A method for predicting the yield of farmed fish
By constructing a nonlinear coupling function and a differential dynamics model, combined with data assimilation and optimization algorithms, the problems of factor interaction and time lag effects in fish yield prediction of traditional models are solved, and high-precision, adaptive prediction of farmed fish yield is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FRESHWATER FISHERIES RES INSITUTE OF JIANGSUPROVINCE
- Filing Date
- 2026-04-20
- Publication Date
- 2026-07-14
AI Technical Summary
Existing technologies struggle to accurately predict fish yields, especially in high-density polyculture scenarios. Traditional models cannot effectively account for the dynamic impact of environmental factors and the energy losses caused by fish competing for space and food, resulting in high prediction uncertainty. Furthermore, machine learning models lack biological interpretability.
A nonlinear coupling function for multi-source time series data is constructed, a differential dynamic model is established, the model parameters are calibrated through optimization algorithms, dynamically updated by combining data assimilation methods, and numerical integration is performed using the fourth-order Runge-Kutta method to output prediction results with confidence intervals, thereby achieving online optimization and adaptation of the model.
It improves the accuracy of fish production forecasting, can predict growth delay effects caused by environmental degradation in advance, reduces average relative error, and provides adaptive forecasting capabilities to adapt to changes in different aquaculture species, regions, and seasons.
Smart Images
Figure CN122390147A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent aquaculture technology, and in particular to a method for predicting the yield of farmed fish. Background Technology
[0002] Aquaculture is one of the fastest-growing food production sectors globally, providing humans with a high-quality source of animal protein. According to the Food and Agriculture Organization of the United Nations, aquaculture production has surpassed that of wild-caught fisheries, becoming the main source of aquatic products. With population growth and rising consumption levels, the aquaculture industry is developing towards intensification, large-scale production, and intelligent manufacturing.
[0003] However, the uncertainty of aquaculture yield has always been a key factor restricting the improvement of industry efficiency. Currently, the main methods for yield prediction in aquaculture include empirical formulas, single-factor regression models, multi-factor linear regression models, and traditional bioenergetics models. Because fish growth is significantly affected by environmental factors such as water temperature, dissolved oxygen, and ammonia nitrogen, and there is a clear time lag effect—that is, environmental changes on a given day often only manifest in growth rates several days later—this complex nonlinear relationship makes it difficult for traditional empirical methods to accurately predict the final yield.
[0004] In recent years, with the development of artificial intelligence, various predictive methods utilizing machine learning algorithms such as deep learning and support vector machines have been introduced. These methods train on large amounts of historical data to obtain the mapping relationship between inputs (environment, feeding) and outputs (yield). However, because machine learning models are black-box models, their prediction process lacks biological interpretability, making it difficult to understand the mechanistic causes of the prediction results. Furthermore, a few studies have attempted to use differential equations to describe fish growth, such as the Logistic growth model and the Gompertz model. These models can describe the S-shaped characteristics of the growth curve, but they typically only consider density constraints, ignoring the dynamic influence of environmental factors. Most models are designed only for a single species and do not consider the additional energy losses caused by fish competing for space and food under high-density conditions. In particular, the combined effect of this energy loss and environmental stress makes it difficult to extend to polyculture scenarios.
[0005] In summary, to address the shortcomings of the existing technologies, a method for predicting farmed fish yield is provided. Summary of the Invention
[0006] The main objective of this invention is to provide a method for predicting the yield of farmed fish, which can effectively solve the problems in the background art.
[0007] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for predicting the yield of farmed fish includes the following steps: S1: Collect multi-source time-series data of water bodies in the aquaculture area and preprocess it. The multi-source time-series data includes water quality environment data, aquaculture operation data, biological growth data and spatial parameter data. S2: The water quality environmental data is coupled using a nonlinear coupling function to construct a comprehensive environmental load factor at time t. ; S3: Utilizing the aforementioned comprehensive environmental load factor Establish a differential dynamics model to describe the growth process of fish; S4: Based on historical aquaculture data, an optimization algorithm is used to calibrate the unknown parameters in the differential dynamics model; S5: Using real-time acquired data, the state of the differential dynamics model is dynamically updated through a data assimilation method; S6: Set up a future prediction scenario, and use the fourth-order Runge-Kutta method to numerically integrate the differential dynamics model to predict the average body weight and total biomass of fish at future times. S7: Output the prediction results and generate aquaculture management suggestions.
[0008] S8: Quantify the uncertainty of the prediction results and calculate the confidence interval of the prediction results using the ensemble forecasting method. Specifically, this includes: making future predictions using the ensemble members after data assimilation, calculating the 5%, 50%, and 95% quantiles of the prediction results of each member, and generating prediction results with confidence intervals.
[0009] S9: After each breeding cycle, compare the actual harvest data with the model prediction results, calculate the prediction error, and use the newly obtained complete breeding cycle data to update and optimize the model parameters online, thereby improving the model's prediction accuracy for future breeding cycles.
[0010] Furthermore, in step S1, The water quality environmental data includes the water temperature at time t. Dissolved oxygen ammonia nitrogen concentration ; The aquaculture operation data includes daily feed intake. Aerator start-up and shutdown time, water exchange volume; The biological growth data includes the average body weight of the sample. Sampling tails; The spatial parameter data includes water volume and initial stocking number of fish. .
[0011] Furthermore, in step S2, the comprehensive environmental load factor Defined as: ; in, Let the water temperature influence function be defined as: , For optimal water temperature, Water temperature tolerance range; The dissolved oxygen effect function is defined as follows: , It is the half-saturation constant; The ammonia nitrogen influence function is defined as follows: , This represents the ammonia nitrogen toxicity threshold.
[0012] Furthermore, in step S3, the differential dynamics model is defined as: ,in, Let t be the average body weight of the fish at time t; V represents the total biomass at time t; V represents the water volume. The intrinsic growth rate; This is the theoretical maximum specification; These are parameters related to allometric growth. For time delay parameters; This is the stress consumption coefficient; is the group interaction loss coefficient; where .
[0013] Furthermore, the time delay parameter Determined using the following methods: Calculate the comprehensive factor of historical environmental load Fish body growth rate sequence Correlation coefficients at different lag times; The lag time corresponding to the maximum correlation coefficient is selected as the time lag parameter. The value of .
[0014] Furthermore, in step S4, the optimization algorithm is a genetic algorithm or a particle swarm optimization algorithm; The objective function for calibrating the unknown parameters in the differential dynamics model is to minimize the weighted sum of squared errors between the simulated and measured values: ,in, The set of parameters to be calibrated; These are the model simulation values for daily average body weight and total biomass, respectively. These are the historical measured values of average daily body weight and total biomass, respectively. These are the weighting coefficients.
[0015] Furthermore, in step S5, the data assimilation method is ensemble Kalman filtering, specifically including: S51: Generate the initial set of states for the parameters to be calibrated. This represents the initial state of the i-th parameter to be calibrated; The number of members in the set of parameters to be calibrated; S54: Perform a prediction step derivation for each set member: This is the model's state transition function; The input vector; This is process noise; The set of parameters to be calibrated; S53: When new observation data arrives, calculate the Kalman gain and update each set member: Kalman gain; For observation vectors; To observe the noise, For observation functions.
[0016] Furthermore, in step S6, the setting of future prediction scenarios includes one or more of the following: a baseline prediction scenario, an optimal scenario, a risk scenario, and a management decision simulation scenario; wherein, The baseline prediction scenario is used to maintain the current environmental trend and feed according to plan; The optimal scenario is used to set environmental parameters to remain within an ideal range; The risk scenarios are designed to assume extreme weather or equipment failure. The management decision simulation scenario is used to set different oxygenation, water exchange, or feeding schemes.
[0017] Furthermore, in step S7, the prediction result includes: average daily weight. Curve, total biomass curve, survival rate Curves, environmental load early warning curves, as well as predicted final yield, predicted total number of harvested tails, predicted average size, and predicted feed coefficient.
[0018] The present invention has the following beneficial effects: Compared with existing technologies, this solution accurately characterizes the nonlinear coupling effects of three key environmental factors—water temperature, dissolved oxygen, and ammonia nitrogen—by constructing a comprehensive environmental load factor, thus overcoming the limitation of traditional linear models in handling interactions between factors. Simultaneously, by determining time-delay parameters through cross-correlation analysis and introducing the delayed effect of environmental stress on growth, it addresses the fundamental deficiency of traditional models that assume instantaneous response. Compared to traditional prediction models, the average relative error of yield prediction in this invention is significantly reduced, and prediction accuracy is significantly improved.
[0019] Compared to existing technologies, this solution quantifies the energy loss caused by environmental degradation through a constructed differential dynamics model. Combined with a time-lag mechanism, it can predict in advance the delayed effect of current environmental degradation leading to growth slowdown several days later. In events such as excessive ammonia nitrogen, compared to traditional threshold alarm systems, it can issue warnings 3-7 days in advance, giving farmers valuable intervention time and thus avoiding unnecessary yield losses.
[0020] Compared to existing technologies, this solution utilizes a genetic algorithm parameter calibration and online update mechanism, enabling the model to adapt to varying conditions across different aquaculture species, regions, and seasons. After completing a breeding cycle, the model automatically optimizes parameters using new data, gradually accumulating a parameter knowledge base for specific conditions, achieving "increasing accuracy with use." Furthermore, when applied across regions, data calibration can be achieved in a short time, reaching the accuracy of locally optimized models. Attached Figure Description
[0021] Figure 1 This is a flowchart illustrating a method for predicting the yield of farmed fish according to the present invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0023] See Figure 1 The flowchart shown is a method for predicting the yield of farmed fish according to the present invention. The following is a further explanation of the solution with reference to specific implementation examples.
[0024] Example 1: When the species of farmed fish in the aquaculture area is only one, the implementation process of the technical solution of this application specifically includes the following stages and steps: Phase 1: System Deployment and Data Acquisition Step 1: Setting up an IoT monitoring system Deploy the following monitoring equipment in the target fishponds: Water quality sensor array: Water temperature sensor, dissolved oxygen sensor and ammonia nitrogen sensor are installed at two depths of 50cm and 150cm below the water surface, respectively, and data is automatically collected every 30 minutes.
[0025] Intelligent feeder: Records the time and amount of each feeding, and the data is automatically uploaded to the cloud.
[0026] Oxygenation controller: Records the start-up and shutdown times and operating power of the oxygenation equipment.
[0027] Edge computing gateway: Deployed at the edge of the pond, responsible for collecting all sensor data and performing initial time alignment.
[0028] Step 2: Basic Data Entry Before starting the breeding process, the following fixed parameters are manually entered: Water volume V: Calculated by measuring the length, width and average depth of the fishpond.
[0029] Initial stocking count N0: Records the actual count of fish fry at the time of stocking.
[0030] Initial average weight W(0): 100 fish fry were randomly sampled and their weights were taken as the average.
[0031] Basic parameters for the species: Input prior knowledge such as the optimal water temperature range and low oxygen tolerance threshold of the cultured fish species.
[0032] Step 3: Periodic manual sampling Weekly sampling: At a fixed time each week (e.g., before feeding on Monday morning), randomly catch 30-50 fish using a net, weigh them, and record the average weight W. obs (t), which serves as the observation data for model calibration.
[0033] Manual verification of water quality: The sensor data is calibrated and verified once a week using a portable water quality analyzer.
[0034] Phase Two: Construction of Environmental Load Factors Step 4: Real-time calculation of environmental load factors The edge computing gateway performs the following calculations every 30 minutes: Step 4.1: Read the water temperature T(t), dissolved oxygen DO(t), and ammonia nitrogen concentration NH3(t) at the current time t.
[0035] Step 4.2: Calculate the water temperature influence function: Among them, the optimal water temperature and water temperature tolerance range Preset the water temperature based on the biological characteristics of the farmed species (e.g., the optimal water temperature for tilapia is 28℃, and its tolerance range is 5℃). Step 4.3: Calculate the dissolved oxygen influence function: The half-saturation constant is typically taken as 2.0 mg / L. Step 4.4: Calculate the ammonia nitrogen influence function: The ammonia nitrogen toxicity threshold is set according to the aquaculture species, and is usually 0.1 mg / L. Step 4.5: Calculate the comprehensive environmental load factor: This value ranges from 0 to 1, with the closer it is to 1 indicating a more suitable environment for fish growth.
[0036] Step 4.6: Store the calculated comprehensive environmental load factor in the historical database and maintain a sliding window of 30 days for subsequent time-delay analysis.
[0037] Phase 3: Time Delay Parameter Identification Step 5: Cross-correlation analysis to determine time delay After accumulating sufficient historical data on the 30th day of the breeding process, a time-delay parameter identification is performed. Specifically: Step 5.1: Extract the daily average environmental load sequence for the past 30 days. and daily average growth rate sequence The daily average growth rate is calculated by interpolation based on weekly sampled data.
[0038] Step 5.2: Calculate the cross-correlation function. In one possible implementation, the Pearson correlation coefficient can be used for calculation. Specifically: This represents the average daily environmental load over the past 30 days. This represents the average daily growth rate over the past 30 days. Step 5.3: Find the value of τ that maximizes R(τ), which is the time lag parameter. For example, if the calculation shows that the correlation coefficient is the largest when τ=3 days, it means that the current growth is most affected by the environmental conditions 3 days ago.
[0039] Step 5.4: Fix this τ value as a model parameter for subsequent differential equation solving. Repeat this analysis every 30 days to accommodate the time lag changes at different growth stages of the fish.
[0040] Phase 4: Model Parameter Calibration Step 6: Genetic Algorithm Parameter Estimation After a certain period of time (e.g., day 45) of aquaculture, and after obtaining at least three sampling data points, the first comprehensive parameter calibration is performed. Specifically: Step 6.1: Determine the set of parameters to be calibrated: Θ={r,W max ,θ,μ,α,m0,β,E threshold}; where r and W max θ, μ, and α are the core growth parameters, and m0, β, and E are the core growth parameters. threshold These are parameters related to mortality.
[0041] Step 6.2: Set the genetic algorithm parameters, such as: population size: 100; number of generations: 200; crossover probability: 0.8; mutation probability: 0.1; parameter value range: set the boundaries according to biological common sense (e.g., r∈[0.01,0.1], W... max ∈[500,5000] etc.).
[0042] Step 6.3: Define the fitness function: where the objective function is to minimize the weighted sum of squared errors between the simulated and measured values. .
[0043] Step 6.4: Run the genetic algorithm and iteratively optimize until the termination condition is met (reaching the maximum number of generations or fitness convergence).
[0044] Step 6.5: Output the optimal parameter set Store it in the model parameter library.
[0045] Phase 5: Data Assimilation and State Update Step 7: Initialize the ensemble Kalman filter Step 7.1: Set the number of members N in the set. e =100.
[0046] Step 7.2: Generate the initial state set. Using the current time t... k Best estimated state x k =[W k B k E k N k ] T Centered on a point, Gaussian noise is added to generate 100 perturbation states: ;in This is the process noise covariance matrix, set according to the sensor accuracy.
[0047] Implementation Step 8: Daily Forecast Step Perform a prediction once for each set member at 0:00 every day: Step 8.1: Read the average environmental data for the past 24 hours. .
[0048] Step 8.2: For each set member i, integrate the core differential equation using the fourth-order Runge-Kutta method for 24 hours to obtain the predicted state: .
[0049] Step 8.3: Add process noise This yields the final set of predicted states.
[0050] Step 9: Update steps weekly New observational data are obtained through manual sampling every Monday morning. Then, perform the update: Step 9.1: Calculate the ensemble mean of the predicted states. Covariance Matrix .
[0051] Step 9.2: Calculate the Kalman gain: H is the observation matrix (the mapping from state to observation), and R is the observation noise covariance matrix (set according to the weighing error).
[0052] Step 9.3: For each set member, generate observation noise perturbation. Then update: .
[0053] Step 9.4: The updated set mean is the optimal state estimate at the current moment. The dispersion of the set reflects the uncertainty of the state estimate.
[0054] Phase 6: Scenario Prediction Implementation Step 10: Set up forecast scenarios, including a baseline forecast scenario, an optimal scenario, a risk scenario, and a management decision simulation scenario. The baseline forecast scenario is used to maintain current environmental trends and feed according to plan; the optimal scenario is used to keep environmental parameters within ideal ranges; the risk scenario assumes extreme weather or equipment failure; and the management decision simulation scenario is used to set different aeration, water exchange, or feeding schemes. Select one or more forecast scenarios based on management needs, such as: Scenario A: Baseline Forecast (for routine management) Environmental variables: It is assumed that water temperature, dissolved oxygen, and ammonia nitrogen will maintain the average trend of the past 7 days over the next 30 days; Feeding plan: Continue to implement the current feeding plan; Forecasting objective: To understand the expected output under normal management conditions.
[0055] Scenario B: Optimal Scenario (for capacity assessment) Environmental variables: The water temperature is set to be constant at the optimal water temperature, dissolved oxygen is maintained above 6 mg / L, and ammonia nitrogen is maintained below 0.05 mg / L; Feeding plan: Implemented according to the theoretically optimal feeding rate; Forecast objective: To assess the maximum production potential under current fish population conditions.
[0056] Scenario C: Risk Scenario (for emergency response plans) Environmental variables: Assume that the temperature drops by 5°C in the next 3 days, or that the dissolved oxygen drops to 2 mg / L for 12 hours due to a malfunction of the oxygenation equipment; Feeding program: Automatically reduce feed intake by 50% based on environmental degradation; The purpose of the forecast is to assess the impact of extreme events on production and to develop contingency plans.
[0057] Scenario D: Management Decision Simulation (for Solution Optimization) Option 1: Increase the frequency of water changes, changing 10% of the water daily; Option 2: Open the micropore oxygenation system to maintain dissolved oxygen at a level above 5 mg / L; Option 3: Adjust the feeding strategy to small, frequent meals; Purpose of prediction: To compare the expected effects of different management measures and select the optimal solution.
[0058] Implementation Step 11: Numerical Integration with Time Delay For each selected scenario, perform the following integral calculation: Step 11.1: Set integration parameters The integration step size Δ = 0.1 days (2.4 hours); Total time for scoring: from the current moment to the expected harvest date (e.g., 60 days).
[0059] Step 11.2: Initialize the historical state queue Create a queue of length τ+10 and fill it with the environmental load values E(t) of the past τ+10 days; During the integration process, at each step forward, the new E(t) value is added to the queue, and the oldest value is removed.
[0060] Step 11.3: For each time step t n Perform the fourth-order Runge-Kutta integral: Here, f is the core differential equation function, which requires retrieving the environmental load value at the corresponding time from the historical queue during calculation.
[0061] Step 11.4: Record W(t) and B(t) for each time step to generate the predicted time series.
[0062] Phase 7: Outcome Output and Decision Support Implementation Step 12: Generate Prediction Results The system can automatically generate visual charts including: growth curve prediction, total biomass prediction, environmental load early warning, and survival rate prediction. For growth curve prediction: X-axis: Time (days) Y-axis: Average weight (g / tail) Plot three curves: baseline forecast, optimal scenario, and risk scenario. The measured points for weekly sampling are marked for comparison.
[0063] For total biomass prediction: X-axis: Time (days) Y-axis: Total biomass (kg) Use different colors to distinguish different scenarios Mark the target yield reference line.
[0064] For environmental load early warning: X-axis: Time (days) Y-axis: Environmental load E(t) Draw the warning threshold line E threshold (e.g., 0.6) When the predicted value E(t) is lower than the threshold, an early warning point is automatically marked.
[0065] For survival rate prediction: X-axis: Time (days) Y-axis: Survival rate (%) It demonstrates the potential risk of death due to environmental degradation.
[0066] Implementation Step 13: Uncertainty Quantification Step 13.1: Use the 100 set members updated in step 9 as the initial state.
[0067] Step 13.2: For each set member, repeat the numerical integration in step 11 to obtain 100 different predicted trajectories.
[0068] Step 13.3: For each future time point, calculate the 5%, 50% (median), and 95% quantiles of the 100 trajectories.
[0069] Step 13.4: Add shaded areas to the growth curve to represent the 90% confidence interval (between the 5% and 95% quantiles).
[0070] Implementation Step 14: Generate Management Decision Recommendations Based on the prediction results, the system automatically generates text-based management suggestions: Example output 1 (normal case), such as: Based on baseline forecasts, the average weight of the fish is expected to reach 650g / fish in 60 days, with a total yield of 2850kg. The current environmental load is stable at 0.85, and it is recommended to maintain the existing management practices. Expected harvest date: June 15, 2025.
[0071] Example output 2 (warning situation), such as: Forecasts indicate that the environmental load will continue to decline over the next 10 days, potentially dropping to 0.55 on day 7, below the warning threshold of 0.6. Recommendations: ① Increase water exchange frequency to 15% per day starting from day 3; ② Prepare emergency aeration equipment in advance; ③ Automatically adjust feeding amounts based on the degree of environmental degradation. Without intervention, a 12% yield loss is expected.
[0072] Example output 3 (decision comparison), such as: Comparing three management plans: Plan A (increased water exchange) is expected to yield 2980 kg; Plan B (enhanced aeration) is expected to yield 3050 kg; and Plan C (optimized feeding) is expected to yield 2920 kg. Plan B is recommended as it offers the highest expected return, a narrower confidence interval, and lower risk.
[0073] Phase 8: Iterative Model Optimization Implementation Step 15: Harvest Verification After the breeding cycle is completed, perform the following verification steps: Step 15.1: Record actual harvest data: Actual total output B actual ; Actual average specification W actual ; Actual number of surviving tails N actual ; Actual total feeding amount ∑F(t).
[0074] Step 15.2: Calculate the prediction error: Relative error in production forecast: |B pred -B actual | / B actual ×100%; Specification prediction relative error: |W pred -W actual | / W actual ×100%; Feed coefficient prediction error.
[0075] Step 15.3: Error Source Analysis: Analyze the main reasons why the error is greater than 10%; Record unusual events (such as typhoons and disease outbreaks) for subsequent model improvements.
[0076] Implementation Step 16: Online Parameter Update Step 16.1: Use the complete data of the entire breeding cycle (including environmental time-series data, feeding records, weekly sampling, and final harvest) as a new training sample.
[0077] Step 16.2: Re-execute the genetic algorithm parameter calibration in step 6, but use the optimized parameters from the previous cycle for the initial population with a small perturbation.
[0078] Step 16.3: Obtain the updated parameter set Replace the original parameters.
[0079] Step 16.4: Link and store the new parameters with information such as aquaculture species, season, and region to form a parameter knowledge base. In future aquaculture cycles under the same conditions, similar parameters can be directly called as initial values, shortening the calibration time.
[0080] Example 2: When there are more than one species of farmed fish in the aquaculture area, the implementation process of the technical solution in this application specifically includes the following stages and steps: Phase 1 Adjustment: Data Acquisition (corresponding to steps 1-3 in Example 1) It should be noted that: Adjustment to step 1: Multi-variety data collection With the IoT monitoring system remaining unchanged, the manual sampling process is adjusted as follows: Species-specific sampling: During weekly sampling, different species are randomly caught and weighed. For example, if grass carp, silver carp, and crucian carp are raised together, the results are recorded separately as follows: ; Counting by Variety: Record the number of surviving tails for each variety: ; Total biomass calculation: Sum of biomass for each variety: , where m is the number of species (in this example, grass carp, silver carp and crucian carp are mixed, so m=3).
[0081] Adjustments to step 2: Basic data entry Enter the initial number of animals released for each species. and initial average weight ; Enter the basic biological parameters (optimal water temperature, low oxygen tolerance threshold, etc.) for each species. These parameters may vary.
[0082] Second-stage adjustment: Environmental load factor (corresponding to step 4 in Example 1) Adjustment to step 4: Environmental load calculation by product type The environmental factor E(t) remains a unified water quality environmental indicator, but because different species respond differently to environmental stress, it is necessary to calculate the environmental load perceived by each species separately: The parameters of each influencing function (such as optimal water temperature and ammonia nitrogen threshold) are set separately according to the biological characteristics of different species. For example: Grass carp (bottom-dwelling fish): has a relatively weak tolerance to low oxygen levels. The value is relatively large; Silver carp (floating fish): has a relatively strong tolerance to low oxygen levels. The value is relatively small.
[0083] Third stage adjustment: Time delay parameters (corresponding to step 5 in Example 1) Adjustment to step 5: Variety-specific time lag identification Cross-correlation analysis was performed on different varieties separately: ; Obtain the time delay parameters for each variety. .For example: Grass carp have a large appetite, a fast metabolism, and respond quickly to environmental changes, so they are set... sky; Silver carp filter-feed on plankton, so their response is relatively slow; therefore, set... sky.
[0084] Fourth stage adjustment: Model parameter calibration (corresponding to step 6 in Example 1) Adjustments to step 6: Variety-specific parameter matrix The original parameter set Θ is expanded into a parameter matrix: ; The objective function is adjusted to: Note the total biomass B sim (t) is the sum of biomass of all varieties, compared with the measured total biomass.
[0085] Adjustments to the genetic algorithm: Parameters for each variety can be optimized individually or in combination; Considering the interactions between varieties (such as competition), it may be necessary to add cross-coupling parameters.
[0086] Fifth stage adjustment: Data assimilation (corresponding to steps 7-9 in Example 1) Adjustment to step 7: State vector expansion Original state vector x k =[W k B k E k N k ] T Expanded to: ; Number of members in the set N e It remains unchanged, but the dimension of each set member increases.
[0087] Adjustment to step 8: Variety-specific forecasting The core differential equation is expanded into a system of equations: For each variety j: ; It should be noted that E(t) in the population interaction consumption term takes the minimum environmental load of all varieties, indicating that the most sensitive variety determines the overall interaction intensity. The total biomass in the denominator is the sum of all varieties, reflecting resource competition.
[0088] Adjustment to step 9: Multi-varietal observation update The observation vector is expanded to: ; The observation matrix H is expanded accordingly, the Kalman filter update formula remains unchanged, but it handles a higher-dimensional state space.
[0089] Sixth stage adjustment: Scenario prediction (corresponding to steps 10-11 in Example 1) Adjustments to step 10: Setting up multiple product scenarios The definitions for each scenario remain unchanged, but the interactions between varieties must be considered: Benchmark forecast: Forecast separately based on the current trend of each variety; Optimal Scenario: Environmental parameters for each species are set to their respective optimum values. However, in actual water bodies, it is impossible to simultaneously meet the optimum conditions for all species, necessitating a compromise. For example: That is, the target water temperature is set according to the biomass weighted average.
[0090] Risk scenarios: Assess the differences in the impact of extreme events on different varieties, for example: Hypoxia events: Varieties with weak tolerance to low oxygen levels have a higher mortality rate; Excessive ammonia nitrogen levels: Sensitive varieties experience more severe growth stagnation.
[0091] Adjustment to step 11: Numerical integration of the system of equations For each time step, m differential equations need to be solved simultaneously: ; The fourth-order Runge-Kutta method in vector form is used to update the state of all varieties simultaneously at each time step.
[0092] The historical state queue needs to store the data for each variety. Because the time delay parameters are different, they need to be maintained separately.
[0093] Seventh stage adjustment: Result output (corresponding to steps 12-14 in Example 1) Adjustment to step 12: Visualization of results by variety Generate the following chart: Variety-specific growth curves: The weight prediction curves for each variety are plotted on the same coordinate system and distinguished by different colors; Total biomass composition diagram: The stacked area diagram shows the contribution of each variety to the total yield; Species-specific environmental load map: The environmental load curves perceived by each species can be used to compare the stress levels of different species under the same water quality conditions; Variety Proportion Change Chart: Shows the time evolution of the biomass proportion of each variety in a mixed-culture structure.
[0094] Adjustment to step 13: Quantification of multi-product uncertainty Confidence intervals were calculated for each variety, and the covariance between varieties was considered in the total biomass prediction.
[0095] Adjustments to step 14: Recommendations for multi-variety management Management recommendations considering species characteristics are generated, such as: In the current polyculture structure, grass carp are growing normally, but silver carp are greatly affected by ammonia nitrogen (E value 0.65). Recommendations: ① Increase water exchange to reduce ammonia nitrogen; ② Appropriately increase the amount of silver carp-specific floating feed; ③ It is estimated that after 60 days, grass carp will reach 800g, silver carp will reach 500g, and the total yield will be 3.2 tons, of which grass carp will account for 60%, silver carp will account for 30%, and crucian carp will account for 10%.
[0096] Eighth stage adjustment: Iterative optimization of the model (corresponding to steps 15-16 in Example 1) Adjustment to step 15: Harvest verification by variety Record the actual harvest data for each variety separately and calculate the prediction error for each variety.
[0097] Regarding the adjustment in step 16: update the parameters by variety. By utilizing complete growth data for each variety, parameters for each variety were optimized. Simultaneously, through the accumulation of data over multiple periods, interaction coefficients between varieties could be identified, further optimizing the expression of population interaction terms.
[0098] 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 illustrative of the 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 present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A method for predicting the yield of farmed fish, characterized in that, Includes the following steps: S1: Collect multi-source time-series data of water bodies in the aquaculture area and preprocess it. The multi-source time-series data includes water quality environment data, aquaculture operation data, biological growth data and spatial parameter data. S2: The water quality environmental data is coupled using a nonlinear coupling function to construct a comprehensive environmental load factor at time t. ; S3: Utilizing the aforementioned comprehensive environmental load factor Establish a differential dynamics model to describe the growth process of fish; S4: Based on historical aquaculture data, an optimization algorithm is used to calibrate the unknown parameters in the differential dynamics model; S5: Using real-time acquired data, the state of the differential dynamics model is dynamically updated through a data assimilation method; S6: Set up a future prediction scenario, and use the fourth-order Runge-Kutta method to numerically integrate the differential dynamics model to predict the average body weight and total biomass of fish at future times. S7: Output the prediction results and generate aquaculture management suggestions.
2. The method for predicting the yield of farmed fish according to claim 1, characterized in that, In step S1, The water quality environmental data includes the water temperature at time t. Dissolved oxygen ammonia nitrogen concentration ; The aquaculture operation data includes daily feed intake. Aerator start-up and shutdown time, water exchange volume; The biological growth data includes the average body weight of the sample. Sampling tails; The spatial parameter data includes water volume and initial stocking number of fish. .
3. The method for predicting the yield of farmed fish according to claim 1, characterized in that, In step S2, the comprehensive environmental load factor Defined as: ; in, Let the water temperature influence function be defined as: For optimal water temperature, Water temperature tolerance range; The dissolved oxygen effect function is defined as follows: It is the half-saturation constant; The ammonia nitrogen influence function is defined as follows: This represents the ammonia nitrogen toxicity threshold.
4. The method for predicting the yield of farmed fish according to claim 1, characterized in that, In step S3, the differential dynamics model is defined as: ,in, Let t be the average body weight of the fish at time t; V represents the total biomass at time t; V represents the water volume. The intrinsic growth rate; This is the theoretical maximum specification; These are parameters related to allometric growth. For time delay parameters; α is the stress consumption coefficient; α is the group interaction loss coefficient; wherein, the time delay parameter It is determined using the following methods: Calculate the comprehensive factor of historical environmental load Fish body growth rate sequence Correlation coefficients at different lag times; The lag time corresponding to the maximum correlation coefficient is selected as the time lag parameter. The value of .
5. The method for predicting the yield of farmed fish according to claim 1, characterized in that, In step S4, the optimization algorithm is a genetic algorithm or a particle swarm optimization algorithm; The objective function for calibrating the unknown parameters in the differential dynamics model is to minimize the weighted sum of squared errors between the simulated and measured values: ,in, The set of parameters to be calibrated; These are the model simulation values for daily average body weight and total biomass, respectively. These are the historical measured values of average daily body weight and total biomass, respectively. These are the weighting coefficients.
6. The method for predicting the yield of farmed fish according to claim 1, characterized in that, In step S5, the data assimilation method is ensemble Kalman filtering, specifically including: S51: Generate the initial set of states for the parameters to be calibrated. This represents the initial state of the i-th parameter to be calibrated; The number of members in the set of parameters to be calibrated; S54: Perform a prediction step derivation for each set member: This is the model's state transition function; The input vector; This is process noise; The set of parameters to be calibrated; S53: When new observation data arrives, calculate the Kalman gain and update each set member: Kalman gain; For observation vectors; To observe the noise, For observation functions.
7. The method for predicting the yield of farmed fish according to claim 1, characterized in that, In step S6, the setting of future prediction scenarios includes one or more of the following: a baseline prediction scenario, an optimal scenario, a risk scenario, and a management decision simulation scenario; wherein, The baseline prediction scenario is used to maintain the current environmental trend and feed according to plan; The optimal scenario is used to set environmental parameters to remain within an ideal range; The risk scenarios are designed to assume extreme weather or equipment failure. The management decision simulation scenario is used to set different oxygenation, water exchange, or feeding schemes.
8. The method for predicting the yield of farmed fish according to claim 1, characterized in that, In step S7, the prediction results include: average daily weight Curve, total biomass curve, survival rate Curves, environmental load early warning curves, as well as predicted final yield, predicted total number of harvested tails, predicted average size, and predicted feed coefficient.
9. The method for predicting the yield of farmed fish according to claim 1, characterized in that, Also includes: The uncertainty of the prediction results is quantified, and the confidence interval of the prediction results is calculated by ensemble forecasting method. Specifically, this includes: making future predictions using the ensemble members after data assimilation, calculating the 5%, 50% and 95% quantiles of the prediction results of each member, and generating prediction results with confidence intervals.
10. The method for predicting the yield of farmed fish according to claim 1, characterized in that, Also includes: After each breeding cycle, the actual harvest data is compared with the model prediction results to calculate the prediction error. The model parameters are then updated and optimized online using newly obtained complete breeding cycle data to improve the model's prediction accuracy for future breeding cycles.