Lightweight digital twin modeling method for deepwater net cage culture environment

Abstract

The present application relates to the technical field of digital twinning and edge computing, and discloses a lightweight digital twinning modeling method for deepwater net cage culture environment, comprising: performing lightweight pretreatment based on physical constraints on multi-source environmental data; constructing a lightweight hydrodynamic digital twinning model through a residual physical information neural operator; performing multi-parameter coupling fusion through a variational Bayesian integrated state estimator; constructing a growth prediction lightweight model through teacher-student knowledge distillation; and performing counterfactual feeding strategy deduction through a structural causal model and realizing virtual-real closed-loop synchronous correction.

Description

Technical Field

[0001] This invention relates to the fields of digital twin and edge computing technology, and in particular to a lightweight digital twin modeling method for deep-sea cage aquaculture environments. Background Technology

[0002] As the global marine aquaculture industry expands into deep-sea areas, deep-sea cage culture has become a crucial approach to addressing nearshore aquaculture space saturation and insufficient environmental carrying capacity. However, the complexity of deep-sea cage culture environments far exceeds that of traditional nearshore pond aquaculture. In deep-sea environments, key water quality parameters such as water temperature, dissolved oxygen, pH, and salinity are highly spatiotemporally non-stationary due to the coupled influence of multiple factors, including ocean currents, tidal variations, seasonal thermoclines, and wind and wave disturbances. Complex interactions exist between different environmental parameters; for example, increased water temperature leads to decreased dissolved oxygen saturation, while changes in salinity affect the metabolic levels and feeding behavior of farmed fish by altering osmotic pressure. The interplay of these factors presents a significant challenge to the accurate modeling of deep-sea cage culture environments.

[0003] Chinese invention application CN119442140A discloses a method for monitoring the growth status of seedlings in aquaculture. This method constructs a time-series dataset of hydrological conditions by acquiring real-time monitoring data of water temperature, dissolved oxygen, and salinity in the aquaculture farm. It then establishes a correlation model of hydrological parameters using Pearson correlation coefficient and a generalized additive model, and builds a seedling growth process model using a Gompertz growth model and a multilayer perceptron regression model. Furthermore, it uses an attention-enhanced long short-term memory network to predict seedling growth, and finally optimizes the feeding strategy using the NSGA-II multi-objective optimization algorithm. While this method achieves a certain degree of modeling of the relationship between the aquaculture environment and seedling growth, it has the following shortcomings.

[0004] First, this method employs a purely data-driven modeling paradigm, relying entirely on historical monitoring data for model training. It fails to incorporate prior knowledge of physical mechanisms such as hydrodynamics and thermodynamics. When extreme conditions arise in the aquaculture environment that exceed the historical data distribution range, the model's extrapolation and generalization capabilities are significantly insufficient, and it is difficult to guarantee that the prediction results meet basic physical conservation constraints. Second, the model architecture of this method includes multi-layer LSTM networks and attention mechanism modules, resulting in a high number of model parameters and computational complexity. This makes it difficult to deploy directly on edge computing nodes with limited computing power in deep-sea scenarios. Furthermore, deep-sea cage aquaculture platforms typically lack stable high-bandwidth communication links, making it impossible to transmit all raw data back to the shore-based cloud for processing in real time. Third, this method only establishes a one-way environment-growth mapping relationship, lacking counterfactual inference capabilities in virtual simulation space. It cannot simulate and compare the causal effects of different feeding strategies under given specific environmental conditions. Its feeding optimization is essentially still a parameter search based on correlation rather than a strategy inference based on causal inference. Fourth, this method does not construct a digital twin framework that maps the virtual and real worlds, and lacks a dynamic synchronization correction mechanism between physical and virtual spaces. As the aquaculture environment continues to change, the model accuracy will gradually degrade and become unable to be updated adaptively.

[0005] In summary, there is an urgent need for a lightweight digital twin modeling method for deep-sea cages that integrates physical mechanism models and data-driven approaches, targeting edge computing nodes with limited computing power, to achieve low-cost, high-precision simulation of aquaculture environments and optimization of feeding strategies. Furthermore, existing water quality prediction models generally lack the ability to quantify model prediction uncertainties. When data quality varies across multiple sensor channels, the fusion weights cannot be dynamically allocated based on the reliability of each channel, making the fusion results susceptible to contamination by low-quality sensor data. Simultaneously, in deep-sea environments, sensors are constantly immersed in high salinity, high humidity, and strong currents, leading to widespread sensor drift due to electrode aging and biofouling. Existing methods lack cross-parameter cross-validation based on physical prior knowledge to identify such progressive failures. Therefore, it is necessary to develop a lightweight digital twin modeling method that can integrate physical mechanism constraints, support multi-parameter fusion with uncertainty perception, operate in real-time at the edge, and possess virtual-real closed-loop correction capabilities. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a lightweight digital twin modeling method for deep-sea cage aquaculture environments. By integrating physical mechanism constraints with data-driven learning, a lightweight digital twin framework that can be deployed on edge computing nodes is constructed, supporting counterfactual feeding strategy deduction and synchronous correction of virtual and real closed loops.

[0007] The lightweight digital twin modeling method for deep-sea cage aquaculture environments provided by this invention includes the following steps:

[0008] Real-time environmental data, including water temperature, dissolved oxygen concentration, pH value, and salinity, collected by multi-source sensors in deep-sea cage aquaculture environment is acquired. Physically constrained edge-side lightweight preprocessing is performed on the real-time environmental data. Physical consistency verification rules between parameters are constructed using the seawater thermodynamic equation of state. Abnormal data is marked and repaired, and an environmental state tensor verified by physical constraints is generated.

[0009] Using the environmental state tensor as input, a lightweight hydrodynamic digital twin model at the deep-sea cage scale is constructed through residual physical information neural operators. The simplified Navier-Stokes equations are embedded as soft constraints into the loss function, and structured channel pruning is used to compress the model, resulting in a lightweight hydrodynamic field prediction model that meets the computing power constraints of edge computing nodes. The output is a digital twin of the cage's internal environment, including velocity distribution, dissolved oxygen diffusion field, and feed dispersion trajectory.

[0010] The environmental digital twin is fused with the environmental state tensor, and multi-parameter coupled state estimation is performed through a variational Bayesian ensemble state estimator. The physical field output by the hydrodynamic digital twin model is used as a priori constraint, and the fusion weights of each environmental parameter channel are dynamically adjusted to output a fused environmental state vector with uncertainty quantification.

[0011] Based on the fusion of environmental state vectors, a lightweight prediction model of growth-environment response is constructed through a teacher-student knowledge distillation framework with physical regularization. During the distillation process, physical consistency regularization loss is introduced to constrain the output of the student model to satisfy the bioenergetics conservation relationship. The predicted values ​​of growth rate and feed conversion efficiency of the target fish species under given environmental conditions are output.

[0012] Based on growth rate predictions, feed conversion efficiency predictions, and environmental digital twins, a counterfactual feeding strategy is deduced using a structural causal model. The growth benefit response surface under different feeding schemes is simulated in a virtual digital twin space to select the optimal feeding strategy. Based on the deviation signal between the counterfactual deduction results and actual aquaculture data, the physical parameters of the hydrodynamic digital twin model and the fusion weights of the state estimator are updated in reverse to achieve synchronous correction of virtual and real closed loop.

[0013] The beneficial effects of this invention include:

[0014] First, by embedding physical information constraints into the training process of neural networks, the constructed digital twin model can still maintain physical consistency under data sparsity and extreme working conditions, effectively overcoming the problem of insufficient extrapolation generalization of pure data-driven methods.

[0015] Secondly, through a two-layer lightweight strategy combining structured channel pruning and teacher-student knowledge distillation, the constructed hydrodynamic digital twin model and growth prediction model can be deployed on edge computing nodes, eliminating the dependence on onshore cloud computing centers and meeting the real-time inference requirements of low latency and low power consumption in deep-sea scenarios.

[0016] Third, by introducing the counterfactual inference mechanism of structural causal model, the traditional correlation-based feeding parameter search is upgraded to a causal inference-based strategy inference. This allows for the systematic evaluation of the causal effects of different feeding schemes in virtual space, providing more reliable simulation support for accurate feeding decisions.

[0017] Fourth, through a virtual-real closed-loop synchronous correction mechanism, the model parameters are continuously corrected using the deviation signal between actual aquaculture data and digital twin predictions, ensuring that the digital twin model maintains dynamic consistency with the real aquaculture environment during long-term operation.

[0018] Fifth, uncertainty-aware fusion of multiple parameter channels is achieved through a variational Bayesian ensemble state estimator. The fusion weights of each channel can be adaptively adjusted according to the real-time accuracy of the sensors, effectively avoiding contamination of the fusion results by inferior sensor data and improving the reliability of the fusion environment state. Overall, this invention constructs a complete closed-loop framework of perception-modeling-fusion-prediction-inference-correction through five deeply coupled core steps. The coupling relationship between each step is not only reflected in the step-by-step transmission of forward data flow but also in the reverse feedback of backward correction signals, thereby enabling the entire digital twin system to have the ability to continuously evolve and adaptively optimize. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the lightweight digital twin modeling method for deep-sea cage aquaculture environment provided in this embodiment of the invention.

[0020] Figure 2 This is a detailed flowchart of the edge-side lightweight preprocessing steps based on physical constraints provided in an embodiment of the present invention.

[0021] Figure 3 This is a detailed flowchart of the steps for constructing a lightweight hydrodynamic digital twin model using the residual physical information neural operator provided in this embodiment of the invention.

[0022] Figure 4 This is a detailed flowchart of the variational Bayesian integrated state estimation and teacher-student knowledge distillation steps provided in this embodiment of the invention.

[0023] Figure 5 This is a detailed flowchart of the counterfactual feeding strategy deduction and virtual-real closed-loop synchronous correction steps provided in the embodiments of the present invention. Detailed Implementation

[0024] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustrative and explanatory purposes only and are not intended to limit the scope of the invention. The technical features involved in the various embodiments of the present invention can be combined with each other without conflicting with each other.

[0025] like Figure 1 As shown, this invention provides a lightweight digital twin modeling method for deep-sea cage aquaculture environments. This method is designed for deep-sea edge computing nodes with limited computing power, integrating physical mechanism models and data-driven learning to construct a virtual-real mapped cage aquaculture simulation environment. It supports virtual extrapolation and counterfactual analysis of feeding-growth benefits. Preferably, the method comprises five core steps, forming a deeply coupled closed-loop collaborative process. The output of the previous step serves as the key input for the next step, while the correction results of subsequent steps inversely influence the model parameters of the previous steps, thereby achieving continuous adaptive optimization of the overall framework. It should be noted that the deep-sea cages described in this invention are applicable to, but not limited to, cylindrical high-density polyethylene cages and square steel structure cages. The aquaculture species are applicable to, but not limited to, common deep-sea cage-cultured fish such as large yellow croaker, grouper, and golden pomfret. The following embodiments use large yellow croaker aquaculture as a preferred example for illustration.

[0026] Step S1: Multi-source environmental data edge acquisition and lightweight preprocessing with physical constraints. For example... Figure 2 As shown, in one embodiment of the present invention, the implementation process of step S1 is as follows.

[0027] First, real-time environmental data collected by multi-source sensors deployed in the deep-sea cage aquaculture environment is acquired. Preferably, the multi-source sensors include, but are not limited to, temperature sensors, dissolved oxygen sensors, pH sensors, and salinity sensors. These sensors are distributed and deployed at different depths and spatial locations within the deep-sea cage to capture the spatial heterogeneity of environmental parameters within the cage. In a preferred embodiment of the invention, the sampling period for the sensors is set to 5 to 15 minutes. The water temperature sensor has a range of -2°C to 40°C and a resolution of 0.01°C; the dissolved oxygen sensor has a range of 0 mg / L to 25 mg / L and a resolution of 0.01 mg / L; the pH sensor has a range of 0 to 14 and a resolution of 0.01; and the salinity sensor has a range of 0‰ to 50‰ and a resolution of 0.01‰. The sensors aggregate the raw collected data to an edge computing gateway on the cage platform via an underwater RS-485 bus or a low-power wireless communication link. It is worth noting that the above sensor parameters are preferred configurations and can be adaptively adjusted according to the specific environmental characteristics and monitoring accuracy requirements of the aquaculture area in practical applications.

[0028] Next, the real-time environmental data undergoes edge-side lightweight preprocessing based on physical constraints. In this invention, this preprocessing differs from traditional purely statistical anomaly detection methods. Its core innovation lies in introducing the seawater thermodynamic equation of state as a physical constraint to achieve cross-parameter consistency verification. Specifically, firstly, range boundary checks are performed on the raw data of each channel. The effective range for water temperature is set to 2℃ to 35℃, the effective range for dissolved oxygen concentration is set to 0.5mg / L to 20mg / L, the effective range for pH value is set to 6.0 to 9.5, and the effective range for salinity is set to 5‰ to 42‰. Data exceeding these ranges are directly marked as sensor malfunctions and discarded.

[0029] Furthermore, in one embodiment of the present invention, a physical consistency verification rule is constructed using the thermodynamic relationship between dissolved oxygen saturation in seawater and water temperature and salinity. This verification is based on the following physical principle: under specific water temperature and salinity conditions, there is a theoretical upper limit to the saturation concentration of dissolved oxygen in seawater, which can be calculated using the Weiss empirical formula for dissolved oxygen saturation. This formula is expressed as:

[0030] ,

[0031] in: This refers to the dissolved oxygen saturation concentration, expressed in mg / L. It is the ratio of absolute temperature to 100, i.e. , Water temperature, in °C; Salinity, expressed in ‰; , , , The empirical constants related to dissolved oxygen solubility are taken as -173.4292, 249.6339, 143.3483 and -21.8492, respectively. The above values ​​are derived from the international oceanographic standard data and are the fitting results under the conditions of water temperature 0℃ to 40℃ and salinity 0‰ to 40‰. , , The salinity correction coefficients are set to -0.033096, 0.014259, and -0.001700, respectively. The technical advantage of this formula lies in providing a cross-parameter cross-validation method based on thermodynamic principles. This allows the detection of anomalies in dissolved oxygen data to no longer rely solely on the statistical distribution of a single parameter, but rather to combine the synergistic constraints of water temperature and salinity for physical consistency judgment.

[0032] When the measured dissolved oxygen concentration When it exceeds 1.3 times the theoretical saturation value under the corresponding water temperature and salinity conditions, that is... The data point is determined to be a physically inconsistent anomaly. Preferably, the threshold coefficient of 1.3 is selected based on the following: considering that factors such as the operation of aeration equipment and photosynthesis in deep-sea cage environments may cause the local dissolved oxygen concentration to temporarily exceed the saturation value, but usually does not exceed 1.2 to 1.3 times the saturation value, using 1.3 times as the discrimination boundary can effectively distinguish between normal oversaturation fluctuations and false readings caused by sensor drift or malfunction.

[0033] For data points marked as anomalous, a repair strategy based on spatiotemporal neighborhood weighted interpolation is adopted. Specifically, taking the timestamp of the anomalous point as the center, valid data from the same sensor within three sampling time steps before and after it are taken, along with valid data from other spatially adjacent sensors within the same time step. The valid data within the aforementioned spatiotemporal neighborhood are then interpolated and repaired using a distance-weighted average. In a preferred embodiment of the invention, the weight decay coefficient of the time dimension... The weight is set to 0.8, meaning that the weight decreases as the time step is further away from the outlier; this is the weight decay coefficient for the spatial dimension. The weighting is set to 0.6, meaning that the weight of a spatial location farther from the anomaly sensor decreases. The selection of the weighting attenuation coefficient is based on the spatiotemporal correlation analysis results of water quality parameters in deep-sea cage environments. The autocorrelation length of water quality parameters in the time dimension is typically 15 to 30 minutes, and the correlation length in the spatial dimension is typically 3 to 8 meters. Therefore, the attenuation coefficient can make the interpolation results biased towards effective observations that are closer in time and space and have stronger correlations.

[0034] Finally, the multi-source environmental data, after undergoing the aforementioned physical constraint verification and anomaly repair, is organized into an environmental state tensor. In one embodiment of the present invention, this environmental state tensor is a four-dimensional tensor structure. Their dimensions correspond to the length of the time window, respectively. Number of spatial sampling points Number of environmental parameter channels and the number of statistical characteristics Preferably, the time window length The sampling time steps are set to be continuous from 48 to 168, with 72 sampling time steps being preferred. A 10-minute sampling cycle corresponds to an observation window of 8 to 28 hours (or 12 hours under the preferred 72 sampling time steps). The lower limit of 48 sampling time steps corresponds to the minimum effective observation period identified in the autocorrelation analysis of the deep-sea cage environmental parameters, while the upper limit of 168 sampling time steps corresponds to the complete hydrological cycle under the superposition of diurnal variation and tidal cycle of the deep-sea cage. The number of spatial sampling points... The number of sensors deployed within the cage depends on the number of sensors, preferably 8 to 16; the number of environmental parameter channels... The value is 4, corresponding to water temperature, dissolved oxygen concentration, pH value, and salinity, respectively; the number of statistical features. The value is 5, including the mean, standard deviation, maximum, minimum, and rate of change within the window. This environmental state tensor serves as the core input for the subsequent step S2, constructing the hydrodynamic digital twin model. Its four-dimensional structure allows the model to simultaneously capture multi-scale features of environmental parameters across time, space, and channel dimensions. Notably, the preprocessing in this step is executed entirely locally on the edge computing node, independent of network connections to the shore-based cloud. The computational complexity of the preprocessing is O(n log n). Under the aforementioned optimized configuration, a single preprocessing cycle takes approximately 15ms, far less than the sampling period, ensuring the real-time nature of the data stream. Furthermore, the thermodynamic parameter table used for physical consistency verification is pre-stored in the non-volatile storage of the edge nodes, eliminating the need for online queries to external databases and thus guaranteeing the system's independent operation capability even under communication link interruption conditions. The spatiotemporal distribution information of abnormal data points detected during preprocessing is recorded in the local log and uploaded to the shore-based management system after communication is restored for maintenance personnel to analyze the sensor's health status. This log information also serves as an auxiliary reference for the virtual-real synchronization correction in step S5, used to determine whether model deviations stem from environmental changes or sensor degradation.

[0035] Step S2: Construction of a lightweight hydrodynamic digital twin with embedded physical information, such as Figure 3 As shown, the environment state tensor output in step S1 As input, this step constructs a lightweight hydrodynamic digital twin model at the scale of a deep-sea cage using a residual physical information neural operator. Preferably, this step is one of the core components of the technical solution of this invention, and its key innovation lies in embedding simplified fluid dynamics control equations into the training process of the neural network in the form of soft constraints. While ensuring that the prediction results satisfy the basic physical conservation laws, the model size is reduced to the order of magnitude that can be handled by edge computing nodes through structured compression technology.

[0036] First, a simplified Navier-Stokes equation system suitable for deep-water cages is constructed. Preferably, considering that typical deep-water cages are cylindrical or square structures with a diameter of 15m to 30m and a depth of 10m to 20m, the Reynolds number of the internal flow field is usually within a certain range. to The flow velocity is on the order of magnitude of turbulent flow, but the drag effect of the netting inside the cage makes the velocity inside the cage significantly lower than the external ocean current; the typical average velocity inside the cage is 0.05 m / s to 0.3 m / s. Based on these characteristics, in one embodiment of the present invention, a simplified form of the Reynolds-averaged Navier-Stokes (RANS) equation is used as the physical constraint. The continuity equation is expressed as:

[0037] ,

[0038] in: For flow velocity vectors, , , These represent the velocity components along the length, width, and depth of the wire mesh cage, respectively, all in m / s. This is the divergence operator. The equation expresses the mass conservation condition of the fluid, that is, under the assumption of incompressible fluid, the flow rate into the control volume is equal to the flow rate out of the control volume.

[0039] The simplified form of the momentum equation is:

[0040] ,

[0041] in: Time, in seconds; The density of seawater is expressed in kg / m³. In a preferred embodiment of the present invention, the value ranges from 1020 kg / m³ to 1030 kg / m³. The specific value is obtained by calculating the seawater state equation based on the measured water temperature and salinity. Pressure, unit is Pa; The effective eddy viscosity coefficient, expressed in m² / s, is derived from the molecular viscosity coefficient. With turbulent eddy viscosity The sum of them constitutes, that is ,in Approximately m² / s, Estimated using an algebraic turbulence model, the value range in the preferred embodiment of this invention is [value range missing]. m² / s to m² / s; The term represents the resistance source of the netting, measured in m / s², and its physical meaning is the damping effect of the netting on the water flow.

[0042] Preferably, the mesh resistance source item The Morison equation is used in a simplified, parameterized form:

[0043] ,

[0044] in: The resistance coefficient of the mesh is dimensionless and ranges from 1.0 to 2.5. The specific value is related to the mesh size, the diameter of the mesh wire, and the thickness of the attached organism. In a preferred embodiment of the present invention, the initial value is 1.5. This represents the projected area of ​​the mesh fabric, in m². This refers to the effective volume of the net cage, in m³. Let be the magnitude of the velocity vector, in m / s. The technical advantage of the above parameterized expression is that it simplifies the complex drag effect of the mesh on the flow field into a volumetric force term proportional to the square of the velocity, while retaining the core physical characteristics of the mesh drag and significantly reducing the computational complexity of the numerical simulation.

[0045] Next, a Residual Physics-Informed Neural Operator (RPNO) is constructed. The network architecture of this operator consists of an encoder, several residual Fourier layers, and a decoder. Preferably, the encoder is a fully connected layer that maps the input environment state tensor to a 128-dimensional latent space feature vector. The number of residual Fourier layers is set to four, each containing a Fourier transform module and a fully connected feedforward module. The Fourier transform module retains the first 16 modes in the frequency domain to filter out high-frequency noise, and the fully connected feedforward module has a hidden layer dimension of 256. Gradients are directly propagated between the residual Fourier layers through residual connections. The decoder is a fully connected layer that maps the latent space features to the output space. Preferably, the output space contains three physical field channels: a velocity field... (3 velocity components), dissolved oxygen diffusion field and feed dispersion concentration field .

[0046] RPNO's loss function is designed as a data-driven loss term. With physical constraint loss term Weighted combination:

[0047] ,

[0048] in: and These are the weighting coefficients for data-driven loss and physical constraint loss, respectively, with initial values ​​set to 1.0 for both. The mean square error between the network prediction and the sensor measurement is... ,in The number of sensor observation points, and The first Predicted and measured values ​​for each observation point; It is obtained by calculating the sum of squared residuals of the network output on the continuity equation and the momentum equation, i.e. ,in The preferred number of physical constraint points is... to A spatial point evenly distributed within the computational domain of the net cage. and These are the residuals of the continuity equation and the momentum equation, respectively.

[0049] Preferably, the weighting coefficient and Dynamically adjusted using an adaptive balancing strategy based on the gradient norm. Specifically, after each training iteration, the following is calculated: and Gradient norm of each network parameter and When the ratio of the two deviates from 1.0 by more than the preset threshold of 0.1, the weights are adjusted according to the following rules: If Then increase Decrease Conversely, the same applies. The technical effect of this adaptive strategy is to ensure that data-driven learning and physical constraint learning remain in balance during training, avoiding a situation where one type of loss dominates training while the other is neglected.

[0050] After RPNO training is complete, structured channel pruning is used to compress the model to meet the computational constraints of edge computing nodes. Preferably, the compression ratio is set to 0.4 to 0.7, that is, retaining 40% to 70% of the original model's channel count. The specific pruning strategy is as follows: calculate the L1 norm of each channel in each convolutional or fully connected layer, remove the channel with the smallest L1 norm according to the preset compression ratio, and then fine-tune the model to restore accuracy. In a preferred embodiment of the present invention, the total number of parameters of the original RPNO is approximately After structured pruning with a compression ratio of 0.5, the number of parameters was reduced to approximately [number missing]. The number of samples, representing 39.3% of the original, was reduced. The accuracy of the hydrodynamic field prediction after compression decreased by approximately 5.2% compared to the original, which is below the allowable degradation threshold of 8%. The compressed model achieved a single-step inference time of approximately 85ms on an edge computing node equipped with an ARM Cortex-A72 processor and 4GB of memory, meeting the real-time inference requirements under a 10-minute sampling period.

[0051] Step S3: Bayesian-guided multi-parameter coupled state estimation and adaptive fusion, such as... Figure 4 As shown, this step fuses the environmental digital twin output from step S2 with the environmental state tensor output from step S1, and performs multi-parameter coupled state estimation through a variational Bayesian ensemble state estimator. The key innovation of this step lies in the fact that, unlike traditional data fusion methods that only perform simple weighted averaging at the data level, this invention uses the physical field output from the hydrodynamic digital twin model as a priori constraint, transforming the fusion problem of physical field prediction and sensor observation into a Bayesian posterior inference problem. This enables adaptive dynamic adjustment of the fusion weights for each environmental parameter channel and synchronous quantification of uncertainties.

[0052] In one embodiment of the present invention, the variational Bayesian integrated state estimator operates as follows. Let at time... The digital twin output in step S2 is the first... One environmental parameter channel ( (corresponding to water temperature, dissolved oxygen, pH, and salinity respectively) in the... The predicted value of each spatial sampling point is The actual sensor values ​​corresponding to the environmental state tensor output in step S1 are The variational Bayesian framework will merge the environmental states. Treating it as a random variable, its posterior distribution is obtained through the following variational inference process:

[0053] ,

[0054] in: It is the optimal variational distribution; The distribution family is variational; in this invention, the Gaussian distribution family is selected. The Kullback-Leibler divergence; This represents the true posterior distribution. Under the Gaussian assumption, the closed-form solution of the above variational inference is expressed in weighted mean form:

[0055] ,

[0056] in: and Digital twin prediction and sensor observation are respectively used in the first half of the year. The fusion weights on each channel are equal to 1; these weights are inversely proportional to their respective prediction variances. , ,in The variance of the digital twin prediction. Let Variance be the variance of sensor observations. Intuitively, sources with smaller prediction variance receive greater weight in the fusion, which aligns with the basic principle of Bayesian optimal estimation.

[0057] Preferably, the initial fusion weights of each environmental parameter channel are determined based on the nominal accuracy class of the sensor. In a preferred embodiment of the invention, the initial observation weights of the water temperature channel are... The initial observation weight for the dissolved oxygen channel is set to 0.25 to 0.35, preferably 0.30. The initial observation weight for the pH channel is set to 0.25 to 0.35, preferably 0.30. The initial observation weight for the salinity channel is set to 0.15 to 0.25, preferably 0.20. The initial observation weight range is set to 0.10 to 0.20, with 0.15 being preferred. The initial observation weight range for each channel is determined as follows: the nominal accuracy of water temperature and dissolved oxygen sensors is typically better than ±1% of the range in deep water environments, corresponding to a weight range of 0.25 to 0.35; the pH sensor experiences electrode drift under long-term immersion conditions, causing its relative accuracy to fall within ±2% of the range, corresponding to a weight range of 0.15 to 0.25; the salinity sensor is most significantly affected by biofouling, with a weight range of 0.10 corresponding to severe biofouling conditions and 0.20 corresponding to clean conditions. The selection of these initial weights reflects the relative reliability of different types of sensors in deep water environments: water temperature and dissolved oxygen sensors typically have high stability and accuracy, thus receiving higher initial weights; pH sensors are prone to zero-point drift due to electrode aging under long-term immersion conditions, thus receiving slightly lower weights; the salinity sensor is significantly affected by biofouling, thus receiving the lowest initial weight.

[0058] Within each estimation period, the aforementioned fusion weights are adaptively updated based on the variance of the prediction residuals for each channel. Specifically, for the ... Each channel calculates the nearest... time steps (preferred) The residual sequence between the digital twin predicted values ​​and the sensor measured values ​​within a 2-hour observation window. ,in Then calculate the variance of the residual sequence. As The online estimate of this parameter. The larger the variance, the less accurate the digital twin model's prediction of this parameter, and the lower its weight in the fusion.

[0059] The fused environment state vector It also includes uncertainty quantification information, namely the fused variance. The fusion variance is always less than the variances of the two individual input sources, which is the core advantage of Bayesian fusion over a single data source. The uncertainty quantification result will be used as the basis for confidence assessment in the counterfactual deduction in step S5, ensuring that the final output feeding strategy has controllable risk.

[0060] Step S4: Lightweight modeling of growth-environment response driven by teacher-student knowledge distillation, such as... Figure 4 As shown, this step, based on the fused environment state vector output in step S3, constructs a lightweight prediction model for growth-environment response using a physically regularized teacher-student knowledge distillation framework. This step complements the lightweight strategy of step S2: step S2 compresses the physical simulation part of the hydrodynamic model through structured pruning, while this step compresses the data-driven part of growth prediction through knowledge distillation. Together, they constitute a two-layer lightweight strategy for edge deployment.

[0061] First, a teacher model is constructed. In one embodiment of the invention, the teacher model is a deep network containing complete physical information constraints, and its architecture consists of two parts connected in series: a temporal feature extraction module and a growth prediction module. The temporal feature extraction module uses a 4-layer bidirectional gated recurrent unit (Bi-GRU), with each layer containing 128 hidden neurons. The input is the time series of the fused environment state vector output in step S3. ,in To determine the number of observation time steps, the optimal method is... This corresponds to a 28-hour continuous observation window. The growth prediction module is a 3-layer fully connected network with hidden layer dimensions of 256, 128, and 64 respectively, and an output layer dimension of 2, corresponding to the daily average weight gain rate (in g / d) and feed conversion efficiency (dimensionless ratio), respectively. The total number of parameters in the teacher model is approximately On a shore-based GPU server (such as an NVIDIA RTX 3060 with 12GB of video memory), the single-step inference time is approximately 12ms.

[0062] The training loss function of the teacher model incorporates bioenergetics conservation constraints. Specifically, the energy budget of fish can be expressed as:

[0063] ,

[0064] in: Energy intake is expressed in kJ / d and is calculated by multiplying the feed amount by the feed energy density. The energy used for body tissue growth is expressed in kJ / d and is directly proportional to the average daily weight gain rate. The proportionality coefficient is the energy density of the fish body tissue, preferably ranging from 5.0 kJ / g to 7.0 kJ / g. It is the energy consumption of basal metabolism and activity metabolism, expressed in kJ / d, and is significantly affected by water temperature. It can be parameterized using the Arrhenius temperature dependence. To excrete lost energy, the preferred method is to... 8% to 12%; The energy consumed for a specific dynamic action, namely the energy required to digest the food itself, is preferably selected. 10% to 15%. Physical constraint loss of the teacher model. Defined as the deviation between the sum of the energy components predicted by the network and the energy intake:

[0065] ,

[0066] in: The optimal number of samples in the training batch is [to be determined]. This constraint ensures that the teacher model's predictions are consistent at the energy level, avoiding physically impossible growth rate predictions.

[0067] Next, a student model is constructed and knowledge distillation is performed. In one embodiment of the invention, the student model is a shallow network deployed at the edge, its architecture consisting of a 1-layer unidirectional GRU (64 hidden neurons) and a 2-layer fully connected network (128 and 32 hidden layers), with a total number of parameters of approximately The number of students is only 14.1% of that of the teacher model. The distillation training loss function of the student model consists of a weighted sum of three parts:

[0068] ,

[0069] in: , and The weight coefficients for soft label distillation loss, hard label supervision loss, and physical consistency regularization loss are respectively, with preferred values ​​of 0.5, 0.3, and 0.2. The basis for the above weight allocation is that the soft labels of the teacher model contain the richest dark knowledge during the distillation process, and therefore are given the highest weight.

[0070] Soft label distillation loss The results were obtained by calculating the KL divergence between the output distributions of the teacher and student models after temperature scaling.

[0071] ,

[0072] in: Temperature is a parameter used to control the smoothness of the output distribution; preferably... Higher temperature values ​​make the distribution smoother, thus revealing the relative relationships between different categories in the teacher model output (i.e., dark knowledge). It is the softmax function; and The outputs are the logits for the teacher model and the student model, respectively. The role of the pre-factor is to compensate for the reduction effect of temperature scaling on the gradient magnitude.

[0073] Physical consistency regularization loss The growth rate predictions of the student model are obtained by constraining them to satisfy the asymptotic upper bound condition of the von Bertalanffy growth equation. The von Bertalanffy growth equation is expressed as:

[0074] ,

[0075] in: for The fish's body weight at any given time, in grams; This refers to the asymptotic weight, which is the theoretically achievable maximum weight, expressed in grams. The value is determined based on the biological characteristics of the fish species, such as the large yellow croaker. Preferably, it is 800g to 1200g; This is the growth coefficient, in d. The preferred value range is 0.002 to 0.008; The theoretical initial age is expressed in days (d). This equation represents the average daily weight gain rate of the fish. There exists an upper bound related to the current weight, and the physical regularization loss... A penalty will be imposed on student model predictions that violate this upper bound:

[0076] ,

[0077] in: For the student model to the first The predicted average daily weight gain rate for each sample; This represents the theoretical maximum daily weight gain rate under the corresponding body weight conditions, calculated based on the von Bertalanffy equation. The ReLU function ensures that only predictions exceeding the upper bound are penalized without affecting predictions within a reasonable range. The effect of this regularization technique is that when the distilled, lightweight student model runs at edge nodes, its predictions still follow basic biological growth patterns and do not output outlier predictions that exceed physical feasibility.

[0078] After distillation, the student model's single-step inference time on the edge computing node is approximately 8ms, which is 33.3% less than that of the teacher model. Meanwhile, the prediction accuracy on the test set (measured by the R² index of the daily weight gain rate) decreased from 0.913 for the teacher model to 0.887, with an accuracy loss of only 2.8%, which is far better than the accuracy of directly training a small model of the same size (R² = 0.821).

[0079] Step S5: Counterfactual feeding strategy deduction and synchronous correction of virtual and real closed loop, such as... Figure 5As shown, this step is the decision output and closed-loop feedback link of the entire lightweight digital twin framework. Based on the predicted growth rate and feed conversion efficiency output from step S4, and the environmental digital twin output from step S2, a counterfactual feeding strategy is deduced through a structural causal model. The deviation signal between the deduction result and the actual data is used to correct the model parameters of the preceding steps, achieving virtual-real closed-loop synchronization. This step forms a bidirectional information flow with steps S2 to S4: the forward flow is the reasoning path from environmental state to feeding strategy, and the reverse flow is the correction path from deviation signal to model parameters. The synergy of the two constitutes the core feature that distinguishes the digital twin framework of this invention from static simulation models.

[0080] First, a structural causal model (SCM) for optimizing feeding strategies is constructed. In one embodiment of the invention, the causal graph of the SCM defines the following set of variables and their causal relationships: exogenous environmental variables. These correspond to water temperature, dissolved oxygen concentration, pH value, and salinity, respectively. These variables are determined by natural environmental processes and are considered fixed conditions in the deduction of feeding strategies; intervention variables These correspond to daily feeding amount (kg / d), feeding frequency (times / d), and feeding time (discrete variables, taking values ​​of morning, noon, or evening); outcome variable. These correspond to the average daily weight gain rate (in g / d) and the feed conversion ratio (dimensionless, defined as the ratio of feed intake to weight gain; the smaller the value, the higher the feed utilization efficiency).

[0081] The structural equation model (SEM) in SCM defines the causal transmission mechanism among the variables. In a preferred embodiment of the invention, the structural equation for the daily weight gain rate is expressed as:

[0082] ,

[0083] in: The growth-environment-feed response function learned by the student model in step S4. This is the set of parameters for the function. The term represents exogenous noise, which follows a mean of 0 and a variance of . Gaussian distribution, The uncertainty can be estimated from the uncertainty quantification results output in step S3.

[0084] Counterfactual inference is achieved by applying hypothetical intervention operations to the intervention variables using the do operator. Specifically, given the currently observed environmental conditions... For any set of candidate feeding schemes Its counterfactual effect is calculated as follows:

[0085] ,

[0086] The above formula represents, under fixed current environmental conditions, if the feeding plan is changed from the actual implementation plan... Replace with candidate solutions The expected daily weight gain rate will become Preferably, candidate feeding schemes are generated using a grid search strategy: daily feeding amount. Discretize the feeding frequency within a range of 1% to 5% of the total fish weight at a step size of 0.5%. Take an integer value within the range of 2 times / day to 6 times / day for the feeding period. Taking 3 discrete values, the total number of candidate solutions generated is: One. Calculate the corresponding daily weight gain rate for each candidate plan. and feed conversion ratio This allows for the construction of a growth benefit response surface in the feeding scheme space.

[0087] Next, a Pareto front search is performed on the response surface to screen for the optimal feeding strategy. Preferably, the Pareto front search is performed with the dual objectives of maximizing daily weight gain and minimizing feed costs to obtain a set of non-dominated solutions. In a preferred embodiment of the invention, the schemes satisfying the following constraints are further screened from the set of non-dominated solutions as the final output: feed conversion ratio. Not higher than the preset conversion threshold Preferred That is, the amount of feed required for every 1g increase in body weight should not exceed 1.8g; daily feeding amount The amount of feed should not exceed 4% of the fish's body weight to avoid overfeeding, which can lead to water quality deterioration and feed waste. The Pareto optimal plan that meets the above constraints is the recommended optimal feeding strategy. This plan includes specific recommendations on daily feed amount, feeding frequency, and feeding time, which can be used as a reference for aquaculture managers or directly executed by automatic feeding systems.

[0088] Finally, a virtual-real closed-loop synchronous correction is performed. This step is the essential difference between the digital twin framework of this invention and traditional static models, ensuring that the digital twin in virtual space can continuously track the dynamic changes of the real aquaculture environment. In one embodiment of this invention, the triggering and execution process of the virtual-real synchronous correction is as follows: In each correction cycle... (Preferably within 7 days) collect the average daily weight gain rate measured during the actual breeding process. and feed conversion ratio The root mean square bias is calculated by comparing it with the counterfactual inference results of digital twins within the same time period. ,in This is for the number of days within the correction period.

[0089] when Exceeding the preset deviation threshold for two consecutive correction cycles (Preferred) When the weight gain rate reaches approximately 15% to 20% of the normal daily weight gain rate (exceeding this threshold indicates a significant deviation between the digital twin model and the actual system), an online parameter update process is triggered. Preferably, this update process includes two aspects of parameter correction:

[0090] Firstly, the physical parameters of the lightweight hydrodynamic digital twin model in step S2 are incrementally corrected. A parameter identification method based on the Extended Kalman Filter (EKF) is employed, using measured environmental data from the most recent calibration period as the observation input, to refine the eddy viscosity coefficient in the hydrodynamic model. and mesh resistance coefficient The state variables to be identified are jointly estimated. The prediction step of the extended Kalman filter uses the discretized form of the hydrodynamic model for state extrapolation, and the update step uses the residuals (i.e., the innovation sequence) between sensor observations and model predictions to correct the state estimates. Preferably, the eddy viscosity coefficient... The process noise variance is set to 5% of the current estimate, and the mesh resistance coefficient is... The process noise variance is set to 3% of the current estimate. This noise variance reflects the expected drift rate of physical parameters due to environmental changes (such as biofouling growth, net aging, etc.). After one calibration cycle of online EKF identification, the updated physical parameters will take effect in the next inference cycle, enabling the hydrodynamic digital twin model to adapt to physical parameter drift caused by factors such as seasonal changes, net washing, and sea state changes.

[0091] Secondly, the fusion weights of the variational Bayesian integrated state estimator in step S3 are updated. When the variance of the prediction residual of a certain environmental parameter channel shows a continuous increasing trend during the correction period, the weight of the digital twin prediction in that channel is reduced and the weight of the sensor observation is increased, and vice versa. This weight update mechanism enables the fusion process to adaptively cope with situations such as sensor degradation or a decrease in the local accuracy of the digital twin model.

[0092] Through the aforementioned virtual-real closed-loop synchronous correction mechanism, the lightweight digital twin framework of this invention achieves a complete closed loop of perception → modeling → fusion → prediction → inference → correction. The correction result in step S5 not only updates the model parameters in steps S2 and S3, but also indirectly improves the input data quality of the growth prediction model in step S4 in subsequent inference, thereby enabling the entire framework to maintain a high degree of consistency with the real aquaculture environment during long-term operation.

[0093] To verify the technical effectiveness of the method proposed in this invention, a 90-day field verification experiment was conducted at a deep-sea cage aquaculture base in the South China Sea in one embodiment of the invention. The experimental environment consisted of a cylindrical HDPE deep-sea cage with a diameter of 20m and a depth of 15m. The cultured species was large yellow croaker (Larimichthys crocea), with an initial stocking density of 5000 fish and an initial average weight of approximately 150g. Twelve multi-parameter water quality sensor nodes were deployed inside the cage, distributed across three depth layers (surface 2m, middle layer 7m, bottom layer 12m) and four azimuth angles (0°, 90°, 180°, 270°), with a sampling period of 10 minutes. The edge computing nodes employed industrial-grade embedded computing modules equipped with an ARM Cortex-A72 processor (1.5GHz), 4GB DDR4 memory, and 16GB eMMC storage.

[0094] Regarding the accuracy of hydrodynamic digital twins, the RPNO model, after structured channel pruning compression, has a root mean square error (RMSE) of 0.032 m / s for velocity field prediction at edge nodes. This represents a degradation of approximately 14.3% compared to the uncompressed full model (RMSE = 0.028 m / s), but is significantly better than the traditional pure data-driven LSTM model (RMSE = 0.067 m / s), which exhibits significantly lower accuracy in velocity prediction without embedded physical constraints. The RMSE for dissolved oxygen diffusion field prediction is 0.41 mg / L, a degradation of 17.1% compared to the full model (0.35 mg / L), but still better than the LSTM model (0.83 mg / L). These results indicate that embedding physical information effectively improves the model's robustness in sparse data regions, while structured pruning achieves significant model compression within an acceptable range of accuracy loss.

[0095] In terms of growth prediction accuracy, the student model, after knowledge distillation, achieved an R² of 0.887 for daily weight gain, compared to 0.913 for the teacher model and 0.821 for the directly trained small model of the same size. During the 90-day validation period, the cumulative predicted weight bias of the student model was 3.7% of the actual measured value, compared to 2.1% for the teacher model and 8.9% for the pure data-driven LSTM baseline model. These results validate the effectiveness of teacher-student knowledge distillation in significantly reducing model complexity while maintaining prediction accuracy, and the technical effectiveness of physical regularization in avoiding outlier predictions.

[0096] Regarding feeding strategy optimization, the counterfactual reasoning module recommended 12 feeding strategy adjustments during the 90-day validation period, 10 of which were adopted and implemented by aquaculture managers. The net cages adopting the recommended strategies achieved an average fish weight of 487g at the end of the validation period, with a daily weight gain of 3.74g / d and a feed conversion ratio of 1.52. In contrast, the control net cages using the empirical feeding strategy employed by aquaculture managers achieved an average fish weight of 451g, a daily weight gain of 3.34g / d, and a feed conversion ratio of 1.71 during the same period. Compared to the empirical feeding strategy, the method of this invention improved the daily weight gain by approximately 12.0%, improved the feed conversion ratio by approximately 11.1%, and reduced feed costs per net cage by approximately 8.5% during the 90-day validation period.

[0097] Regarding the virtual-real synchronous correction, four online updates of physical parameters were triggered during the 90-day validation period. After the updates, the prediction bias of the digital twin model recovered to below the bias threshold within one correction cycle (7 days), demonstrating the effectiveness of the closed-loop synchronous correction mechanism. In particular, during the period from day 45 to day 55, the experimental sea area experienced a sudden change in sea state due to the impact of a typhoon. The prediction bias of the digital twin model briefly increased after the typhoon, but after online identification and updating of the eddy viscosity coefficient and net drag coefficient by EKF, the prediction bias recovered to the normal level on day 58, demonstrating the adaptive recovery capability of the method of this invention under extreme conditions.

[0098] In terms of edge computing performance, the lightweight digital twin framework's overall inference pipeline (including the serial execution of five steps: data preprocessing, hydrodynamic field prediction, state fusion, growth prediction, and feeding strategy deduction) has an end-to-end inference time of 312ms on the edge node. Specifically, the hydrodynamic field prediction step S2 takes approximately 85ms, the growth prediction step S4 takes approximately 8ms, and the remaining steps take approximately 219ms. This inference time is significantly less than the 10-minute data sampling period, meeting the requirements for real-time inference. The peak memory usage of the edge node is 1.87GB, and the average power consumption is 5.2W, both within the resource budget of an industrial-grade embedded computing module.

[0099] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A lightweight digital twin modeling method for deep-water cage aquaculture environments, characterized in that, The method includes: Real-time environmental data, including water temperature, dissolved oxygen concentration, pH value and salinity, collected by multi-source sensors in deep-sea cage aquaculture environment, are obtained. Physical consistency verification rules between parameters are constructed using the seawater thermodynamic equation of state. Abnormal data are marked and repaired, and an environmental state tensor verified by physical constraints is generated. Using the environmental state tensor as input, a lightweight hydrodynamic digital twin model is constructed through the residual physical information neural operator. The simplified Navier-Stokes equation is used as the soft constraint embedding loss function, and structured channel pruning is used to compress the model. The output is a digital twin of the internal environment of the net cage containing the velocity distribution, dissolved oxygen diffusion field and feed dispersion trajectory. The environmental digital twin is fused with the environmental state tensor, and multi-parameter coupled state estimation is performed through a variational Bayesian integrated state estimator. The physical field output by the hydrodynamic digital twin model is used as a priori constraint to dynamically adjust the fusion weight of each parameter channel, and the fused environmental state vector with uncertainty quantification is output. Based on the fused environment state vector, a lightweight prediction model for growth-environment response is constructed through a physical regularized teacher-student knowledge distillation framework. The teacher model is a deep network with physical information constraints, and the student model is a shallow network deployed at the edge. The physical consistency regularization loss constraint introduced in the distillation satisfies the bioenergetics conservation and outputs the predicted growth rate and feed conversion efficiency. Based on the predicted values ​​and the environmental digital twin, a counterfactual feeding strategy is performed through a structural causal model to simulate the growth benefit response surface under different combinations of feeding amounts, frequencies, and time periods. The optimal feeding strategy is selected, and the physical parameters of the hydrodynamic digital twin model and the fusion weights of the state estimator are updated in reverse according to the deviation between the simulation results and the actual data, so as to achieve synchronous correction of virtual and real closed loop.

2. The method according to claim 1, characterized in that, The physical consistency verification rules include: setting the effective range of water temperature to 2℃ to 35℃, the effective range of dissolved oxygen concentration to 0.5mg / L to 20mg / L, the effective range of pH value to 6.0 to 9.5, and the effective range of salinity to 5‰ to 42‰; based on the thermodynamic relationship between dissolved oxygen saturation in seawater and water temperature and salinity, when the measured dissolved oxygen concentration exceeds 1.3 times the theoretical saturation value under the corresponding water temperature and salinity conditions, the data point is determined to be a physical inconsistency anomaly.

3. The method according to claim 1, characterized in that, The compression ratio of the structured channel pruning is set to 0.4 to 0.7, the number of model parameters of the residual physical information neural operator after compression does not exceed 40% of that before compression, and the prediction accuracy of the hydrodynamic field after compression deteriorates by no more than 8% compared with the root mean square error before compression.

4. The method according to claim 1, characterized in that, In the variational Bayesian integrated state estimator, the initial fusion weights of each environmental parameter channel are determined based on the nominal accuracy level of the sensor. The initial weights for the water temperature channel are 0.25 to 0.35, the initial weights for the dissolved oxygen channel are 0.25 to 0.35, the initial weights for the pH channel are 0.15 to 0.25, and the initial weights for the salinity channel are 0.10 to 0.

20. The fusion weights are adaptively updated in each estimation period based on the variance of the prediction residuals of each channel.

5. The method according to claim 1, characterized in that, The loss function of the residual physical information neural operator includes a weighted combination of a data-driven loss term and a physical constraint loss term. The physical constraint loss term is obtained by calculating the sum of squared residuals of the network output on each control equation of the Navier-Stokes equations. The data-driven loss term is the mean square error between the network prediction and the sensor measurement. The weight coefficients of the two losses are dynamically adjusted by an adaptive balancing strategy based on the gradient norm, so that the two losses remain on the same order of magnitude during training.

6. The method according to claim 1, characterized in that, In the physical regularization framework for teacher-student knowledge distillation, the distillation loss function is composed of a weighted sum of three parts: soft-label distillation loss, hard-label supervision loss, and physical consistency regularization loss. The soft-label distillation loss is calculated using the KL divergence between the output distributions of the teacher model and the student model after temperature scaling. The physical consistency regularization loss is obtained by constraining the predicted growth rate of the student model to satisfy the asymptotic upper bound condition of the von Bertalanffy growth equation.

7. The method according to claim 1, characterized in that, In the counterfactual feeding strategy simulation, the structural causal model uses water temperature, dissolved oxygen concentration, pH value, and salinity as exogenous environmental variables, feeding amount and feeding frequency as intervention variables, and average daily weight gain rate and feed conversion ratio as outcome variables. The do operator applies hypothetical intervention operations to the intervention variables, and calculates the causal effects of different intervention schemes on the outcome variables under the condition of fixed exogenous environmental variables.

8. The method according to claim 1, characterized in that, The environmental state tensor is a four-dimensional tensor structure, whose dimensions correspond to the time window length, the number of spatial sampling points, the number of environmental parameter channels, and the number of statistical features, respectively. The time window length is 48 to 168 consecutive sampling time steps, and the statistical features include the mean, standard deviation, maximum value, minimum value, and rate of change within the window.

9. The method according to claim 1, characterized in that, The virtual-real closed-loop synchronous correction includes: within each correction cycle, calculating the root mean square deviation between the predicted daily average weight gain rate in the counterfactual inference results and the actual measured daily average weight gain rate; when the root mean square deviation continuously exceeds a preset deviation threshold, triggering the online update process of the physical parameters of the lightweight hydrodynamic digital twin model; the online update of physical parameters adopts a parameter identification method based on extended Kalman filtering, using the measured data in the most recent correction cycle as the observation input, and incrementally correcting the eddy viscosity coefficient and net drag coefficient in the hydrodynamic model.

10. The method according to claim 1, characterized in that, The process of selecting the optimal feeding strategy includes: on the growth benefit response surface, with the dual objectives of maximizing the average daily weight gain rate and minimizing feed costs, obtaining a set of non-dominated solutions through Pareto front search, and selecting from the set of non-dominated solutions the feeding scheme that satisfies the feed conversion ratio not lower than a preset conversion threshold and the daily feeding amount not exceeding a preset proportion of the fish body weight as the optimal feeding strategy.

Images