Intelligent energy-saving water supply method and system for burrowing shellfish cultivation pool

Abstract

The present application relates to an intelligent energy-saving water supply method and system for infaunal shellfish cultivation ponds, belonging to the field of cross technology of intelligent aquaculture and artificial intelligence. The system comprises a sensing network, an edge computing control unit, a variable frequency execution unit and a cloud collaborative platform. The present application takes the suitable interval of chlorophyll a concentration as the control basis, and only starts water supply when the algae deviates from the suitable interval or the water quality deteriorates. When the algae concentration is suitable, invalid water exchange is avoided. The present application uses the ST-Attention-BiLSTM model to realize multi-step water quality prediction, which can predict risks such as water body anoxia, ammonia nitrogen exceeding standard, abnormal algae concentration and water temperature mutation ahead of time for several hours, reducing the stress loss of fry and the secondary pollution induced by sediment false fecal accumulation. By setting a cost function with bidirectional punishment constraint of chlorophyll a, both the dilution of bait in low-algae environment and the false fecal pollution and secondary ammonia nitrogen increase in high-algae environment can be prevented, realizing the dynamic balance between the filtering demand of fry and the safety of water quality.

Description

Technical Field

[0001] This invention relates to the field of intelligent aquaculture and artificial intelligence, specifically to an intelligent energy-saving water supply method and system for burial shellfish cultivation ponds. Background Technology

[0002] Buried shellfish (such as razor clams, hard clams, and small clams) spend their seedling stage buried in the bottom mud for a long time, and their growth and survival are highly dependent on environmental factors such as dissolved oxygen, ammonia nitrogen, algal feed concentration, temperature, and pH at the bottom mud-water interface. In the process of large-scale seedling cultivation, maintaining stable water quality at the interface and controlling the appropriate concentration of algal feed are key to ensuring the survival rate and growth rate of the seedlings.

[0003] Currently, seedling production generally employs constant flow or timed intermittent water supply modes, both of which are open-loop controls and cannot adaptively adjust the water supply based on dynamic changes in chlorophyll a concentration and shellfish filter-feeding activity. This approach has the following drawbacks: Firstly, when algae concentration is low, continuous water changes dilute natural food, rendering the water pump ineffective, wasting energy and hindering seedling feeding. Secondly, when algae concentration is too high, excessive filter-feeding by shellfish easily produces large amounts of pseudofeces deposited in the bottom sediment. The decomposition of this pseudofeces leads to increased ammonia nitrogen levels, directly threatening seedling survival. Furthermore, factors such as light and temperature affect algae proliferation, pH fluctuates with algal photosynthesis, and water temperature changes directly impact shellfish metabolic rates and dissolved oxygen saturation. These environmental factors are coupled with water supply control, making coordinated regulation difficult with traditional methods.

[0004] Existing closed-loop water supply control schemes using single-parameter feedback of dissolved oxygen or ammonia nitrogen can partially improve the problems of water hypoxia and excessive ammonia nitrogen, but they neglect the active regulation of chlorophyll a feed concentration and cannot solve problems such as ineffective water supply and secondary pollution caused by fecal decomposition. Model predictive control (MPC) can achieve multi-objective coordinated regulation in advance through multi-step prediction, but its application to the burrowing stage of shellfish seedling cultivation still faces many challenges: water quality prediction models need to simultaneously capture the complex nonlinear temporal dynamics between water physicochemical factors, algae concentration, and water temperature; optimized control needs to simultaneously consider energy consumption, dissolved oxygen guarantee, ammonia nitrogen control, chlorophyll a suitable concentration range constraint, and the linkage effect of temperature and pH, and there is a lack of adaptive regulation strategies adapted to shellfish seedling cultivation technology.

[0005] Therefore, the industry urgently needs an intelligent and energy-saving water supply method and system that is geared towards the seedling cultivation stage, uses chlorophyll a and water temperature as core predictive indicators, and comprehensively considers multiple environmental factors such as temperature and pH. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent and energy-saving water supply method and system for burial shellfish cultivation ponds. By using the ST-Attention-BiLSTM model to simultaneously predict the dynamic changes of dissolved oxygen, ammonia nitrogen, chlorophyll a, and water temperature, and combining temperature-linked water quality thresholds and pH safety constraints, an on-demand variable frequency water supply mechanism adapted to the filter feeding needs of burial shellfish seedlings is constructed. This minimizes water supply energy consumption while ensuring the safety of the bottom sediment-water interface and maintaining a suitable concentration of algae feed.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] In a first aspect of the present invention, a smart energy-saving water supply method for a burrowing shellfish cultivation pond is disclosed, the method comprising the following steps:

[0009] S1. Collect multidimensional time-series data through a sensor network deployed in each culture pool, and preprocess the multidimensional time-series data.

[0010] S2. Input the preprocessed multidimensional time-series data into the locally deployed water quality prediction model, and output the water quality prediction sequence of each culture pond in the future prediction time domain; the water quality prediction model is a spatiotemporal attention bidirectional long short-term memory network.

[0011] S3. Based on the real-time ambient temperature, retrieve the dynamic water quality safety threshold related to temperature from the local knowledge base;

[0012] S4. Combining the water quality prediction sequence of each cultivation pond with the dynamic water quality safety threshold, a multi-constraint rolling optimization cost function is constructed. With the minimum total energy consumption as the optimization objective, the optimal control sequence of the water pump and each branch pipe electric valve is solved, and the first term of the optimal control sequence is extracted as a real-time control command and sent to the frequency converter and the corresponding electric regulating valve.

[0013] S5. Adjust the total water supply and the branch water supply flow of each cultivation tank according to the control command, and record the current water supply flow data;

[0014] S6. Enter the next control cycle, re-collect water quality and environmental measurement data, and repeat the above steps to achieve closed-loop control.

[0015] Furthermore, the multidimensional time-series data in step S1 includes at least dissolved oxygen, water temperature, ammonia nitrogen, chlorophyll a, pH, light intensity, ambient temperature, and the actual water supply in the previous cycle.

[0016] Further, in step S2, the water quality prediction model includes a bidirectional LSTM encoding layer, a temporal attention branch, a spatial attention branch, and a fully connected output layer; the temporal attention branch is used to adaptively allocate weights to historical time-series information, and the spatial attention branch is used to adaptively allocate weights to multi-dimensional water quality features; the water quality prediction model takes a sliding temporal window matrix containing T consecutive historical time steps as input, and the feature vector of a single time step contains at least eight types of sensor features: dissolved oxygen, water temperature, ammonia nitrogen, chlorophyll a, pH, light intensity, ambient temperature, and historical water supply; the output is a water quality prediction sequence for the next M steps that contains at least dissolved oxygen, ammonia nitrogen, chlorophyll a, and water temperature.

[0017] Furthermore, the dynamic water quality safety thresholds mentioned in step S3 include at least the lower limit of dissolved oxygen, the upper limit of ammonia nitrogen, the upper and lower limits of suitable chlorophyll a concentration, and the pH safety range.

[0018] Further, step S4 specifically includes:

[0019] S41. Construct a multi-constraint rolling optimization cost function, simultaneously incorporating temperature-linked water quality constraints, water body pH safety range constraints, and bidirectional range constraints for suitable chlorophyll a concentration; the cost function takes minimizing total water supply energy consumption as the optimization objective, and includes pump power consumption terms, dissolved oxygen and ammonia nitrogen exceeding standard penalty terms, chlorophyll a deviation from suitable concentration range penalty terms, pump frequency change suppression terms, and branch pipe valve opening change suppression terms; the cost function and all constraints together constitute a nonlinear programming optimization problem.

[0020] S42. The nonlinear programming optimization problem is solved by numerical optimization algorithm to obtain the optimal sequence of global pump frequency and the optimal sequence of electric valve opening degree of each cultivation tank in the prediction time domain. According to the rolling optimization principle of model predictive control, only the first frequency value of the current time of the sequence is sent to the frequency converter, and the first valve opening degree command of each tank is sent to the corresponding electric regulating valve to drive the pump and valve to execute according to the optimization scheme.

[0021] S43. Wait for the next control cycle, collect a new round of measured water quality data to update the historical time series window, and repeatedly execute the water quality prediction and optimization solution process to form a closed-loop control.

[0022] Furthermore, the expression for the rolling optimization cost function is: ;

[0023] The expressions for each water quality penalty index within the rolling optimization cost function are as follows:

[0024] , ;

[0025] , ;

[0026] , ;

[0027] in, To predict the total cost within the time domain; To optimize the objective; Indicates the discrete time step number within the prediction time domain; This is the weighting coefficient corresponding to the water pump energy consumption item; This is a function of the water pump's power consumption. This refers to the total operating frequency of the water pump. The weighting coefficients for dissolved oxygen and ammonia nitrogen water quality violations; This is the weighting coefficient for the penalty term of chlorophyll a deviating from the suitable range; Number the cultivation ponds; This represents the penalty index for defaulting on dissolved oxygen levels in pool i. This is the lower limit of dissolved oxygen safety. The model output is the predicted dissolved oxygen value for the i-th pool. , This is a default indicator for dissolved oxygen; , This represents the penalty index for ammonia nitrogen violation in pool i; This represents the upper limit of safe toxicity for ammonia nitrogen. The predicted ammonia nitrogen value for the i-th pool; , This is a default indicator for ammonia nitrogen, meaning that penalties are only incurred if the predicted ammonia nitrogen level exceeds the upper limit. , This represents the penalty index for chlorophyll a deviating from the interval in the i-th pool; This represents the lower limit of the optimal concentration for chlorophyll a. This represents the upper limit of the optimal concentration of chlorophyll a. This represents the predicted value of chlorophyll a in the i-th pool; , This is a penalty for chlorophyll a deviating from the optimal range; This indicates the weight of the pump frequency fluctuation suppression term; This represents the optimal frequency of the water pump at step k. Indicates the pump frequency in the previous control step. This is a frequency change suppression term; This indicates the weight of the valve opening fluctuation suppression term; This indicates the opening degree of the electric regulating valve in the k-th step of the i-th pool. This indicates the valve opening degree of the i-th pool in the previous step. This is a valve action inhibition term.

[0028] Furthermore, the constraint conditions are as follows:

[0029] (1) Hardware physical constraints: , , For water pump frequency, This is the upper limit of the water pump frequency. This refers to the opening degree of a single-pool electric valve;

[0030] (2) Pipeline hydraulic constraints: The total water supply flow rate is equal to the sum of the flow rates of each pool, and satisfies the pump characteristic curve and the pipeline hydraulic relationship;

[0031] (3) pH range constraint: The predicted pH value range for each culture tank is as follows: ,in, Indicates the lower limit of the safe pH level in water; Indicates the upper limit of the safe pH level of the water body;

[0032] (4) Water quality relaxation constraints: The predicted water quality values ​​for each step in the rearing pond must meet the following requirements:

[0033] ;

[0034] ;

[0035] ;

[0036] ;

[0037] in, For the current control moment, To predict discrete time steps in the time domain; , , , ε represents the predicted values ​​for dissolved oxygen, ammonia nitrogen, chlorophyll a, and pH, respectively; ε is the allowable small slack variable.

[0038] Furthermore, the method also includes:

[0039] S6, Real-time Abnormal Operating Condition Monitoring

[0040] S61. Anomaly Detection and Control Mode Switching: Real-time detection of sensor data quality and water quality prediction model output confidence level; if sensor data quality is abnormal, or the water quality prediction model output confidence level is lower than a preset threshold, the system executes a smooth switching strategy, switching from MPC optimized control mode to either dynamic threshold-based expert rule control mode or PI proportional-integral feedback control mode; the expert rule control mode uses the dynamic water quality safety threshold corresponding to the ambient temperature as the control setpoint, and starts water supply regulation when the measured water quality approaches the safety threshold; the PI proportional-integral feedback control mode uses the deviation between the dynamic safety threshold and the real-time measured water quality value as input, and outputs the water pump supply frequency through proportional-integral calculation;

[0041] S62. Water quality index estimation during malfunctions: Using a soft sensor model trained with historical normal aquaculture data, water quality indexes measured by malfunctioning sensors are estimated based on the associated sensor data of water temperature, ammonia nitrogen, and chlorophyll a.

[0042] S63, Abnormal Alarm Push: When the control mode switch is triggered, the system sends alarm information to the outside through the communication network;

[0043] S64. Automatic recovery of control mode: When the sensor data returns to normal and the model prediction confidence rises above the preset threshold, the system automatically switches back to the MPC optimized control mode.

[0044] Furthermore, the method also includes:

[0045] S7, Cloud Collaboration and Models

[0046] S71. Federated Learning Global Model Update: The edge computing and control units of each breeding pond use local recent aquaculture data to calculate the parameter update gradient of the water quality prediction model. After desensitization and encryption, the gradient is uploaded to the cloud management platform. The cloud management platform uses the federated averaging algorithm to aggregate the gradients of each node, update the global water quality prediction model, and distribute the updated model parameters to each edge computing and control unit.

[0047] S72, Meta-learning New Batch Quick Adaptation

[0048] When starting a new batch of buried shellfish seedlings for cultivation, the cloud management platform generates a model initialization parameter package based on the model training trajectory of similar seedling varieties, similar cultivation batches, or similar environmental conditions in the past, using a meta-learning algorithm. After the parameter package is sent to the edge computing and control unit of the corresponding cultivation pond, the edge computing and control unit fine-tunes the online data to make the water quality prediction model reach a usable prediction accuracy.

[0049] S73, Knowledge Base Continuous Optimization

[0050] The cloud management platform aggregates abnormal water quality data and seedling growth data from each cultivation pond, continuously optimizes the dynamic water quality threshold parameter surface related to temperature, and regularly distributes the updated knowledge base to each edge computing and control unit. The abnormal water quality data includes events where chlorophyll a deviates from the suitable range and water quality threshold violation events. The seedling growth data includes seedling survival rate, growth rate, and growth uniformity. The cloud management platform performs statistical analysis on the abnormal water quality data and seedling growth data to optimize the parameter surfaces for the lower safe limit of dissolved oxygen, the upper safe limit of ammonia nitrogen, the upper and lower limits of suitable chlorophyll a concentration, and the safe pH range.

[0051] In a second aspect of the invention, an intelligent energy-saving water supply system for a burial shellfish cultivation pond is disclosed. The system includes a sensor network, an edge computing and control unit, a variable frequency water supply execution unit, and a cloud-based collaborative management platform.

[0052] The sensor network is deployed inside each cultivation tank to periodically collect multivariate time-series water quality and environmental data. The collected indicators include at least dissolved oxygen concentration, water temperature, ammonia nitrogen concentration, chlorophyll a concentration, pH, light intensity, ambient temperature, and historical water supply flow. The sensor network adopts a vertical gradient combined with representative planar points for deployment. In the vertical dimension of the cultivation tank, sensors are placed at least at a preset height above the sediment-water interface, and anti-siltation guide covers are provided on the outside of the sensors at this location. The data collected by the sensor network is transmitted to the edge computing and control unit in real time.

[0053] The edge computing and control unit is an industrial PLC or an embedded edge gateway, including a data preprocessing module, an ST-Attention-BiLSTM prediction model, a DA-MPC rolling optimization solution module, and an abnormal operating condition processing module, and a knowledge base for local storage of temperature-related dynamic water quality safety thresholds.

[0054] The variable frequency water supply execution unit is controlled by the edge computing and control unit and includes a frequency converter, a water supply pump, and multiple water supply branch pipes. Each water supply branch pipe corresponds to a cultivation tank, and each branch pipe is equipped with an electric regulating valve. The frequency converter adjusts the total water output of the water pump according to the frequency command issued by the edge computing and control unit, and each electric regulating valve adjusts the water inflow of the corresponding cultivation tank according to the opening command. The variable frequency water supply execution unit transmits the water pump operating frequency, valve opening, sub-tank flow, and total flow back to the edge computing and control unit for archiving and storage in real time.

[0055] The cloud-based collaborative management platform is deployed on a cloud server and communicates with multiple edge computing and control units; the platform has built-in federated learning operation unit, MAML meta-learning operation unit, and dynamic threshold iterative optimization unit.

[0056] Compared with existing technologies, this invention achieves the following beneficial effects by constructing a closed-loop control framework of "water quality prediction - rolling optimization - on-site feedback":

[0057] (1) This invention abandons the traditional uninterrupted constant flow and fixed period intermittent water supply mode and adopts on-demand predictive water supply. It takes the appropriate range of chlorophyll a concentration as the core control basis and starts water supply only when algae deviate from the appropriate range or water quality deteriorates. When the algae concentration is appropriate, it avoids ineffective water exchange, thereby achieving an energy saving rate of more than 30%.

[0058] (2) The present invention uses the ST-Attention-BiLSTM model to achieve multi-step water quality prediction for up to 6 hours. It can predict risks such as water hypoxia, ammonia nitrogen exceeding the standard, abnormal algae concentration and water temperature change several hours in advance, and take mild and gradual water supply regulation in advance to eliminate the lag defects of traditional post-event water replenishment, significantly reducing seedling stress loss and secondary pollution induced by the accumulation of false feces in bottom mud.

[0059] (3) This invention adopts a coordinated regulation of feed supply and water safety to achieve a two-way balance. By setting a cost function with a two-way penalty constraint of chlorophyll a, it can avoid the dilution of feed by water supply in a low-algae environment, and prevent false fecal pollution and secondary ammonia nitrogen increase in a high-algae environment, thus achieving a dynamic balance between the filter feeding needs of seedlings and water quality safety.

[0060] (4) This invention adopts a cloud-edge-device architecture. When the network is interrupted, the edge computing and control unit can run independently locally to complete the entire closed-loop control, which has strong robustness. The anomaly detection mechanism can ensure seamless degradation to safe mode when the sensor fails. The water quality prediction and DA-MPC optimization core algorithms used in this invention are independently packaged and can be deployed on edge hardware with different computing power levels through software. Only the threshold and model parameters need to be fine-tuned to adapt to other burrowing shellfish species and benthic aquaculture scenarios.

[0061] (5) The present invention equips each cultivation tank with an independent electric regulating valve under the same water pump group, and allocates the flow rate according to the real-time predicted demand. The water inflow is independently adjusted by the electric valve, and the water supply is allocated on demand under the premise of minimizing the total water supply, which effectively solves the problem of flow distribution imbalance caused by overflow in the front tank and lack of oxygen in the end tank. Attached Figure Description

[0062] Figure 1 This is a schematic diagram of the algorithm architecture and data flow of the intelligent energy-saving water supply system for the burial shellfish cultivation pond of the present invention;

[0063] Figure 2 This is a schematic diagram of the input and output of the ST-Attention-BiLSTM prediction model in this invention;

[0064] Figure 3This is a schematic diagram of the rolling optimization principle of Domain Adaptive Model Predictive Control (DA-MPC) in this invention;

[0065] Figure 4 This is the main flowchart of the intelligent water supply control algorithm executed at the edge end in this invention;

[0066] Figure 5 The simulation curves show the comparison of water quality and energy consumption between the intelligent energy-saving water supply method described in this invention and the traditional constant frequency water supply method during a typical 24-hour operating cycle.

[0067] Figure 6 This is a simulation curve comparing the cumulative water supply energy consumption of the intelligent energy-saving water supply method described in this invention with that of the traditional constant frequency constant flow water supply method during a 45-day cultivation period. Detailed Implementation

[0068] To provide a better understanding of the structural features and effects achieved by the present invention, a detailed description is provided below, accompanied by preferred embodiments and accompanying drawings:

[0069] This invention belongs to the field of intelligent aquaculture and artificial intelligence interdisciplinary technology. Specifically, it relates to a variable frequency water supply method and system for burial shellfish cultivation ponds that combines spatiotemporal deep learning prediction and domain adaptive optimal control. It is particularly suitable for the intensive planar flow cultivation stage of burial shellfish such as razor clams from attached juvenile shellfish to seedlings of about 2cm.

[0070] Example 1

[0071] like Figure 1 The system shown is an intelligent energy-saving water supply system for a burial shellfish seedling cultivation pond. The system adopts a three-layer collaborative architecture of cloud-edge-device, including four main parts: sensor network, edge computing and control unit, variable frequency water supply execution unit, and cloud collaborative management platform.

[0072] The sensor network, deployed within each rearing pond, periodically collects multivariate time-series water quality and environmental data reflecting the ecological state of the aquatic body. The collected indicators include at least dissolved oxygen concentration, water temperature, ammonia nitrogen concentration, chlorophyll a concentration, pH, light intensity, ambient temperature, and historical water supply flow. This sensor network employs a deployment strategy combining vertical gradient placement and selection of representative planar points. Vertically, sensors are positioned at a predetermined height above the sediment-water interface. Furthermore, the sensors positioned at the sediment-water interface are equipped with anti-siltation guide covers to prevent sediment from burying the sensor probes, ensuring continuous water flow through the sensing area and guaranteeing that the collected data accurately reflects the water quality status of the seedling habitat. The collected data is transmitted in real-time to the edge computing and control unit.

[0073] The edge computing and control unit, employing an industrial PLC or embedded edge gateway, serves as the system's local intelligent computing core. It incorporates a data preprocessing module, an ST-Attention-BiLSTM prediction model, a DA-MPC rolling optimization solution module, and an abnormal operating condition handling module. It also locally stores a knowledge base of temperature-related dynamic water quality safety thresholds. The edge computing and control unit possesses independent local operation capabilities, enabling it to complete the entire closed-loop control process without cloud intervention even in offline conditions. The specific functions of the edge computing and control unit include: (1) receiving data collected in real time from the sensor network and performing caching and preprocessing, including filtering, time alignment, missing value filling, outlier removal, and standardization; (2) running a lightweight ST-Attention-BiLSTM prediction model and outputting future multi-step prediction sequences of dissolved oxygen, ammonia nitrogen, chlorophyll a, and water temperature; (3) constructing and solving a nonlinear programming optimization problem with multiple constraints, i.e., performing chlorophyll adaptive MPC rolling optimization solution, outputting pump frequency conversion control commands and valve opening commands for each pool, and coordinating the collaborative water supply logic of multiple cultivation pools within the same water supply group; (4) monitoring sensor failures, insufficient model confidence, and other anomalies in real time, triggering degraded control and local and cloud-based synchronous alarms.

[0074] The variable frequency water supply execution unit, controlled by an edge computing and control unit, consists of a frequency converter, a water supply pump, and multiple water supply branch pipes. Each water supply branch pipe corresponds to a cultivation tank within the same water supply group, and each branch pipe is equipped with an electric regulating valve. The frequency converter receives frequency commands from the edge computing and control unit to adjust the total water output of the pump, while each electric regulating valve receives opening commands to adjust the water inflow to its corresponding cultivation tank, thereby achieving differentiated water supply through shared pumping and separate tank control. During equipment operation, the pump operating frequency, valve opening, separate tank flow rate, and total flow rate are transmitted back to the edge computing and control unit in real time for storage and archiving.

[0075] The cloud-based collaborative management platform is deployed on a cloud server and establishes communication connections with multiple edge computing and control units. The platform has built-in federated learning operation units, MAML meta-learning operation units, and dynamic threshold iterative optimization units. The functions of the platform are as follows: (1) Aggregate historical aquaculture data from each breeding pond and use the federated learning framework to regularly train and optimize the global ST-Attention-BiLSTM prediction model; (2) Use the meta-learning (MAML) algorithm to generate model initialization parameters that can be quickly adapted for the breeding ponds of new batches of seedlings; (3) Maintain and continuously update a knowledge base containing dynamic water quality safety thresholds, which stores the lower limit of dissolved oxygen, the upper limit of ammonia nitrogen, the suitable concentration range of chlorophyll a, and the safe pH range; (4) Have remote monitoring, abnormal alarm, and aquaculture data visualization functions.

[0076] Example 2

[0077] This embodiment discloses an intelligent and energy-saving water supply method for a burial shellfish seedling cultivation pond, such as... Figure 4 As shown, the method includes the following steps:

[0078] S1. Multidimensional time-series data of each cultivation pond are collected synchronously through a sensor network and preprocessed; the multidimensional time-series data includes at least dissolved oxygen, water temperature, ammonia nitrogen, chlorophyll a, pH, light intensity, ambient temperature and actual water supply in the previous cycle.

[0079] The water supply system is set with a fixed control cycle, preferably 15 minutes. Within each control cycle, the edge computing and control unit synchronously collects multi-dimensional time-series data from all cultivation tanks within the same water supply group via a sensor network. The collected data includes at least dissolved oxygen, water temperature, ammonia nitrogen, chlorophyll a, pH, light intensity, ambient temperature, and the actual water supply volume from the previous cycle. The specific functions of each collected indicator are as follows:

[0080] (1) Dissolved oxygen concentration (DO): Characterizes the oxygen supply capacity of water bodies and is crucial for the respiration of buried shellfish seedlings.

[0081] (2) Water temperature (Temp): affects the metabolic rate of seedlings and changes the dissolved oxygen saturation point of the water.

[0082] (3) Ammonia nitrogen concentration (NH3): The main excretion of seedlings during metabolism. When the concentration exceeds the standard, it will produce toxicity.

[0083] (4) Chlorophyll a concentration (Chla): Characterizes the amount of algal food in the water body and is the core control object for regulating the filter feeding state of seedlings in this invention.

[0084] (5) pH: reflects the acidity or alkalinity of the water body, and the value fluctuates significantly with the photosynthesis of algae.

[0085] (6) Light intensity: drives algae to carry out photosynthesis and achieve proliferation.

[0086] (7) Ambient temperature (EnvTemp): used to characterize the seasonal temperature changes in the aquaculture environment.

[0087] (8) Historical water supply flow: Records the actual water supply conditions of the pumps and valves in the previous cycle, which can be calculated from the operating frequency fed back by the frequency converter.

[0088] After data acquisition, the raw sensor data is preprocessed, including: filtering the raw sensor data; unifying the timestamps of all sensor data to achieve time-series alignment of multiple data streams; filling missing data using adjacent time-time interpolation or regression estimation based on associated sensors; removing anomalous jump data based on rate of change thresholds and data stationarity detection; and standardizing all numerical features to eliminate dimensional differences between various indicators and form a standardized time-series sliding window matrix.

[0089] S2. Input the preprocessed multidimensional time-series data into the locally pre-deployed water quality prediction model, and output the water quality prediction sequence for each cultivation pond in the future prediction time domain. The water quality prediction sequence includes dissolved oxygen, ammonia nitrogen, chlorophyll a, and water temperature. The water quality prediction model is a spatiotemporal attention bidirectional long short-term memory network (ST-Attention-BiLSTM) model.

[0090] The edge computing and control unit establishes a signal connection with the sensor network. In each control cycle, it receives multidimensional heterogeneous time-series data from all cultivation ponds in the same water supply group. The data types include at least dissolved oxygen, water temperature, ammonia nitrogen, chlorophyll a, pH, light intensity, ambient temperature, and the actual water supply in the previous cycle.

[0091] like Figure 2 As shown, after uniformly filtering, aligning, and standardizing the collected raw data, the standardized time-series data is input into a locally pre-deployed multi-step water quality prediction model. This model is a spatiotemporal attention bidirectional long short-term memory network (ST-Attention-BiLSTM). The model outputs water quality prediction sequences for each cultivation pond in the future prediction time domain. Each prediction sequence contains four types of time-series prediction values: dissolved oxygen, ammonia nitrogen, chlorophyll a, and water temperature, thus completing multi-step water quality prediction. The model's operating logic is as follows:

[0092] 1. Model Input

[0093] The model input is a historical time series matrix X within a sliding time window, containing T consecutive historical time steps, preferably T=288, corresponding to 24 hours of data collection; the multidimensional feature vector of a single time step contains at least 8 types of sensor features, including dissolved oxygen, water temperature, ammonia nitrogen, chlorophyll a, pH, light intensity, ambient temperature, and historical water supply.

[0094] 2. Internal working mechanism of the model

[0095] The water quality prediction model incorporates a spatiotemporal attention mechanism, which can adaptively allocate attention weights to historical time series moments and various sensor feature dimensions, fully extract nonlinear correlation information of water quality time series context, accurately capture the water quality lag change patterns caused by factors such as temperature and light, and complete multi-step water quality prediction.

[0096] (1) Bidirectional LSTM coding layer

[0097] A bidirectional LSTM coding layer is used to perform forward and backward bidirectional temporal coding on the historical sequence, simultaneously capturing the past and future bidirectional contextual dependencies at each time step, and outputting the temporal hidden state sequence.

[0098] (2) Spatiotemporal attention layer

[0099] The spatiotemporal attention layer comprises two branches: a temporal attention unit and a spatial attention unit. The temporal attention unit uses a trainable feedforward network to calculate the weights of each historical moment in relation to the future prediction target, enabling the model to adaptively focus on the historical time period most relevant to the prediction target. For example, when predicting the lowest dissolved oxygen level in the early morning, the model adaptively assigns higher attention weights to oxygen consumption data from the first half of the night; when predicting algae concentration, it adaptively strengthens the weights of historical light and water temperature data.

[0100] (3) Spatial attention unit

[0101] The spatial attention unit weights the importance of multi-dimensional sensor features such as dissolved oxygen, water temperature, ammonia nitrogen, chlorophyll a, and pH, and dynamically identifies the dominant factors of current water quality changes.

[0102] (4) Feature fusion and output mapping

[0103] The context vector obtained through spatiotemporal attention weighting is directly passed to the fully connected output layer to complete feature mapping, outputting a complete water quality prediction sequence for the next M steps. The water quality prediction model used in this invention uniformly predicts the single metabolic characteristics of burrowing shellfish larvae during their seedling stage, eliminating the need for additional growth stage identifiers. Preferably, M=24 is set, corresponding to the prediction time domain of the next 6 hours. The fully connected output layer simultaneously outputs four types of prediction results: dissolved oxygen prediction sequence. Ammonia nitrogen prediction sequence Chlorophyll a predicted sequence and water temperature prediction sequence .

[0104] 3. Model Output

[0105] The output of the multi-step water quality prediction model is multiple sets of water quality prediction values ​​for the next M steps, namely, the predicted sequence of dissolved oxygen concentration, the predicted sequence of ammonia nitrogen concentration, the predicted sequence of chlorophyll a concentration, and the predicted sequence of water temperature.

[0106] S3. Obtain the temperature-related dynamic water quality safety threshold.

[0107] The dynamic water quality safety thresholds include at least the lower limit of dissolved oxygen, the upper limit of ammonia nitrogen, the upper and lower limits of suitable chlorophyll a concentration, and the safe pH range. These dynamic water quality safety thresholds are generated by a cloud-based knowledge base based on historical aquaculture data and the experience of aquaculture experts, and the cloud can continuously iterate and optimize based on on-site aquaculture operation results, and periodically issue updates. This threshold set includes four types of parameters that change in conjunction with ambient temperature T: the lower limit of dissolved oxygen... Upper limit of ammonia nitrogen Upper limit of optimal chlorophyll a concentration Lower limit of optimal concentration of chlorophyll a and pH safe range .

[0108] The objective function constructed for the optimization problem includes four weighted cost components: the cumulative power consumption of the water pump at each time point in the prediction time domain, the penalty for exceeding the limits of dissolved oxygen and ammonia nitrogen, the penalty for chlorophyll a concentration exceeding the appropriate range, and the suppression of changes in adjacent control variables.

[0109] Specifically, the optimizer reads the real-time ambient temperature T of each culture tank and queries the dynamic water quality safety threshold at the corresponding temperature from the knowledge base stored locally or distributed in the cloud.

[0110] The built-in mapping rules of the knowledge base are:

[0111] Ambient temperature T → Dissolved oxygen safety lower limit upper limit of ammonia nitrogen safety Lower limit of optimal concentration of chlorophyll a and upper limit pH safe range , ]}.

[0112] in, This represents the minimum dissolved oxygen level at which the seedlings can safely survive under the current temperature. This represents the highest ammonia nitrogen level at which seedlings can safely survive under the current temperature. This represents the lower limit of the suitable concentration range for algae feed. This represents the upper limit of the suitable concentration range for algae feed; below this range... Supplying water at certain times will further dilute the feed, rendering the water pump ineffective; above this level... Excessive filter feeding by seedlings can easily lead to the formation of false feces, which, after sedimentation and decomposition, can cause an increase in ammonia nitrogen in the water; , [This represents the safe range for the acidity and alkalinity of the water body.]

[0113] The entire threshold knowledge base is generated in the cloud through offline aquaculture data analysis and fitting with the experience of aquaculture experts. It can continuously optimize parameters based on the actual aquaculture results on site and distribute them to the edge computing and control unit for local storage and retrieval.

[0114] S4. Combining the water quality prediction sequence of each cultivation pond with the dynamic water quality safety threshold, the prediction sequence is compared with the threshold to construct a rolling optimization cost function with multiple constraints. The optimal frequency sequence of the water pump and the opening sequence of each branch pipe electric valve are solved by rolling optimization to satisfy all the constraints of the cultivation pond and the total energy consumption. The current first frequency value and opening value in the sequence are extracted as control commands and sent to the frequency converter and the corresponding electric regulating valve respectively.

[0115] Specifically, the edge computing and control unit incorporates an adaptive model predictive control (MPC) module. Based on the predicted sequences for each pool, dynamic water quality safety thresholds, and pump power consumption models output in step S2, it constructs and solves a finite-time domain optimization problem with the objective of minimizing total water supply energy consumption and considering constraints across multiple pools within the group. The optimization problem incorporates the following constraints: temperature-linked water quality safety constraints, water pH safety range constraints, and bidirectional range constraints for suitable chlorophyll a concentration. Solving the optimization problem yields the total pump frequency sequence and the opening command sequence of the branch pipe electric valves for each group's cultivation pools within the predicted time domain. Following the MPC rolling optimization rule, the first value of the frequency sequence is selected as the total pump frequency conversion control command and sent to the frequency converter. The first value of the valve opening sequence for each pool is also sent to the corresponding cultivation pool electric regulating valve.

[0116] Further, step S4 specifically includes:

[0117] like Figure 3 As shown, after obtaining the multi-step water quality prediction sequence for each culture pond, the MPC optimizer performs the following steps:

[0118] S41. Constructing a finite-time constrained optimization problem

[0119] Within each control cycle, the MPC optimizer combines the water quality prediction sequences of each cultivation tank with the temperature-related dynamic water quality safety threshold to establish a finite-time-domain constrained nonlinear programming problem with the objective of minimizing the total prediction time-domain cost. The inputs to this optimization problem are the multi-step water quality prediction sequences output in step S2 and the temperature-related dynamic water quality safety threshold obtained in step S3. Both types of data serve as the penalty term and the criterion for determining water quality constraints.

[0120] The rolling optimization cost function aims to minimize the total water supply energy consumption and includes pump power consumption, penalties for excessive dissolved oxygen and ammonia nitrogen, penalties for chlorophyll a deviation from the suitable concentration range, pump frequency change suppression, and branch valve opening change suppression. During optimization, the predicted pH value of each cultivation tank is simultaneously constrained to be within a preset safe range.

[0121] Specifically, the rolling optimization cost function is designed as follows: ;

[0122] The water quality penalty indicators within the rolling optimization cost function are defined as follows:

[0123] , ;

[0124] , ;

[0125] , ;

[0126] in, To predict the total cost in the time domain, the optimization objective is to minimize the total cost in the predicted time domain. To minimize this, the cost function consists of four weighted components, which correspond to three mutually constraining optimization objectives: energy consumption, water quality safety, and control stability. To optimize the objective, calculate the total cost in the time domain for prediction. The minimum value; This represents the summation over each time step k in the prediction time domain; k represents the discrete time step number in the prediction time domain. This is the weighting coefficient corresponding to the water pump energy consumption item; This is a function of the water pump's power consumption. The total operating frequency of the water pump is approximately equal to the frequency. cubed ; The weighting coefficients for dissolved oxygen and ammonia nitrogen water quality violations; This is the weighting coefficient for the penalty term of chlorophyll a deviating from the suitable range; This is to sum the results of all culture ponds i within the same water supply group; i is the culture pond number. This represents the penalty index for defaulting on dissolved oxygen levels in pool i. This is the lower limit of dissolved oxygen safety. The model output is the predicted dissolved oxygen value for the i-th pool. , This is a default indicator for dissolved oxygen, meaning that penalties are only incurred if dissolved oxygen is predicted to fall below the lower limit. , This represents the penalty index for ammonia nitrogen violation in pool i; This represents the upper limit of safe toxicity for ammonia nitrogen. The predicted ammonia nitrogen value for the i-th pool; , This is a default indicator for ammonia nitrogen, meaning that penalties are only incurred if the predicted ammonia nitrogen level exceeds the upper limit. , This represents the penalty index for chlorophyll a deviating from the interval in the i-th pool; This represents the lower limit of the optimal concentration for chlorophyll a. This represents the upper limit of the optimal concentration of chlorophyll a. This represents the predicted value of chlorophyll a in the i-th pool; , This is a penalty term for chlorophyll a deviating from the optimal range. The penalty is triggered when the predicted concentration is below the lower limit or above the upper limit, and when the predicted concentration is within […]. [Internal Time] ; This indicates the weight of the pump frequency fluctuation suppression term; This represents the optimal frequency of the water pump at step k. Indicates the pump frequency in the previous control step. This is a frequency variation suppression term to prevent frequent start-stop of the water pump; This indicates the weight of the valve opening fluctuation suppression term; This indicates the opening degree of the electric regulating valve in the k-th step of the i-th pool. This indicates the valve opening degree of the i-th pool in the previous step. This is a valve actuation suppression measure to protect the actuator's lifespan; , These are the corresponding suppression weights.

[0127] Furthermore, the rolling optimization constraint also includes the requirement that the predicted pH value of each culture tank falls within a certain range. Inside. Indicates the lower limit of the safe pH level in water; This indicates the upper limit of the safe pH level in the water.

[0128] The power consumption function of the water pump is approximately equal to the frequency. cubed .

[0129] The constraints of the nonlinear programming problem include:

[0130] (1) Hardware physical constraints: The pump frequency must satisfy the following within the physical range: ;

[0131] (2) The opening degree of a single-pool electric valve within the physical range satisfies: ;

[0132] (3) Pipeline hydraulic constraints: The total water supply flow rate is equal to the sum of the flow rates of each pool, and satisfies the pump characteristic curve and the pipeline hydraulic relationship;

[0133] (4) Water quality relaxation constraints: Introducing small relaxation variables By relaxing the water quality boundary, a feasible solution is ensured for optimization, and the predicted water quality values ​​for each pool at each step satisfy the following:

[0134] ;

[0135] ;

[0136] ;

[0137] ;

[0138] Here, ε represents an allowable small slack variable to ensure the feasibility of the optimization problem; the predicted pH value for each culture tank must fall within [a certain range]. Inside, This indicates the lower limit of the safe pH level in water. This indicates the upper limit of the safe pH level of the water body. Preferably, the value of ε is 3% of the median value of the normal range of each indicator.

[0139] S42. Solve the optimization problem and issue control commands.

[0140] The optimizer employs lightweight numerical optimization algorithms suitable for embedded systems, such as Sequential Quadratic Programming (SQP) or the interior-point method, to solve the nonlinear programming optimization problem and obtain the globally optimal frequency sequence of the water pump in the prediction time domain. and the optimal opening sequence of electric valves in each culture tank Based on the rolling optimization principle of model predictive control, only the first frequency value of the sequence at the current time step is selected. The command is sent to the frequency converter, which then sends the first valve opening command for each pool to the corresponding electric regulating valve, driving the water pump and valves to execute according to the optimized scheme. In this embodiment, not all optimization parameters in the prediction time domain are executed; only the first optimal value at the current time step is extracted and sent. This is a key feature that distinguishes the model predictive control region from static optimization. The frequency converter receives the frequency command to adjust the total water supply, and the electric valves of each branch pipe receive the opening command, realizing differentiated water supply for each pool through a shared pump.

[0141] S43, Rolling Feedback Correction

[0142] Waiting for the next control cycle, the edge computing and control unit uses the newly collected actual water quality data to update the historical time series window, and cyclically executes the water quality prediction and optimization solution process to form a closed-loop control.

[0143] After a single water supply cycle is completed, the system waits for the next 15-minute control cycle. The edge computing and control unit collects a new round of measured water quality data, slides to update the timing input window, and re-executes S1 data preprocessing → S2 water quality prediction → S3 dynamic threshold retrieval → S41 optimization problem construction → S42 solution distribution, forming a closed-loop structure of prediction → optimization → execution → feedback → re-prediction. This enables the system to continuously use the latest measured information to eliminate the accumulated error of the ST-Attention-BiLSTM prediction model, while resisting external random disturbances such as water temperature, light, and pipeline water pressure. It has strong robustness to model uncertainty, environmental changes, and hydraulic condition fluctuations.

[0144] S5. The variable frequency water supply execution unit receives control commands from the frequency converter and electric regulating valve, adjusts the total water supply and the branch water supply flow of each cultivation tank, and records and updates the actual total water supply flow and branch flow of each tank in the current period.

[0145] S6. Wait until the next preset control cycle, return to step S1, collect a new round of environmental and water quality measurement data to complete rolling feedback correction, and repeat the entire water supply control process from S1 to S5.

[0146] Furthermore, the method also includes: during the execution of steps S1 to S6, the system monitors abnormal operating conditions in real time throughout the process, and the abnormal operating conditions are handled as follows:

[0147] The system has a built-in multi-layered safety protection mechanism. If the quality of sensor data deteriorates (such as disconnection, continuous stagnation, abnormal rate of change, etc.), or if the output confidence of the ST-Attention-BiLSTM water quality prediction model is lower than the preset threshold, the system will execute a smooth switching strategy: automatically and seamlessly switch from the MPC-based optimization control mode to either the dynamic threshold-based expert rule control mode or the PI (proportional integral) feedback control mode.

[0148] (1) Expert rule control mode: The dynamic water quality safety threshold corresponding to the current ambient temperature is used as the control set value. When the measured water quality is close to the safety threshold, the water supply is started.

[0149] (2) PI proportional integral feedback control mode: The deviation between the dynamic safety threshold and the real-time measured water quality is used as input, and the output water pump supply frequency is calculated according to the proportional integral law to stabilize the water quality.

[0150] Simultaneously, the system can employ a soft sensor model trained based on historical normal aquaculture data, utilizing data from other relevant sensors such as water temperature, ammonia nitrogen, and chlorophyll a to estimate the indicators measured by faulty sensors, maintaining basic operational continuity. Upon triggering a mode switch, the system sends alarm information to the cloud management platform and aquaculture administrators via the communication network. Once the anomaly is resolved and data returns to normal—that is, sensor data returns to normal and model prediction confidence rises above the threshold—the system automatically switches back to MPC optimized control mode for continuous operation. Furthermore, when sensor data is abnormal or the prediction model confidence falls below the threshold, the system automatically switches from model predictive control mode to proportional-integral feedback control mode based on a dynamic safety threshold to maintain basic water supply.

[0151] Furthermore, the method also includes a cloud-based collaboration and model evolution step, which is a background offline periodic task that is automatically executed at fixed times each day and does not consume computing power for single-cycle real-time adjustment. Specifically, this step includes:

[0152] (1) Federated learning global model update

[0153] Each edge computing and control unit in each rearing pool calculates the parameter update gradients of the ST-Attention-BiLSTM model locally using recent aquaculture data. After anonymization and encryption, the gradients are uploaded to the cloud management platform. The cloud platform uses a federated averaging (FedAvg) algorithm to aggregate the gradients of each node, update the global prediction model, and distribute the updated model parameters to each edge computing and control unit. This process is executed periodically, allowing the global model to continuously learn from data from multiple pools, gradually improving prediction accuracy and generalization ability, while protecting the data privacy of each node.

[0154] (2) Meta-learning new batch quick adaptation

[0155] When initiating a new batch of burrowing shellfish seedlings for aquaculture, the cloud management platform generates a personalized model initialization parameter package based on the model training trajectories of similar varieties, batches, or environmental conditions from the past, using meta-learning algorithms (such as MAML, Model-Agnostic Meta-Learning). After this parameter package is distributed to the corresponding edge computing and control unit, the edge computing and control unit only needs a small amount of online data for fine-tuning to bring the ST-Attention-BiLSTM model to a usable prediction accuracy, significantly shortening the adaptation period required for training from scratch.

[0156] (3) Continuous optimization of the knowledge base

[0157] The cloud-based management platform aggregates water quality anomalies and seedling growth data from each rearing pond, continuously optimizes the temperature-related dynamic water quality threshold parameter surface, and regularly distributes the updated knowledge base to each edge computing and control unit. Specifically, the cloud-based management platform gathers chlorophyll deviation events, water quality threshold violation events, and seedling growth performance data (such as survival rate, growth rate, and uniformity) from each rearing pond, and continuously optimizes the data through statistical analysis. , , (T), (T) and the parameter surface of dynamic thresholds such as pH safety range. The optimized knowledge base is periodically distributed to each edge computing and control unit, thereby continuously improving the system's chlorophyll adaptive decision-making capability.

[0158] The water supply system is pre-set with a fixed 15-minute control cycle, and each cycle executes the closed-loop control process from steps S1 to S6. Real-time anomaly monitoring is activated synchronously throughout the entire cycle, and the system executes a background cloud-based collaborative model evolution task daily.

[0159] To clearly explain the underlying computational logic corresponding to steps S2 and S4 in the above water supply method, the two core algorithm modules and the periodic cloud-based collaborative evolution mechanism are explained below:

[0160] (1) Water quality multi-step prediction module

[0161] This module is built on a spatiotemporal attention bidirectional long short-term memory network (ST-Attention-BiLSTM), with a focus on its input-output interface and prediction performance.

[0162] ① Module Input

[0163] Historical time series matrix within a single sliding time series window ; where the vector corresponding to each discrete time step It is a multi-dimensional sensing feature vector, including: dissolved oxygen at the bottom layer, water temperature, ammonia nitrogen concentration, chlorophyll a concentration, pH, light intensity, ambient temperature, and water supply flow rate of the previous cycle (the water supply flow rate value can be calculated from the operating frequency of the frequency converter).

[0164] ②Prediction Mechanism

[0165] The model utilizes a BiLSTM structure to encode the bidirectional dependencies of historical time-series data. Then, through a spatiotemporal attention mechanism, it adaptively allocates weights and dynamically focuses on the most critical historical moments and sensor dimensions for predicting future water quality. For example, to predict dissolved oxygen at 5 AM the following day, the model adaptively assigns higher attention weights to the previous night's data; when predicting algae concentration, it adaptively strengthens the weights of historical light and temperature periods. By explicitly predicting water temperature, the model can capture the lagged effects of temperature changes on dissolved oxygen saturation, algae growth, and seedling metabolism in advance, making optimal control more forward-looking.

[0166] ③ Module Output

[0167] The output prediction time domain length is Four sets of water quality time-series prediction sequences for the step:

[0168] Dissolved oxygen prediction series ;

[0169] Ammonia nitrogen prediction series ;

[0170] Chlorophyll a Predicted Value Sequence ;

[0171] Water temperature prediction series .

[0172] The prediction step length and total prediction time domain length can be flexibly configured according to the system's control cycle and water exchange rate. For example, during peak aquaculture water use, a prediction step can be taken every 15 minutes, predicting the next 6 hours in advance, for a total of 24 prediction steps.

[0173] (2) Chlorophyll Adaptive Rolling Optimization Solution Module

[0174] This module is a standard model predictive control (MPC) framework. In each control cycle, it transforms the prediction results (i.e., water quality prediction time series) into control commands for pumps and valves that can be issued and executed. It also introduces an adaptive mechanism based on the suitable concentration range of chlorophyll a.

[0175] ① Module Input

[0176] 1. Multi-step prediction sequences of dissolved oxygen, ammonia nitrogen, chlorophyll a, and water temperature output by the prediction model;

[0177] 2. Current ambient temperature T;

[0178] 3. Dynamic safety thresholds indexed by T from the knowledge base: Dissolved oxygen safety lower limit upper limit of ammonia nitrogen safety Lower limit of optimal concentration of chlorophyll a and upper limit and pH safe range This threshold table is generated in the cloud through offline analysis and expert experience and is updated regularly.

[0179] ②Optimize problem construction

[0180] The optimization objective is to find a set of pump frequency sequences in the prediction time domain. and the valve opening sequence of each culture tank branch pipe This makes the total cost in the prediction time domain... Minimum.

[0181] The cost function consists of four weighted components:

[0182] (a) Pump energy consumption cost: Energy consumption is proportional to the cube of the pump operating frequency. The summation is performed over all time steps in the prediction time domain.

[0183] (b) Penalty for breach of dissolved oxygen and ammonia nitrogen: For any prediction step in which dissolved oxygen (DO) falls below the lower limit or ammonia nitrogen falls below the lower limit. If the toxicity level exceeds the upper limit, a penalty will be imposed.

[0184] (c) Penalty for chlorophyll a deviating from the optimal range: when the predicted chlorophyll a concentration is lower than the optimal range. or higher The design incorporates a penalty clause to address two types of risks: insufficient food due to low algae concentration, and excessive ammonia nitrogen levels caused by high algae concentration. This penalty clause reduces ineffective water exchanges when algae concentration is appropriate, and only activates or increases water supply when algae levels deviate from the optimal range.

[0185] (d) Control change suppression cost: The penalty term is proportional to the square of the difference in pump frequency and the square of the difference in valve opening between adjacent control steps, which limits the drastic adjustment of the actuator and extends the service life of the frequency converter and electric regulating valve.

[0186] The mathematical expression for the rolling optimization cost function is:

[0187] ; in, For water pump frequency, For the opening degree of the branch pipe valve of the i-th cultivation pool,

[0188] ;

[0189] ;

[0190] ;

[0191] The water pump power consumption meets ; These are the weighting coefficients for energy consumption, conventional water quality violations, and chlorophyll deviation, respectively. These are the weights corresponding to the frequency and valve opening fluctuation suppression terms, respectively.

[0192] ③ Optimization constraints

[0193] When solving optimization problems, all constraints must be satisfied simultaneously:

[0194] 1. Hydraulic flow constraint: The total system water supply flow rate equals the sum of the flow rates of all branch lines in the cultivation tanks. The total flow rate is related to the pump frequency. and the opening degree of each valve The hydraulic relationship between them must be satisfied by the pipeline network;

[0195] 2. Hardware physical boundary constraints: , ;

[0196] 3. Soft constraints on water quality with slack variables: Introducing small slack variables Relaxing the water quality boundary ensures that a feasible solution exists for the optimization problem.

[0197] The predicted water quality values ​​at each step meet the following requirements:

[0198] DO predictions at each step: ;

[0199] Predicted value: ;

[0200] Chlorophyll a predicted value: ;

[0201] pH prediction value: .

[0202] ④ Solving and Execution:

[0203] A high-efficiency nonlinear programming solver adapted to embedded edge computing and control units is used to solve the above-mentioned constrained optimization problem, outputting the complete predicted optimal frequency sequence in the time domain and the valve opening sequence for each pool. Based on the MPC rolling principle, only the first frequency value is used. and the first valve opening value The analog output module sends signals to the water pump inverter and the corresponding branch pipe's electric regulating valve. Waiting for the next control cycle, a new round of measured water quality data is collected to update the timing window. Based on the new measured data, the entire prediction-optimization process is repeated to form feedback correction.

[0204] (3) Cloud-based collaboration and model evolution mechanism

[0205] This mechanism is a periodic, long-term background optimization process that does not consume real-time adjustment computing power in a single cycle. It is completed collaboratively by the edge computing and control unit and the cloud management platform, and includes three functional parts:

[0206] 1. Federated Learning Global Prediction Model Update

[0207] Each rearing pond's corresponding edge computing and control unit calculates the gradient of the ST-Attention-BiLSTM model daily based on local aquaculture time-series data. The gradient data is then anonymized, encrypted, and uploaded to the cloud management platform. The cloud uses a federated learning framework to aggregate data from all aquaculture nodes and update the global prediction model.

[0208] 2. Rapid adaptation of new breeding batches to basic learning methods

[0209] When a new batch of juvenile oysters begins to be cultured, an initial model parameter package generated based on meta-learning is distributed, enabling the prediction accuracy to reach a usable level in a very short time.

[0210] 3. Continuous iterative optimization of the dynamic threshold knowledge base

[0211] The cloud-based knowledge base continuously collects data on chlorophyll deviation events, water temperature changes, and seedling growth performance from various ponds, and dynamically optimizes [the data]. , , (T), (T) threshold surface; the optimized knowledge base is regularly distributed and synchronized to the local storage of each edge computing and control unit to continuously improve the accuracy of the system's chlorophyll adaptive water supply decision.

[0212] To verify the effectiveness of the technical solution of this invention, a system simulation model was built using MATLAB / Simulink based on typical operating parameters of razor clam cultivation at a coastal seedling base, and compared with a constant-frequency constant-flow water supply mode. The simulation conditions were set as follows: the cultivated objects were razor clam seedlings from attached juveniles to 2 cm in size; the single water supply group included 6 cultivation ponds, each with an area of ​​20 m² and a water depth of 30 cm; the system control cycle was 15 min, the model prediction time domain was 6 h (M=24 steps), and the historical time window was 24 h (T=288 steps); the ambient temperature range was 22~26℃; the water quality constraints were: the suitable concentration range of chlorophyll a was [8, 25] μg / L, the lower limit of dissolved oxygen was 4.0 mg / L, the upper limit of ammonia nitrogen was 0.5 mg / L, and the safe pH range was [7.6, 8.6]. The control group used a constant-frequency 50 Hz continuous water supply with a daily water exchange rate of 200%.

[0213] Figure 5 The curves show a comparison of water quality and energy consumption between two water supply modes during a typical 24-hour operating cycle. Figure 5 (a) It can be seen that under constant frequency and constant flow water supply, the chlorophyll a concentration fluctuates significantly within the range of 334 μg / L, frequently deviating from the suitable concentration range [8, 25] μg / L. This invention, by introducing a two-way penalty constraint mechanism for chlorophyll a, can stably maintain the algal concentration within the range of 9~23 μg / L, effectively avoiding the risk of false fecal accumulation caused by excessive water exchange diluting feed during low algae periods and by algal excess during high algae periods. Figure 5 (b) It is known that under the constant-frequency constant-flow water supply mode, dissolved oxygen drops to 3.4 mg / L in the early morning, below the safety limit of 4.0 mg / L, posing a risk of hypoxia stress to seedlings. This invention, utilizing the precise predictive capabilities of the ST-Attention-BiLSTM model, can gradually adjust the water supply before the early morning dissolved oxygen trough, maintaining the lowest dissolved oxygen level throughout the day above 4.2 mg / L, thus eliminating the potential for hypoxia in the early morning. Figure 5 (c) It can be seen that the traditional fixed-frequency water supply method operates the water pump at a constant 50Hz, outputting full power at all times, resulting in high energy consumption; while the present invention can dynamically adjust the water pump operating frequency according to the real-time water quality demand, maintaining the water pump frequency at 24~42 Hz throughout the day, operating at low load and energy-saving at night, and moderately replenishing water during the afternoon water quality fluctuation period, avoiding ineffective energy consumption and achieving good energy-saving effect.

[0214] Figure 6The simulation results show a comparison of the cumulative energy consumption of two water supply modes during a complete 45-day cultivation cycle. The simulation results indicate that the traditional constant-frequency water supply mode has an average daily energy consumption of 18.7 kWh and a cumulative energy consumption of 842.4 kWh over 45 days. Under the control strategy of this invention, the system's average daily energy consumption is only 11.9 kWh, and the cumulative energy consumption over 45 days is 535.8 kWh, resulting in a cumulative energy saving of 306.6 kWh compared to the traditional mode, with a comprehensive energy saving rate of 36.4%. The experimental results fully verify that the on-demand predictive water supply strategy based on predictive optimization of this invention can significantly reduce cultivation energy consumption while ensuring water quality safety, demonstrating outstanding energy-saving advantages.

[0215] The main performance indicators of the two water supply modes are compared in Table 1.

[0216] Table 1 Comparison of main performance indicators of the two water supply modes

[0217] Average daily water pump energy consumption (kWh) 18.7 11.9 Down 36.4% 45-day cumulative energy consumption (kWh) 842.4 535.8 Save 306.6 Chlorophyll a fluctuation range (μg / L) 3~34 9~23 Volatility narrowed by 62%. Minimum dissolved oxygen level (mg / L) 3.4 4.2 Improved by 0.8 Average number of times the water pump starts and stops per day 24 7 Decrease of 70.8%

[0218] As shown in Table 1, compared with the constant frequency and constant flow water supply mode, the intelligent water supply mode adopted in this invention can reduce the average daily energy consumption by 36.4%, reduce the average daily start-up and shutdown frequency of water pumps by 70.8%, narrow the fluctuation range of chlorophyll a concentration by 62%, and increase the minimum dissolved oxygen in the water by 0.8 mg / L. Table 1 demonstrates that this invention can reduce water supply energy consumption, extend equipment life, optimize the aquatic environment, reduce stress on razor clam seedlings, and balance energy conservation goals with aquaculture water quality safety.

[0219] In summary, this invention, while maintaining water quality safety, effectively avoids ineffective water exchange through the synergistic effect of chlorophyll a bidirectional penalty constraint and MPC rolling optimization, achieving significant energy-saving effects, while eliminating the risk of hypoxia in the early morning and reducing seedling stress losses.

[0220] Example 3

[0221] An electronic device includes: at least one processor; and a memory storing instructions that, when executed by the at least one processor, cause the at least one processor to perform the above-described intelligent energy-saving water supply method for burial shellfish cultivation ponds.

[0222] In this embodiment, the electronic device may include, but is not limited to: personal computer, server computer, workstation, desktop computer, laptop computer, notebook computer, mobile computing device, smartphone, tablet computer, cellular phone, personal digital assistant (PDA), handheld system, messaging device, wearable computing device, consumer electronic device, etc.

[0223] Example 4

[0224] A machine-readable storage medium storing executable instructions that, when executed, cause the machine to perform the above-described intelligent energy-saving water supply method for burial shellfish cultivation ponds.

[0225] Specifically, a system or system equipped with a readable storage medium storing software program code that implements the functions of any of the embodiments described above, and enabling the computer or processor of the system or system to read and execute instructions stored in the readable storage medium. In this case, the program code read from the readable medium itself can implement the functions of any of the embodiments described above, and therefore the machine-readable code and the readable storage medium storing the machine-readable code constitute a part of this specification. Embodiments of readable storage media include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RAM, DVD-RW, DVD-RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, the program code can be downloaded from a server computer or the cloud via a communication network. Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The present invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, produce a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A system specifying functions within one or more boxes. These computer program instructions may also be stored in a computer-readable storage medium capable of directing a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including an instruction system implemented in a process. Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0226] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0227] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A smart and energy-saving water supply method for a burial shellfish cultivation pond, characterized in that, The method includes the following steps: S1. Collect multidimensional time-series data through a sensor network deployed in each culture pool, and preprocess the multidimensional time-series data. S2. Input the preprocessed multidimensional time-series data into the locally deployed water quality prediction model, and output the water quality prediction sequence of each culture pond in the future prediction time domain; the water quality prediction model is a spatiotemporal attention bidirectional long short-term memory network. S3. Based on the real-time ambient temperature, retrieve the dynamic water quality safety threshold related to temperature from the local knowledge base; S4. Combining the water quality prediction sequence of each cultivation pond with the dynamic water quality safety threshold, a multi-constraint rolling optimization cost function is constructed. With the minimum total energy consumption as the optimization objective, the optimal control sequence of the water pump and each branch pipe electric valve is solved, and the first term of the sequence is extracted as a real-time control command and sent to the frequency converter and the corresponding electric regulating valve. S5. Adjust the total water supply and the branch water supply flow of each cultivation tank according to the control command, and record the current water supply flow data; S6. Enter the next control cycle, re-collect water quality and environmental measurement data, and repeat the above steps to achieve closed-loop control.

2. The intelligent energy-saving water supply method for burial shellfish cultivation ponds according to claim 1, characterized in that, The multidimensional time-series data in step S1 includes at least dissolved oxygen, water temperature, ammonia nitrogen, chlorophyll a, pH, light intensity, ambient temperature, and the actual water supply in the previous cycle.

3. The intelligent energy-saving water supply method for burial shellfish cultivation ponds according to claim 1, characterized in that, In step S2, the water quality prediction model includes a bidirectional LSTM encoding layer, a temporal attention branch, a spatial attention branch, and a fully connected output layer. The temporal attention branch is used to adaptively allocate weights to historical time-series information, and the spatial attention branch is used to adaptively allocate weights to multi-dimensional water quality features. The water quality prediction model takes a sliding temporal window matrix containing T consecutive historical time steps as input, and the feature vector of a single time step contains at least eight types of sensor features: dissolved oxygen, water temperature, ammonia nitrogen, chlorophyll a, pH, light intensity, ambient temperature, and historical water supply. The output is a water quality prediction sequence for the next M steps that contains at least dissolved oxygen, ammonia nitrogen, chlorophyll a, and water temperature.

4. The intelligent energy-saving water supply method for burial shellfish cultivation ponds according to claim 1, characterized in that, The dynamic water quality safety thresholds mentioned in step S3 include at least the lower limit of dissolved oxygen, the upper limit of ammonia nitrogen, the upper and lower limits of suitable chlorophyll a concentration, and the pH safety range.

5. The intelligent energy-saving water supply method for burial shellfish cultivation ponds according to claim 1, characterized in that, Step S4 specifically includes: S41. Construct a multi-constraint rolling optimization cost function, simultaneously incorporating temperature-linked water quality constraints, water body pH safety range constraints, and bidirectional range constraints for suitable chlorophyll a concentration; the cost function takes minimizing total water supply energy consumption as the optimization objective, and includes pump power consumption terms, dissolved oxygen and ammonia nitrogen exceeding standard penalty terms, chlorophyll a deviation from suitable concentration range penalty terms, pump frequency change suppression terms, and branch pipe valve opening change suppression terms; the cost function and all constraints together constitute a nonlinear programming optimization problem. S42. The nonlinear programming optimization problem is solved by numerical optimization algorithm to obtain the optimal sequence of global pump frequency and the optimal sequence of electric valve opening degree of each cultivation tank in the prediction time domain. According to the rolling optimization principle of model predictive control, only the first frequency value of the current time of the sequence is sent to the frequency converter, and the first valve opening degree command of each tank is sent to the corresponding electric regulating valve to drive the pump and valve to execute according to the optimization scheme. S43. Wait for the next control cycle, collect a new round of measured water quality data to update the historical time series window, and repeatedly execute the water quality prediction and optimization solution process to form a closed-loop control.

6. The intelligent energy-saving water supply method for burial shellfish cultivation ponds according to claim 5, characterized in that, The expression for the rolling optimization cost function is: ； The expressions for each water quality penalty index within the rolling optimization cost function are as follows: ； ； ； in, To predict the total cost within the time domain; To optimize the objective; Indicates the discrete time step number within the prediction time domain; This is the weighting coefficient corresponding to the water pump energy consumption item; This is a function of the water pump's power consumption. This refers to the total operating frequency of the water pump. The weighting coefficients for dissolved oxygen and ammonia nitrogen water quality violations; This is the weighting coefficient for the penalty term of chlorophyll a deviating from the suitable range; Number the cultivation ponds; This represents the penalty index for defaulting on dissolved oxygen levels in pool i. This is the lower limit of dissolved oxygen safety. The model output is the predicted dissolved oxygen value for the i-th pool. This is a default indicator for dissolved oxygen; This represents the penalty index for ammonia nitrogen violation in pool i; This represents the upper limit of safe toxicity for ammonia nitrogen. The predicted ammonia nitrogen value for the i-th pool; This is a default indicator for ammonia nitrogen, meaning that penalties are only incurred if the predicted ammonia nitrogen level exceeds the upper limit. This represents the penalty index for chlorophyll a deviating from the interval in the i-th pool; This represents the lower limit of the optimal concentration for chlorophyll a. This represents the upper limit of the optimal concentration of chlorophyll a. This represents the predicted value of chlorophyll a in the i-th pool; Penalty for chlorophyll a deviating from the optimal range This indicates the weight of the pump frequency fluctuation suppression term; This represents the optimal frequency of the water pump at step k. Indicates the pump frequency in the previous control step. This is a frequency change suppression term; This indicates the weight of the valve opening fluctuation suppression term; This indicates the opening degree of the electric regulating valve in the k-th step of the i-th pool. This indicates the valve opening degree of the i-th pool in the previous step. This is a valve action inhibition term.

7. The intelligent energy-saving water supply method for burial shellfish cultivation ponds according to claim 5, characterized in that, The constraints are as follows: (1) Hardware physical constraints: , , For water pump frequency, This is the upper limit of the water pump frequency. This refers to the opening degree of a single-pool electric valve; (2) Pipeline hydraulic constraints: The total water supply flow rate is equal to the sum of the flow rates of each pool, and satisfies the pump characteristic curve and the pipeline hydraulic relationship; (3) pH range constraint: The predicted pH value range for each culture tank is as follows: ,in, Indicates the lower limit of the safe pH level in water; Indicates the upper limit of the safe pH level of the water body; (4) Water quality relaxation constraints: The predicted water quality values ​​for each step in the rearing pond must meet the following requirements: ； ； ； ； in, For the current control moment, To predict discrete time steps in the time domain; , , , ε represents the predicted values ​​for dissolved oxygen, ammonia nitrogen, chlorophyll a, and pH, respectively; ε is the allowable small slack variable.

8. The intelligent energy-saving water supply method for burial shellfish cultivation ponds according to claim 1, characterized in that, The method also includes: S7, Real-time Abnormal Operating Condition Monitoring S71. Anomaly Detection and Control Mode Switching: Real-time detection of sensor data quality and water quality prediction model output confidence level; if sensor data quality is abnormal, or the water quality prediction model output confidence level is lower than a preset threshold, the system executes a smooth switching strategy, switching from MPC optimized control mode to either dynamic threshold-based expert rule control mode or PI proportional-integral feedback control mode; the expert rule control mode uses the dynamic water quality safety threshold corresponding to the ambient temperature as the control setpoint, and starts water supply regulation when the measured water quality approaches the safety threshold; the PI proportional-integral feedback control mode uses the deviation between the dynamic safety threshold and the real-time measured water quality value as input, and outputs the water pump supply frequency through proportional-integral calculation; S72. Water quality index estimation during malfunctions: Using a soft sensor model trained with historical normal aquaculture data, water quality indexes measured by malfunctioning sensors are estimated based on the associated sensor data of water temperature, ammonia nitrogen, and chlorophyll a. S73, Abnormal Alarm Push: When the control mode switch is triggered, the system sends alarm information to the outside through the communication network; S74. Automatic recovery of control mode: When the sensor data returns to normal and the model prediction confidence rises above the preset threshold, the system automatically switches back to the MPC optimized control mode.

9. The intelligent energy-saving water supply method for burial shellfish cultivation ponds according to claim 1, characterized in that, The method also includes: S8, Cloud Collaboration and Model Evolution S81. Federated Learning Global Model Update: The edge computing and control units of each breeding pond use local recent aquaculture data to calculate the parameter update gradient of the water quality prediction model. After desensitization and encryption, the gradient is uploaded to the cloud management platform. The cloud management platform uses the federated averaging algorithm to aggregate the gradients of each node, update the global water quality prediction model, and distribute the updated model parameters to each edge computing and control unit. S82, Meta-learning New Batch Quick Adaptation When starting a new batch of buried shellfish seedlings for cultivation, the cloud management platform generates a model initialization parameter package based on the model training trajectory of similar seedling varieties, similar cultivation batches, or similar environmental conditions in the past, using a meta-learning algorithm. After the parameter package is sent to the edge computing and control unit of the corresponding cultivation pond, the edge computing and control unit fine-tunes the online data to make the water quality prediction model reach a usable prediction accuracy. S83, Knowledge Base Continuous Optimization The cloud management platform aggregates abnormal water quality data and seedling growth data from each cultivation pond, continuously optimizes the dynamic water quality threshold parameter surface related to temperature, and regularly distributes the updated knowledge base to each edge computing and control unit. The abnormal water quality data includes events where chlorophyll a deviates from the suitable range and water quality threshold violation events. The seedling growth data includes seedling survival rate, growth rate, and growth uniformity. The cloud management platform performs statistical analysis on the abnormal water quality data and seedling growth data to optimize the parameter surfaces for the lower safe limit of dissolved oxygen, the upper safe limit of ammonia nitrogen, the upper and lower limits of suitable chlorophyll a concentration, and the safe pH range.

10. A smart energy-saving water supply system for a burial shellfish cultivation pond, characterized in that, The system includes a sensor network, edge computing and control unit, variable frequency water supply execution unit, and cloud-based collaborative management platform; The sensor network is deployed inside each cultivation tank to periodically collect multivariate time-series water quality and environmental data. The collected indicators include at least dissolved oxygen concentration, water temperature, ammonia nitrogen concentration, chlorophyll a concentration, pH, light intensity, ambient temperature, and historical water supply flow. The sensor network adopts a vertical gradient combined with representative planar points. In the vertical dimension of the cultivation tank, sensors are placed at least at a preset height above the sediment-water interface, and anti-siltation guide covers are provided on the outside of the sensors at this location. The data collected by the sensor network is transmitted to the edge computing and control unit in real time. The edge computing and control unit is an industrial PLC or an embedded edge gateway, including a data preprocessing module, an ST-Attention-BiLSTM prediction model, a DA-MPC rolling optimization solution module, and an abnormal operating condition processing module, and a knowledge base for local storage of temperature-related dynamic water quality safety thresholds. The variable frequency water supply execution unit is controlled by the edge computing and control unit and includes a frequency converter, a water supply pump, and multiple water supply branch pipes. Each water supply branch pipe corresponds to a cultivation tank, and each branch pipe is equipped with an electric regulating valve. The frequency converter adjusts the total water output of the water pump according to the frequency command issued by the edge computing and control unit, and each electric regulating valve adjusts the water inflow of the corresponding cultivation tank according to the opening command. The variable frequency water supply execution unit transmits the water pump operating frequency, valve opening, sub-tank flow rate, and total flow rate back to the edge computing and control unit for archiving and storage in real time. The cloud-based collaborative management platform is deployed on a cloud server and communicates with multiple edge computing and control units; the platform has built-in federated learning operation unit, MAML meta-learning operation unit, and dynamic threshold iterative optimization unit.

Images