Agrometeorological decision service system and method based on artificial intelligence large model

Abstract

The invention relates to the technical field of machine learning, in particular to an agricultural meteorological decision service system and method based on an artificial intelligence large model, and the system comprises a meteorological coupling analysis module, an abnormal traceability positioning module, an influence path modeling module, a dynamic scoring matrix module and a response decision generation module. According to the method, meteorological parameters are divided through a time sliding window, crop growth deviation is quantified in combination with a long and short-term memory network, dynamic modeling of meteorological fluctuation and growth response is achieved, 72-hour meteorological data are collected through reverse-sequence window backtracking, a forest is isolated to detect extreme value offset characteristics, and crop index sudden drop inducements are positioned. A lagging mapping relation is established between lagging correlation coefficients and random forest weight sorting, the time sequence influence of key factors is determined, minimum and maximum normalized superposition hierarchical clustering coding and multi-dimensional scoring are converted into hierarchical decision basis, causal association and decision timeliness of meteorological risk early warning are enhanced, and the extreme weather misjudgment probability is reduced.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to an agricultural meteorological decision-making service system and method based on an artificial intelligence large model. Background Art

[0002] Machine learning, centered around data-driven algorithms and model training, achieves prediction, classification, and other functions by learning data patterns. It encompasses supervised learning, unsupervised learning, and reinforcement learning models, and involves data preprocessing, feature extraction, model evaluation, and optimization. It is widely used in industries such as agriculture, healthcare, and finance, playing a particularly critical role in multi-source information and dynamic environments.

[0003] The agricultural meteorological decision-making service system uses machine learning to analyze historical meteorological data and crop parameters to assess and predict meteorological risks. The system includes mechanisms such as data collection, crop growth period classification, and meteorological factor sensitivity matching. It builds a rule-driven and sample-optimized early warning database to provide time recommendations and meteorological warnings for agricultural activities.

[0004] Existing technologies rely on static thresholds and one-way reasoning, making them difficult to cope with real-time weather fluctuations. They cannot accurately warn when extreme weather events exceed historical data. Models also lack the ability to trace anomalies, making it difficult to track sudden data changes, hindering precise adjustments. Rule matching ignores the lag in weather impacts, making it easy to misjudge the timing of interventions. Fixed classification methods are also difficult to adapt to complex weather environments. Summary of the Invention

[0005] The purpose of the present invention is to solve the shortcomings of the existing technology and propose an agricultural meteorological decision-making service system and method based on an artificial intelligence large model.

[0006] In order to achieve the above objectives, the present invention adopts the following technical solutions: The agricultural meteorological decision-making service system based on the artificial intelligence large model includes: The meteorological coupling analysis module is used to obtain meteorological time series and observation index data during the crop growth period. The daily average temperature, relative humidity, daily precipitation, and solar radiation intensity are input into the time sliding window partitioning algorithm to divide the equal-length segments. The long-short-term memory network is used to calculate the deviation sequence between meteorological parameters and daily crop growth within the window. The sliding window scoring data is generated and passed to the dynamic scoring matrix module and the response decision generation module. An anomaly tracing and positioning module is used to trigger reverse window backtracking based on the sudden drop in crop indicators, collect 72 hours of meteorological data, call the isolation forest algorithm to detect the extreme value offset, extract meteorological anomaly feature data and pass it to the impact path modeling module; An impact path modeling module is used to receive the meteorological anomaly characteristic data, calculate the lag time correlation coefficient, input the lag term weight factor into the random forest model ranking contribution, generate path mapping data and pass it to the response decision generation module; The dynamic scoring matrix module is used to receive the sliding window scoring data, perform minimum and maximum normalization processing, call the hierarchical clustering algorithm for discretization coding, generate hierarchical coding data and pass it to the response decision generation module.

[0007] As a further solution of the present invention, the sliding window scoring data includes the temperature change rate, humidity fluctuation coefficient, radiation accumulation and precipitation distribution discreteness within the window; the meteorological anomaly characteristic data specifically includes the anomaly intensity index, offset period identifier, and mutation meteorological parameter type; the path mapping data includes the lag time step, contribution ranking list, and path correlation matrix; the level coding data specifically refers to the normalized scoring interval, cluster discrete level, and coding mapping table.

[0008] As a further solution of the present invention, the meteorological coupling analysis module includes: The meteorological segmentation processing submodule obtains the daily average temperature, relative humidity, daily precipitation, and solar radiation intensity. Based on the time sliding window partitioning algorithm, it sets a fixed-length interval and divides the meteorological parameters into segments of the same number of days in a continuous time sequence. The multiple meteorological parameters in the segment are serialized and aligned with the corresponding date to generate a meteorological sequence of equal length. The growth deviation calculation submodule calls the four meteorological parameters of the equal-length meteorological sequence on a daily basis, combines the long-short-term memory network hidden layer state, calculates the daily crop growth forecast value, extracts the actual crop growth value in the observation data, calculates the difference between the forecast value and the actual value on a daily basis, takes the absolute value of the difference sequence and accumulates and sums them to generate the daily growth deviation; The window score generation submodule extracts the daily growth deviation, sums the deviations in a fixed period, sets the deviation interval between the lower limit and the upper limit, and compares the deviation sum with the lower limit and the upper limit of the interval in turn. If the sum is within the interval, the corresponding score is output; if it is lower than the lower limit or higher than the upper limit, the corrected score is output. The multi-period score results are arranged in chronological order to generate sliding window score data.

[0009] As a further solution of the present invention, the abnormality tracing and positioning module includes: The abnormal trigger backtracking submodule monitors the rate of change of crop index values ​​and calculates the average rate of change over three consecutive days. When the single-day drop exceeds a set multiple of the average rate of change, it is marked as a sudden drop node. A fixed-length window is defined in reverse order from the sudden drop node as the end point. At the same time, hourly temperature, humidity, and wind speed data within the window are collected to generate an abnormal backtracking window. The offset detection and analysis submodule calls the hourly temperature, humidity, and wind speed data in the abnormal backtracking window, calculates the deviation of the multi-parameter value from the mean within the window, marks the time period where the deviation exceeds the set threshold as a candidate offset point, calculates the time distribution density of the candidate offset points, selects the time period where the distribution density exceeds the benchmark value, and generates the offset extreme value point; The feature extraction and transfer submodule calls the offset extreme point, extracts the difference between the maximum and minimum temperature fluctuations as the fluctuation range, counts the total number of time periods in which the humidity parameter continuously exceeds the deviation threshold, calculates the number of sudden changes in wind speed that exceed the benchmark value in adjacent hours, and arranges the three feature values ​​in chronological order, matches the deviation range with the sudden drop node time, and generates meteorological anomaly feature data.

[0010] As a further solution of the present invention, the impact path modeling module includes: The lag correlation submodule obtains the temperature deviation value and precipitation accumulation in the meteorological anomaly characteristic data, calculates the Pearson correlation coefficient between them and the probability of disaster occurrence within a predefined lag time range, arranges them into a row and column structure matrix according to the time dimension, and generates a lag correlation matrix; The weight factor ranking submodule calls the lagged correlation matrix and performs range normalization on each lagged weight factor using the formula: ; The spatiotemporal correction weight values ​​are obtained by calculation and input into the random forest model to perform iterative feature sorting of the Gini coefficient to generate a contribution sequence; in, Representative meteorological characteristics In the lag time The weight factor, Meteorological characteristics The abnormal fluctuation range, Lag time The Pearson correlation coefficient, is the time window benchmark length, Lag time In the space node The dispersion of the distribution, Characterized by At the node the intensity of cross-influence; The path integration submodule calls the contribution sequence, screens the lag term weight factors whose contributions are higher than the dynamic screening threshold, maps them to geographic grid coordinates, overlays the multi-dimensional meteorological feature contribution distribution, and generates path mapping data.

[0011] As a further solution of the present invention, the dynamic scoring matrix module includes: The score normalization submodule obtains the sliding window score data, extracts the maximum and minimum scores in multiple windows, and calculates the dynamic volatility based on the average of the absolute values ​​of the differences between adjacent scores in the window, using the formula: ; Normalize the denominator and dynamically compensate the numerator of each scoring item, integrate the results of all scoring items in the window, and generate normalized scoring data; in, Representative Window No. The standardized results of the item scores, is the original score value, is the maximum score within the window, is the minimum score, is the mean of the absolute values ​​of the adjacent score differences within the window, is the window score mean.

[0012] The clustering submodule calls the normalized score data, calculates the Euclidean distance between score items, constructs a symmetric distance matrix, sets the inter-cluster merging threshold according to the average difference value of the scores of adjacent windows, completes hierarchical clustering by successively merging the clusters with the closest distance, and generates a clustering result; The discrete coding submodule calculates the interquartile range and median of the score distribution in each cluster based on the clustering grouping results, assigns grade coding numbers in ascending order of the median size, maps the original scores to the corresponding coding categories, and generates grade coding data.

[0013] As a further embodiment of the present invention, the system further comprises: a response decision generation module, configured to receive the sliding window scoring data, the grade coding data, and the path mapping data, splice a multi-dimensional feature vector into a multi-objective optimization model, and output a decision instruction set; The decision instruction set includes optimization weight configuration, priority instruction queue, and execution time node.

[0014] As a further solution of the present invention, the response decision generation module includes: The feature splicing and integration submodule calls the period score mean of the sliding window scoring data, the priority value of the grade coding data, and the node connectivity of the path mapping data, extracts the fields with overlapping timestamps in the three data items, adjusts the field values ​​to the same dimensional range, and splices them horizontally into a three-dimensional matrix according to the time dimension to generate a multidimensional feature vector; The optimization target parsing submodule allocates scoring weights, priority coefficients, and connectivity influencing factors based on the multidimensional feature vector according to preset rules, superimposes the three weights to calculate a comprehensive score, defines the upper limit of resource consumption as the maximum value of the comprehensive score, and the lower limit of node coverage as the minimum value of the comprehensive score, and screens the extreme score points that meet the upper and lower limit constraints to generate the optimization target weight; The instruction generation output submodule calls the optimization target weight, compares the weight distribution with the matching degree of the instruction triggering conditions in the policy library, verifies whether the weight fluctuation range in the time window is within the allowable threshold of the policy execution cycle, outputs a set of instruction numbers that meet the matching degree and fluctuation range, and generates a decision instruction set.

[0015] The agricultural meteorological decision-making service method based on the artificial intelligence big model is executed based on the agricultural meteorological decision-making service system based on the artificial intelligence big model, and includes the following steps: S1: Use agricultural meteorological monitoring equipment to obtain daily average temperature, relative humidity, daily precipitation, and solar radiation intensity data. Input these four meteorological parameters into a time sliding window partitioning algorithm to divide the time into equal-length segments. Based on the long-short-term memory network, the deviation sequence between the meteorological parameters and the daily crop growth in the window is modeled to generate sliding window scoring data. S2: Using crop indicator monitoring devices to obtain observation indicator data during the growth period, detecting sudden drop nodes triggers reverse backtracking, collecting meteorological data for the previous 72 hours, and using the isolation forest algorithm to identify anomalies such as temperature offsets, humidity mutations, and radiation drops in the meteorological data. The system extracts temperature fluctuation characteristics, humidity mutation characteristics, and radiation anomaly characteristics to generate meteorological anomaly feature data. S3: Based on the meteorological anomaly characteristic data, calculating the lag time correlation coefficients between the temperature fluctuation characteristics, humidity sudden change characteristics, and radiation anomaly characteristics and the crop growth indicators, inputting the lag term weight factors into the random forest model to rank the contribution of meteorological parameters, and generating path mapping data; S4: performing minimum and maximum normalization on the sliding window score data, calling a hierarchical clustering algorithm to discretize the normalized result into five level intervals, and generating level-coded data; S5: The sliding window scoring data, the grade coding data and the path mapping data are spliced ​​into a multi-dimensional feature vector, which is input into a multi-objective optimization model to perform strategy optimization on the irrigation amount, shading duration and fertilization frequency, and a decision instruction set is output.

[0016] Compared with the prior art, the advantages and positive effects of the present invention are: In the present invention, meteorological parameters are divided through a time sliding window, and crop growth deviations are quantified in combination with a long-short-term memory network to achieve dynamic modeling of meteorological fluctuations and growth responses. 72 hours of meteorological data are collected retrospectively through a reverse window, extreme value offset characteristics are detected through isolation forest, the causes of sudden drops in crop indicators are located, a lagged mapping relationship is established between the lagged correlation coefficient and the random forest weight ranking, the temporal influence of key factors is clarified, the minimum and maximum normalization is superimposed on the hierarchical clustering coding, and the multidimensional score is converted into a hierarchical decision-making basis, thereby enhancing the causal relationship and decision-making timeliness of meteorological risk warnings and reducing the probability of misjudgment of extreme weather. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a system flow chart of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0019] Example 1 See also Figure 1 The present invention provides a technical solution: an agricultural meteorological decision-making service system based on an artificial intelligence large model includes: The meteorological coupling analysis module is used to obtain meteorological time series and observation index data during the crop growth period. The daily average temperature, relative humidity, daily precipitation, and solar radiation intensity are input into the time sliding window partitioning algorithm to divide the equal-length segments. The long-short-term memory network is used to calculate the deviation sequence between meteorological parameters and daily crop growth within the window. The sliding window scoring data is generated and passed to the dynamic scoring matrix module and the response decision generation module. Daily mean temperature: the arithmetic mean of all daily observation temperatures, derived from meteorological observation station data; relative humidity: the ratio of the actual water content in the atmosphere to the saturated water content at the same temperature, expressed as a percentage; daily precipitation: the depth of precipitation received by the ground in one day, in millimeters; solar radiation intensity: the solar radiation energy received per unit area per unit time, in watts per square meter, and the data is generally obtained from meteorological stations or satellite remote sensing products. Time sliding window partitioning algorithm: Common algorithms such as the fixed-step sliding window method or the time scale-based segmentation method cut long time series data into equal-length intervals for analysis. Long short-term memory network: a structure of artificial neural network, particularly suitable for processing and predicting long time series data, which can capture long-term dependencies in the data; daily growth deviation series: the difference series between the actual measured daily growth of crops and the predicted value of the crop growth model, used to evaluate the impact of meteorological conditions on crop growth.

[0020] The anomaly tracing and positioning module is used to trigger reverse window backtracking based on the sudden drop in crop indicators, collect 72 hours of meteorological data, use the isolation forest algorithm to detect extreme value deviations, extract meteorological anomaly feature data, and pass it to the impact path modeling module; Isolation Forest Algorithm: An unsupervised anomaly detection algorithm based on a tree model. It rapidly identifies outliers by randomly selecting attributes and split values ​​to construct a tree structure. It is suitable for anomaly detection tasks in high-dimensional or time-series data. Extreme value excursions: Extreme value excursions in meteorological observation data that significantly deviate from historical normal ranges, such as drastic temperature fluctuations or sudden increases or decreases in precipitation, are detected using Isolation Forest Detection.

[0021] The impact path modeling module is used to receive meteorological anomaly feature data, calculate the lag time correlation coefficient, input the lag term weight factor into the random forest model ranking contribution, generate path mapping data, and pass the path mapping data to the response decision generation module; Lag Time Correlation Coefficient: This describes the time delay between the occurrence of a meteorological anomaly and the response of crop indicators. This is typically calculated using a correlation coefficient or cross-correlation function to determine the extent of the lagged impact of meteorological variables on crop indicators. Random Forest Model: This is an ensemble learning method comprised of multiple decision trees. It calculates the contribution of each input factor to the forecast to determine its importance and is used to identify the most critical meteorological factors in the impact path.

[0022] The dynamic scoring matrix module is used to receive the sliding window scoring data, perform minimum and maximum normalization processing, call the hierarchical clustering algorithm for discretization coding, generate level-coded data, and pass the level-coded data to the response decision generation module; Minimum-maximum normalization processing: One of the data preprocessing methods, which normalizes the original data to a specific interval through linear transformation to reduce the data differences between different dimensions; Hierarchical clustering algorithm: An unsupervised clustering algorithm that merges or splits data level by level by building a tree-like hierarchy, thereby achieving data classification or discrete coding and generating hierarchical and discrete level data.

[0023] The response decision generation module is used to receive sliding window scoring data, grade coding data, path mapping data, splice multi-dimensional feature vectors into the multi-objective optimization model, and output a decision instruction set.

[0024] Multi-objective optimization model: A mathematical decision-making model that simultaneously optimizes multiple conflicting objectives. Common methods include non-dominated sorting algorithms or strategies based on the best frontier, and ultimately output a set of optimized decision instructions to guide crop cultivation management strategies.

[0025] The sliding window scoring data includes the temperature change rate, humidity fluctuation coefficient, radiation accumulation and precipitation distribution dispersion within the window. The meteorological anomaly characteristic data specifically includes the anomaly intensity index, offset period identifier, and mutation meteorological parameter type. The path mapping data includes the lag time step, contribution ranking list, and path association matrix. The level coding data specifically refers to the normalized scoring interval, cluster discrete level, and coding mapping table. The decision instruction set includes optimization weight configuration, priority instruction queue, and execution time node.

[0026] The meteorological coupling analysis module includes: The meteorological segmentation processing submodule first obtains the meteorological data within a specific monitoring period. The specific operation is to retrieve the daily meteorological records of the target plot (plot number A01) from October 1, 2024 to December 31, 2024, a total of 92 days, from the meteorological monitoring database. The acquired data items include daily average temperature (unit: °C), daily average relative humidity (unit: %), daily cumulative precipitation (unit: mm), and daily average solar radiation intensity (unit: W / m²). Then, the time sliding window partitioning algorithm is applied to set a fixed time window length of 7 days. This window length is set based on the response cycle of crops (such as rice) to short-term meteorological changes and the trade-off between data processing efficiency. The 92 days of continuous meteorological data are divided in chronological order to generate The first segment covers the 1st to 7th day (2024-10-01 to 2024-10-07), the second segment covers the 2nd to 8th day (2024-10-02 to 2024-10-08), and so on, until the last segment covers the 86th to 92nd day (2024-12-25 to 2024-12-31). Within each 7-day segment, the four daily meteorological parameters are strictly serialized and aligned with the corresponding date to ensure that the record of each day contains the date and the corresponding four meteorological parameter values, forming a sequence with a unified structure. The data example of the first segment (2024-10-01 to 2024-10-07) is shown in Table 1 below.

[0027] Table 17-day sliding window meteorological data example table As shown in Table 1, the daily records of the four meteorological parameters in the first 7-day segment (2024-10-01 to 2024-10-07) are displayed. Through such processing, 86 meteorological sequences of equal length, each of which is 7 days in length, are finally generated.

[0028] The growth deviation calculation submodule calls the equal-length meteorological sequence generated in the previous step. Specifically, it extracts the daily meteorological data within each 7-day segment. Taking the data of the first segment (2024-10-01 to 2024-10-07) shown in Table 1 as an example, the four daily meteorological parameters in the sequence (such as 15.2°C, 65%, 0.0mm, 180.5W / m² on 2024-10-01; 16.1°C, 68%, 0.0mm, 190.2W / m² on 2024-10-02) and the corresponding date information are input into the pre-trained long-term short-term memory network structure on a daily basis. Combined with the hidden state of the network transmitted in the previous time step (the preset initial state is used on the first day of the sequence), the predicted value of the crop (set as rice) growth on that day is calculated through the input gate, forget gate, output gate operation and cell state update within the network. Here, the daily plant height growth (unit: Taking the prediction results as an example, the prediction results are as follows: 0.80 cm is predicted for 2024-10-01, 0.85 cm for 2024-10-02, 0.82 cm for 2024-10-03, 0.90 cm for 2024-10-04, 0.95 cm for 2024-10-05, 0.75 cm for 2024-10-06, and 0.70 cm for 2024-10-07. At the same time, the actual daily growth values ​​of rice plant height measured by field automation sensors on the corresponding dates (2024-10-01 to 2024-10-07) are retrieved from the crop growth monitoring database: 0.70 cm, 0.90 cm, 0.80 cm, 0.90 cm, 1.00 cm, 0.65 cm, and 0.60 cm. Next, the difference between the predicted value and the actual observed value (growth deviation) is calculated day by day: , , , , , , Centimeters, get the difference sequence [0.10, -0.05, 0.02, 0.00, -0.05, 0.10, 0.10]. Take the absolute value of each value in the difference sequence to get the absolute deviation sequence [0.10, 0.05, 0.02, 0.00, 0.05, 0.10, 0.10]. Finally, accumulate the absolute deviation values ​​for these 7 days and calculate the total daily growth deviation for this section (the first 7-day section): cm, generating the daily growth deviation for the window.

[0029] The window score generation submodule extracts the cumulative value of the daily growth deviation of each sliding window calculated in the previous step, and then calculates the deviation of the first window calculated as 0.42 cm. The deviations of the next four windows are calculated as 0.75 cm, 0.50 cm, 0.90 cm, and 0.80 cm, respectively, to form a sequence containing five window deviations [0.42, 0.75, 0.50, 0.90, 0.80]. The deviation sum is calculated once every five consecutive windows at a fixed period for periodic evaluation. The deviation sum of the first evaluation period is Centimeters, set an interval for evaluating the total deviation, and its lower and upper limits are determined based on historical data and expert experience. Specifically, the interval is set based on the statistical distribution of growth simulation deviations of similar crops (rice) at similar growth stages (tillering stage) during the same historical period (nearly five years). A range covering 80% of the historical deviation data is selected, and combined with the acceptable error set by agricultural experts of ±1.0 cm / evaluation cycle (corresponding to a total deviation of 2.0 cm), adjustments are made. Finally, the lower limit is determined to be 2.0 cm and the upper limit is 4.0 cm. The calculated total deviation (3.37 cm) is compared with this interval to determine whether 3.37 is greater than or equal to 2.0 and less than or equal to 4.0. Since 3.37 cm falls within the interval of [2.0,4.0] cm, the corresponding score is output according to the preset scoring rule. The scoring rule is designed as follows: the total deviation is within [2.0 ,4.0] cm interval is scored as 80 points; if the total deviation is less than 2.0 cm (for example, the total calculated in the second evaluation cycle is 1.8 cm), it indicates that the model prediction is too close to the actual or the growth is abnormally slow, and the output correction score is 60 points; if the total deviation is greater than 4.0 cm (for example, the total calculated in the third evaluation cycle is 4.5 cm), it indicates that the model prediction deviation is too large, and the output correction score is 50 points. This scoring rule is designed to reward prediction performance close to the historical average deviation level (80 points), distinguish between cases where the deviation is too small (60 points) and cases where the deviation is too large (50 points), and arrange the scores calculated in each evaluation cycle in chronological order to form a scoring sequence. The score of the first cycle is 80 points, the second cycle is 60 points, and the third cycle is 50 points, generating the sliding window score data [80,60,50,…] (subsequent scores are similar).

[0030] The abnormality tracing and positioning module includes: The abnormal trigger backtracking submodule continuously monitors key crop physiological indicators, selecting the leaf area index (LAI) as the monitoring object. The average LAI value of the rice canopy in plot A01 is collected daily, and the average LAI change rate for three consecutive days is calculated. The LAI values ​​on days t-3, t-2, and t-1 are 3.10, 3.15, and 3.20, respectively. The average daily change rate for these three days is calculated: , The average change rate of the three consecutive days is , that is, an average daily growth of 1.60%. The trigger multiple threshold for a single-day decline is set at 3 times. This multiple is based on the research results of known stress responses of this crop (rice) (such as acute water loss and low temperature stress). Severe stress usually causes the daily decline of growth indicators to exceed 3 times the standard deviation of the normal physiological fluctuation range. Here, it is simplified to 3 times the average change rate as a quick judgment basis to calculate the trigger threshold: On day t (2025-03-15), the LAI value was monitored to drop significantly from 3.20 on the previous day to 3.00. The daily drop was calculated as , which is 6.25%, comparing the single-day drop (6.25%) with the trigger threshold (4.80%), because , if the trigger condition is met, the tth day (2025-03-15) is marked as a sudden drop node, and the sudden drop node (2025-03-15 00:00) is used as the time end point. A fixed 72-hour lookback window is defined in reverse order, and the lookback time range is determined to be 2025-03-12 00:00 to 2025-03-14 23:59. At the same time, the hourly temperature (°C), relative humidity (%), and wind speed (m / s) data of the corresponding station of plot A01 within the 72-hour lookback window are accurately collected and sorted from the meteorological database to form a A time series dataset with data points is used to generate abnormal lookback window data.

[0031] The offset detection and analysis submodule calls 72 sets of hourly meteorological data (temperature, humidity, wind speed) within the abnormal backtracking window (2025-03-12 00:00 to 2025-03-14 23:59). First, the time average of each parameter in the 72-hour window is calculated: the average temperature is calculated to be 8.5°C, the average relative humidity is 75%, and the average wind speed is 2.1m / s. Then, for each hourly data point in the window, the deviation (absolute difference) between its parameter value and the corresponding parameter window mean is calculated. The data at the time point 2025-03-13 14:00 is selected for illustration. The measured temperature in this hour is 3.2°C, the humidity is 95%, and the wind speed is 5.5m / s. The deviation is calculated: the temperature deviation °C, humidity deviation , wind speed deviation m / s, set the deviation amplitude threshold for the three parameters. The setting of the threshold refers to the statistical fluctuation characteristics of the parameter in the same period of history. The specific setting process is: the temperature deviation threshold is set to 2 times the standard deviation (1.5°C) of the temperature at the specific hour (14:00) in the past five years in the month of the abnormal backtracking window (March). °C; the humidity deviation threshold is set to 2 times the humidity standard deviation (7.5%) of the corresponding period, and the calculated The wind speed deviation threshold is set to 2.08 times the standard deviation of wind speed (1.2m / s) in the corresponding period, and the value is 2.5m / s to better capture potential strong wind events. The hours with deviations exceeding the set thresholds are marked as candidate offset points for the corresponding parameters. For 2025-03-13 14:00: the temperature deviates from 5.3°C>3.0°C, the humidity deviates from 20%>15%, and the wind speed deviates from 3.4m / s>2.5m / s. Therefore, this hour is marked as a candidate offset point for temperature, humidity, and wind speed at the same time. Then, the distribution density of all candidate offset points in the time dimension is counted, and 6 hours is used as the statistical unit. The number of occurrences of candidate offset points of various parameters in each consecutive 6 hours is calculated, and the basis of the time distribution density is set. The benchmark value is the candidate offset points of the same parameter that appear at least three times within six consecutive hours. This benchmark value is set based on the characteristic that meteorological anomalies usually have a certain degree of persistence. It aims to filter out isolated offset points that may be noise. It was found that in the six-hour time period from 12:00 to 17:59 on March 13, 2025, the temperature candidate offset point appeared four times, the humidity candidate offset point appeared five times, and the wind speed candidate offset point appeared three times, all reaching or exceeding the set benchmark value three times. These time periods with high distribution density were screened out, and the time points or time periods with the most severe offset and the most concentrated simultaneous offset of multiple parameters were determined to generate offset extreme point information, and the key offset period was determined to be from 13:00 to 16:00 on March 13, 2025.

[0032] The feature extraction and transfer submodule uses the excursion extreme point information determined in the previous step, focusing on the key excursion period from 2025-03-13 13:00 to 16:00. During this 4-hour period, it extracts the maximum temperature record (5.5°C) and the minimum temperature record (3.0°C), and calculates the difference between the two as the temperature fluctuation range: °C, count the cases where the humidity parameter continuously exceeds its deviation threshold (15%), query the hourly humidity deviation data of this period, and find that the humidity deviation amplitude exceeds 15% in the three hours of 13:00, 14:00, and 15:00, but does not exceed 15% at 16:00. Therefore, there is a continuous threshold exceeding period, the total number of which is 1, and the duration is 3 hours. Calculate the absolute value of the change in wind speed in adjacent hours and count the number of times it exceeds the benchmark value. Set the benchmark value of wind speed hourly change to 1.5m / s. This benchmark value is based on the standard deviation of hourly wind speed changes in the same period in history (March) (0.7m / s) multiplied by a multiple (about 2.14) to identify rapid and significant changes in wind speed. Check the wind speed changes in adjacent hours in this period: the change from 13:00 (5.5m / s) to 14:00 (4.8m / s) m / s; Change from 14:00 (4.8m / s) to 15:00 (5.1m / s) m / s; Change from 15:00 (5.1m / s) to 16:00 (3.0m / s) m / s, and compare these changes with the baseline value of 1.5m / s: 0.7 < 1.5, 0.3 < 1.5, 2.1 > 1.5. Only the last change exceeds the baseline value, so the number of wind speed mutations is 1. Arrange the three extracted feature values ​​in the order of (extreme temperature fluctuation, duration of continuous humidity exceeding the threshold, number of wind speed mutations) to form a feature vector [2.5, 3, 1]. This feature vector and its corresponding time information (the center point of the abnormal period is recorded as 2025-03-13) are used to calculate the wind speed mutation frequency. The time deviation range of the previously marked crop index sudden drop node (2025-03-15) is matched, and it is confirmed that the abnormal meteorological event occurred within 1 to 3 days before the sudden drop node (2025-03-13 14:30 is about 1.4 days before 2025-03-15, which meets the condition). Finally, the meteorological anomaly feature data is generated: {time: 2025-03-13 14:30, feature: [2.5, 3, 1], associated node: 2025-03-15}.

[0033] The impact path modeling module includes: The lagged correlation submodule obtains specific values ​​from the meteorological anomaly feature data generated in the previous step, selects the temperature fluctuation extreme value of 2.5°C, and combines the meteorological records to obtain the precipitation accumulation within the abnormal period (2025-03-13 13:00-16:00), which is determined to be 15mm. At the same time, the historical occurrence probability data of specific disasters related to plot A01 (such as frost and bacterial blight) within the predefined lag time range (set to 0 days to 10 days) are retrieved from the disaster history database and risk assessment model. For the two features of temperature fluctuation extreme value (feature 1) and precipitation accumulation value (feature 2), the Pearson correlation coefficient between them and the probability sequence of disaster occurrence in the next 0 days, 1 day, and up to 10 days is calculated item by item. The calculation process involves collecting all identified similar feature values ​​in the historical data (such as the historical temperature fluctuation extreme value sequence) and their occurrence on the jth day ( ) corresponding to the probability sequence of disaster occurrence, the Pearson correlation coefficient calculation method is used to calculate the correlation coefficient for each pair (feature i, lag j days) , the calculation results are shown in Table 2 below.

[0034] Table 2. Lag correlation coefficients between meteorological characteristics and disaster probability As shown in Table 2, the Pearson correlation coefficients of the two characteristics of temperature fluctuation extremes and precipitation accumulation with the probability of disaster occurrence within the lag period of 0 to 10 days are listed. These coefficients are arranged according to the structure of characteristics (rows) and lag days (columns) to form a lag correlation matrix with 2 rows and 11 columns. The data of this matrix will be used for subsequent weight calculations.

[0035] The weight factor sorting submodule calls the lagged correlation matrix generated in the previous link (such as the data in Table 2) and calculates each meteorological feature according to the following formula In the lag time The spatiotemporal correction weight value of : In this formula, the detailed description and value assignment of each parameter are as follows: Representative meteorological characteristics The abnormal fluctuation amplitude of this time is obtained °C (extreme temperature fluctuations), obtain mm (accumulated precipitation). Represents lag time The Pearson correlation coefficient of is directly read from the corresponding row and column in Table 2. Represents the time window benchmark length for analyzing abnormal events, which is set to 72 hours to ensure that it is consistent with the lag time. (Unit: Day) unit consistency requires conversion processing, and the conversion rule is set as: Days = Hours / 24. This rule is based on standard time measurement, so This conversion ensures a unified calculation basis for time-related terms in the formula. Represents lag time The influence of a specific spatial node The distribution dispersion on plot A01 reflects the spatial heterogeneity of the lag effect. It is set based on the statistical analysis of the geographical distribution of the locations of specific disasters (disasters associated with feature i and lag period j) within a 5 km radius of plot A01 and its surroundings in the past five years. The calculated spatial distribution standard deviation is normalized to the interval [0,1] and is set , . Representative characteristics For specific analysis objects The cross-influence intensity of the wheat in the jointing stage on plot A01 reflects the sensitivity of the crop or environment to this characteristic and its interaction with other factors. It is set based on the simulation results of the crop growth model and the assessment of the response degree of the current crop growth stage to specific meteorological stress by agricultural experts. , .

[0036] Specific calculation example: Calculate the temperature fluctuation range ( ) after a lag of 2 days ( ) : Calculate the precipitation accumulation ( ) after a lag of 4 days ( ) : The benefit of the formula is that it integrates the influence of abnormal fluctuation amplitude and direct time correlation through the square root term, and adjusts the scale with the length of the time window. At the same time, the summation term introduces corrections to the spatial distribution discreteness and the sensitivity of specific objects, so that the weights can more comprehensively reflect the comprehensive potential impact of meteorological characteristics in a specific time and space background.

[0037] All features calculated at all lag times The values ​​(forming a weight matrix with the same dimensions as Table 2) are input as features into a trained random forest model, and the model is used to evaluate the importance of features. The specific execution process is: for each decision tree in the model, calculate each feature (that is, each ) is used to reduce the Gini Impurity when it is used for node splitting. The results of all trees are averaged to obtain the average Gini coefficient reduction of each feature. This value is the contribution of the feature. The model uses this to calculate the contribution of all features. The features are arranged in descending order to generate a contribution sequence. For example, the feature sequence with the highest contribution is: [( ,contribution=0.15),( ,contribution=0.12),( ,contribution=0.10),( ,contribution=0.08),…].

[0038] The path integration submodule calls the contribution sequence generated in the previous step, namely [( ,Gini=0.15),( ,Gini=0.12),( ,Gini=0.10),( ,Gini=0.08),…], set a dynamic screening threshold to select the most important influencing factors. The threshold is set by calculating the average of all positive contribution values ​​in the contribution sequence, and then multiplying it by an adjustment coefficient of 1.2. The average contribution of all features in the current sequence is calculated to be 0.07 (this is the actual value calculated). The dynamic screening threshold is , the lag item weight factors and their corresponding characteristics with contribution higher than 0.084 were screened, and the screening results were: {temperature fluctuation range @ lag 2 days (contribution 0.15), precipitation accumulation @ lag 4 days (contribution 0.12), temperature fluctuation range @ lag 3 days (contribution 0.10)}, and these high-contribution lag items and their corresponding meteorological characteristics (temperature fluctuation range, precipitation accumulation) and contribution values ​​were mapped to the pre-divided geographic grid coordinate system. Block A01 and its surrounding areas were divided into 10m x 10m grid cells. Based on the meteorological The site location information and spatial interpolation algorithm (such as the inverse distance weighted method) are used to assign the impact of each screened feature (quantifying the intensity with its contribution value) to the corresponding grid cell. Subsequently, the spatial distribution layers of the contribution of meteorological characteristics of different dimensions (temperature, precipitation, etc.) are superimposed. The final value of each grid cell represents the degree to which the location is affected by the combined influence of multiple high-contribution meteorological factors. Through weighted summation (the weight is the contribution of each factor) or other spatial analysis methods, a geographic spatial data layer showing high-risk impact paths and key impact areas is generated, namely path mapping data.

[0039] The dynamic scoring matrix module includes: The score normalization submodule obtains the sliding window score data sequence. It uses a complete sequence of 20 period scores: [80, 50, 60, 75, 85, 70, 65, 55, 78, 82, 90, 68, 72, 88, 60, 50, 70, 80, 85, 75]. First, it extracts the maximum score in the sequence (considered as an analysis window). and minimum value , calculate the arithmetic mean of all scores in the window : , then calculate the average absolute value of the difference between adjacent scoring items in the window , as the dynamic volatility: Then, use the following formula to evaluate each scoring item in the window: ( is the window number, is the item number) to transform: In this formula, Is the rating The transformation result of , , , , The parameter values ​​are calculated as above. The formula takes into account the score range and volatility through the denominator to adjust the scale, and the numerator adds a factor that is proportional to the window volatility ( ) and the degree of deviation of the maximum value from the mean ( ) related dynamic compensation items. Calculate the first score in the window The transformation value of : Calculate the score of the second item in the window The transformation value of : The formula is beneficial because it not only reflects the absolute level of the score but also, through dynamic compensation and a comprehensive normalization factor, reflects the relative performance and stability of the score within a specific window. This calculation is performed on all 20 rated items within the window, resulting in a transformed score sequence [6.60, 5.90, …, Z1, 20], generating normalized score data. This result indicates that an item with an original score of 80 will have a transformed score of 6.60 after accounting for the extreme values, mean, and volatility of the window. This value will be used in subsequent cluster analysis and represents a comprehensive measure of the status of the rated item within this window.

[0040] The clustering submodule calls the normalized (transformed) score sequence Z=[6.60,5.90,Z1,3,…,Z1,20] generated in the previous step (all 20 values ​​need to be calculated) and calculates the Euclidean distance between each score item in the sequence. For example, the first item is calculated. With the second Distance between: , for all Calculate the Euclidean distance for the score items and construct a 20x20 symmetric distance matrix D, where the diagonal elements , off-diagonal elements , set the distance threshold for cluster merging. The threshold is set based on the average distance statistics between classes of such transformed score data calculated in multiple historical windows, and combined with the desired clustering granularity (aimed at distinguishing sets of cycles with significantly different score dynamics), it is determined to be 5.0. The average link method (AverageLinkage) in the agglomerative hierarchical clustering algorithm is used. Initially, each score item (i.e., each score item in the sequence) is ) is regarded as an independent cluster, and then the following steps are iteratively performed: find a pair of clusters with the minimum distance between all current clusters (the inter-cluster distance is defined as the average of all point-pair distances between the two clusters), merge the pair of clusters into a new cluster, recalculate the distance between the new cluster and all other clusters, and repeat this merging process until the minimum distance between all clusters is greater than the set merging threshold of 5.0, or the preset cluster number limit is reached (no cluster number limit is set in this example). Through this process, the 20 scoring items are divided into different groups, and the specific clustering results are: Cluster A={scoring item number 1,4,5,8,9,10,11,13,14,18,19,20}, Cluster B={scoring item number 2,7,15,16}, Cluster C={scoring item number 3,6,12,17}, and the clustering results are generated.

[0041] The discrete coding submodule extracts the original score values ​​corresponding to the score items contained in each cluster based on the cluster grouping results (Cluster A, Cluster B, Cluster C) generated in the previous step. (rather than the transformed ), statistical analysis was performed on the original score distribution within each cluster, and the interquartile range (Q1, Q3) and median (Q2) were calculated: for cluster A (containing 12 scoring items), its original score values ​​are {80, 75, 85, 55, 78, 82, 90, 72, 88, 80, 85, 75}, and after sorting, they are {55, 72, 75, 75, 78, 80, 80, 82, 85, 85, 88, 90}. The median Q2 is calculated to be (80+80) / 2=80, the first quartile Q1 is (75+75) / 2=75, the third quartile Q3 is (85+85) / 2=85, and the interquartile range is [75, 85]. For cluster B (containing 4 scoring items), its original scoring values ​​are {50, 65, 60, 50}, and after sorting, they are {50, 50, 60, 65}. The calculated median Q2 is (50+60) / 2=55, the first quartile Q1 is 50, the third quartile Q3 is (60+65) / 2=62.5, and the interquartile range is [50, 62.5]. For cluster C (containing 4 scoring items), its original scoring values ​​are {60, 70, 68, 70}, and after sorting, they are {60, 68, 70, 70}. The calculated median Q2 is (68+70) / 2=69, the first quartile Q1 is (60+68) / 2=64, the third quartile Q3 is 70, and the interquartile range is [64, 70].

[0042] Arrange the clusters in ascending order according to the size of their medians: Cluster B (median 55) < Cluster C (median 69) < Cluster A (median 80). Assign a rank coding number to each cluster in this order. Cluster B with the lowest median is assigned code 1, Cluster C with the middle median is assigned code 2, and Cluster A with the highest median is assigned code 3. Finally, map the original 20 score values ​​one by one to the rank coding corresponding to the cluster to which they belong. The coding sequence corresponding to the original score sequence [80, 50, 60, 75, 85, 70, 65, 55, 78, 82, 90, 68, 72, 88, 60, 50, 70, 80, 85, 75] is: [3, 1, 2, 3, 3, 2, 1, 3, 3, 3, 3, 2, 3, 3, 1, 1, 2, 3, 3, 3], generating rank coding data.

[0043] The response decision generation module includes: The feature splicing and integration submodule calls data from three sources: one is the sliding window period score mean , the second is the level-coded data sequence [3,1,2,3,3,2,1,3,3,3,3,2,3,3,1,1,2,3,3,3], from which the average priority value of the window is calculated: Third, in the path mapping data, the comprehensive connectivity index value of the grid node most closely associated with the current analysis plot A01 is extracted, which is 0.18. Since these three data (period score mean 72.4, average priority value 2.4, node connectivity 0.18) are all calculated based on the same analysis time window or closely related time period, they are naturally overlapping or associated in timestamps. In order to enable the eigenvalues ​​of different sources and scales to be effectively integrated and compared in subsequent models, they need to be adjusted to a unified comparable range. Here, the maximum and minimum normalization method is used to map each eigenvalue to the [0,1] interval. The reference range of each feature is set based on historical statistics or theoretical limits: the range of the mean score is [50,100], the range of the average priority code value is [1,3], and the range of the node connectivity is [0,0.5]. The normalization conversion rule is: normalized value = (original value - corresponding minimum value) / (corresponding maximum value - corresponding minimum value). Normalized calculation: score mean normalized value: Average priority normalized value: Normalized value of node connectivity: ; These three normalized values ​​are horizontally concatenated in a fixed dimensional order of (mean score, priority, and connectivity) to form a three-dimensional feature vector: [0.448, 0.700, 0.360]. This vector comprehensively reflects the model fit, risk level, and criticality of potential impact spread of crop growth in the current analysis period, generating a multidimensional feature vector.

[0044] The optimization target parsing submodule is based on the multidimensional feature vector [0.448, 0.700, 0.360] generated in the previous step. It assigns weights to each dimension in the vector according to the preset management strategy rules. These weights reflect the importance of different aspects in the current decision-making scenario. The weight setting is based on the current management objectives (such as prioritizing the control of high-risk areas and improving overall growth efficiency) and the expert knowledge system. The score weight is determined to be 0.3, the priority coefficient is 0.5, and the connectivity impact factor is 0.2. This weight distribution scheme (0.3+0.5+0.2=1.0) reflects the current stage of focusing on rapid response to potential high-risk events (reflected by high priority values) while taking into account the risk space correlation (connectivity) and the basic growth model performance (mean score). The priority is given the highest weight of 0.5. Then, the value of each dimension in the feature vector is multiplied by its corresponding weight and accumulated to calculate the comprehensive score: Comprehensive score = (mean score normalized value) Scoring weight) + (average priority normalized value Priority coefficient) + (normalized value of node connectivity Connectivity impact factor) comprehensive score = (0.448 0.3)+(0.700 0.5)+(0.360 0.2)=0.1344+0.3500+0.0720=0.5564 The optimization target is determined based on the calculated comprehensive score of 0.5564 and the set of comprehensive scores corresponding to other options (calculated in parallel in an actual application scenario involving multiple alternative management options). In this process, the "optimization target weight" does not refer to the weight of adjusting the comprehensive score, but rather to the set of weight vectors [0.3, 0.5, 0.2] used to calculate the score. It defines the current optimization evaluation criteria and focus, and directly guides the subsequent selection of appropriate intervention measures from the strategy library to generate optimization target weight information (i.e., the weight vector [0.3, 0.5, 0.2]).

[0045] The instruction generation output submodule calls the optimization target weight vector [0.3, 0.5, 0.2] determined in the previous step and matches it with a pre-built strategy library containing a variety of management or intervention measures. Each instruction in the strategy library is associated with a set of clear trigger conditions, which are usually based on the threshold of the optimization target weight (or other related indicators). The following are examples of several instructions in the strategy library and their trigger conditions: Instruction 'IRG-001' (Measures: Strengthen irrigation in high-risk areas): The trigger condition is the priority coefficient 0.4 and connectivity impact factor 0.15, additional condition: the weight combination of the triggering conditions must remain stable in the past 5 analysis cycles, and the measurement standard is that the standard deviation of each component of the weight vector is less than 0.05. Instruction 'FER-003' (Measure: Targeted foliar fertilizer supplementation): The triggering condition is the scoring weight 0.35 and priority coefficient 0.45, additional condition: the volatility of the weighted combination must be controllable in the past 5 periods, with a standard deviation of less than 0.08. Instruction 'MON-002' (Measure: Increase the number of field monitoring points): The trigger condition is the connectivity factor 0.3, no strict stability requirements.

[0046] Compare the current optimization target weight vector [0.3, 0.5, 0.2] with the trigger conditions of the above instructions one by one: For 'IRG-001': priority coefficient 0.5 0.4 (satisfied), connectivity factor 0.2 0.15 (satisfied), the basic trigger conditions are met, check the stability conditions, retrieve the weight data of the past 5 periods and calculate the standard deviation to be 0.03, which is less than the threshold of 0.05, and the stability conditions are met. For 'FER-003': score weight 0.3 0.35 (satisfied), priority coefficient 0.5 0.45 (satisfied), the basic trigger condition is met, check the stability condition, the calculated standard deviation is 0.06, which is less than the threshold of 0.08, and the stability condition is met. For 'MON-002': connectivity factor 0.2 0.3 (not satisfied), the trigger condition is not met.

[0047] Finally, all instructions that meet all basic trigger conditions and additional stability conditions are screened out, and a set of numbers of these instructions is output. In this example, the instructions that meet the conditions are 'IRG-001' and 'FER-003', generating a decision instruction set {'IRG-001', 'FER-003'}.

[0048] The agricultural meteorological decision-making service method based on the artificial intelligence big model is executed based on the agricultural meteorological decision-making service system based on the artificial intelligence big model, and includes the following steps: S1: Use agricultural meteorological monitoring equipment to obtain daily average temperature, relative humidity, daily precipitation, and solar radiation intensity data. Input these four meteorological parameters into a time sliding window partitioning algorithm to divide the time into equal-length segments. Based on the long-short-term memory network, the deviation sequence between the meteorological parameters and the daily crop growth in the window is modeled to generate sliding window scoring data. S2: Using crop indicator monitoring devices to obtain observation indicator data during the growth period, detecting sudden drop nodes triggers reverse backtracking, collecting meteorological data for the previous 72 hours, and using the isolation forest algorithm to identify anomalies such as temperature offsets, humidity mutations, and radiation drops in the meteorological data. The system extracts temperature fluctuation characteristics, humidity mutation characteristics, and radiation anomaly characteristics to generate meteorological anomaly feature data. S3: Based on the meteorological anomaly characteristic data, the lag time correlation coefficients of temperature fluctuation characteristics, humidity sudden change characteristics, and radiation anomaly characteristics with crop growth indicators are calculated. The lag term weight factors are input into the random forest model to rank the contribution of meteorological parameters and generate path mapping data. S4: Perform minimum and maximum normalization on the sliding window score data, call the hierarchical clustering algorithm to discretize the normalized results into five level intervals, and generate level-coded data; S5: The sliding window scoring data, grade coding data and path mapping data are spliced ​​into a multi-dimensional feature vector, which is input into the multi-objective optimization model to optimize the strategy for irrigation amount, shading duration and fertilization frequency, and output a decision instruction set.

[0049] The above are merely preferred embodiments of the present invention and do not limit the present invention in any other form. Any technician familiar with the profession may use the technical content disclosed above to change or modify it into an equivalent embodiment with equivalent changes and apply it to other fields. However, any simple modification, equivalent change and modification made to the above embodiment based on the technical essence of the present invention without departing from the content of the technical solution of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. The agricultural meteorological decision-making service system based on artificial intelligence large model is characterized by: The system comprises: The meteorological coupling analysis module is used to obtain meteorological time series and observation index data during the crop growth period. The daily average temperature, relative humidity, daily precipitation, and solar radiation intensity are input into the time sliding window partitioning algorithm to divide the time into equal length segments. The long short-term memory network is used to calculate the deviation sequence between meteorological parameters and daily crop growth within the window. The sliding window scoring data is generated and passed to the dynamic scoring matrix module and the response decision generation module. An anomaly tracing and positioning module is used to trigger reverse window backtracking based on the sudden drop in crop indicators, collect 72 hours of meteorological data, call the isolation forest algorithm to detect the extreme value offset, extract meteorological anomaly feature data and pass it to the impact path modeling module; An impact path modeling module is used to receive the meteorological anomaly characteristic data, calculate the lag time correlation coefficient, input the lag term weight factor into the random forest model ranking contribution, generate path mapping data and pass it to the response decision generation module; The dynamic scoring matrix module is used to receive the sliding window scoring data, perform minimum and maximum normalization processing, call the hierarchical clustering algorithm for discretization coding, generate hierarchical coding data and pass it to the response decision generation module.

2. The agricultural meteorological decision-making service system based on the artificial intelligence large model according to claim 1 is characterized in that: The sliding window scoring data includes the temperature change rate, humidity fluctuation coefficient, radiation accumulation and precipitation distribution discreteness within the window; the meteorological anomaly characteristic data specifically includes the anomaly intensity index, offset period identifier, and mutation meteorological parameter type; the path mapping data includes the lag time step, contribution ranking list, and path association matrix; the level coding data specifically refers to the normalized scoring interval, cluster discrete level, and coding mapping table.

3. The agricultural meteorological decision-making service system based on the artificial intelligence large model according to claim 2 is characterized in that: The meteorological coupling analysis module includes: The meteorological segmentation processing submodule obtains the daily average temperature, relative humidity, daily precipitation, and solar radiation intensity. Based on the time sliding window partitioning algorithm, it sets a fixed-length interval and divides the meteorological parameters into segments of the same number of days in a continuous time sequence. The multiple meteorological parameters in the segment are serialized and aligned with the corresponding date to generate a meteorological sequence of equal length. The growth deviation calculation submodule calls the four meteorological parameters of the equal-length meteorological sequence on a daily basis, combines the long-short-term memory network hidden layer state, calculates the daily crop growth forecast value, extracts the actual crop growth value in the observation data, calculates the difference between the forecast value and the actual value on a daily basis, takes the absolute value of the difference sequence and accumulates and sums them to generate the daily growth deviation; The window score generation submodule extracts the daily growth deviation, sums the deviations in a fixed period, sets the deviation interval between the lower limit and the upper limit, and compares the deviation sum with the lower limit and the upper limit of the interval in turn. If the sum is within the interval, the corresponding score is output; if it is lower than the lower limit or higher than the upper limit, the corrected score is output. The multi-period score results are arranged in chronological order to generate sliding window score data.

4. The agricultural meteorological decision-making service system based on the artificial intelligence large model according to claim 3 is characterized in that: The abnormality tracing and positioning module includes: The abnormal trigger backtracking submodule monitors the rate of change of crop index values ​​and calculates the average rate of change over three consecutive days. When the single-day drop exceeds a set multiple of the average rate of change, it is marked as a sudden drop node. A fixed-length window is defined in reverse order from the sudden drop node as the end point. At the same time, hourly temperature, humidity, and wind speed data within the window are collected to generate an abnormal backtracking window. The offset detection and analysis submodule calls the hourly temperature, humidity, and wind speed data in the abnormal backtracking window, calculates the deviation of the multi-parameter value from the mean within the window, marks the time period where the deviation exceeds the set threshold as a candidate offset point, calculates the time distribution density of the candidate offset points, selects the time period where the distribution density exceeds the benchmark value, and generates the offset extreme value point; The feature extraction and transfer submodule calls the offset extreme point, extracts the difference between the maximum and minimum temperature fluctuations as the fluctuation range, counts the total number of time periods in which the humidity parameter continuously exceeds the deviation threshold, calculates the number of sudden changes in wind speed that exceed the benchmark value in adjacent hours, and arranges the three feature values ​​in chronological order, matches the deviation range with the sudden drop node time, and generates meteorological anomaly feature data.

5. The agricultural meteorological decision-making service system based on the artificial intelligence large model according to claim 4 is characterized in that: The impact path modeling module includes: The lag correlation submodule obtains the temperature deviation value and precipitation accumulation in the meteorological anomaly characteristic data, calculates the Pearson correlation coefficient between them and the probability of disaster occurrence within a predefined lag time range, arranges them into a row and column structure matrix according to the time dimension, and generates a lag correlation matrix; The weight factor ranking submodule calls the lagged correlation matrix and performs range normalization on each lagged weight factor using the formula: The spatiotemporal correction weight values ​​are obtained by calculation and input into the random forest model to perform iterative feature sorting of the Gini coefficient to generate a contribution sequence; Among them, w ij represents the weight factor of meteorological feature i at lag time j, M i is the abnormal fluctuation amplitude of meteorological feature i, τ j is the Pearson correlation coefficient of lag time j, Δt is the time window reference length, δ js is the distribution dispersion of lag time j at spatial node s, κ ic is the cross-influence strength of feature i at node c; The path integration submodule calls the contribution sequence, screens the lag term weight factors whose contributions are higher than the dynamic screening threshold, maps them to geographic grid coordinates, overlays the multi-dimensional meteorological feature contribution distribution, and generates path mapping data.

6. The agricultural meteorological decision-making service system based on the artificial intelligence large model according to claim 5 is characterized in that: The dynamic scoring matrix module includes: The score normalization submodule obtains the sliding window score data, extracts the maximum and minimum scores in multiple windows, and calculates the dynamic volatility based on the average of the absolute values ​​of the differences between adjacent scores in the window, using the formula: Normalize the denominator and dynamically compensate the numerator of each scoring item, integrate the results of all scoring items in the window, and generate normalized scoring data; Among them, Z pq represents the normalized result of the qth item score in the pth window, R pq is the original score value, R max is the maximum score within the window, R min is the minimum score value, γ is the mean of the absolute values ​​of the adjacent score differences within the window, is the window score mean. The clustering submodule calls the normalized score data, calculates the Euclidean distance between score items, constructs a symmetric distance matrix, sets the inter-cluster merging threshold according to the average difference value of the scores of adjacent windows, completes hierarchical clustering by successively merging the clusters with the closest distance, and generates a clustering result; The discrete coding submodule calculates the interquartile range and median of the score distribution in each cluster based on the clustering grouping results, assigns grade coding numbers in ascending order of the median size, maps the original scores to the corresponding coding categories, and generates grade coding data.

7. The agricultural meteorological decision-making service system based on the artificial intelligence large model according to claim 6 is characterized in that: The system further comprises: a response decision generation module, configured to receive the sliding window scoring data, the grade coding data, and the path mapping data, splice a multi-dimensional feature vector into a multi-objective optimization model, and output a decision instruction set; The decision instruction set includes optimization weight configuration, priority instruction queue, and execution time node.

8. The agricultural meteorological decision-making service system based on the artificial intelligence large model according to claim 7 is characterized in that: The response decision generation module includes: The feature splicing and integration submodule calls the period score mean of the sliding window scoring data, the priority value of the grade coding data, and the node connectivity of the path mapping data, extracts the fields with overlapping timestamps in the three data items, adjusts the field values ​​to the same dimensional range, and splices them horizontally into a three-dimensional matrix according to the time dimension to generate a multidimensional feature vector; The optimization target parsing submodule allocates scoring weights, priority coefficients, and connectivity influencing factors based on the multidimensional feature vector according to preset rules, superimposes the three weights to calculate a comprehensive score, defines the upper limit of resource consumption as the maximum value of the comprehensive score, and the lower limit of node coverage as the minimum value of the comprehensive score, and screens the extreme score points that meet the upper and lower limit constraints to generate the optimization target weight; The instruction generation output submodule calls the optimization target weight, compares the weight distribution with the matching degree of the instruction triggering conditions in the policy library, verifies whether the weight fluctuation range in the time window is within the allowable threshold of the policy execution cycle, outputs a set of instruction numbers that meet the matching degree and fluctuation range, and generates a decision instruction set.

9. The agricultural meteorological decision-making service method based on artificial intelligence large model is characterized by: The method is used to implement the agricultural meteorological decision-making service system based on the artificial intelligence large model according to any one of claims 1 to 8, comprising the following steps: S1: Use agricultural meteorological monitoring equipment to obtain daily average temperature, relative humidity, daily precipitation, and solar radiation intensity data. Input these four meteorological parameters into a time sliding window partitioning algorithm to divide the time into equal-length segments. Based on the long-short-term memory network, the deviation sequence between the meteorological parameters and the daily crop growth in the window is modeled to generate sliding window scoring data. S2: Using crop indicator monitoring devices to obtain observation indicator data during the growth period, detecting sudden drop nodes triggers reverse backtracking, collecting meteorological data for the previous 72 hours, and using the isolation forest algorithm to identify anomalies such as temperature offsets, humidity mutations, and radiation drops in the meteorological data. The system extracts temperature fluctuation characteristics, humidity mutation characteristics, and radiation anomaly characteristics to generate meteorological anomaly feature data. S3: Based on the meteorological anomaly characteristic data, calculating the lag time correlation coefficients between the temperature fluctuation characteristics, humidity sudden change characteristics, and radiation anomaly characteristics and the crop growth indicators, inputting the lag term weight factors into the random forest model to rank the contribution of meteorological parameters, and generating path mapping data; S4: performing minimum and maximum normalization on the sliding window score data, calling a hierarchical clustering algorithm to discretize the normalized result into five level intervals, and generating level-coded data; S5: The sliding window scoring data, the grade coding data and the path mapping data are spliced ​​into a multi-dimensional feature vector, which is input into a multi-objective optimization model to perform strategy optimization on the irrigation amount, shading duration and fertilization frequency, and a decision instruction set is output.

Images