Method for forecasting effective rainfall in rice field based on physical mechanism-data driven hybrid model

Abstract

This invention discloses a method for forecasting effective rainfall in paddy fields based on a physical mechanism-data-driven hybrid model. First, it acquires historical long-series datasets of the target paddy field area and public weather forecast data for the future forecast period. Then, it constructs a time-series set of historical feature parameters, aligning it with measured effective rainfall to form training samples. A machine learning model is used to train the mapping relationship between "physical feature parameters and effective rainfall" to obtain a paddy field effective rainfall forecasting model. Finally, during the forecast period, based on public weather forecast data, the paddy field surface water layer conditions on the forecast start date, and the corresponding growth period information, a future feature vector is constructed, input into the forecast model, and corrected by physical rationality constraints to output the daily forecast value of effective rainfall in the paddy field. This invention introduces the constraint of paddy field water process mechanism into data-driven modeling, improving the stability and cross-scenario applicability of effective rainfall forecasts, and providing effective rainfall forecasting services for digital twin irrigation districts or smart irrigation platforms.

Description

Technical Field

[0001] This invention belongs to the field of agricultural hydrological process forecasting, smart irrigation and digital twin irrigation district technology, and in particular relates to a method for forecasting effective rainfall in paddy fields based on a physical mechanism-data driven hybrid model. Background Technology

[0002] In irrigation management and scheduling at the irrigation district scale, rainfall is one of the most important natural sources of replenishment. For paddy fields, due to the presence of water layers on the field surface, the constraints of field ridges, and continuous seepage and evapotranspiration, rainfall is not equivalent to the "effective rainfall" that can be used to replenish the water balance of paddy fields. Effective rainfall is closely related to factors such as the water storage capacity of the field surface, the timing of rainfall, seepage conditions, evapotranspiration water requirements at the crop growth stage, and potential runoff outflow.

[0003] In current practices, irrigation management often uses empirical reduction factors or general empirical formulas to convert rainfall into effective rainfall. These methods typically do not adequately consider key processes such as the initial water layer state on the field surface, differences in crop growth stages, seepage losses, and runoff gating, making it difficult to meet the needs of digital twin irrigation districts for "short-term forecasting—rolling updates—fine-grained scheduling."

[0004] On the other hand, public weather forecasts can provide meteorological elements such as daily rainfall for the next few days. However, due to the coexistence of forecast errors and the nonlinear amplification effect of paddy field processes, directly using forecast rainfall as a substitute for effective rainfall will lead to significant bias. If only a pure data-driven model is used for fitting, although a certain accuracy can be achieved on local samples, the model is prone to learning statistical relationships specific to a particular site or year, and the generalization is unstable when crossing years, growth stages, or management scenarios (different field ridge heights, different seepage levels, different initial water layers).

[0005] Therefore, there is an urgent need for a technology that can stably predict effective rainfall in paddy fields, driven by public weather forecast data. Summary of the Invention

[0006] The purpose of this invention is to overcome the problems of insufficient characterization of physical processes, sensitivity to public weather forecast errors, and weak cross-scenario generalization ability in existing methods for estimating / forecasting effective rainfall in paddy fields. This invention provides a method for forecasting effective rainfall in paddy fields based on a physical mechanism-data-driven hybrid model. This method can both use physical mechanisms such as water balance and seepage / evapotranspiration to constrain the problem and use machine learning to comprehensively correct complex nonlinearities and multi-source errors, thereby meeting the operational application requirements of digital twin irrigation districts.

[0007] To achieve the above objectives, the present invention adopts the following technical solution: A method for forecasting effective rainfall in paddy fields based on a physical mechanism-data-driven hybrid model is provided, comprising the following sub-steps:

[0008] (1) Data acquisition and preprocessing: acquire historical long-series datasets of the target paddy field area, acquire public weather forecast data for the future forecast period; and perform missing data processing, anomaly removal, time scale unification and feature encoding processing on the historical long-series datasets and public weather forecast data;

[0009] (2) Calculation of physical characteristic parameters and construction of dataset: Based on the historical long series dataset, calculate the physical characteristic parameters that characterize the water consumption and water storage status of paddy fields; construct a time series set of historical characteristic parameters together with the physical characteristic parameters, historical meteorological data and rice growth period data; and align the time series set of historical characteristic parameters with the corresponding measured effective rainfall of paddy fields on a daily basis to form a model training sample set;

[0010] (3) Construction and training of hybrid forecast model: The model is trained by taking the feature parameters in the training sample set as input and the measured effective rainfall in paddy fields as output target, and the trained effective rainfall forecast model in paddy fields is obtained.

[0011] (4) Forecast of future effective rainfall: For the future target forecast period, based on the public weather forecast data, the growth period information corresponding to the forecast period and the water depth of the paddy field on the forecast start date, the corresponding future feature vector is calculated; the future feature vector is input into the trained paddy field effective rainfall forecast model, so as to output the daily paddy field effective rainfall forecast value for the future target forecast period.

[0012] Specifically, in step (2), the historical meteorological data is obtained from the official meteorological data network, which is the daily historical meteorological observation data of the corresponding stations, including the daily maximum temperature, minimum temperature, average temperature, wind speed, sunshine duration and average relative humidity; this data is used to calculate the reference crop evapotranspiration.

[0013] Specifically, in step (1), the rice growth period data is determined based on agricultural meteorological observation standards and local planting history or accumulated temperature conditions, including sowing period, seedling emergence period, three-leaf stage, transplanting period, greening period, tillering period, jointing period, booting period, heading period, milk stage and maturity period. The time range of each growth stage is determined based on local rice planting history or accumulated temperature conditions, and the corresponding growth period code is constructed.

[0014] Specifically, in step (2), the measured effective rainfall in the paddy field is obtained through paddy field water balance observation at the irrigation test station, and the water balance equation is:

[0015] ;

[0016] In the formula: P e For effective rainfall, For rainfall, Surface runoff, This refers to the depth of leakage. This represents the change in water level in the field.

[0017] Furthermore, the physical characteristic parameters in step (2) are composed of five physical characteristic components, specifically including: rainfall component, actual evapotranspiration component, deep infiltration component, field surface water storage capacity component, and growth stage component; the rainfall component adopts historical measured daily rainfall, the actual evapotranspiration component adopts actual evapotranspiration calculated based on historical meteorological data, the deep infiltration component is derived or estimated based on historical field surface water layer conditions and soil parameters, the field surface water storage capacity component is obtained by subtracting the effective height of the historical field ridge from the measured water layer depth on the corresponding date, and the growth stage component adopts historical real growth stage coding.

[0018] Further, in step (2), the actual evapotranspiration calculated based on historical meteorological data is obtained by multiplying the reference crop evapotranspiration by the Penman-Monteith formula and then by the corresponding rice crop coefficient.

[0019] Specifically, in step (3), the machine learning model is a gradient boosting decision tree regression model, which is XGBoost, LightGBM or CatBoost.

[0020] Furthermore, in step (3), to ensure the water balance mechanism, the predicted output value is adjusted during the training of the machine learning model. Apply nonnegativity constraint penalties, or perform physical plausibility constraint corrections during the post-processing phase of model inference: And mandatory satisfaction ;in, The final output is the effective rainfall forecast value. Enter the total rainfall value for the corresponding date into the rainfall component.

[0021] Furthermore, in the future forecast stage of step (4), according to the calculation rules of the five isomorphic physical feature components of the forecast stage defined in step S2, the rainfall component, actual evapotranspiration component, deep infiltration component, field surface water storage capacity component, and growth stage component corresponding to the future target forecast period are determined respectively, and a future feature vector is constructed accordingly, wherein:

[0022] The rainfall component uses the future forecast rainfall, which is obtained from the median of the rainfall range corresponding to the weather type in the public weather forecast;

[0023] The actual evapotranspiration component uses the actual evapotranspiration calculated based on future meteorological forecast data.

[0024] The deep seepage component is derived from the seepage rate based on the hydrological conditions of the forecast start date;

[0025] The water storage capacity of the field surface is obtained by subtracting the effective height of the field ridges and the depth of the water layer on the field surface on the forecast start date.

[0026] The reproductive stage component uses the reproductive period code corresponding to the forecast period.

[0027] Specifically, the actual evapotranspiration calculated based on future meteorological forecast data is obtained by multiplying the reference crop evapotranspiration by the Hargreaves-Samani method and then by the corresponding rice crop coefficient.

[0028] The present invention has the following beneficial effects:

[0029] 1) Enhanced Mechanism Constraints: By introducing physical process parameters such as evapotranspiration, seepage, and field water storage, the model learns a more reasonable mapping relationship under the meaning of water balance, reducing the risk of overfitting by "relying solely on statistical fitting".

[0030] 2) Improved forecast availability: Driven by public weather forecasts, it outputs the effective daily rainfall for the future, which can directly serve short-term irrigation forecasting and scheduling, and is compatible with the rolling update mechanism of digital twin irrigation districts.

[0031] 3) Enhanced generalization across scenarios: Information such as reproductive stage, initial water layer and water storage capacity of field ridges are explicitly encoded as input features, and the model can maintain more stable forecast performance in different years, different reproductive stages and different management conditions.

[0032] 4) Easy to expand and deploy: The machine learning model can use various algorithms such as gradient boosting trees and neural networks. It can be deployed on site or on platform system. It can also expand the deviation correction and rolling update mechanism for public weather forecasts, which is convenient for business deployment. Attached Figure Description

[0033] Figure 1 This is a flowchart of the steps in the method for forecasting effective rainfall in paddy fields according to the present invention;

[0034] Figure 2 This is the overall technical roadmap for the method of forecasting effective rainfall in paddy fields according to the present invention. Detailed Implementation

[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are only for explaining the present invention and are not intended to limit the scope of protection of the present invention. Those skilled in the art can make equivalent substitutions for the order of steps, feature sets, model forms, and parameter settings without departing from the principles of the present invention, and all such substitutions should fall within the scope of protection of the present invention.

[0036] This invention provides a method for forecasting effective rainfall in paddy fields based on a physical mechanism-data-driven hybrid model. The steps are as follows: Figure 1 and Figure 2 As shown, the specific implementation steps are as follows:

[0037] S1. Data Acquisition and Preprocessing

[0038] Obtain a long-term historical dataset of the target paddy field area, which includes at least measured effective rainfall in paddy fields, historical meteorological data, and rice growth period data.

[0039] The historical meteorological data includes daily maximum temperature, minimum temperature, average temperature, wind speed, sunshine duration, and average relative humidity. It is obtained from the China Meteorological Data Network (http: / / data.cma.gov.cn) and is used to calculate the reference crop evapotranspiration ET0.

[0040] The rice growth period data are determined according to the "Agricultural Meteorological Observation Standard for Rice", including the sowing period, seedling emergence period, three-leaf stage, transplanting period, greening period, tillering period, jointing period, booting period, heading period, milk stage, and maturity period. The morphological characteristics of the main growth periods of rice are shown in Table 1. The time range of each growth stage is determined according to the local rice planting history or accumulated temperature conditions, and the corresponding growth period codes are constructed.

[0041] Table 1 ; The aforementioned long-term historical data series and future forecast public weather data underwent unified preprocessing, including missing data imputation, outlier removal, time scale standardization, and data alignment. Missing data was imputed using interpolation between adjacent time periods, historical averages, or regression estimation methods based on relevant meteorological elements. Outlier data was identified and removed through physical rationality constraints and water balance consistency checks. Subsequently, the multi-source data were uniformly converted to a daily scale, and data alignment was completed using dates as indexes.

[0042] Furthermore, discontinuous variables such as the rice growth period are transformed into numerical features to meet the input requirements of machine learning models. Since public weather forecast data is typically provided in the form of qualitative weather types (such as "light rain" or "moderate to heavy rain"), this invention adopts the Predicted Default Rainfall (PDR) method. Referring to the China Meteorological Administration standard (China Meteorological Administration, 2012), the median of the rainfall intervals corresponding to various rainfall levels is used as the quantitative rainfall input, thereby constructing a daily rainfall sequence for the future forecast period, providing driving data for subsequent effective rainfall prediction in paddy fields. Specifically:

[0043] Select the target paddy field experimental area (which can be a standard field or a group of representative fields in an irrigation district experimental station) as the research object, and obtain daily data for at least 3 irrigation seasons, including: daily rainfall, temperature, humidity, wind speed, sunshine / radiation, air pressure and other meteorological elements; division of rice growth period and key dates (transplanting, tillering, jointing, heading, maturity, etc.).

[0044] The water balance of paddy fields was monitored to obtain daily surface runoff, deep infiltration, and changes in the field water layer; then, the water balance equation was applied. The daily effective rainfall was calculated. ;in, For rainfall, Surface runoff, This refers to the depth of leakage. The change in field water level is defined as follows: preferably, the surface runoff can be obtained by measuring water at the field outlet or monitoring the outflow from the drainage gate; the deep infiltration can be obtained by a seepage meter, seepage bucket, or based on water balance inversion; the change in field water level can be observed by a water level gauge, pressure water level sensor, or manual water gauge. To ensure consistency, all the above quantities are uniformly converted to equivalent water layer depth (mm).

[0045] The above preprocessing of public weather forecast data, namely the quantitative conversion of weather type to rainfall, yields daily quantitative rainfall forecasts. As input for the subsequent step S4.

[0046] S2. Calculation of physical characteristic parameters and construction of historical characteristic parameter time series sets

[0047] To ensure the consistency of the input interface of the machine learning model and to inject prior physical knowledge, this invention defines an input feature vector consisting of five physical feature components. The specific quantitative calculation and construction methods are as follows:

[0048] (1) Rainfall components This represents the dynamic term of water input. During the training phase, historical measured daily rainfall is used for back-substitution; during the forecasting phase, the forecasted rainfall obtained from step S1 is used. .

[0049] (2) Actual evaporation amount ET c This represents crop physiological water consumption. Considering that historical meteorological observation data is relatively comprehensive, while public weather forecasts generally only include weather type and maximum and minimum temperatures, this method employs a dual-mechanism calculation:

[0050] Training phase: Calculate the reference crop evapotranspiration ET0 using the Penman–Monteith method, and assign a dynamic crop coefficient K based on the growth stage.c The actual evapotranspiration of rice, ET, was obtained. c = ET0×K c Among them, K c It can be given in segments according to the reproductive stage or determined by continuous interpolation on a daily scale;

[0051] ;

[0052] in: _PM The reference crop evapotranspiration (mm·d⁻¹) is calculated using the Penman–Monteith method; Δ is the slope of the saturated vapor pressure curve (kPa·°C⁻¹). γ is the net radiation of the crop canopy (MJ·m⁻²·d⁻¹); G is the soil heat flux (MJ·m⁻²·d⁻¹), which is taken as 0 on a daily scale; γ is the wettable constant (kPa·°C⁻¹); T is the daily average air temperature (°C). The wind speed (m·s⁻¹) is at a height of 2m above the ground. The saturated vapor pressure is (kPa). This represents the actual water vapor pressure (kPa). The meteorological elements required for the calculation include daily maximum temperature, minimum temperature, average temperature, wind speed, sunshine duration, and average relative humidity.

[0053] Forecasting phase: The commonly used Hargreaves-Samani (HS) method is used, multiplied by the crop coefficient K. c The calculations are as follows, where the HS formula is used:

[0054] ;

[0055] Where: ET 0_HS The ET0 value is calculated using the HS method; the FAO (Food and Agriculture Organization of the United Nations) recommended values ​​for parameters C and E are 0.0023 and 0.5, respectively; R a Zenithal radiation (ZHV) can be calculated based on the geographical latitude and day number of the site, in MJ·m. -2 / d;T max and T min These are the highest and lowest temperatures, respectively, in °C.

[0056] (3) Component of deep seepage in paddy fields L th Determination of: Characterizing groundwater loss. It is derived using Darcy's law or local empirical formulas based on soil infiltration characteristics and surface water depth. Specifically, the deep infiltration rate in paddy fields can be calculated using Darcy's law or engineering empirical formulas depending on the available parameters: When detailed hydrogeological parameters such as soil saturated hydraulic conductivity, hydraulic head difference, and water flow path length are available, Darcy's law is used, and the expression is:

[0057] ;

[0058] In the formula: q is the leakage flux (mm / d). Δh is the soil saturated hydraulic conductivity (mm / d), Δh is the hydraulic head difference (mm), and L is the water flow path length (mm).

[0059] When detailed soil hydrological parameters are lacking and only rapid engineering calculations are required, empirical formulas are used for calculation. The empirical expression for the seepage rate is as follows:

[0060] ;

[0061] In the formula: F represents the leakage amount (mm) over a period of time. Let be the stable soil infiltration rate (mm / d), t be the leakage duration (d), and α be the groundwater depth correction coefficient; the calculation parameters include at least: soil saturated hydraulic conductivity, field water holding capacity, surface water layer depth, soil texture type, and groundwater depth; wherein, the stable soil infiltration rate The values ​​were selected according to the current national standard "Design Standard for Irrigation and Drainage Engineering" (GB 50288-2018), with values ​​ranging from 2 to 8 mm / d. Smaller values ​​are preferred for cohesive soils and larger values ​​are preferred for sandy soils. The values ​​were determined based on regional irrigation test data. This range has been verified in the field as the optimal range. Deviating from this range will lead to a significant increase in the error of effective rainfall calculation.

[0062] (4) Determination of the field surface water storage capacity (S) component: The field surface water storage capacity S is one of the model input features, used to characterize the field's ability to accommodate new rainfall under the initial forecast conditions. Based on the effective height h of the field ridge... b Calculate the surface water storage capacity based on the predicted water depth h0 on the initial date of the forecast. This is used as a characteristic of "field surface water storage capacity" to represent the remaining water storage space on the field surface. If the rainfall input on a future day exceeds the stated field surface water storage capacity and also exceeds the usable storage capacity released by evapotranspiration and seepage on that day, the excess portion is more likely to be converted into field surface overflow and discharge. Specifically, when the rainfall on that day is P, if... If so, it is determined that surface overflow and discharge occurred on that day; if If the overflow does not occur on that day, it is determined that no overflow will occur.

[0063] (5) The growth stage component is used to characterize the crop's water requirement at different growth stages. The growth stages include sowing, emergence, three-leaf stage, transplanting, greening, tillering, jointing, booting, heading, milk stage, and maturity stage, and the stage to which the target date belongs is determined according to the "Agricultural Meteorological Observation Standard for Rice". Preferably, the stage is coded using a stage number, with the sowing to maturity stage coded sequentially as 1 to 11; the training stage uses the actual growth stage code corresponding to the historical date, and the forecast stage uses the family planning stage code corresponding to the future date, as one of the model input features.

[0064] During the model training preparation phase, using historical date t as the time anchor, the five physical characteristic components mentioned above are calculated and arranged chronologically on a daily basis. Among them, the rainfall component uses historically measured daily rainfall, the actual evapotranspiration component uses actual evapotranspiration calculated based on historical meteorological data, the deep seepage component uses seepage derived or estimated based on historical field surface water layer conditions and soil parameters, the field surface water storage capacity component uses the water storage capacity calculated based on the historical effective height of the field ridge and the corresponding date's measured field surface water layer depth, and the growth period component uses the actual growth stage code corresponding to the historical date.

[0065] Furthermore, the five components corresponding to each historical date are aligned by day and arranged in chronological order to construct a historical feature parameter time series set. The historical feature parameter time series set can be represented as a time series set composed of input feature vectors X(t) on consecutive dates, where the input feature vector at each time t includes rainfall, actual evapotranspiration, deep infiltration, surface water storage capacity, and growing season components.

[0066] Simultaneously, based on the paddy field water balance observation data in step S1, a label for the measured effective rainfall in the paddy field is constructed. Specifically, based on daily rainfall, surface runoff, deep infiltration, and changes in the field water layer, the measured effective rainfall in the paddy field corresponding to each historical date is calculated according to the paddy field water balance relationship, and this is used as a supervised learning label sequence. In the supervised learning label sequence, the measured effective rainfall in the paddy field corresponding to each historical date is denoted as P. e .

[0067] Subsequently, the historical feature parameter time series set and the measured effective rainfall label sequence in paddy fields are aligned daily to form supervised learning sample pairs, thereby constructing the model training sample set. The specific correspondence between model input and output data is shown in Table 2.

[0068] Table 2 .

[0069] S3, Training and Validation of Hybrid Forecasting Model

[0070] S31. Model Training: After obtaining the model training sample set constructed in step S2, the input feature vector in the historical feature parameter time series set is used as the model input, and the measured effective rainfall in paddy fields on the corresponding date is used as the model output target to construct and train the effective rainfall forecast model for paddy fields.

[0071] The machine learning model is a regression model, preferably a gradient boosting decision tree regression model, which can be implemented using any one of XGBoost, LightGBM or CatBoost.

[0072] During model training, leave-one-out-of-year validation or cross-validation is used for parameter optimization and generalization ability assessment. Preferably, root mean square error, mean absolute error, correlation coefficient, and bias are used as model evaluation indicators to comprehensively evaluate the model's fitting accuracy and stability for effective rainfall in paddy fields.

[0073] S32. Physical Rationality Constraints: To ensure that the forecast results conform to the laws of agricultural hydrophysical systems, the forecast output values ​​can be constrained in the loss function during the training process. Apply a non-negative penalty term, or perform a physical plausibility correction during post-processing, with the correction formula as follows:

[0074] ;

[0075] And simultaneously satisfy:

[0076] ;

[0077] In the formula: The final output is the effective rainfall forecast value. This is the total rainfall value entered into the rainfall component for the corresponding date. This constraint mandates that the effective rainfall intercepted by the paddy field cannot exceed the actual total rainfall in any physical scenario.

[0078] S4, Future Effective Rainfall Forecast

[0079] For future target forecast periods (e.g., the next 1-7 days): obtain public weather forecast data for the future forecast period, information on the corresponding fertility period, and the water depth of the field surface on the forecast start date.

[0080] Following the calculation rules for the five isomorphic physical feature components of the forecast stage defined in step S2, the rainfall component, actual evapotranspiration component, deep infiltration component, field surface water storage capacity component, and growth stage component corresponding to the future target forecast period are determined respectively, and the future feature vector X is constructed accordingly. foreThe predicted future rainfall is obtained from the median of the rainfall range corresponding to the weather type in the public weather forecast. If the public weather forecast only provides rainfall information and lacks temperature information, the missing temperature information is supplemented based on the historical average meteorological data for the same period, and then the calculation of the future actual evapotranspiration component and the future feature vector X are completed. fore The construction.

[0081] The future feature vector X fore The data is input into the paddy field effective rainfall forecast model trained in step S3 to obtain the original predicted values ​​of effective rainfall for each date in the future target forecast period. Subsequently, physical rationality constraint corrections are applied to the original predicted values: when the original predicted value is less than 0, it is corrected to 0; when the original predicted value is greater than the total rainfall input into the rainfall component for the corresponding date, it is corrected to the total rainfall for the corresponding date; when the original predicted value is between 0 and the total rainfall for the corresponding date, the original predicted value remains unchanged. After the physical rationality constraint corrections, the final daily effective rainfall forecast values ​​are obtained, and the daily effective rainfall forecast results for the future target forecast period are output.

[0082] In summary, this invention, through a hybrid paradigm of "explicit construction of physical parameters + machine learning fitting," achieves predictable, interpretable, and coordinated scheduling of effective rainfall in paddy fields, making it suitable for platform-based deployment in digital twin irrigation districts. The method of this invention can be deployed at the single-plot level or in parallel computation at the irrigation district scale using plot groups as units, enabling coordinated forecasting of effective rainfall and irrigation scheduling across multiple plots.

[0083] The present invention has been described in detail with reference to the accompanying drawings and specific embodiments. However, the scope of protection of the present invention is not limited to the above embodiments. After reading this specification, those skilled in the art can make various equivalent substitutions, modifications, or improvements to the technical solutions of the present invention without departing from the scope of the present invention. All such equivalent substitutions, modifications, or improvements should fall within the scope of protection of the present invention.

Claims

1. A method for forecasting effective rainfall in paddy fields based on a physical mechanism-data-driven hybrid model, characterized in that, Includes the following steps: (1) Data acquisition and preprocessing: acquire historical long-series datasets of the target paddy field area, acquire public weather forecast data for the future forecast period; and perform missing data processing, anomaly removal, time scale unification and feature encoding processing on the historical long-series datasets and public weather forecast data; (2) Calculation of physical characteristic parameters and construction of dataset: Based on the historical long series dataset, calculate the physical characteristic parameters that characterize the water consumption and water storage status of paddy fields; The physical characteristic parameters, historical meteorological data, and rice growth period data are used to construct a time series set of historical characteristic parameters; and the time series set of historical characteristic parameters is aligned with the corresponding measured effective rainfall in paddy fields on a daily basis to form a model training sample set. (3) Construction and training of hybrid forecast model: The model is trained by taking the feature parameters in the training sample set as input and the measured effective rainfall in paddy fields as output target, and the trained effective rainfall forecast model in paddy fields is obtained. (4) Forecast of future effective rainfall: For the future target forecast period, based on the public weather forecast data, the growth period information corresponding to the forecast period and the water depth of the paddy field on the forecast start date, the corresponding future feature vector is calculated; the future feature vector is input into the trained paddy field effective rainfall forecast model, so as to output the daily paddy field effective rainfall forecast value for the future target forecast period.

2. The method according to claim 1, characterized in that, In step (2), the historical meteorological data is obtained from the official meteorological data network, specifically including the daily highest temperature, lowest temperature, average temperature, wind speed, sunshine duration and average relative humidity; this data is used to calculate the evapotranspiration of the reference crop.

3. The method according to claim 1, characterized in that, In step (1), the rice growth period data is determined based on agricultural meteorological observation standards and local planting history or accumulated temperature conditions, including sowing period, seedling emergence period, three-leaf stage, transplanting period, greening period, tillering period, jointing period, booting period, heading period, milk stage and maturity period. The time range of each growth stage is determined based on local rice planting history or accumulated temperature conditions, and the corresponding growth period code is constructed.

4. The method according to claim 1, characterized in that, In step (2), the measured effective rainfall in the paddy field is obtained through paddy field water balance observation at the irrigation experimental station, and the water balance equation is: ； In the formula: P e For effective rainfall, For rainfall, Surface runoff, This refers to the depth of leakage. This represents the change in water level in the field.

5. The method according to claim 1, characterized in that, The physical characteristic parameters in step (2) are composed of five physical characteristic components, specifically including: rainfall component, actual evapotranspiration component, deep infiltration component, field surface water storage capacity component, and growth stage component; the rainfall component adopts historical measured daily rainfall, the actual evapotranspiration component adopts actual evapotranspiration calculated based on historical meteorological data, the deep infiltration component is derived or estimated based on historical field surface water layer conditions and soil parameters, the field surface water storage capacity component is obtained by subtracting the effective height of the historical field ridge from the measured water layer depth on the corresponding date, and the growth stage component adopts historical real growth stage coding.

6. The method according to claim 5, characterized in that, In step (2), the actual evapotranspiration calculated based on historical meteorological data is obtained by multiplying the reference crop evapotranspiration by the Penman-Monteith formula and then by the corresponding rice crop coefficient.

7. The method according to claim 1, characterized in that, In step (3), the machine learning model is a gradient boosting decision tree regression model, which is XGBoost, LightGBM or CatBoost.

8. The method according to claim 1, characterized in that, In step (3), to ensure the water balance mechanism, the predicted output value is adjusted during the training of the machine learning model. Apply nonnegativity constraint penalties, or perform physical plausibility constraint corrections during the post-processing phase of model inference: And mandatory satisfaction ;in, The final output is the effective rainfall forecast value. Enter the total rainfall value for the corresponding date into the rainfall component.

9. The method according to claim 5, characterized in that, In the future forecast stage of step (4), according to the calculation rules of the five isomorphic physical feature components of the forecast stage defined in step S2, the rainfall component, actual evapotranspiration component, deep infiltration component, field surface water storage capacity component, and growth stage component corresponding to the future target forecast period are determined respectively, and a future feature vector is constructed accordingly, where: The rainfall component uses the future forecast rainfall, which is obtained from the median of the rainfall range corresponding to the weather type in the public weather forecast; The actual evapotranspiration component uses the actual evapotranspiration calculated based on future meteorological forecast data. The deep seepage component is derived from the seepage rate based on the hydrological conditions of the forecast start date; The water storage capacity of the field surface is obtained by subtracting the effective height of the field ridges and the depth of the water layer on the field surface on the forecast start date. The reproductive stage component uses the reproductive period code corresponding to the forecast period.

10. The method according to claim 9, characterized in that, The actual evapotranspiration calculated based on future meteorological forecast data is obtained by multiplying the reference crop evapotranspiration calculated using the Hargreaves-Samani method by the corresponding rice crop coefficient.

Images