A smart agriculture fusion system based on internet of things and digital twinning

Abstract

This invention discloses a smart agriculture integrated system based on the Internet of Things (IoT) and digital twins, relating to the field of information processing technology for agricultural production, operation, and management. The system includes a data acquisition module, a state assimilation module, a scenario generation module, a strategy valuation module, and a decision output module. By combining real-time IoT-sensed data with a crop digital twin model, it generates predicted quality values ​​and commodity value depreciation indices for future harvest times. Based on a market price fluctuation model, it generates transaction price paths, forming a set of simulated business profit and loss flows. A reverse-engineering method is used to calculate the commercial option value of the "continue pruning" strategy, which is compared with the definite net profit of immediate harvesting, and harvesting decision instructions are automatically output. This invention achieves economic optimization decision-making for the harvesting timing of facility crops, solving the problem of difficulty in quantifying and weighing the economic benefits of two strategies within the harvestable window.

Description

Technical Field

[0001] This invention relates to the field of information processing technology for agricultural production, operation and management; more specifically, it relates to a smart agriculture integrated system based on the Internet of Things and digital twins. Background Technology

[0002] Once high-value greenhouse crops enter their harvest window, fruit quality dynamically changes with weather conditions, and wholesale market prices fluctuate daily. Growers face the choice of whether to harvest each day. Early harvesting can lock in current profits but may miss out on the premium brought by improved quality. Late harvesting may yield higher quality and unit price, but faces the risk of market price decline and fruit overripe loss.

[0003] Currently, management methods in facility agriculture primarily rely on growers' experience or fixed harvest calendars for decision-making, lacking comprehensive quantitative analysis tools for crop quality changes and market price fluctuations. Some smart agriculture systems monitor environmental data and crop growth status through IoT sensors and use digital twin models to predict suitable harvest dates. These systems only provide harvesting suggestions from a crop growth perspective, without incorporating market price fluctuations into the decision-making process. Growers still need to manually weigh the agronomic advice provided by the system against their own subjective market judgment. Furthermore, existing systems lack quantitative assessment methods for the risk of overripe fruit associated with continued pruning, failing to automatically reflect potential economic losses due to overripe or cracked fruit in expected return calculations. These shortcomings mean that harvesting decisions within the harvestable window still rely on experience, making it difficult to reliably guarantee economic benefits. Therefore, a smart agriculture integrated system based on the Internet of Things and digital twins is proposed to address the above problems. Summary of the Invention

[0004] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide a smart agriculture integration system based on the Internet of Things and digital twins. The problem to be solved is that, within the harvestable window of facility crops, the prior art cannot simultaneously combine dynamic changes in crop quality and random fluctuations in market prices to automatically generate economically optimal harvesting decisions.

[0005] To achieve the above objectives, the present invention provides the following technical solution: This invention proposes a smart agriculture integration system based on the Internet of Things and digital twins. At each harvest decision point, the system automatically calculates the certain net income of the immediate harvest strategy and the commercial option value of the continue-on-the-branch strategy. By comparing the two, it outputs a harvest decision instruction, thereby solving the problem that growers of greenhouse crops find it difficult to quantify and weigh the economic benefits of the two strategies during the harvest window.

[0006] The system consists of a data acquisition module, a state assimilation module, a scenario generation module, a strategy evaluation module, and a decision output module.

[0007] The data acquisition module acquires the current marketability data of crops within the facility, the current transaction price data of the wholesale market, and the weather forecast data for a preset period in the future. It then sends the marketability data and the current transaction price data to the status assimilation module and the weather forecast data to the scenario generation module.

[0008] The state assimilation module receives commodity state data and generates multiple probabilistic state samples representing the current quality state of the crop, thus reflecting the uncertainty of the crop's current state. Simultaneously, this module receives current transaction price data, updates the parameters of the market price fluctuation model, and obtains a market fluctuation parameter set to capture the dynamic characteristics of market price changes.

[0009] The state assimilation module sends the probability state samples and market fluctuation parameter set to the scenario generation module.

[0010] Furthermore, the state assimilation module includes a crop state assimilation unit and a market parameter update unit. The crop state assimilation unit uses marketability state data as an observation vector, employs ensemble Kalman filtering to update the state variables of the crop digital twin model, and outputs a set of posterior state particles as probabilistic state samples. By fusing real-time IoT observation data to correct the model state, the crop state estimation can remain synchronized with the actual field conditions.

[0011] The market parameter update unit uses the daily transaction price in the current transaction price data as the observation value, and adopts the Bayesian online update algorithm to correct the posterior distribution of the parameters of the market price fluctuation model, thereby obtaining the market fluctuation parameter set, so that the price model can adapt to the latest market conditions.

[0012] The crop digital twin model includes state variables and state transition equations. State variables include fruit sugar content, fruit firmness, and accumulated temperature. The state transition equations describe the diurnal variation of state variables driven by temperature, light, and water stress indices. By simulating the crop growth and quality formation process, it provides a basis for quality prediction at various future time points for harvesting decisions.

[0013] The scenario generation module uses probabilistic state samples as the initial state of the crop digital twin model, and combines them with meteorological forecast data to generate predicted values ​​of quality indicators and commodity value depreciation index at each future harvest decision point.

[0014] Among them, the commodity value loss index is used to quantify the potential economic losses caused by overripe or cracked fruit. Its introduction enables the system to automatically include potential future fruit losses in economic accounting when evaluating the strategy of continuing to hang the fruit on the branch.

[0015] This module also generates multiple transaction price paths at each future harvest decision point based on the market fluctuation parameter set, pairs the predicted value of the quality indicator with the transaction price path, and calculates the expected net income of each paired path at each harvest decision point according to the mapping rule between quality and transaction price, forming a set of business profit and loss simulation streams, which is then sent to the strategy valuation module.

[0016] Furthermore, the commodity value depreciation index is arrive The values ​​between these ranges are calculated by the scenario generation module based on temperature and humidity data from future weather forecasts and the current accumulated temperature. This index represents the expected degree of decline in the proportion of marketable fruit due to overripeness or cracking; the closer the value is to the expected value, the higher the probability of a decrease. This indicates that the higher the risk of over-ripening and the greater the expected loss, the value loss caused by the over-ripening risk will be automatically deducted from the expected return of the continued hanging branch strategy.

[0017] The commodity value depreciation index is calculated in the following way: The high-temperature stress factor is determined based on the highest daily temperature in the future, reflecting the accelerating effect of high temperature on fruit overripening; The high humidity stress factor is determined based on the future daily average humidity, reflecting the inducing effect of high humidity environment on fruit cracking and disease; The larger value between the high temperature stress factor and the high humidity stress factor is taken to reflect the actual situation that either high temperature or high humidity can lead to severe over-ripening loss. The commodity value depreciation index is obtained by multiplying the larger value by the normalized cumulative temperature value, where the normalized cumulative temperature value is the ratio of the current cumulative temperature to the total cumulative temperature required for physiological over-ripening of the variety.

[0018] As crops gradually approach their physiological overripe limit, the normalized value of accumulated temperature gradually increases, and the commodity value depreciation index also increases accordingly.

[0019] The mapping rule between quality and transaction price consists of two parts: First, the relationship between the quality grades classified by fruit sugar content or fruit firmness and the premium of wholesale prices is used to transform the physical quality indicators of crops into price advantages in market transactions. Second, the adjustment rule of lowering the average transaction price based on the commodity value depreciation index reflects the economic loss caused by the decline in the marketability of the fruit due to overripeness in the transaction price.

[0020] The adjustment rule specifically involves multiplying the average transaction price by an adjustment factor, where the adjustment factor is... Subtracting the commodity value depreciation index, the adjustment factor is smaller when the over-maturity risk is higher, and the expected transaction price is lower. This automatically reflects the negative impact of over-maturity risk on returns in the profit and loss calculation.

[0021] The strategy valuation module receives a set of simulated business profit and loss streams and uses a reverse solution method to calculate the business option value of the continued hanging strategy, thereby quantifying the expected economic return of continuing to wait.

[0022] Simultaneously, the module calculates the certain net profit of the immediate harvest strategy based on the current quality status and current transaction price data, thereby quantifying the certain economic benefit of immediate harvest. The module then sends the commercial option value and the certain net profit to the decision output module.

[0023] Furthermore, when calculating the value of commercial options, the strategy valuation module treats the continuing to wait strategy as an American option and uses the least squares Monte Carlo method to solve backwards from the cutoff date of the harvestable window to the current decision date. By simulating the optimal execution time under various possible future scenarios, the economic value of the flexible decision-making right to continue waiting is quantified.

[0024] When performing regression calculations during the reverse solution process, the commodity value depreciation index is used. Predicted values ​​of quality indicators and their product As independent variables of the regression basis function.

[0025] The product term is introduced as an interaction term, which allows the estimate of continued holding value to reflect the combined impact of overripe risk and fruit quality on expected returns.

[0026] The decision output module compares the commercial option value with the determined net profit. When the determined net profit is greater than or equal to the commercial option value, a harvest execution instruction is generated; when the determined net profit is less than the commercial option value, an instruction to postpone harvesting and proceed to the next decision cycle is generated. In this way, the system automatically makes a harvesting decision that maximizes profit at each decision point.

[0027] Commodity status data includes environmental parameters and crop physiological parameters collected by IoT sensing devices installed within the facility.

[0028] Environmental parameters include air temperature, air humidity, light intensity, and CO2. At least one of the concentrations reflects the impact of climatic conditions within the facility on crop growth.

[0029] Crop physiological parameters include at least one of stem microstrain, leaf temperature, and canopy spectral reflectance, reflecting the crop's own water status and nutrient level.

[0030] Transaction price data is obtained through the wholesale market data interface, and weather forecast data is obtained through the external weather service interface.

[0031] The technical effects and advantages of this invention are as follows: (1) This system collects commodity status data of crops in the facility, transaction price data of the wholesale market, and weather forecast data through the data acquisition module. The state assimilation module uses ensemble Kalman filtering to fuse real-time observation data, generating multiple probabilistic state samples that reflect the uncertainty of the current quality status of crops, and uses a Bayesian online update algorithm to correct the posterior distribution of parameters of the market price fluctuation model, so that the price model can adapt to the latest market conditions. The scenario generation module uses the probabilistic state samples as the initial state of the crop digital twin model, and extrapolates the predicted values ​​of quality indicators and commodity value depreciation index at each future harvest time. Combined with the transaction price path generated by the Monte Carlo method, the system calculates the expected net income of each path at each time point according to the mapping rule between quality and transaction price, forming a set of simulated commercial profit and loss flows. The strategy valuation module calculates the commercial option value of the continue-hanging strategy and the certain net income of the immediate harvest strategy, respectively. The decision output module compares the results and automatically outputs the harvest instruction. Through the above process, the system integrates the crop growth status and market conditions into economic accounting at each decision time point and automatically provides a harvest suggestion that maximizes profits.

[0032] (2) The system introduces a commodity value depreciation index to quantify the potential economic losses caused by overripe or cracked fruit. This index is obtained by multiplying the larger of the high-temperature stress factor determined by the future daily maximum temperature and the high-humidity stress factor determined by the future daily average humidity, with the normalized value of accumulated temperature. As the crop gradually approaches the physiological overripe limit, the normalized value of accumulated temperature gradually increases, and the depreciation index increases accordingly. The mapping rule between quality and transaction price converts physical quality indicators such as fruit sugar content or fruit firmness into market price premiums, while reflecting the depreciation impact of the commodity value depreciation index on the average transaction price through adjustment factors. The above design enables the system to automatically include the value loss caused by overripe risk in the expected return calculation when evaluating the strategy of continuing to hang the fruit on the branch, comprehensively considering the potential returns of quality improvement and the potential losses of overripe risk.

[0033] (3) When calculating the value of commercial options, the system treats the continued holding strategy as an American option and uses the least squares Monte Carlo method to solve it backward from the harvestable window deadline to the current decision date. By simulating the optimal execution time under various possible future scenarios, the system quantifies the economic value of the flexible decision-making right to continue waiting. In the regression calculation, the commodity value depreciation index, the predicted value of the quality index, and their product are used as independent variables of the regression basis function. The product term is introduced as an interaction term, so that the estimation of the continued holding value can reflect the comprehensive impact of overripe risk and fruit quality on expected returns. This method combines financial option valuation theory with agricultural digital twin technology, providing a quantitative economic analysis tool for harvesting decisions of facility crops. Attached Figure Description

[0034] Figure 1 This is a schematic diagram of the system module composition of the present invention; Figure 2 This is a flowchart illustrating the overall implementation of the system according to the present invention; Figure 3 This is a flowchart of the overall decision-making cycle of the present invention. Detailed Implementation

[0035] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] Example 1 As attached Figures 1 to 3 The system shown is a smart agriculture integration system based on the Internet of Things and digital twins. At each harvest decision point, the system automatically calculates the certain net income of the immediate harvest strategy and the commercial option value of the continue-on-the-branch strategy. By comparing the two, it outputs a harvest decision instruction to solve the problem that growers of greenhouse crops find it difficult to quantify and weigh the economic benefits of the two strategies during the harvest window.

[0037] This system includes a data acquisition module, a state assimilation module, a scenario generation module, a strategy valuation module, and a decision output module. The data acquisition module obtains three types of data from external data sources, sends commodity status data and current transaction price data to the state assimilation module, and sends weather forecast data to the scenario generation module.

[0038] After receiving the data, the state assimilation module generates a set of probabilistic state samples and a set of market fluctuation parameters through an assimilation algorithm, and sends them to the scenario generation module.

[0039] The scenario generation module takes the probability state sample set and the market fluctuation parameter set as input, and sequentially performs quality path deduction, price path generation and profit and loss path synthesis to generate a set of business profit and loss simulation streams, which are then sent to the strategy valuation module.

[0040] The strategy valuation module calculates and determines the net income and commercial option value based on a set of simulated business profit and loss flows, and sends both to the decision output module. The decision output module compares the two values ​​and generates a harvesting decision instruction.

[0041] The system automatically executes the above process at the daily decision-making time. When a temporary harvesting instruction is generated, it enters the next decision cycle to continue execution, forming a rolling decision-making closed loop.

[0042] Data acquisition module: The data acquisition module is responsible for acquiring the three types of data required for system operation from external data sources.

[0043] The first category of data is commodity status data, specifically including environmental parameters and crop physiological parameters. Environmental parameters are collected by IoT sensing devices installed within the facility, including air temperature, air humidity, and light intensity. Crop physiological parameters include stem microstrain and leaf temperature. Stem microstrain is collected by microstrain sensors installed on the main stem of the crop, and leaf temperature is collected by infrared temperature sensors aligned with the plant canopy.

[0044] Each sensor operates at a frequency of once every 10 minutes. The data acquisition module takes the arithmetic mean of the six data points collected within each hour as the hourly value. If a sensor experiences data loss within an hour, the hourly value from the previous hour is used to fill the gaps. If data loss occurs for several consecutive hours, the most recent valid hourly value is used until data acquisition resumes normally.

[0045] As an alternative implementation, when the sensor system installed in the facility is more comprehensive, the environmental parameters may also include Concentration, crop physiological parameters can also include canopy spectral reflectance acquired by multispectral sensor scanning. In this case, the data acquisition module uses red band reflectance... and near-infrared reflectivity Calculate the normalized vegetation index This index, as a supplementary dimension to the observation vector, is sent to the state assimilation module along with the commodity status data.

[0046] The second type of data is the current transaction price data, which is obtained by the data acquisition module through the wholesale market data interface, including the average transaction price of the day and the price difference information of different quality grades.

[0047] The third type of data is meteorological forecast data, which is obtained by the data acquisition module through an external meteorological service interface. This includes hourly temperature, humidity and solar radiation forecasts for the next 72 hours, as well as forecasts for the township-level grid points where the facility is located.

[0048] After completing data collection and aggregation, the data acquisition module sends commodity status data and current transaction price data to the status assimilation module, and sends weather forecast data to the scenario generation module.

[0049] State assimilation module: The state assimilation module receives commodity status data and current transaction price data from the data acquisition module. This module includes a crop state assimilation unit and a market parameter update unit, and a crop digital twin model is pre-built within the module.

[0050] Crop digital twin models are the foundation for quality path extrapolation, and they include crop quality state variables and state transition equations. State variables include fruit sugar content. Fruit firmness and accumulated temperature The state transition equation describes the diurnal variation of state variables under environmental driving forces, including temperature. ,illumination and water stress index The state transition equation is in the form of:

[0051]

[0052]

[0053] In the above equation, , , , , For model parameters, The baseline temperature for crop development is set at 10℃.

[0054] It is a temperature response function. When the temperature is in the range of 20℃ to 28℃, The value is 1; when the temperature is below 20℃, The temperature decreases linearly from 1 to 0 at 10℃; when the temperature is above 28℃... The temperature decreases linearly from 1 to 0 at 35°C. When the temperature is below 10°C or above 35°C... Keep it at 0.

[0055] This is the water stress response function. When When it exceeds 60% of field capacity, Take 1; when When it is between 40% and 60% Decrease linearly to 0; when When it is below 40%, Keep it at 0.

[0056] It is calculated from air humidity. Specifically, the conversion method is as follows: when the humidity is above 70%... Take 1, when the humidity is below 40%. Take 0, when the humidity is between 40% and 70%. It is obtained through linear interpolation.

[0057] Model parameters were determined by fitting historical planting data of the same variety. During fitting, historical data from at least one complete growing season of the same variety were selected, using daily measured data. , , The values ​​were used as the fitting target, calculated from the daily average temperature, total daily sunshine, and daily average air humidity of the corresponding date. As input, the least squares method is used to solve for the optimal values ​​of each parameter. The entire fitting process is completed offline before system deployment, and the fitted parameter values ​​are pre-set in the system as variety parameters.

[0058] As an alternative implementation, when historical planting data is insufficient to accurately fit the mechanistic model parameters, the state transition equation can employ a data-driven model based on a long short-term memory network. This model takes the environmental data sequence and state variable sequence of the past 7 consecutive days as input and outputs the predicted values ​​of the state variables for the next day. After training, the model is loaded into the crop state assimilation unit in the form of a weight file.

[0059] The crop state assimilation unit receives hourly values ​​of marketability state data and organizes them into an observation vector. The observation vector consists of hourly average values ​​of stem microstrain and leaf temperature. When the marketability state data also includes canopy spectral reflectance data, the observation vector also includes the normalized vegetation index (NVI). .

[0060] The crop state assimilation unit is based on the crop digital twin model prediction at the current time. , , The value is used as the background field to generate a background set containing 50 state particles. Each particle interacts with the background field... , , Gaussian perturbations are applied to obtain the standard deviation of the perturbation, which is 5% of the current value of each state variable. If the current value of a state variable is 0, then 5% of the historical mean of that variable is taken as the standard deviation of the perturbation.

[0061] After generating the background set, the crop state assimilation unit uses ensemble Kalman filtering for data assimilation.

[0062] The first step is to calculate the 3×3 covariance matrix among the 50 particles. This matrix reflects the correlation between the state variables and the uncertainty of the background field.

[0063] The second step involves mapping the state variables of each particle to predicted observations using the observation operator matrix. The observation operator matrix is ​​determined by the linear regression coefficients between state variables and observations in historical data, and is calibrated once before the system runs. Specifically, it is constructed by using multiple sets of historical data... , , The values ​​are independent variables, and the corresponding observations such as stem microstrain and leaf temperature are dependent variables. Multiple linear regression is performed, and the coefficients obtained from the regression constitute the rows of the observation operator matrix.

[0064] The third step is to calculate the Kalman gain matrix. The Kalman gain is determined by the background covariance matrix, the observation operator matrix, and the observation error covariance matrix. The observation error covariance matrix is ​​determined based on the statistical results of sensor measurement errors and is in the form of a diagonal matrix, where each element on the diagonal represents the measurement variance of the corresponding observation.

[0065] The fourth step involves multiplying the Kalman gain by the difference between the actual observed vector and the predicted observed vector to obtain the state correction. This correction is then added to the state of each particle. , , In terms of value, the output is the set of posterior state particles. This set of posterior state particles is the set of probabilistic state samples.

[0066] The market parameter update unit internally maintains a market price fluctuation model. This model employs a mean regression model, and its parameters include long-term mean parameters. Mean regression speed parameter and volatility parameters The discretized recursive form of the model is:

[0067] in It is the time step. It is a random variable that follows a standard normal distribution.

[0068] The market parameter update unit extracts the average transaction price of the day from the current transaction price data, uses it as the latest observation, and inputs it into the Bayesian online update algorithm to correct the posterior distribution of the model parameters.

[0069] The first step is to... (The sentence is incomplete and ends abruptly.) , , The posterior distribution is used as the prior distribution for this update. This prior distribution is described by the posterior mean and the posterior covariance matrix.

[0070] The second step is to calculate the price forecast based on the prior mean. (Forecast value) ,in , It refers to the average transaction price on the day of the previous decision.

[0071] The third step is to calculate the prediction error. Prediction error = average transaction price of the day - predicted value.

[0072] The fourth step involves using the prediction error combined with the Kalman gain to... , , The posterior mean is corrected, and the posterior covariance matrix is ​​updated simultaneously. The Kalman gain is calculated from the prior covariance matrix and the observation noise variance, where the observation noise variance is taken as the variance of the historical prediction error.

[0073] After the update is complete, the Bayesian online update algorithm outputs a set of market fluctuation parameters, which includes... , , The posterior mean and posterior covariance matrix.

[0074] After completing the above processing, the state assimilation module sends the set of probability state samples and the set of market fluctuation parameters to the scenario generation module.

[0075] Scenario generation module: The scenario generation module receives the set of probability state samples and market fluctuation parameters sent by the state assimilation module, as well as the weather forecast data sent by the data acquisition module. After receiving the data, the module sequentially performs quality path deduction, price path generation, and profit and loss path synthesis.

[0076] The quality path deduction is carried out according to the following process.

[0077] Each state particle in the probability state sample set is used as the initial state of the crop digital twin model, and the corresponding state of that particle is taken. , , The value is used as the initial value.

[0078] The hourly temperature, humidity, and solar radiation values ​​from the weather forecast data are aggregated daily. Specifically, temperature is calculated as the arithmetic mean of 24-hour values, solar radiation is calculated as the sum of 24-hour values, and humidity is calculated as the arithmetic mean of 24-hour values.

[0079] The crop digital twin model is recursively pushed forward in 1-day increments. Each recursion uses the temperature obtained from the aggregation of that day. Solar radiation and the moisture stress index converted from humidity. Substitute the equations into the state transition equations. Use Euler's integral method to calculate the next day's... , , The value is set to 0.1d for the integration step size, meaning that the daily recursion consists of 10 integration steps.

[0080] The recursion begins on the current day and continues until the harvestable window expires. The harvestable window expires as follows: during the recursion process, when a certain day's... The value first reaches or exceeds the total accumulated temperature required for physiological over-ripening of the variety. Once that date is reached, it is designated as the cutoff date, and subsequent dates are no longer pushed back. The system automatically matches varieties based on the grower's selection, or the grower can manually input the variety.

[0081] Each recursive date represents a harvest decision point, with 8:00 AM as the decision time each day. The recursive result generates predicted fruit sugar content values ​​for each future harvest decision point. Fruit firmness prediction value ,in It is a point-in-time index. , It represents the total number of decision-making points.

[0082] At each point in the recursive harvesting decision process, the commodity value depreciation index is calculated simultaneously. . It is a dimensionless numerical value that takes a value between 0 and 1, and its calculation formula is:

[0083] in It is a high-temperature stress factor. It is a high humidity stress factor. It is the cumulative temperature value at the current point in time.

[0084] The calculation method involves obtaining the future daily maximum temperature corresponding to that point in time from meteorological forecast data. When the future daily maximum temperature does not exceed 32℃, Take 0; when the highest daily temperature in the future is not lower than 35℃, Take 1; when the future daily high temperature is between 32℃ and 35℃, Values ​​are obtained using linear interpolation.

[0085] The calculation method involves obtaining the future daily average humidity at that specific time from meteorological forecast data. When the future daily average humidity does not exceed 85%, Set to 0; when the average daily humidity in the future is not lower than 95%, Take 1; when the future daily average humidity is between 85% and 95%, Values ​​are obtained using linear interpolation.

[0086] Pick and The larger value in the range is used as the comprehensive stress factor, and then it is combined with the normalized value of accumulated temperature. Multiply, we get .

[0087] Here is a specific example. Let's take greenhouse tomatoes as an example. The value is typically between 1200℃ and 1500℃. Assuming a future daily high of 34℃ and an average daily humidity of 80%, then the following can be calculated: , , And let's assume the current situation... ℃, ℃, then .

[0088] As another alternative implementation, when sufficient historical overripeness loss statistics have been accumulated, This can be achieved using a two-dimensional lookup table. Temperature is divided into 5 intervals, and humidity into 4 intervals, resulting in 20 entries. Each entry stores a base value for loss calculated based on historical overripeness loss rates. The table was consulted to obtain the result. Then, combine it with Multiplication, that is .

[0089] After the recursion is completed, each state particle forms a quality path that includes all time points. , , Sequence. All quality paths generated by all state particles constitute a quality path set, and the number of paths in the set is equal to the number of state particles.

[0090] The price path is generated according to the following process.

[0091] Extracting from market volatility parameters , , The posterior mean is used as the parameter input for the Monte Carlo simulation. The Monte Carlo method is used to generate transaction price paths, with the number of generated paths equal to the number of quality paths.

[0092] The recursion is based on the average transaction price of the day. Starting from this point, the recursive formula is:

[0093] in , These are random numbers that follow a standard normal distribution. Each price path generates its own independently. The sequence ensures diversity among paths by setting different random seeds.

[0094] In the recursive formula, the mean regression term Its function is to bring prices closer to the long-run average, and the random disturbance term Its purpose is to introduce daily price fluctuations.

[0095] As an alternative implementation, when it is necessary to reflect the uncertainty of the parameter estimation itself, each price path is independently sampled from the posterior distribution of the parameters at the time of generation. , , The value is determined by sampling from a multivariate normal distribution whose mean is the posterior mean and whose covariance is the posterior covariance matrix.

[0096] All generated transaction price paths constitute a price path set.

[0097] The profit and loss path synthesis is carried out according to the following process.

[0098] Each quality path in the quality path set is paired one-to-one with a transaction price path in the price path set. That is, the first quality path is paired with the first price path, the second quality path is paired with the second price path, and so on.

[0099] For each pair of paths, the expected net revenue is calculated at each harvesting decision point using the following formula:

[0100] in, It is the first The expected transaction price at a given time. It is an estimated harvest volume. This is the preset daily management cost per unit area for continued branch hanging. It is the loss of expected value. This is the benchmark unit price for fruit.

[0101] The method for determining this is to first base it on the mapping rules between quality and transaction price. This is mapped to quality grades and premium coefficients. Here's a concrete example: the quality grades are divided as follows, when... When the premium factor is 1.15, the premium factor is 1.15; when When the premium factor is 1.00; when At that time, the premium coefficient was 0.90. Different crop varieties and different wholesale markets can use different thresholds and coefficients, which are set by growers in the system according to local market conditions.

[0102] After determining the premium factor, the price along the transaction price path will be... Multiply by the premium factor, then multiply by the adjustment factor. ,Right now Premium coefficient Regulatory factors This reflects the impact of over-maturity risk on the transaction price.

[0103] The crop digital twin model outputs based on the current state variables. Specifically, the output method is to... Convert the weight of a single fruit to its value, then multiply by the number of plants per unit area and the average number of fruits per plant to obtain the final value. .in, The conversion relationship between the revalued value of a single fruit and the revalued value of a single fruit is obtained through linear fitting of historical data and is determined and pre-set in the system before system deployment.

[0104] It is a preset fixed value, set by the grower based on actual inputs such as irrigation costs, plant protection costs, and facility depreciation.

[0105] It is the benchmark fruit unit price, which can be the average of the average transaction prices in the wholesale market over the past 7 days, or the average transaction price of the day, which growers can choose in the system.

[0106] The expected net income of all paired paths at each point in time constitutes the business profit and loss simulation flow set. This set is... OK A matrix of columns, where It is the total number of paths. This refers to the number of decision-making points. The scenario generation module sends the set of business profit and loss simulation streams to the strategy valuation module.

[0107] Strategy Valuation Module: The strategy valuation module receives a set of business profit and loss simulation streams sent by the scenario generation module, and then calculates and determines the net income and business option value respectively. These two calculation processes are independent of each other and can be executed in parallel.

[0108] The calculation of net income is performed according to the following procedure.

[0109] Take each particle from the posterior state particle set The average value is used as the current quality status, and the corresponding premium coefficient is determined according to the mapping rule between quality and transaction price.

[0110] Calculate the current commodity value depreciation index In the calculation method and quality path deduction The calculation method is exactly the same, using the current weather forecast data and the individual particles in the posterior state particle set as input. The mean.

[0111] according to To calculate and determine net income. Average transaction price of the day Premium coefficient , It is a preset harvesting cost per unit area, including labor and transportation costs, which is set by the grower.

[0112] As another alternative implementation, the current quality state can also be taken from each particle. The median, not the mean.

[0113] Calculated Determined Net Income As one of the outputs of the strategy valuation module, it is sent to the decision output module.

[0114] The value of a commercial option is calculated using the following process.

[0115] The strategy valuation module will continue to treat the hanging branch strategy as an American-style option. This option is valid from the current decision date until the expiration date of the harvestable window. The method for determining the expiration date is consistent with the method used in the quality path deduction.

[0116] The least squares Monte Carlo method is used for inverse solving. The input to this method is a set of simulated business profit and loss flows. It is assumed that there are a total of... Path, Each harvesting decision point, indexed from 0 to... Where 0 represents the current decision date, The deadline.

[0117] Set an early harvest marker for each path. Initially, all paths are marked as "unmarked".

[0118] At time In other words, as of the cutoff date, the holding value of all paths is directly taken as the expected net return at that point in time. .

[0119] Then from the time point Begin by recursively calculating backwards from time point 0. At each time point... Select all paths where the early harvest markers are still "unmarked".

[0120] Using these selected paths at time points Commodity value depreciation index Predicted values ​​of quality indicators and the product term of the two As independent variables. These paths at time points. The present value of holdings discounted to The value at a given time point is used as the dependent variable. Based on this, a linear regression is performed, and the regression equation is:

[0121] in , , , These are regression coefficients, estimated using the least squares method.

[0122] Each path participating in the regression and Substituting into the regression equation above, we can calculate the conditional expected return for continuing to hold. .

[0123] Then, at time point, this path... Expected net income and Compare them.

[0124] if This indicates at a certain point in time. Harvesting is more advantageous than holding; therefore, the early harvest flag for this path is set to "marked," and its holding value is updated accordingly. .

[0125] if This indicates that continuing to hold is more advantageous. In this case, the early harvest marker for this path remains "unmarked," and its holding value is updated to the time point. The present value of the holding. The discount factor is... ,in , It is the risk-free interest rate. The rate can be taken from the interest rate of government bonds of the same period, or the grower can preset an expected rate of return. The system defaults to the risk-free rate.

[0126] As an alternative implementation, if the number of "unlabeled" paths participating in the regression at a certain time point is lower than a preset minimum sample size, regression calculation is no longer performed; instead, a direct comparison is made. And 0. If If the target is met, harvest and mark the crop; otherwise, continue to hold the crop and update the holding value according to the discounted value.

[0127] The process is repeated in reverse order, point by point, until point 0. At point 0, all values ​​are retrieved. The arithmetic mean of the value held by each path is the commercial option value. .

[0128] Calculated commercial option value As the second output of the strategy valuation module, and They are sent together to the decision output module.

[0129] Decision output module: The decision output module receives the determined net income sent by the strategy valuation module. and commercial option value Compare the two.

[0130] when If the expected return from immediate harvesting is not lower than the expected return from continuing to hold, the decision output module generates a harvesting execution instruction, suggesting a harvesting period of the current day.

[0131] when When the system indicates that continuing to hold offers a higher expected return, the decision output module generates a temporary halt to harvesting instruction. The system sets the next decision cycle to the same time the following day, which is 8:00 AM. When the system clock reaches this time, it automatically triggers the execution of the complete process starting from the data acquisition module.

[0132] As another alternative implementation method, when the volatility parameter If the preset threshold is exceeded, it indicates that the market is in a period of severe volatility. The decision interval is shortened to 12 hours, and the decision times are adjusted to 8:00 AM and 8:00 PM daily. The preset threshold is set to... 1.5 times the historical average, The historical average over the last 30 decision cycles The mean of the posterior means is calculated.

[0133] Before each comparison, the decision output module first determines the performance of each particle in the posterior state particle set. Has the mean been reached? If it has already reached this point, it means the crop has reached the limit of physiological overripeness, at which point... and Based on the comparison results, a harvesting execution instruction is directly generated, and a message is displayed indicating that the physiological overripe limit has been reached.

[0134] During actual system operation, some abnormal situations may be encountered. When sensor data is missing, it is handled by imputing previous values. When the market data interface is unresponsive, the posterior distribution of parameters from the previous period is used. When weather forecast data is delayed, the most recently successfully acquired valid data is used for extrapolation. These anomaly handling mechanisms ensure the continuous and stable operation of the system.

[0135] The different implementation methods described in the above modules can be flexibly combined according to actual facility conditions, data foundation and business needs, and all can realize the technical solution proposed in this invention.

[0136] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A smart agriculture integrated system based on the Internet of Things and digital twins, characterized in that, It includes a data acquisition module, a state assimilation module, a scenario generation module, a strategy evaluation module, and a decision output module; The data acquisition module acquires current marketability data of crops within the facility, current transaction price data from the wholesale market, and weather forecast data for a preset future period. The commodity status data and current transaction price data are sent to the status assimilation module. The weather forecast data is sent to the scenario generation module; The state assimilation module receives the commodity status data and generates multiple probabilistic state samples representing the current quality status of the crop. It receives the current trading price data, updates the parameters of the market price fluctuation model, and obtains a market fluctuation parameter set. The probability state sample and the market fluctuation parameter set are sent to the scenario generation module; The scenario generation module uses the probability state samples as the initial state of the crop digital twin model, and combines them with the weather forecast data to generate predicted values ​​of quality indicators and commodity value loss indexes for each future harvest decision point. Based on the aforementioned market volatility parameter set, multiple transaction price paths are generated for each future harvesting decision point. The predicted value of the quality indicator is paired with the transaction price path. Based on the mapping rule between quality and transaction price, the expected net income of each paired path at each harvest decision point is generated, forming a set of business profit and loss simulation streams, which is then sent to the strategy valuation module. The strategy valuation module receives the set of simulated business profit and loss streams and uses a reverse solution method to calculate the commercial option value of the continued branching strategy. And calculate the definite net profit of the immediate harvest strategy based on the current quality status and current transaction price data. The commercial option value and the determined net profit are sent to the decision output module; The decision output module compares the value of the business option with the determined net return. When the determined net profit is greater than or equal to the commercial option value, a harvest execution instruction is generated. When the determined net profit is less than the commercial option value, an instruction is generated to temporarily suspend harvesting and proceed to the next decision cycle.

2. The smart agriculture integrated system based on the Internet of Things and digital twins according to claim 1, characterized in that, The state assimilation module includes a crop state assimilation unit and a market parameter update unit; The crop state assimilation unit uses the marketability state data as an observation vector, employs ensemble Kalman filtering to update the state variables of the crop digital twin model, and outputs a posterior state particle set as the probability state sample. The market parameter update unit uses the daily transaction price in the current transaction price data as the observation value, and uses a Bayesian online update algorithm to correct the posterior distribution of the parameters of the market price fluctuation model to obtain the market fluctuation parameter set.

3. The smart agriculture integrated system based on the Internet of Things and digital twins according to claim 1 or 2, characterized in that, The crop digital twin model includes state variables and state transition equations. The state variables include fruit sugar content, fruit firmness, and accumulated temperature. The state transition equations describe the diurnal variation of the state variables driven by temperature, light, and water stress indices.

4. The smart agriculture integrated system based on the Internet of Things and digital twins according to claim 1, characterized in that, The commodity value depreciation index is a value between 0 and 1, calculated by the scenario generation module based on temperature and humidity in future weather forecast data and current accumulated temperature. The commodity value depreciation index represents the expected degree of decline in the proportion of marketable fruit due to overripe fruit or fruit cracking.

5. The smart agriculture integration system based on the Internet of Things and digital twins according to claim 4, characterized in that, The commodity value depreciation index is calculated in the following way: The high-temperature stress factor is determined based on the highest daily temperature in the future. Determine the high humidity stress factor based on the future daily average humidity. Take the larger value between the high temperature stress factor and the high humidity stress factor. The commodity value depreciation index is obtained by multiplying the larger value by the normalized cumulative temperature value. The normalized value of the accumulated temperature is the ratio of the current accumulated temperature to the total accumulated temperature required for physiological over-ripening of the variety.

6. The smart agriculture integrated system based on the Internet of Things and digital twins according to claim 1, characterized in that, The mapping rules between quality and transaction price include: The relationship between quality grades based on fruit sugar content or firmness and wholesale price premiums. And the adjustment rules for lowering the average transaction price based on the commodity value depreciation index.

7. The smart agriculture integrated system based on the Internet of Things and digital twins according to claim 6, characterized in that, The adjustment rule is to multiply the average transaction price by an adjustment factor, where the adjustment factor is 1 minus the commodity value depreciation index.

8. The smart agriculture integration system based on the Internet of Things and digital twins according to claim 1, characterized in that, When calculating the value of commercial options, the strategy valuation module treats the continued branching strategy as an American option and uses the least squares Monte Carlo method to solve it in reverse from the harvestable window's cutoff date to the current decision date.

9. The smart agriculture integration system based on the Internet of Things and digital twins according to claim 8, characterized in that, When performing regression calculations in the inverse solution process, the least squares Monte Carlo method uses the commodity value depreciation index, the predicted value of the quality index, and the product of the commodity value depreciation index and the predicted value of the quality index as the independent variables of the regression basis function.

10. The smart agriculture integrated system based on the Internet of Things and digital twins according to claim 1, characterized in that, The commodity status data includes environmental parameters and crop physiological parameters collected by IoT sensing devices installed within the facility. The environmental parameters include at least one of air temperature, air humidity, light intensity, and CO2 concentration. The crop physiological parameters include at least one of stem microstrain, leaf temperature, and canopy spectral reflectance. The transaction price data is obtained through a wholesale market data interface. The weather forecast data is obtained through an external weather service interface.

Images