An artificial intelligence-based intelligent decision system for agricultural irrigation

Abstract

This invention relates to the field of intelligent decision-making technology and discloses an intelligent decision-making system for agricultural irrigation based on artificial intelligence. The system constructs a collaborative architecture comprising a field perception layer, an edge computing layer, a cloud-based decision-making layer, and an irrigation execution layer, forming an irrigation task context through multi-source perception data. Based on this, a meta-reinforcement learning-driven multi-round agent-based search decision-making mechanism is introduced, combined with explicit self-reflection to achieve iterative optimization of the decision-making process. Simultaneously, a preference-aware reward evaluation model is constructed, achieving multi-objective trade-off modeling through latent variable mirror constraints and inverse autoregressive flow transformation, and globally optimizing the strategy using a cross-round benefit attribution mechanism to generate an irrigation scheduling scheme that meets water-saving, crop water requirement, and energy consumption constraints. This invention can improve the accuracy, stability, and adaptability of irrigation decisions.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making technology, and in particular to an intelligent decision-making system for agricultural irrigation based on artificial intelligence. Background Technology

[0002] With the development of smart agriculture and digital agriculture, intelligent irrigation technology based on sensor perception and model calculation has been widely used. Existing technologies typically generate irrigation strategies by combining soil moisture monitoring, meteorological data collection, and crop water requirement models with threshold control, empirical formulas, or single machine learning models. Some solutions further introduce time series prediction models or reinforcement learning methods to optimize and control the irrigation process, thereby improving irrigation automation and decision-making accuracy to a certain extent. However, existing intelligent irrigation technologies are still mainly based on single-round decision-making mechanisms. Their decision-making process is usually based on a one-time input information to directly output an irrigation plan, lacking the ability to accumulate information across rounds and iteratively optimize. This makes it difficult to achieve continuous correction and global optimization of strategies in complex and dynamic environments. At the same time, most methods fail to effectively integrate external agricultural knowledge resources and lack the ability to dynamically retrieve and reason about agronomic rules, historical irrigation cases, and meteorological forecast information, resulting in insufficient adaptability of the decision-making process to environmental changes. In addition, in terms of reward evaluation and strategy optimization, existing technologies mostly use fixed-weight multi-index evaluation functions or simple reward mechanisms, which are difficult to effectively model the nonlinear trade-offs between multiple objectives such as water-saving goals, crop water requirement satisfaction, energy consumption control, and risk constraints. They also lack a structured expression of the differences in preferences for different irrigation strategies, resulting in insufficient stability and limited generalization ability in the strategy learning process. Furthermore, existing systems generally lack explicit reasoning and self-reflection mechanisms for the decision-making process, and cannot dynamically correct the search direction, parameter selection, and constraints during the decision-making process, thus limiting the applicability and decision quality of irrigation strategies in complex and ever-changing agricultural scenarios. Summary of the Invention

[0003] This invention addresses the limitations of existing intelligent irrigation decision-making technologies, such as limited single-round decision-making capabilities, insufficient knowledge utilization, and weak multi-objective trade-off modeling capabilities. It proposes an AI-based intelligent decision-making system for agricultural irrigation, which constructs a multi-round continuous optimization decision-making mechanism centered on meta-reinforcement learning. By introducing a meta-round structure composed of multiple sequentially related internal rounds and embedding an explicit self-reflection mechanism between adjacent rounds, it identifies and corrects search biases, missing evidence, and parameter conflicts in previous decisions, achieving round-by-round optimization and global convergence of the irrigation strategy. Simultaneously, it constructs a proxy-based search decision-making mechanism for agricultural scenarios, dynamically generating search query actions and invoking external agricultural knowledge resources during the inference process, and backfilling the retrieved evidence into the decision trajectory. In this process, a multi-round interactive decision-making process of "reasoning-retrieval-evidence update-re-reasoning" is formed to enhance the system's adaptability to complex environmental changes. Furthermore, a preference-aware reward modeling method is introduced. By constructing dual representations of the original strategy and the preference exchange strategy, a preference representation structure that satisfies mirror constraints is established in the latent variable space. Combined with inverse autoregressive flow transformation, a nonlinear expression of multimodal irrigation strategy preferences is realized, thereby enhancing the modeling ability of multi-objective trade-offs such as water-saving goals, crop water requirement satisfaction, energy consumption control, and risk constraints. In addition, through a cross-round benefit attribution mechanism, the evaluation results of subsequent rounds are applied back to previous rounds to achieve global guidance for strategy updates, enabling the system to generate irrigation decision results with optimal comprehensive performance.

[0004] This invention provides an intelligent decision-making system for agricultural irrigation based on artificial intelligence. The system includes an irrigation information acquisition module, a protocol adaptation module, an agricultural knowledge retrieval module, a meta-reinforcement search decision module, a strategy verification and reward evaluation module, and an irrigation execution scheduling module.

[0005] The irrigation information acquisition module is deployed in the field sensing layer to collect raw irrigation monitoring data;

[0006] The protocol adaptation module is deployed at the edge computing layer and runs on the edge gateway. It performs protocol parsing, time alignment, missing data completion, anomaly removal, spatial partitioning mapping, unit normalization, and device identity binding on the raw irrigation monitoring data to form standardized irrigation status data. The standardized irrigation status data is organized according to the hierarchical relationship of "plot - crop - growth period - irrigation unit - device node" and the current irrigation task context is constructed.

[0007] The agricultural knowledge retrieval module is deployed in the cloud decision layer and runs on the cloud server. It calls external agricultural information sources to perform proxy searches. When the meta-enhanced search decision module issues a search query, the agricultural knowledge retrieval module calls the corresponding external agricultural information source to obtain the tool observation results according to the search query, and returns the tool observation results to the meta-enhanced search decision module, so that the agricultural irrigation decision-making process forms a multi-round interactive trajectory of "thinking - searching - observing - rethinking".

[0008] The meta-enhanced search decision module is deployed in the cloud decision layer. Based on the current irrigation task context, it generates multiple sequentially related internal episodes. In each internal episode, a search query action is generated and the tool observation results returned by the agricultural knowledge retrieval module are received. The tool observation results are introduced into the irrigation reasoning process to generate corresponding candidate irrigation answers. An explicit self-reflection mechanism is introduced between adjacent internal episodes to construct a meta-episode. This allows subsequent internal episodes to revise the irrigation decision round by round based on the search experience and reflection results of previous internal episodes. Through multiple rounds of internal episode iterations, multiple candidate irrigation answers are gathered to determine the target irrigation decision result.

[0009] The strategy verification and reward evaluation module is deployed in the cloud decision layer and runs on the same cloud computing node as the meta-reinforcement search decision module. By introducing a preference-aware strategy verification and reward evaluation method, the candidate irrigation answers output by each irrigation decision episode are verified and scored. By performing irrigation execution element analysis, rule verification, preference latent variable modeling and joint reward evaluation on the candidate irrigation answers, the round rewards of the corresponding internal round episodes are generated and the round reward sequence is formed according to the generation order of the internal round episodes.

[0010] The irrigation execution scheduling module is deployed at the irrigation execution layer to perform structured analysis and task scheduling processing on the target irrigation decision results.

[0011] Furthermore, the process of determining the outcome of the target irrigation decision includes the following steps:

[0012] Step S1: Initialize the current irrigation task context and construct the initial decision input as the starting condition for the first round of irrigation decision search; the current irrigation task context includes target plot identifier, crop type, growth stage, soil moisture status, weather change trend, irrigation equipment operation constraints, historical irrigation records and current irrigation target;

[0013] Step S2: Generate the first irrigation decision round based on the initial decision input, and define the first irrigation decision round as the first internal round episode; in the first internal round episode, the meta-enhanced search decision module outputs the irrigation reasoning content, search query action, tool call request and first round candidate irrigation answer in sequence;

[0014] Step S3: Send the search query action to the agricultural knowledge retrieval module, call the external agricultural information source to perform the proxy search, and receive the tool observation results returned by the agricultural knowledge retrieval module; write the tool observation results into the decision trajectory of the current internal round episode, so that the first internal round episode continues to perform irrigation reasoning after receiving external evidence, completes the first round of irrigation decision process, and obtains the search trajectory of the first internal round episode.

[0015] Step S4: After the first internal round episode ends, perform explicit self-reflection processing on the first internal round episode to identify search biases, missing evidence, parameter conflicts, constraint omissions, and defects in the first round of candidate irrigation answers in the first round of irrigation decision-making; and generate reflection content based on the search trajectory, tool observation results, and first round candidate irrigation answers of the first internal round episode; wherein, the reflection content includes suggestions for subsequent search direction correction, priority knowledge categories for retrieval, range of irrigation parameters to be corrected, and constraints to be supplemented;

[0016] Step S5: Concatenate the search trajectory and reflection content of the first internal round episode, and combine them with the current irrigation task context to form the input context of the second internal round episode; generate the second internal round episode based on the input context of the second internal round episode, so that the second internal round episode, based on inheriting the previous search experience and reflection results, corrects the reasoning direction of the first round of irrigation decision, and regenerates the irrigation reasoning content, search query action, tool call request and updated candidate irrigation answer;

[0017] Step S6: Repeat steps S3 to S5, sequentially performing tool search, evidence backfilling, candidate answer generation, and explicit self-reflection processing on the second internal round episode and its subsequent internal round episodes, and delivering the search trajectory and corresponding reflection content of each preceding internal round episode to the next internal round episode; thereby forming a continuously related meta-episode under the same irrigation task from multiple sequentially generated internal round episodes.

[0018] Step S7: In the meta-episode, extract the candidate irrigation answers corresponding to each internal episode generated under the same irrigation task, and send them to the strategy verification and reward evaluation module to perform round verification, and receive the round reward sequence returned by it.

[0019] Step S8: Based on the round reward sequence, construct the reward propagation relationship according to the generation order of the internal round episode in the meta-episode, and perform a discount accumulation calculation on the round rewards of subsequent episodes to obtain the cross-round benefit attribution results corresponding to the preceding episodes; based on the cross-round benefit attribution results, update the decision strategy of the meta-reinforcement search decision module, so that the strategy will prioritize the generation of search behaviors and irrigation parameter selection strategies with higher decision benefits in subsequent irrigation tasks; determine the target irrigation decision result from the candidate irrigation answers corresponding to multiple internal round episodes.

[0020] Furthermore, a preference-aware strategy verification reward evaluation method is introduced to verify and score the candidate irrigation answers output for each irrigation decision episode, forming a round reward sequence. This process specifically includes the following steps:

[0021] Step B1: Parse the irrigation execution elements from the candidate irrigation answers for each inner round episode in the meta-episode; and align the irrigation execution elements with the tool observation results on which the inner round episode was generated to form the irrigation strategy representation to be verified for the current inner round.

[0022] Step B2: Perform consistency verification, constraint verification, and effectiveness verification on the characterization of the irrigation strategy to be verified according to the preset irrigation verification rules;

[0023] Step B3: For the irrigation strategy representation to be verified in each internal round episode, construct the corresponding preference exchange strategy representation. The preference exchange strategy representation is constructed by exchanging the priority execution parameter direction, resource allocation tendency, partition irrigation order bias, or preferred / inferior selection mark in the candidate irrigation answer. It is used to represent the virtual preference strategy that is opposed to the current candidate irrigation answer. The encoder performs encoding processing on the irrigation strategy representation to be verified and the preference exchange strategy representation, and performs latent variable mapping to generate the original strategy posterior distribution and the exchange strategy posterior distribution, and outputs additional context information.

[0024] Step B4: Perform exchange-guided constraint processing based on the original policy posterior distribution and the exchange policy posterior distribution, and fuse them to obtain the basic latent variable distribution;

[0025] Step B5: After completing the construction of the basic posterior distribution and the exchange guiding constraints, a preference-aware inverse autoregressive flow transformation is performed on the basic latent variables in the basic latent variable distribution. The additional context information output by the encoder is decomposed into exchange-inverted context components and exchange-invariant context components, which are then used to modulate the translation and scaling functions in the inverse autoregressive flow transformation, respectively. Specifically, the translation function is modulated using the exchange-inverted context components to characterize the irrigation preference signal whose direction changes due to the exchange of candidate irrigation preferences. The scaling function is modulated using the exchange-invariant context components to characterize the plot background conditions, irrigation equipment operating status, crop growth stage, and environmental constraint information that do not change with preference exchange. Through the above preference-aware inverse autoregressive flow transformation, the basic latent variables are mapped to a latent variable space that can express the multimodal irrigation preference structure, thereby obtaining the latent variables after the inverse autoregressive flow transformation.

[0026] Step B6: Perform joint reward evaluation on the latent variables after inverse autoregressive transformation and the validation feature vector, calculate the reward for the candidate irrigation strategy of the current internal episode, obtain the round reward of the corresponding internal episode, and form a round reward sequence according to the generation order of the internal episodes.

[0027] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0028] This invention constructs a multi-round iterative decision-making mechanism based on meta-reinforcement learning, realizing the transformation of agricultural irrigation decision-making from traditional single-round static generation to cross-round dynamic optimization. This enables irrigation strategies to be continuously corrected and converged in the continuous decision-making process, significantly improving the global optimization capability and stability of irrigation schemes under complex environmental conditions. It solves the problems of under-irrigation, over-irrigation, and unreasonable resource allocation caused by insufficient information in a single decision in existing intelligent irrigation systems, thereby enhancing the adaptability and decision reliability of the irrigation process under different plot conditions, crop growth stages, and weather changes.

[0029] This invention introduces a proxy-based search and external agricultural knowledge fusion mechanism to achieve dynamic access to multi-source agricultural knowledge and evidence-driven reasoning in the irrigation decision-making process. This enables the system to comprehensively utilize information such as crop water requirement models, agronomic rules, weather forecasts, and historical irrigation cases during the decision-making process, significantly improving the scientific nature and accuracy of irrigation decisions. It solves the problems of insufficient knowledge utilization and delayed environmental response in existing methods, thereby enhancing the system's decision-making accuracy and robustness in the face of complex agricultural scenarios and uncertain conditions.

[0030] This invention achieves refined modeling and global optimization of the multi-objective trade-offs in irrigation strategies by constructing a preference-aware reward modeling mechanism and a cross-round benefit attribution mechanism. This enables the system to achieve coordinated optimization among multiple dimensions such as water-saving efficiency, crop water requirement satisfaction, energy consumption control, and risk constraints, significantly improving the comprehensive performance and generalization ability of irrigation decision results. It solves the problems of insufficient evaluation function expression and unstable strategy optimization in existing intelligent irrigation methods, thereby enhancing the scalability and long-term performance of this invention in different application scenarios. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of a module of an intelligent decision-making system for agricultural irrigation based on artificial intelligence proposed in this invention;

[0032] Figure 2 This is the posterior distribution diagram of the preference mirror latent variable proposed in Example 4. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] Example 1, according to Figure 1 This invention provides an intelligent decision-making system for agricultural irrigation based on artificial intelligence, which can be applied to scenarios such as water-saving irrigation in farmland, irrigation in facility agriculture, precision irrigation in orchards, or zonal irrigation in large fields. It is deployed in an agricultural irrigation operation environment that includes a field perception layer, an edge computing layer, a cloud decision-making layer, and an irrigation execution layer. The system includes an irrigation information acquisition module, a protocol adaptation module, an agricultural knowledge retrieval module, a meta-enhanced search decision module, a strategy verification and reward evaluation module, and an irrigation execution scheduling module.

[0035] This embodiment is applied to a tomato planting scenario in a facility agriculture greenhouse. The target farmland is Zone A of a multi-span greenhouse, with a total area of ​​12.6 acres, divided into 4 independent irrigation zones: Zone Z1, Zone Z2, Zone Z3, and Zone Z4. The crop planted is tomatoes, with Zones Z1 and Z2 in the flowering and fruit-setting stage, and Zones Z3 and Z4 in the fruit expansion stage. The irrigation system adopts an execution structure of "head pump station + main pipeline + zone solenoid valves + drip irrigation tape + Venturi fertilizer applicator". The rated working pressure of the system is 0.23 MPa, and the allowable fluctuation range is ±0.03 MPa. The maximum water consumption for a single irrigation is set at 58 m³, and the irrigation window is limited to two periods: 05:00 to 10:00 and 17:00 to 22:00 on the same day, to avoid increased ineffective water consumption due to high-temperature evaporation at noon.

[0036] The irrigation information acquisition module is deployed in the field sensing layer, set up within the target farmland zone. It collects data on soil moisture, soil temperature, electrical conductivity, pH value, ambient temperature and humidity, light intensity, wind speed and direction, rainfall, evapotranspiration, crop canopy images, leaf temperature, irrigation network pressure, flow rate, pump operating status, valve opening and closing status, and historical irrigation records within the target farmland zone using soil moisture sensors, soil temperature sensors, weather stations, flow meters, pressure sensors, electric valve status detectors, high-definition visible light cameras, and near-infrared imaging devices. This data forms raw irrigation monitoring data with timestamps, plot identifiers, crop type identifiers, and growth stage identifiers. The raw irrigation monitoring data includes volumetric water content data at different depths of the soil profile, estimated surface evaporation data, crop canopy temperature difference data, near-infrared reflectance data, transient flow rate data of the irrigation network, pump power curves, plot elevation information, irrigation zone topology information, and meteorological forecast data for the next 24 to 72 hours.

[0037] The protocol adaptation module is deployed at the edge computing layer and runs on the edge gateway. It performs protocol parsing, time alignment, missing data completion, anomaly removal, spatial partitioning mapping, unit normalization, and device identity binding on the raw irrigation monitoring data to form standardized irrigation status data. The standardized irrigation status data is organized according to the hierarchical relationship of "plot - crop - growth period - irrigation unit - device node" and the current irrigation task context is constructed.

[0038] In this embodiment, the current irrigation task context is obtained. The data collection results show that: the soil volumetric moisture content in zone Z1 (0-20cm) is 18.4%, and in zone Z2 (20-40cm) it is 20.1%; in zone Z2, the values ​​are 17.9% and 19.3%; in zone Z3, the values ​​are 16.8% and 18.7%; and in zone Z4, the values ​​are 17.2% and 18.9%. The suitable soil volumetric moisture content control range for the target crop at the current growth stage is 22.0% to 26.0%. Meanwhile, the 24-hour weather forecast data outside the greenhouse shows that from 12:00 to 16:00 on that day... The highest outside temperature is 31.6℃, and the lowest relative humidity is 41%. There is no effective rainfall expected in the next 18 hours. The reference evapotranspiration (ET0) forecast is 4.7 mm / d. The irrigation equipment operation status shows that pump No. 1 is operating stably with a rated flow rate of 32 m³ / h and a real-time health score of 0.96. Pump No. 2 is in standby mode. The main valve MV-01 of the main pipeline is operating normally, and the branch valves V-Z1, V-Z2, V-Z3, and V-Z4 in each zone are all remotely controllable. The fertilizer tank has a remaining capacity of 240L, and water and fertilizer synergy operations can be performed at present.

[0039] The agricultural knowledge retrieval module is deployed in the cloud decision layer and runs on the cloud server. It calls external agricultural information sources to perform proxy searches. When the meta-enhanced search decision module issues a search query, the agricultural knowledge retrieval module calls the corresponding external agricultural information source to obtain the tool observation results according to the search query, and returns the tool observation results to the meta-enhanced search decision module, so that the agricultural irrigation decision-making process forms a multi-round interactive trajectory of "thinking - searching - observing - rethinking".

[0040] The meta-enhanced search decision module is deployed in the cloud decision layer. Based on the current irrigation task context, it generates multiple sequentially related internal episodes. In each internal episode, a search query action is generated and the tool observation results returned by the agricultural knowledge retrieval module are received. The tool observation results are introduced into the irrigation reasoning process to generate corresponding candidate irrigation answers. An explicit self-reflection mechanism is introduced between adjacent internal episodes to construct a meta-episode. This allows subsequent internal episodes to revise the irrigation decision round by round based on the search experience and reflection results of previous internal episodes. Through multiple rounds of internal episode iterations, multiple candidate irrigation answers are gathered to determine the target irrigation decision result.

[0041] Target irrigation decision results: The irrigation execution mode for this round of irrigation tasks is determined to be a zoned sequential drip irrigation mode; the irrigation execution sequence is determined to be zone Z3 → zone Z4 → zone Z2 → zone Z1; among them, zones Z3 and Z4 are in the fruit expansion stage and the soil moisture is low, so they are given priority for irrigation; zones Z2 and Z1 are in the flowering and fruit setting stage, and it is necessary to control the risk of over-irrigation while ensuring soil moisture recovery, so they are executed later; the target irrigation volume for each zone is determined as follows: zone Z3 14.2 m³, zone Z4 13.6 m³, zone Z2 11.4 m³, zone Z1 10.8 m³, with a total single irrigation volume of 50.0 m³. The target irrigation duration for each zone is as follows: zone Z3 27 min, zone Z4 26 min, zone Z2 22 min, zone Z1 21 min, with a total irrigation duration of 96 min. The system control objectives are: after irrigation, the volumetric moisture content of the 0-20cm soil layer in each zone should be restored to 22.5%-24.0%, and the volumetric moisture content of the 20-40cm soil layer should be restored to 21.5%-23.5%, with the risk of over-irrigation controlled below 0.08 and the risk of under-irrigation controlled below 0.12. In view of the current fertilizer requirements of crops in the fruit expansion stage, the system also determines the water and fertilizer synergistic application strategy as follows: diluted liquid fertilizer is injected simultaneously during irrigation in zones Z3 and Z4, with the total application concentration controlled at EC 1.8mS / cm and the fertilizer flow rate set at 18L / h; only low-concentration maintenance application is implemented in zones Z2 and Z1, with EC controlled at 1.2mS / cm and the fertilizer flow rate set at 10L / h.

[0042] The strategy verification and reward evaluation module is deployed in the cloud decision layer and runs on the same cloud computing node as the meta-reinforcement search decision module. By introducing a preference-aware strategy verification and reward evaluation method, the candidate irrigation answers output by each irrigation decision episode are verified and scored. By performing irrigation execution element analysis, rule verification, preference latent variable modeling and joint reward evaluation on the candidate irrigation answers, the round rewards of the corresponding internal round episodes are generated and the round reward sequence is formed according to the generation order of the internal round episodes.

[0043] The irrigation execution scheduling module is deployed at the irrigation execution layer. It performs structured parsing and task scheduling processing on the target irrigation decision results, converting the target irrigation decision results into irrigation operation scheduling information that can be recognized by irrigation equipment. It then sends the information to the corresponding irrigation equipment management node through the equipment communication interface, so that the irrigation equipment can execute the corresponding irrigation operation based on the scheduling information. The irrigation operation scheduling information includes pump station operation scheduling information, main pipe valve and branch pipe valve opening configuration parameters, zonal irrigation sequence plan, irrigation duration configuration, drip irrigation and sprinkler irrigation mode selection information, and water and fertilizer co-application parameters.

[0044] The irrigation operation scheduling information generated in this embodiment is as follows:

[0045] 1. Main pipe valve and branch pipe valve opening configuration parameters: Main pipe valve MV-01 maintains 92% opening throughout the irrigation cycle; branch pipe valve V-Z3 in zone Z3 opens to 86% opening at 05:40 and remains open for 27 minutes; branch pipe valve V-Z4 in zone Z4 opens to 84% opening at 06:08 and remains open for 26 minutes; branch pipe valve V-Z2 in zone Z2 opens to 79% opening at 06:35 and remains open for 22 minutes; branch pipe valve V-Z1 in zone Z1 opens to 77% opening at 06:58 and remains open for 21 minutes; a 30-second buffer interval is set during the switching between adjacent zones, and the valve in the next zone is opened 10 seconds after the valve in the previous zone is closed, in order to suppress transient pressure shocks in the main pipe.

[0046] 2. Irrigation duration configuration: Zone Z3 duration 27 min, corresponding to a target irrigation volume of 14.2 m³; Zone Z4 duration 26 min, corresponding to a target irrigation volume of 13.6 m³; Zone Z2 duration 22 min, corresponding to a target irrigation volume of 11.4 m³; Zone Z1 duration 21 min, corresponding to a target irrigation volume of 10.8 m³; Each zone is allowed to make dynamic fine adjustments of ±2 min based on pressure feedback during actual execution, but the total irrigation volume deviation shall be controlled within ±3%.

[0047] Example 2 differs from Example 1 in that the strategy verification and reward evaluation module introduces a preference-aware strategy verification and reward evaluation method to verify and score the candidate irrigation answers output by each irrigation decision episode. The strategy verification and reward evaluation module in this example includes the following: It introduces a rule-based deterministic scoring method, performs irrigation execution element analysis and rule verification on the candidate irrigation answers, constructs a deterministic scoring function based on the verification results, comprehensively scores the candidate irrigation strategies, and obtains the round rewards for the corresponding internal episodes. The round rewards corresponding to each internal episode are arranged according to the generation order to form a round reward sequence.

[0048] Example 3, based on Example 1, uses external agricultural information sources including a regional weather forecast platform, a soil moisture database, a crop water requirement model library, agronomic rule library, historical irrigation case library, water price and energy consumption constraint library, pest and disease early warning information source, and a manual maintenance knowledge base. The proxy search process includes the following steps: generating a search query based on the current plot's crop type, growth stage, soil moisture threshold deviation, weather change trend, and irrigation equipment operation constraints; sending the search query to the regional agricultural knowledge base, agronomic rule library, irrigation case library, and weather forecast interface for information retrieval; receiving tool observation results returned by each external agricultural information source, and processing the tool observation results. The process involves parsing and structuring the data; the parsed tool observations are then backfilled into the current episode trajectory as evidence fragments, enabling the meta-enhanced search decision module to update the current irrigation inference process based on the new evidence. The tool observations include recommended irrigation depth ranges, crop water requirement coefficients, extreme weather warnings, historical summaries of irrigation effects on similar plots, water resource allocation constraints, and suggestions for handling abnormal conditions. Based on the tool observations, the meta-enhanced search decision module dynamically adjusts the irrigation inference direction, irrigation parameter range, and irrigation execution strategy in the current or subsequent episodes, thereby achieving multi-round proxy-based search decision-making for agricultural irrigation tasks.

[0049] Example 4, according to Figure 2 This embodiment is based on Embodiment 3. In this embodiment, the process of determining the target irrigation decision result specifically includes the following steps:

[0050] Step S1: Initialize the current irrigation task context and construct the initial decision input as the starting condition for the first round of irrigation decision search; the current irrigation task context includes target plot identifier, crop type, growth stage, soil moisture status, weather change trend, irrigation equipment operation constraints, historical irrigation records and current irrigation target;

[0051] Step S2: Generate the first irrigation decision round based on the initial decision input, and define the first irrigation decision round as the first internal round episode; in the first internal round episode, the meta-enhanced search decision module outputs the irrigation reasoning content, search query action, tool call request and first round candidate irrigation answer in sequence, so that the current irrigation task forms the first round decision trajectory in the manner of "thinking - searching - observing - answering";

[0052] Step S3: Send the search query action to the agricultural knowledge retrieval module, call the external agricultural information source to perform the proxy search, and receive the tool observation results returned by the agricultural knowledge retrieval module; write the tool observation results into the decision trajectory of the current internal round episode, so that the first internal round episode continues to perform irrigation reasoning after receiving external evidence, completes the first round of irrigation decision process, and obtains the search trajectory of the first internal round episode.

[0053] Step S4: After the first internal round episode ends, perform explicit self-reflection processing on the first internal round episode to identify search biases, missing evidence, parameter conflicts, constraint omissions, and defects in the first round of candidate irrigation answers in the first round of irrigation decision-making; and generate reflection content based on the search trajectory, tool observation results, and first round candidate irrigation answers of the first internal round episode; wherein, the reflection content includes suggestions for subsequent search direction correction, priority knowledge categories for retrieval, range of irrigation parameters to be corrected, and constraints to be supplemented;

[0054] Explicit self-reflection includes the following:

[0055] Diagnose whether the search query used in the preceding episode is too broad, too narrow, or deviates from the target crop's water requirements;

[0056] Determine whether there is regional mismatch, reproductive period mismatch, or time lag in the evidence cited in the preceding episode;

[0057] Identify potential risks of over-irrigation, under-irrigation, high energy consumption, or equipment constraint conflicts in the candidate irrigation solutions of the preceding episode;

[0058] Generate reflective suggestions on the knowledge directions to be prioritized for the next subsequent episode, the range of irrigation parameters to be modified, and the execution constraints to be restricted;

[0059] Step S5: Concatenate the search trajectory and reflection content of the first internal round episode, and combine them with the current irrigation task context to form the input context of the second internal round episode; generate the second internal round episode based on the input context of the second internal round episode, so that the second internal round episode, based on inheriting the previous search experience and reflection results, corrects the reasoning direction of the first round of irrigation decision, and regenerates the irrigation reasoning content, search query action, tool call request and updated candidate irrigation answer;

[0060] Step S6: Repeat steps S3 to S5, sequentially performing tool search, evidence backfilling, candidate answer generation, and explicit self-reflection on the second internal round episode and subsequent internal round episodes, and delivering the search trajectory and corresponding reflection content of each preceding internal round episode to the next internal round episode; thus, multiple sequentially generated internal round episodes form a continuously related meta-episode under the same irrigation task; during the generation of the meta-episode, subsequent internal round episodes are modified round by round based on the failure experience, insufficient evidence information, search query bias information, and candidate strategy effect information of the preceding internal round episodes, adjusting the search direction, evidence selection method, irrigation parameter range, and irrigation execution strategy; thus, the originally independent single irrigation search process is transformed into a cross-round continuously optimized meta-reinforcement search process, so as to realize the transfer of the previous exploration results to the subsequent utilization process;

[0061] Step S7: In the meta-episode, extract the candidate irrigation answers corresponding to each internal episode generated under the same irrigation task, and send them to the strategy verification and reward evaluation module to perform round verification. Receive the round reward sequence returned by the module to characterize the comprehensive quality of each round's candidate irrigation answer in terms of soil moisture content compliance, crop water requirement satisfaction, irrigation uniformity, energy consumption level, over-irrigation risk, under-irrigation risk, and rule consistency.

[0062] Step S8: Based on the round reward sequence, construct the reward propagation relationship according to the generation order of the internal round episode in the meta-episode, and perform a discount cumulative calculation on the round rewards of subsequent episodes to obtain the cross-round benefit attribution results corresponding to the previous episode; based on the cross-round benefit attribution results, update the decision strategy of the meta-reinforcement search decision module, so that the strategy will prioritize the generation of search behaviors and irrigation parameter selection strategies with higher decision benefits in subsequent irrigation tasks; determine the target irrigation decision result from the candidate irrigation answers corresponding to multiple internal round episodes; wherein, the target irrigation decision result is the optimal irrigation scheme obtained after multiple rounds of proxy search, self-reflection correction and cross-round benefit attribution, and the target irrigation decision result includes irrigation start time, irrigation volume, irrigation duration, zonal irrigation order, execution mode and water and fertilizer synergy parameters;

[0063] In the cross-round payoff attribution process, the cumulative advantage value of the nth inner round episode is represented as:

[0064] ;

[0065] in, Indicates the first In the meta-episode of the first meta-episode The cumulative advantage value of each internal round episode; Indicates the sequence number of the subsequent internal episode. This indicates the total number of episodes within the meta-round; This represents the cross-round discount factor, used to model the decay of returns in future rounds; the exponential term... This indicates the time interval between the current round and subsequent rounds, thus ensuring that rewards further away from the current round have a smaller impact on the cumulative advantage value; Indicates the first In the first meta-round The relative round reward corresponding to each internal round episode.

[0066] Example 5, based on Example 4, introduces a preference-aware strategy verification reward evaluation method to verify and score the candidate irrigation answers output by each irrigation decision episode, forming a round reward sequence. The specific steps include:

[0067] Step B1: Parse the irrigation execution elements from the candidate irrigation answers for each internal episode in the meta-episode. The irrigation execution elements include irrigation zone, irrigation start time, irrigation duration, target water volume, valve start / stop sequence, pump operation strategy, water and fertilizer synergy parameters, and weather disturbance response strategy. Align the irrigation execution elements with the tool observation results on which the internal episode was generated to form the irrigation strategy representation to be verified for the current internal episode.

[0068] Step B2: Perform consistency verification, constraint verification, and effectiveness verification on the characterization of the irrigation strategy to be verified according to the preset irrigation verification rules. Among them, consistency verification is used to determine whether the candidate irrigation answer is consistent with the soil moisture status, weather forecast information, crop water requirement patterns, and equipment operating conditions; constraint verification is used to determine whether the candidate irrigation answer meets the upper limit of water volume, irrigation time window, equipment load, valve linkage logic, and agronomic safety boundary; effectiveness verification is used to determine the improvement effect of the candidate irrigation answer on the degree of achievement of soil target moisture content, the degree of suppression of under-irrigation risk, the degree of control of over-irrigation risk, water consumption per unit area, energy consumption per unit area, and irrigation uniformity after execution; and generate the original verification score and corresponding verification feature vector of this internal round episode based on the verification results.

[0069] Step B3: For the irrigation strategy representation to be verified in each internal round episode, construct the corresponding preference exchange strategy representation. The preference exchange strategy representation is constructed by exchanging the priority execution parameter direction, resource allocation tendency, partition irrigation order bias, or preferred / inferior selection mark in the candidate irrigation answer. It is used to represent the virtual preference strategy that is opposed to the current candidate irrigation answer. The encoder performs encoding processing on the irrigation strategy representation to be verified and the preference exchange strategy representation, and performs latent variable mapping to generate the original strategy posterior distribution and the exchange strategy posterior distribution, and outputs additional context information.

[0070] The original posterior distribution of the irrigation strategy generated from the characterization of the strategy to be verified is represented as follows:

[0071] ;

[0072] in, Indicates that the parameter is The posterior distribution of the preference coding model output; This represents a latent variable used to characterize the irrigation preference features of the current internal episode; This represents the preference data corresponding to the irrigation strategy to be verified; This represents the mean vector of the posterior distribution of the original policy. Represents the variance vector of the posterior distribution of the original policy;

[0073] The posterior distribution of the exchange policy generated by the preference exchange policy representation is expressed as follows:

[0074] ;

[0075] in, The preference exchange strategy represents the corresponding exchange preference data. Let represent the mean vector of the posterior distribution of the exchange strategy. Represents the variance vector of the posterior distribution of the exchange strategy;

[0076] Step B4: Based on the original strategy posterior distribution and the exchanged strategy posterior distribution, perform exchange-guided constraint processing to ensure that the mean representation in the original strategy posterior distribution satisfies the sign reversal relationship before and after preference exchange, and that the log-variance representation in the original strategy posterior distribution remains unchanged before and after preference exchange. Through this exchange-guided constraint, information related to irrigation preference direction is encoded in a reversible latent representation, and information related to plot background, crop stage, equipment stability, and environmental uncertainty is encoded in an exchange-invariant latent representation. By constructing a preference mirror structure in the latent variable space in the above manner, the phenomenon of latent variable collapse is suppressed, and the ability of latent variables to express the characteristics of irrigation decision preferences is enhanced. This ensures that the latent preference representation generated during the reward evaluation process can stably carry the irrigation strategy tendency, risk bias characteristics, and multi-objective trade-offs of the current internal episode, thereby constructing a basic latent variable distribution that satisfies the preference mirror structure constraint.

[0077] In this embodiment, the representation structure and distribution relationship of the preference mirror latent variable are as follows: Figure 2 As shown; Figure 2 This is a posterior distribution plot of the latent variables mirrored in the figure, used to illustrate the distribution and mirror relationship of the original policy representation and the exchanged policy representation in the latent variable space during the verification and reward evaluation of the preference-aware strategy. The horizontal axis represents latent variable dimension 1, and the vertical axis represents latent variable dimension 2. The original policy posterior sample represents the set of latent variable sampling points generated based on the candidate irrigation strategy, and the exchanged policy posterior sample represents the set of latent variable sampling points generated based on the candidate strategy constructed after preference exchange. The original mean vector represents the center position of the original policy posterior distribution, and the exchanged mean vector represents the center position of the exchanged policy posterior distribution. As can be seen from the figure, the original mean vector and the exchanged mean vector are distributed in opposite directions, with a mean cosine similarity of -1.0000, indicating that they form a strictly opposite directional relationship in the latent variable space. Simultaneously, the mirror error is 0.0707, indicating that they basically satisfy a symmetrical distribution relationship at the amplitude level, with only a small offset. The above results are used to characterize the reversible encoding characteristics of preference information in the latent variable space, and the stable retention characteristics of environmental uncertainty information before and after preference exchange.

[0078] Step B5: After completing the construction of the basic posterior distribution and the exchange guiding constraints, a preference-aware inverse autoregressive flow transformation is performed on the basic latent variables in the basic latent variable distribution. The additional context information output by the encoder is decomposed into exchange-inverted context components and exchange-invariant context components, which are then used to modulate the translation and scaling functions in the inverse autoregressive flow transformation, respectively. Specifically, the translation function is modulated using the exchange-inverted context components to characterize the irrigation preference signal whose direction changes due to the exchange of candidate irrigation preferences. The scaling function is modulated using the exchange-invariant context components to characterize the plot background conditions, irrigation equipment operating status, crop growth stage, and environmental constraint information that do not change with preference exchange. Through the above preference-aware inverse autoregressive flow transformation, the basic latent variables are mapped to a latent variable space that can express the multimodal irrigation preference structure, thereby obtaining the latent variables after the inverse autoregressive flow transformation.

[0079] In the k-th step flow transformation, the preference-aware inverse autoregressive flow is represented as:

[0080] ;

[0081] in, Indicates the first Latent variables after step flow transformation Indicates the first Latent variables after step flow transformation; Indicates the first A preference-perceived inverse autoregressive flow transformation function; This indicates the swapping of the inverted context components. Indicates the exchange of invariant context components; This indicates that the context component is reversed by the exchange. The modulation shift function; Represents the commutative invariant context component The scaling function of the modulation;

[0082] After the K-step flow transformation, the final latent variable posterior distribution satisfies:

[0083] ;

[0084] in, This represents the total number of layers in the inverse autoregressive flow transform. Represents the basic latent variables. Indicates the process The final latent variable after step-inverse autoregressive flow transformation. This represents the posterior logarithmic density of the underlying latent variables. This represents the posterior log density of the final latent variable; The dimension of the latent variable vector. Indicates the first In the step flow transformation, the first Scaling factor for each latent variable dimension Indicates the first Step flow transformation in the first Log-scale variation across each latent variable dimension;

[0085] Step B6: Perform joint reward evaluation on the latent variables after inverse autoregressive transformation and the validation feature vector, calculate the reward for the candidate irrigation strategy of the current internal episode, obtain the round reward of the corresponding internal episode, and form a round reward sequence according to the generation order of the internal episodes.

[0086] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. An intelligent decision-making system for agricultural irrigation based on artificial intelligence, characterized in that, The system includes: The irrigation information acquisition module collects raw irrigation monitoring data; The protocol adaptation module processes the raw irrigation monitoring data, organizes the hierarchical relationships, and constructs the context of the current irrigation task. The agricultural knowledge retrieval module calls external agricultural information sources to perform proxy searches and obtain tool observation results; The meta-enhanced search decision module generates multiple internal episodes based on the current irrigation task context. In each internal episode, a search query action is generated and the tool observation results returned by the agricultural knowledge retrieval module are received. The tool observation results are introduced into the irrigation reasoning process to generate corresponding candidate irrigation answers. An explicit self-reflection mechanism is introduced between internal episodes. Through multiple rounds of internal episode iteration, a meta-episode is formed, and multiple candidate irrigation answers are aggregated to determine the target irrigation decision result. The strategy verification and reward evaluation module introduces a preference-aware strategy verification and reward evaluation method to verify and score multiple candidate irrigation answers, generate round rewards for corresponding internal round episodes, form a round reward sequence, and return it to the meta-reinforcement search decision module. The irrigation execution scheduling module performs structured analysis and task scheduling processing on the target irrigation decision results; The process of introducing a preference-aware strategy to validate reward evaluation methods, validating and scoring multiple candidate irrigation answers, and forming a round-based reward sequence includes the following steps: Step B1: Parse the irrigation execution elements from the candidate irrigation answers for each inner round episode in the meta-episode, and align the irrigation execution elements with the tool observation results on which the inner round episode was generated to form a representation of the irrigation strategy to be verified. Step B2: Perform consistency verification, constraint verification, and effectiveness verification on the representation of the irrigation strategy to be verified according to the preset irrigation verification rules, and generate the original verification score and corresponding verification feature vector of this internal round episode. Step B3: For the irrigation policy representation to be verified in each internal round episode, construct the corresponding preference exchange policy representation; perform encoding processing on the irrigation policy representation to be verified and the preference exchange policy representation through the encoder, perform latent variable mapping, generate the original policy posterior distribution and the exchange policy posterior distribution, and output additional context information. Step B4: Based on the original policy posterior distribution and the exchange policy posterior distribution, perform exchange-guided constraint processing and fuse them to obtain the basic latent variable distribution; Step B5: After completing the exchange guiding constraints, perform a preference-aware inverse autoregressive flow transformation on the basic latent variables in the distribution of basic latent variables, decompose the additional context information into exchange-inverted context components and exchange-invariant context components; and use them to regulate the translation function and scaling function in the inverse autoregressive flow transformation respectively; through the above preference-aware inverse autoregressive flow transformation, map the basic latent variables to obtain the latent variables after the inverse autoregressive flow transformation; Step B6: Combine the latent variables after the inverse autoregressive flow transformation with the validation feature vector to perform joint reward evaluation, obtain the round reward for the corresponding internal episode, and form a round reward sequence.

2. The intelligent decision-making system for agricultural irrigation based on artificial intelligence according to claim 1, characterized in that: The current irrigation task context includes target plot identification, crop type, growth stage, soil moisture status, weather change trends, irrigation equipment operation constraints, historical irrigation records, and current irrigation objectives.

3. The intelligent decision-making system for agricultural irrigation based on artificial intelligence according to claim 1, characterized in that: The process of determining the outcome of a target irrigation decision includes the following steps: Step S1: Initialize the current irrigation task context and construct the initial decision input; Step S2: Generate the first irrigation decision round based on the initial decision input, and define the first irrigation decision round as the first internal round episode; in the first internal round episode, output the irrigation reasoning content, search query action, tool call request and the first round of candidate irrigation answers in sequence; Step S3: Send the search query to the agricultural knowledge retrieval module, execute the proxy search, and receive the tool observation results; write the tool observation results into the decision trajectory of the current internal round episode, continue to execute irrigation reasoning, complete the first round of irrigation decision process, and obtain the search trajectory of the first internal round episode; Step S4: After the first internal round episode ends, perform explicit self-reflection processing on the first internal round episode to identify search bias, missing evidence, parameter conflicts, constraint omissions and defects in the first round of candidate irrigation answers in the first round of irrigation decision-making process; and generate reflection content based on the search trajectory of the first internal round episode, tool observation results and the first round of candidate irrigation answers. Step S5: Combine the search trajectory and reflection content of the first internal round episode, and combine them with the current irrigation task context to form the input context of the second internal round episode; generate the second internal round episode based on the input context, correct the reasoning direction of the first round irrigation decision, and regenerate the irrigation reasoning content, search query action, tool call request, and updated candidate irrigation answer; Step S6: Repeat steps S3 to S5 to perform tool search, evidence backfilling, candidate answer generation and explicit self-reflection processing on the second internal round episode and its subsequent internal round episodes in sequence, and deliver the search trajectory and corresponding reflection content of each preceding internal round episode to the next internal round episode; forming a meta-episode. Step S7: In the meta-episode, extract the candidate irrigation answers corresponding to each internal episode generated under the same irrigation task, and send them to the strategy verification and reward evaluation module to perform round verification processing, and receive the round reward sequence returned by it. Step S8: Based on the round reward sequence, construct the reward propagation relationship, perform discount accumulation calculation, and obtain the cross-round benefit attribution result corresponding to the previous episode; based on the cross-round benefit attribution result, update the strategy and determine the target irrigation decision result from the candidate irrigation answers corresponding to multiple internal round episodes.

4. The intelligent decision-making system for agricultural irrigation based on artificial intelligence according to claim 3, characterized in that: The reflection includes suggestions for revising the subsequent search direction, prioritizing the knowledge categories to be searched, the range of irrigation parameters to be revised, and the constraints to be supplemented.

5. The intelligent decision-making system for agricultural irrigation based on artificial intelligence according to claim 1, characterized in that: Modulate the translation function using the swapped inverted context component; modulate the scaling function using the swapped invariant context component.

Images