water factory farming system control system combining the Internet of Things and AI, for implementing the method of any one of claims 1-6, characterized in that, including: The first unit is configured to collect environmental parameters and biological behavior parameters of the aquaculture water body through the distributed sensing unit, associate and map the environmental parameters and the biological behavior parameters according to the collection time and the spatial position, and establish the space-time state representation data of the aquaculture water body. A second unit is configured to identify a spatial distribution position of a stress factor deviating from a normal range in the aquaculture water body and a concentration gradient of a metabolic load accumulation feature of a local area based on the spatio-temporal state representation data, and calculate an environmental regulation demand of each spatial area of the aquaculture water body in combination with a dissolved oxygen demand threshold and an ammonia nitrogen tolerance threshold of the aquaculture object at different growth stages. A third unit is configured to determine a spatial response priority and a time sequence coordination relationship of each functional unit in the recirculating water system according to the environmental regulation demand, determine the spatial response priority based on the spatial distribution position of the stress factor, determine the time sequence coordination relationship based on the concentration gradient of the metabolic load accumulation feature, and generate a cooperative working parameter of each functional unit. A fourth unit is configured to iteratively optimize the cooperative working parameter by a reinforcement learning model, take a growth rate of the aquaculture object and a water body stability as an optimization target, take an energy consumption constraint and a device load constraint as a boundary condition, and output a cycle regulation instruction sequence.

8. An electronic device, comprising: It comprises: a processor; a memory for storing processor-executable instructions; wherein the processor is configured to invoke the instructions stored in the memory to execute the method of any one of claims 1 to 6.

9. A computer-readable storage medium having stored thereon computer program instructions, wherein, The computer program instructions, when executed by the processor, implement the method of any one of claims 1 to 6.

Images