Parameter tuning method of crop growth simulation model, yield prediction method based on crop growth simulation model, electronic equipment and medium

Abstract

The invention provides a parameter adjusting and optimizing method of a crop growth simulation model, a yield prediction method based on the crop growth simulation model, electronic equipment and a medium. The adjusting and optimizing method comprises the following steps: acquiring environment characteristics of each target area in a target region; based on the environment characteristics, determining a value range of a to-be-adjusted parameter in each target area, and obtaining a plurality of tuning discrete points in the value range to form a first set; removing tuning discrete points which do not conform to a preset rule in the first set to form a second set; according to the environment characteristics, environment similarity of the bordered target areas is obtained through calculation; when the environment similarity is greater than a similarity threshold value, performing cross exchange on the partial tuning discrete points of the second set of the at least two target areas to obtain a third set; and based on a first preset strategy, performing screening in the third set to obtain an optimal tuning discrete point.

Description

Technical Field

[0001] This invention relates to the field of environmental engineering technology, and more specifically, to a method for parameter tuning of a crop growth simulation model, a method for yield prediction based on a crop growth simulation model, electronic equipment, and media. Background Technology

[0002] Existing technologies generally ignore the environmental heterogeneity of different regions within the target area when optimizing crop growth simulation model parameters, and adopt a uniform parameter range or optimal parameters. This results in a mismatch between parameters and the local environment, leading to large simulation errors. The wide parameter range results in a large number of parameter combinations, including many invalid combinations with high errors. Substituting all of them into the model leads to extremely large computational loads and long optimization cycles. Adjacent and environmentally similar regions have excessively different parameters due to independent parameter searches, which does not conform to actual growth patterns and causes "discontinuities" in simulation results between regions. Particle swarm optimization alone is prone to getting trapped in local optima, while Bayesian optimization alone has slow convergence and low efficiency when the search range is too wide, making it impossible to achieve both efficient and accurate searches at the same time. Summary of the Invention

[0003] In view of this, one of the objectives of this application is to provide a parameter tuning method for a crop growth simulation model, which can improve the problems existing in the prior art.

[0004] To achieve the above technical objectives, the technical solution adopted in this application is as follows:

[0005] In a first aspect, embodiments of this application provide a method for parameter tuning of a crop growth simulation model, the method comprising:

[0006] Obtain the environmental characteristics of each target area within the target region;

[0007] Based on the environmental characteristics, the range of values ​​for the parameters to be adjusted in each target region is determined, and multiple optimization discrete points are obtained within the range of values ​​to form a first set;

[0008] In the first set, the optimized discrete points that do not conform to the preset rules are removed to form the second set;

[0009] Based on the environmental characteristics, the environmental similarity of the adjacent target areas is calculated;

[0010] When the environmental similarity is greater than the similarity threshold, a portion of the optimized discrete points of the second set of the at least two target regions are cross-exchanged to obtain a third set;

[0011] Based on the first preset strategy, the optimal tuning discrete points are selected from the third set and used as the configuration parameters of the crop growth simulation model.

[0012] Furthermore, in the first set, optimization discrete points that do not conform to the preset rules are removed to form a second set, including:

[0013] Each of the optimized discrete points in the first set is input into the trained deep learning model, and the trained deep learning model outputs a first prediction error value. The optimized discrete points corresponding to the first prediction error values ​​that are greater than the error threshold are removed to obtain the second set.

[0014] Furthermore, before removing the optimized discrete points that do not conform to the preset rules from the first set to form the second set, the method further includes:

[0015] A fourth set is obtained within the range of values. The fourth set includes several training discrete points and does not overlap with the first set.

[0016] Obtain the first historical driving data and the first historical true result of the target area;

[0017] The training discrete points are deployed on the crop growth simulation model, and the first historical driving data is input to obtain the first prediction result;

[0018] Calculate a second prediction error value between the first prediction result and the first historical true result, so that the second prediction error value is marked on the corresponding training discrete point;

[0019] The labeled training discrete points are input into the deep learning model to obtain the trained deep learning model.

[0020] Furthermore, the step of selecting the optimal discrete points for tuning from the third set based on the first preset strategy includes:

[0021] The third set is input into the particle swarm optimization model, which optimizes all the fine-tuned discrete points of the third set to obtain the fifth set.

[0022] The fifth set is input into the Bayesian algorithm model, and the optimal tuning discrete points are obtained through the Bayesian algorithm model.

[0023] Furthermore, the third set is input into the particle swarm optimization (PSO) model, which optimizes all the fine-tuned discrete points of the third set to obtain a fifth set, including:

[0024] A1: Based on the second preset strategy, obtain the cross-unit learning factor of each optimized discrete point of the third set, and the cross-unit learning factor is used to characterize the applicability of each optimized discrete point of the third set in the adjacent target area.

[0025] A2: Set an encoding for each of the optimization discrete points in the third set, the encoding including environmental features and cross-unit learning factors;

[0026] A3: Deploy the optimized discrete points of the third set on the crop growth simulation model, input the second historical driving data, and obtain the second prediction result;

[0027] A4: Calculate the third prediction error value between the second prediction result and the second historical true result;

[0028] A5: In the third set, the optimized discrete point with the smallest third prediction error value is selected as the first optimal discrete point; among all the optimized discrete points that have undergone the cross-interchange, the optimized discrete point with the smallest third prediction error value is selected as the second optimal discrete point; and among the other optimized discrete points in the third set, excluding all the optimized discrete points that have undergone the cross-interchange, the optimized discrete point with the smallest third prediction error value is selected as the third optimal discrete point.

[0029] A6: Based on the preset speed update function, determine the current speed according to the first optimal discrete point, the second optimal discrete point, the third optimal discrete point and the code;

[0030] A7: Update the encoding based on the current speed;

[0031] A8: Repeat A1-A7 a preset number of times to obtain the optimized discrete points.

[0032] Furthermore, the speed update function is:

[0033]

[0034] Indicates the current speed;

[0035] Indicates the previous current speed;

[0036] Indicates the inertia weighting coefficient;

[0037] These represent self-learning factors, local learning factors, and cross-unit learning factors, respectively.

[0038] , , These represent self-learning random numbers, local learning random numbers, and cross-unit learning random numbers, respectively.

[0039] , and These are the codes for the first optimal discrete point, the second optimal discrete point, and the third optimal discrete point, respectively.

[0040] This represents the current encoding of the optimized discrete points of the third set.

[0041] Secondly, this embodiment proposes a yield prediction method based on a crop growth simulation model, the method comprising:

[0042] Acquire the target area, the environmental characteristics of the target area, and driving data, including predicted climate data;

[0043] Based on the target area and environmental characteristics, and using the above-mentioned optimization method, the optimal optimization discrete point is obtained.

[0044] The optimized discrete points are deployed on the crop growth simulation model to obtain the optimized crop growth simulation model.

[0045] The driving data is input into the optimized crop growth simulation model, and the predicted yield is obtained through the optimized crop growth simulation model.

[0046] Thirdly, this embodiment proposes an electronic device, which includes a processor and a memory coupled to each other. The memory stores a computer program, and when the computer program is executed by the processor, the electronic device performs a tuning method or a prediction method.

[0047] Fourthly, this embodiment proposes a computer-readable storage medium, characterized in that the computer-readable storage medium stores a computer program, which, when run on a computer, causes the computer to execute a tuning method or a prediction method.

[0048] The invention employing the above technical solution has the following advantages:

[0049] In the technical solution provided in this application, the environmental characteristics of each target area are first obtained; then, based on these environmental characteristics, the value range of the parameters to be adjusted is determined for each area, and multiple optimization discrete points are generated within this range to form a first set; subsequently, optimization discrete points that do not conform to preset rules in the first set are removed to form a simplified second set; then, the environmental similarity between adjacent target areas is calculated based on the environmental characteristics. When the similarity is greater than a preset threshold, some optimization discrete points in the second set of these adjacent areas are cross-exchanged to obtain a third set that integrates high-quality parameters across regions; finally, the optimal optimization discrete points are selected from the third set based on a first preset strategy to complete the parameter optimization. This solution ensures that the initial parameter search fits the local environment by defining the range by region, reduces redundancy by filtering invalid points through rules, and achieves synergy between adjacent regions through cross-regional parameter exchange. The final optimal parameters not only fit the environment of a single region but also take into account the consistency between regions, significantly improving the matching degree between the crop growth simulation model parameters and the actual situation of the target area, reducing simulation errors, and achieving accurate optimization while conducting efficient search.

[0050] In the technical solution provided in this application, each optimized discrete point in the first set is input into a trained deep learning model, which outputs a first prediction error value for the corresponding discrete point. Then, using a preset error threshold as a criterion, optimized discrete points with first prediction errors greater than the threshold are removed from the first set, and the remaining optimized discrete points constitute the second set. The trained deep learning model quickly predicts the error of the optimized discrete points, replacing the traditional method of substituting all discrete points into a crop growth simulation model to calculate the error. This significantly reduces the subsequent computational workload of invalid parameter combinations and improves the generation efficiency of the second set. Simultaneously, the accurate selection based on model prediction effectively retains high-quality discrete points with low errors, laying a high-quality foundation for subsequent cross-regional exchange and optimal parameter selection, and shortening the overall optimization cycle.

[0051] In the technical solution provided in this application, the particle swarm optimization algorithm fully considers the cross-regional adaptability of parameters by using cross-unit learning factors and encoding containing environmental information, avoiding isolated parameter optimization. The setting of three optimal discrete points provides multiple references for parameter updates, ensuring that the optimization both approaches the global optimum and retains the characteristics of high-quality cross-regional and local parameters. Multiple rounds of iterative updates enable the tuning discrete points to continuously converge towards the direction of "local adaptation + cross-regional collaboration," resulting in higher quality parameters in the fifth set of outputs. This provides high-quality input for the subsequent precise fine-tuning of the Bayesian algorithm, ultimately improving the scientific rigor and practicality of the entire tuning process. Attached Figure Description

[0052] This application can be further illustrated by the non-limiting embodiments given in the accompanying drawings. It should be understood that the following drawings only illustrate some embodiments of this application and should not be considered as limiting the scope. For those skilled in the art, other related drawings can be obtained from these drawings without any inventive effort.

[0053] Figure 1 The main flowchart provided for Embodiment 1 of this application.

[0054] Figure 2 This is a sub-flowchart of S160 provided in Embodiment 1 of this application.

[0055] Figure 3 The main flowchart provided for Embodiment 2 of this application. Detailed Implementation

[0056] The present application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that similar or identical parts are referred to by the same reference numerals in the drawings or description. Implementations not shown or described in the drawings are forms known to those skilled in the art. In the description of this application, terms such as "first" and "second" are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0057] Example 1

[0058] Please refer to Figure 1 This application provides a method for parameter tuning of a crop growth simulation model. The method for parameter tuning of a crop growth simulation model may include the following steps:

[0059] S110, Obtain the environmental characteristics of each target area in the target region;

[0060] S120, based on the environmental characteristics, determine the value range of the parameter to be adjusted in each target area, and obtain multiple optimization discrete points in the value range to form a first set;

[0061] S130, In the first set, remove the optimized discrete points that do not conform to the preset rules to form a second set;

[0062] S140, Based on the environmental characteristics, calculate the environmental similarity of the adjacent target areas;

[0063] S150, when the environmental similarity is greater than the similarity threshold, the optimized discrete points of the second set of the at least two target regions are cross-exchanged to obtain a third set;

[0064] S160, based on the first preset strategy, the optimal tuning discrete points are selected from the third set.

[0065] In the above implementation, the environmental characteristics of each target region in the target area are first obtained. Based on the environmental characteristics, the range of values ​​for the parameters to be adjusted is determined for each region, and optimization discrete points are generated. At the same time, parameters with less impact on crop simulated growth are fixed, and the remaining key parameters are processed in segments. Then, the trained deep learning model is used to quickly eliminate high-error invalid discrete points. Next, the optimization discrete points of the bordering regions with qualified environmental similarity are cross-exchanged. A cross-unit learning factor is introduced to characterize the cross-regional applicability of the parameters and complete the encoding. Then, the parameter combination after cross-exchanging is coarsely searched using an improved particle swarm optimization algorithm. The parameter update speed and encoding are combined with the first, second, and third optimal discrete points to obtain a parameter set with good convergence. Finally, this set is input into the Bayesian algorithm for precise fine-tuning. The root mean square error and standard root mean square error between the simulated and observed values ​​are used as evaluation conditions throughout the process. Finally, the optimal parameter combination that takes into account regional adaptability, cross-regional synergy, optimization efficiency and accuracy is obtained, effectively reducing the impact of parameter setting deviation on the model results.

[0066] The crop growth simulation model in this embodiment can be the Worst model.

[0067] The following section will elaborate on the steps involved in parameter tuning for crop growth simulation models:

[0068] In S110, it can be understood that the target area may include several target regions, each with different environmental characteristics. Environmental characteristics may include altitude and soil type. Altitude may be divided into <200m, 200-600m, and >600m. Soil type may include purple soil, yellow soil, and paddy soil.

[0069] In this embodiment, based on environmental characteristics, the target area M city is divided into 5 target regions, namely A, B, C, D and E, where A is a low-altitude purple soil region, B is a low-altitude yellow soil region, C is a mid-altitude purple soil region, D is a mid-altitude yellow soil region and E is a high-altitude paddy soil region.

[0070] In this embodiment, since the environmental characteristics of each target area are different, it is necessary to perform targeted parameter optimization design for each target area. Thus, the Worst model can predict the yield based on the target area to be predicted, thereby improving the accuracy of the prediction.

[0071] In S120, this embodiment discards model parameters that have little impact on yield prediction and selects model parameters that have a greater impact on yield prediction as parameters to be adjusted. The range of values ​​for the parameters to be adjusted is determined based on the environmental characteristics of the target area. For example, the harvest index (HI), maximum vernalization days (DPDmax), light energy utilization efficiency (ɛ), accumulated temperature during the grain-filling period (TSUMGF), accumulated temperature during the flowering period (TSUMFL), and minimum soil moisture content threshold (WCmin) can be selected as parameters to be adjusted.

[0072] For low-altitude purple soil areas, assuming an altitude <300m, an average annual temperature of 17-18℃, ample sunshine (1300-1400 hours annually), good water and fertilizer retention, and abundant rainfall (1100-1200 mm annually), then the following parameters are considered: HI (Harvest Index): 0.39-0.42 (sufficient light and heat, high biomass-to-grain conversion efficiency); DPDmax (maximum vernalization days): 26-30 days (short cumulative low temperature period, rapid vernalization process); α (light utilization efficiency): 1.65-1.75 g•MJ. -1 (Sufficient sunlight, high photosynthetic conversion efficiency); TSUMGF (accumulated temperature during grain filling): 400-430℃•d (high temperature, short grain filling period, low accumulated temperature required); TSUMFL (accumulated temperature during flowering): 240-260℃•d (rapid temperature rise, short sowing to flowering cycle); WCmin (minimum soil moisture content threshold): 0.18-0.22cm³ / cm³ (purple soil has good water retention, and the minimum moisture content requirement is relatively high).

[0073] For mid-altitude yellow soil regions, assuming an altitude of 400-600m, an average annual temperature of 15-16℃, moderate sunshine (1100-1200 hours per year), and weaker water and fertilizer retention than purple soil, with slightly less rainfall (1000-1100mm per year), the following parameters are considered: HI (Harvest Index): 0.37-0.40 (light and heat conditions are slightly worse than at lower altitudes, resulting in a slight decrease in conversion efficiency); DPDmax (maximum vernalization days): 31-35 days (increased altitude leads to longer periods of accumulated low temperatures, requiring a longer vernalization period); α (light utilization efficiency): 1.55-1.65 g•MJ. -1 (Reduced sunshine leads to a slight decrease in photosynthetic conversion efficiency); TSUMGF (accumulated temperature during grain filling): 430-460℃•d (lower temperature leads to a longer grain filling period and an increase in required accumulated temperature); TSUMFL (accumulated temperature during flowering): 260-280℃•d (slower temperature rise leads to a longer sowing-to-flowering cycle); WCmin (minimum soil moisture content threshold): 0.15-0.19cm³ / cm³ (yellow soil has weak water retention capacity, and crops are adapted to lower minimum moisture content).

[0074] In this embodiment, Latin hypercube sampling can be used to extract multiple values ​​from the range of each parameter to be adjusted, and then these values ​​can be freely combined to obtain many discrete points. A portion of these discrete points are used as training discrete points, and another portion are used as tuning discrete points. All the tuning discrete points are then integrated to obtain the first set.

[0075] In S130, a deep learning model can be used to remove discrete points that do not conform to preset rules. Therefore, S130 can specifically be:

[0076] Each of the optimized discrete points is input into the trained deep learning model, and the trained deep learning model outputs a first prediction error value. The optimized discrete points corresponding to the first prediction error values ​​that are greater than the error threshold are removed to obtain the second set.

[0077] Therefore, the preset rule can be understood as follows: the fine-tuning discrete points whose first error value is greater than the error threshold obtained from the trained deep learning model should be removed, while the fine-tuning discrete points whose first error value is less than or equal to the error threshold should be retained.

[0078] Therefore, the training method for deep learning models is as follows:

[0079] A fourth set is obtained within the range of values. The fourth set includes several training discrete points, and the fourth set does not overlap with the first set. First historical driving data and first historical true results for the target region are obtained. The training discrete points are deployed on the crop growth simulation model, and the first historical driving data is input to obtain a first prediction result. A second prediction error value between the first prediction result and the first historical true result is calculated so that the second prediction error value is labeled on the corresponding training discrete points. The labeled training discrete points are input into the deep learning model to obtain the trained deep learning model.

[0080] For example, based on the established value ranges of core parameters to be adjusted in this region, such as HI (0.39-0.42) and DPDmax (26-30 days), Latin hypercube sampling is used to generate 30 sets of training discrete points that do not overlap with the first set (optimization discrete points), forming the fourth set. Then, the first historical driving data (including meteorological data such as daily average temperature and sunshine hours, soil data such as purple soil bulk density, and management data such as sowing period and fertilizer application) and the corresponding first historical real results (field data such as measured yield and flowering period) for this mid-altitude yellow soil region from 2019 to 2022 are obtained. Next, each set of training discrete points in the fourth set is deployed to... In the WOFOST crop growth simulation model, the first historical driving data is input to obtain the first prediction result. By calculating the NRMSE (yield error) and RMSE (growth period error) between the first prediction result and the first historical true result, a comprehensive second prediction error value is obtained and labeled on the corresponding training discrete points. Finally, the 30 labeled training discrete points are divided into training set and validation set in a 7:3 ratio and input into the FCNN deep learning model for training until the R² of the prediction error and the true error of the validation set is ≥0.9. Finally, a trained deep learning model that can quickly predict the error of any combination of parameters is obtained, which can support the subsequent selection of discrete points for the first set of optimization.

[0081] In S140, the core environmental characteristics affecting crop growth in the bordering target areas (such as the low-altitude purple soil area and the mid-altitude yellow soil area of ​​rapeseed in Chongqing) are first identified, and five key indicators (covering the core dimensions of climate, soil, and topography) are selected: altitude, soil type, average annual temperature, annual sunshine hours, and annual precipitation. Then, the quantitative factors (altitude, average annual temperature, annual sunshine hours, and annual precipitation) are standardized using a min-max method, mapping values ​​of different units and magnitudes to a uniform 0-1 range (e.g., the altitude standardization formula is: Standardized value = (Target area altitude - Minimum altitude of all bordering areas) / (Maximum altitude of all bordering areas - Minimum altitude of all bordering areas)). Qualitative factors (soil type...) are then... The similarity of the two regions is quantified according to the rule of "1 for consistent types and 0 for inconsistent types". Then, for each standardized factor, the single-factor similarity between the bordering regions is calculated (quantitative factor similarity = 1 - |standardized value of region A - standardized value of region B|, qualitative factor similarity is directly taken from the quantification result). Then, according to the weight of each factor's influence on crop growth (e.g., soil type and annual average temperature are each set to 0.3, altitude and annual sunshine hours are each set to 0.15, annual precipitation is set to 0.1, and the total weight is 1), the single-factor similarity is weighted and summed to finally obtain the environmental similarity of the bordering target regions (the result ranges from 0 to 1, the closer to 1, the smaller the environmental difference between the two regions, and the closer to 0, the greater the difference).

[0082] In this embodiment, the premise for calculating environmental similarity is that the regions are adjacent. This is because the geographical space of adjacent areas is continuous, and their environmental characteristics such as climate (e.g., temperature, light, and water gradients), soil type, and topography usually change gradually rather than abruptly. As a result, crop growth patterns have stronger correlation and consistency, providing a natural basis for cross-regional collaborative optimization of parameters. At the same time, independent parameter tuning in adjacent areas can easily lead to excessive differences in the optimal parameters of adjacent areas, resulting in "discontinuities" in crop growth simulation results at regional boundaries (e.g., sudden jumps in simulated values ​​of yield and growth period in transitional zones). Limiting the parameters to adjacent areas ensures that the tuning discrete points of cross-exchange have practical adaptation value, avoiding extremely low cross-regional applicability of parameters and the introduction of invalid parameters to interfere with tuning due to abrupt differences in environmental characteristics in non-adjacent areas (e.g., non-adjacent low-altitude plains and high-altitude mountains). In addition, the environmental similarity of adjacent areas is more likely to meet the threshold requirements. After parameter cross-exchange, it can ensure that the parameters of each region conform to their own environment while achieving logical coherence between parameters in different regions.

[0083] In S150, the first step is to preset an environmental similarity threshold (determined by crop type and regional characteristics; for example, the threshold for rapeseed planting areas is set to 0.6, with a larger value indicating that only adjacent areas with extremely similar environments are allowed to exchange points). If the calculated environmental similarity of the adjacent areas exceeds this threshold, the cross-exchange process is initiated. The first step is to screen interchangeable, optimized discrete points: from the second set of each target region, select discrete points with "high-quality error performance" (i.e., after pre-screening by the deep learning model, the first prediction error value is lower than the average error of the second set by 20%-30%), ensuring that the exchanged parameters have potential adaptation value and avoiding the introduction of low-quality parameters. The second step is to determine the cross-exchange ratio: set the number of exchanges based on the particle swarm size (e.g., when each second set contains 50 sets of discrete points, extract 3-5 sets of high-quality points at a 1:1 ratio, with the total not exceeding 10% of the single set size, balancing the dominance of local parameters with cross-regional diversity). The third step is to perform a cross-exchange: the high-quality discrete points selected from region A are added to the second set of region B, and the high-quality discrete points selected from region B are added to the second set of region A. After the exchange, the discrete points that were not selected in the original second set are retained. Finally, each region forms a merged set of "local high-quality points + neighboring high-quality points", which is the third set of that region.

[0084] For example, taking the low-altitude purple soil region and the mid-altitude yellow soil region as examples, the environmental similarity between the two regions is calculated to be 0.75 (> the threshold of 0.6), satisfying the conditions for cross-exchange. Assume that the second set of data for the low-altitude purple soil region contains 50 sets of low-error optimized discrete points (all errors < 8%), and the second set of data for the mid-altitude yellow soil region contains 45 sets of low-error optimized discrete points (all errors < 9%). The exchange ratio is set to 20%: ① Screening of exchange points—The low-altitude purple soil region selects the 10 sets of discrete points with the smallest error from the second set (e.g., (HI=0.40, DPDmax=28 days), with an average error of 5.2%), and the mid-altitude yellow soil region selects the 9 sets of discrete points with the smallest error from the second set (e.g., (HI=0.38, DPDmax=32 days), with an average error of 6.1%); ② Adding labels and factors—Labeling the 10 sets of points in the low-altitude purple soil region with "..." Source: Low-altitude purple soil region, cross-unit learning factor = 0.72 (medium potential for adaptation to mid-altitude yellow soil region), and label the 9 sets of points in the mid-altitude yellow soil region as "Source: Mid-altitude yellow soil region, cross-unit learning factor = 0.68 (medium potential for adaptation to low-altitude purple soil region); ③ Merge to form the third set - the third set of low-altitude purple soil region = the remaining 40 sets of points in the original second set of low-altitude purple soil region + 9 sets of interchange points in mid-altitude yellow soil region, the third set of mid-altitude yellow soil region = the remaining 36 sets of points in the original second set of mid-altitude yellow soil region + 10 sets of interchange points in low-altitude purple soil region. Finally, the third sets of the two regions retain the local adaptation basis and introduce the high-quality parameter experience of the bordering regions.

[0085] In S160, such as Figure 2 As shown, the specific steps may include the following:

[0086] S161: The third set is input into the particle swarm optimization model, which optimizes all the fine-tuning discrete points of the third set to obtain the fifth set;

[0087] S162: Input the fifth set into the Bayesian algorithm model, and obtain the optimal tuning discrete points through the Bayesian algorithm model.

[0088] In S161, the following steps are repeated a preset number of times to obtain the optimized discrete points. The steps include:

[0089] Based on the second preset strategy, a cross-unit learning factor is obtained for each optimized discrete point in the third set. This cross-unit learning factor characterizes the applicability of each optimized discrete point in the third set within adjacent target areas. An encoding is set for each optimized discrete point in the third set, including environmental features and the cross-unit learning factor. The optimized discrete points in the third set are deployed on a crop growth simulation model, and the second historical driving data is input to obtain a second prediction result. A third prediction error value is calculated between the second prediction result and the second historical true result. The third prediction error value is then selected from the third set. The smallest optimized discrete point is selected as the first optimal discrete point; among all the optimized discrete points obtained through the cross-interchange, the optimized discrete point with the smallest third prediction error value is selected as the second optimal discrete point; and among the other optimized discrete points in the third set excluding all the optimized discrete points obtained through the cross-interchange, the optimized discrete point with the smallest third prediction error value is selected as the third optimal discrete point; based on a preset speed update function, the current speed is determined according to the first optimal discrete point, the second optimal discrete point, the third optimal discrete point, and the code; and the code is updated based on the current speed.

[0090] In this embodiment, the calculation method for cross-unit learning is as follows: First, the optimized discrete points to be exchanged (high-quality points in the second set of the source region) are substituted into the crop growth simulation model of the target region. The historical driving data (meteorological, soil, and management data) of the target region are input to obtain the prediction results. The prediction results are compared with the historical true results of the target region to calculate the cross-regional error (such as the combined value of NRMSE and RMSE). Then, the local error of the discrete point in the source region (already marked in the second set) is extracted. The error relationship is transformed into the fitness degree in the 0-1 interval by the formula "basic fitness coefficient = 1 - (cross-regional error / (local error + cross-regional error))". Finally, the basic fitness coefficient is multiplied by the calculated environmental similarity of the two adjacent regions to obtain the preliminary factor. The result is then constrained to the 0-1 interval (if it exceeds the range, the boundary value is taken), which is the final cross-unit learning factor.

[0091] For example, step 1: Obtain the cross-unit learning factor of each optimized discrete point in the third set. The cross-unit learning factor is calculated based on "the adaptability of parameters in the target region + environmental similarity". The result is directly related to the adaptability potential of the discrete point: local high-quality points (such as point A: HI=0.40, DPDmax=28 days): no cross-region source, cross-unit learning factor=0.92 (indicating local adaptability and high applicability to region 2); exchanged high-quality points (such as point B: HI=0.38, DPDmax=32 days, from region 2): cross-unit learning factor=0.68 (previously calculated by "cross-region error + local error + environmental similarity", with a moderate degree of adaptability to region 1); all 49 groups of points in the third set obtain the cross-unit learning factor (range 0.55-0.95) according to this rule.

[0092] Step 2: Set up an encoding for the discrete points. The encoding is a vector (3 dimensions, adapting 2 parameters + 1 factor) of "environmental feature quantification value + cross-unit learning factor". The environmental feature quantification is based on the core attribute of Region 1: Point A (local high-quality point) encoding: [0.1 (altitude normalization value, altitude of Region 1 < 300m), 1.0 (soil type quantification value, purple soil = 1), 0.92 (cross-unit learning factor)]; Point B (exchange high-quality point) encoding: [0.1 (altitude normalization value of Region 1), 1.0 (purple soil quantification value), 0.68 (cross-unit learning factor)]; All discrete points are encoded in this format to ensure that the particles carry the dual information of "environmental adaptation + cross-region applicability".

[0093] Step 3: Deploy discrete points to the model and obtain the second prediction results. Substitute the 49 sets of discrete points from the third set into the WOFOST model one by one. Input the second historical driving data of region 1 in 2023: After deployment of point A, the output (second prediction result) is: simulated yield 241 kg / mu, simulated flowering period March 10; after deployment of point B, the output (second prediction result) is: simulated yield 239 kg / mu, simulated flowering period March 12; the remaining 47 sets of points have all completed model deployment and obtained the corresponding second prediction results.

[0094] Step 4: Calculate the third prediction error value. Using "yield NRMSE + fertility period RMSE" as the comprehensive error index, compare the second prediction result with the second historical actual result: The third prediction error value for point A: NRMSE = |241-243| / 243×100%≈0.82%, RMSE = 1 day, comprehensive error = (0.82+1) / 2≈0.91; The third prediction error value for point B: NRMSE = |239-243| / 243×100%≈1.65%, RMSE = 3 days, comprehensive error = (1.65+3) / 2≈2.32; The third prediction error value (range 0.85-3.20) was calculated for all 49 groups of points.

[0095] Step 5: Select three optimal discrete points and classify them according to the principle of minimum error to identify the optimal representatives of discrete points from different sources: First optimal discrete point (optimal of the third set as a whole): Point C (local high-quality point, HI=0.405, DPDmax=27 days), third prediction error value=0.85 (minimum); Second optimal discrete point (optimal among exchange points): Point D (from region 2, HI=0.39, DPDmax=31 days), third prediction error value=1.23 (minimum among exchange points); Third optimal discrete point (optimal among non-exchange points): i.e., point C (optimal among local high-quality points, overlapping with the first optimal point, but local points have better adaptability).

[0096] Step 6: Determine the current speed based on the speed update function. Taking point B (parameters corresponding to the current encoding: HI=0.38, DPDmax=28 days; old speed v_old=[0.01,0.5]) as an example, substitute the parameters of the three optimal discrete points: First optimal (point C) parameters: [0.405,27]; Second optimal (point D) parameters: [0.39,31]; Third optimal (point C) parameters: [0.405,27]; Calculate the current speed (calculate separately for HI and DPDmax parameters): Current speed of HI: 0.8×0.01+2×0.5×( 0.405-0.38)+2×0.5×(0.39-0.38)+2×0.5×(0.405-0.38)=0.008+0.025+0.01+0.025=0.068; Current velocity of DPDmax: 0.8×0.5+2×0.5×(27-32)+2×0.5×(31-32)+2×0.5×(27-32)=0.4-5-1-5=-10.6; Current velocity vector: [0.068,-10.6] (HI needs to be increased by 0.068, DPDmax needs to be decreased by 10.6 days).

[0097] Step 7: Update the encoding based on the current speed. The core of the encoding update is to adjust the parameter values ​​(the environmental feature quantization values ​​in the encoding are fixed, only the implicit values ​​corresponding to the parameters are updated, and the cross-unit learning factor is dynamically adjusted according to the parameter adaptability): New parameter values ​​for point B: HI=0.38+0.068=0.448 (since the HI value range of region 1 is 0.39-0.42, the constraint is 0.42); DPDmax=32+(-10.6)=21.4 (the constraint is 26 days, which meets the lower limit of the value range of region 1); New cross-unit learning factor: recalculated based on the new parameters (HI=0.42, DPDmax=26 days), resulting in 0.75 (the degree of adaptation to region 1 is improved); Updated encoding of point B: [0.1,1.0,0.75], completing the update of this round of particles (optimized discrete points). The above steps will be repeated a preset number of times (e.g., 20 generations), and the encoding of all discrete points will be updated according to this logic in each generation, finally obtaining the fifth set with good convergence.

[0098] The speed update function in this embodiment is:

[0099]

[0100] Indicates the current speed;

[0101] Indicates the previous current speed;

[0102] Indicates the inertia weighting coefficient;

[0103] These represent self-learning factors, local learning factors, and cross-unit learning factors, respectively.

[0104] , , These represent self-learning random numbers, local learning random numbers, and cross-unit learning random numbers, respectively.

[0105] , and These are the codes for the first optimal discrete point, the second optimal discrete point, and the third optimal discrete point, respectively.

[0106] This represents the current encoding of the optimized discrete points of the third set.

[0107] Example 2

[0108] This embodiment proposes a yield prediction method based on a crop growth simulation model, such as... Figure 3 As shown, the prediction method specifically includes the following steps:

[0109] S210: Acquire the target area, the environmental characteristics of the target area, and driving data, wherein the driving data includes predicted climate data;

[0110] S220: Based on the target area and environmental characteristics, and using the method described in Example 1, obtain the optimal tuning discrete point;

[0111] S230: Deploy the optimized discrete points on the crop growth simulation model to obtain the optimized crop growth simulation model;

[0112] S240: Input the driving data into the optimized crop growth simulation model, and obtain the predicted yield through the optimized crop growth simulation model.

[0113] In this embodiment, the optimized crop growth simulation model obtained in Example 1 is used for yield prediction. This scheme achieves accurate crop yield prediction through a closed-loop process of "data preparation - parameter tuning - model deployment - yield prediction". The specific implementation logic is as follows: First, step S210 is executed to identify the specific target area for yield prediction. The system collects the environmental characteristics of the area (such as core indicators such as altitude, soil type, and annual average temperature) and corresponding driving data (including predicted climate data for future simulation, such as predicted values ​​of daily average temperature, sunshine hours, and precipitation for the next year). Then, step S220 is executed. Based on the spatial attributes of the target area and the collected environmental characteristics, the system uses the "regional range determination, deep learning pre-screening, parameter exchange in border areas, particle swarm optimization, and other methods" from Example 1. The complete parameter tuning method of "Bayesian cascade optimization" selects the optimal tuning discrete point with the smallest error and that meets physiological rules from the tuning discrete points that conform to the regional environment. Then, in step S230, the optimal tuning discrete point (i.e., the optimal parameter combination adapted to the target area) is deployed to the crop growth simulation model (such as WOFOST) to complete the targeted configuration of the model parameters and form the tuned crop growth simulation model. Finally, step S240 is executed to input the driving data containing predicted climate data into the tuned crop growth simulation model. Based on the parameters adapted to the target area and the predicted environmental conditions, the model simulates the entire growth process of crops from sowing to harvest and finally outputs the predicted crop yield of the target area, realizing accurate yield prediction supported by "environmentally adapted parameters + accurate climate prediction".

[0114] In this embodiment, the processing module can be an integrated circuit chip with signal processing capabilities. The processing module can be a general-purpose processor. For example, the processor can be a Central Processing Unit (CPU), a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0115] The storage module can be, but is not limited to, random access memory, read-only memory, programmable read-only memory, erasable programmable read-only memory, electrically erasable programmable read-only memory, etc.

[0116] This application provides an electronic device that may include a processing module and a storage module. The storage module stores a computer program, which, when executed by the processing module, enables the electronic device to perform the methods described in Embodiment 1 or the corresponding steps in the methods described in Embodiment 2.

[0117] It should be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the electronic device described above can be referred to the corresponding steps in the aforementioned method, and will not be elaborated further here.

[0118] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the prediction method or method as described in the above embodiments.

[0119] Based on the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product. This software product can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, electronic device, or network device, etc.) to execute the methods described in the various implementation scenarios of this application.

[0120] In the embodiments provided in this application, it should be understood that the disclosed methods can also be implemented in other ways. The method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of methods and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, program segment, or part of code, which includes one or more executable instructions for implementing a specified logical function. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions. Furthermore, the functional modules in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0121] The above description is merely an embodiment of this application and is not intended to limit the scope of protection of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of protection of this application.

Claims

1. A method for parameter tuning of a crop growth simulation model, characterized in that, The method includes: Obtain the environmental characteristics of each target area within the target region; Based on the environmental characteristics, the range of values ​​for the parameters to be adjusted in each target region is determined, and multiple optimization discrete points are obtained within the range of values ​​to form a first set; In the first set, the optimized discrete points that do not conform to the preset rules are removed to form the second set; Based on the environmental characteristics, the environmental similarity of the adjacent target areas is calculated; When the environmental similarity is greater than the similarity threshold, a portion of the optimized discrete points of the second set of the at least two target regions are cross-exchanged to obtain a third set; Based on the first preset strategy, the optimal tuning discrete points are selected from the third set and used as the configuration parameters of the crop growth simulation model.

2. The method according to claim 1, characterized in that, The step of removing optimization discrete points that do not conform to preset rules from the first set to form a second set includes: Each of the optimized discrete points in the first set is input into the trained deep learning model, and the trained deep learning model outputs a first prediction error value. The optimized discrete points corresponding to the first prediction error values ​​that are greater than the error threshold are removed to obtain the second set.

3. The method according to claim 2, characterized in that, Before removing optimization discrete points that do not conform to preset rules from the first set and forming the second set, the method further includes: A fourth set is obtained within the range of values. The fourth set includes several training discrete points and does not overlap with the first set. Obtain the first historical driving data and the first historical true result of the target area; The training discrete points are deployed on the crop growth simulation model, and the first historical driving data is input to obtain the first prediction result; Calculate a second prediction error value between the first prediction result and the first historical true result, so that the second prediction error value is marked on the corresponding training discrete point; The labeled training discrete points are input into the deep learning model to obtain the trained deep learning model.

4. The method according to claim 1, characterized in that, The step of selecting the optimal tuning discrete points from the third set based on the first preset strategy includes: The third set is input into the particle swarm optimization model, and all the fine-tuning discrete points of the third set are optimized by the particle swarm optimization model to obtain the fifth set; The fifth set is input into the Bayesian algorithm model, and the optimal tuning discrete points are obtained through the Bayesian algorithm model.

5. The method according to claim 4, characterized in that, The third set is input into the particle swarm optimization (PSO) model, which optimizes all the fine-tuned discrete points of the third set to obtain a fifth set, including: A1: Based on the second preset strategy, obtain the cross-unit learning factor of each optimized discrete point of the third set, and the cross-unit learning factor is used to characterize the applicability of each optimized discrete point of the third set in the adjacent target area. A2: Set an encoding for each of the optimization discrete points in the third set, the encoding including environmental features and cross-unit learning factors; A3: Deploy the optimized discrete points of the third set on the crop growth simulation model, input the second historical driving data, and obtain the second prediction result; A4: Calculate the third prediction error value between the second prediction result and the second historical true result; A5: In the third set, the optimized discrete point with the smallest third prediction error value is selected as the first optimal discrete point; among all the optimized discrete points that have undergone the cross-interchange, the optimized discrete point with the smallest third prediction error value is selected as the second optimal discrete point; and among the other optimized discrete points in the third set, excluding all the optimized discrete points that have undergone the cross-interchange, the optimized discrete point with the smallest third prediction error value is selected as the third optimal discrete point. A6: Based on the preset speed update function, determine the current speed according to the first optimal discrete point, the second optimal discrete point, the third optimal discrete point and the code; A7: Update the encoding based on the current speed; A8: Repeat A1-A7 a preset number of times to obtain the optimized discrete points.

6. The method according to claim 5, characterized in that, The speed update function is: ; Indicates the current speed; Indicates the previous current speed; Indicates the inertia weighting coefficient; These represent self-learning factors, local learning factors, and cross-unit learning factors, respectively. , , These represent self-learning random numbers, local learning random numbers, and cross-unit learning random numbers, respectively. , and These are the codes for the first optimal discrete point, the second optimal discrete point, and the third optimal discrete point, respectively. This represents the current encoding of the optimized discrete points of the third set.

7. A yield prediction method based on a crop growth simulation model, characterized in that, The method includes: Acquire the target area, the environmental characteristics of the target area, and driving data, including predicted climate data; Based on the target area and environmental characteristics, the optimal tuning discrete point is obtained according to the method described in any one of claims 1-6; The optimized discrete points are deployed on the crop growth simulation model to obtain the optimized crop growth simulation model. The driving data is input into the optimized crop growth simulation model, and the predicted yield is obtained through the optimized crop growth simulation model.

8. An electronic device, characterized in that, The electronic device includes a processor and a memory coupled together, the memory storing a computer program that, when executed by the processor, causes the electronic device to perform the method as described in any one of claims 1-6, or to perform the method as described in claim 7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when run on a computer, causes the computer to perform the method as described in any one of claims 1-6, or to perform the method as described in claim 7.

Images