An intelligent irrigation decision method with a look-ahead period
By combining multi-step weather forecasting and Bayesian probabilistic decision-making risk correction with deep learning technology based on physical constraints, the problem of insufficient predictability in irrigation decisions in large and medium-sized irrigation districts has been solved, achieving efficient and scientific irrigation management and improving water resource utilization efficiency and the level of intelligence in irrigation districts.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GANFU PLAIN WATER CONSERVANCY ENG ADMINISTRATION BUREAU OF JIANGXI PROVINCE (JIANGXI PROVINCIAL IRRIGATION TEST CENT STATION)
- Filing Date
- 2026-02-05
- Publication Date
- 2026-06-23
AI Technical Summary
Existing irrigation decision-making methods struggle to achieve multi-step predictive decision-making when dealing with complex large and medium-sized irrigation areas. They rely heavily on the accuracy of weather forecasts and lack physical constraints, leading to deviations in state simulations from physical reality during the forecast period. This makes them unable to effectively address weather uncertainties and water waste.
By employing multi-step weather forecasting, Bayesian probabilistic decision risk correction, and deep learning state extrapolation techniques under physical constraints, a rolling optimization mechanism for irrigation decisions is established. Combined with reinforcement learning algorithms and long short-term memory (LSTM) network models, dynamic simulation and coordinated regulation of irrigation decisions are carried out.
It has enabled timely irrigation decision support, improved the scientific and systematic nature of irrigation water use in large and medium-sized irrigation areas, enhanced the rational allocation and utilization efficiency of water resources, reduced resource waste, and promoted the intelligent upgrading and sustainable development of irrigation areas.
Smart Images
Figure CN122264342A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of agricultural irrigation technology, specifically relating to an intelligent irrigation decision-making method with a predictive timeframe. Background Technology
[0002] Traditional agricultural irrigation management often relies on static data and empirical rules for decision-making, making it difficult to effectively address dynamic environmental changes. This approach frequently fails to provide timely irrigation forecasts, leading to water waste or crop water shortages. With the increasing severity of climate change and water scarcity, there is an urgent need for a smart irrigation decision-making method that can flexibly respond to environmental changes. This method, through accurate forecasting and dynamic adjustments, improves irrigation efficiency, ensures appropriate water supply to crops at different growth stages, and promotes sustainable agriculture.
[0003] Currently, patent CN11084539 discloses an irrigation decision-making learning method that can accurately formulate irrigation decisions for the current decision-making cycle. However, this method can only make decisions on a single day. Given the complex canal systems, long water conveyance distances, and high difficulty in allocation and scheduling in large and medium-sized irrigation districts, the irrigation water supply after the daily decision is often insufficient to meet the requirements of "timely and appropriate irrigation, and irrigation on demand." Therefore, irrigation decisions with a forecast period are needed, fully considering factors such as water supply availability in the irrigation district, multi-source water distribution, and optimized canal system water allocation. However, existing multi-step forecast-based decision-making relies heavily on the accuracy of weather forecasts. In actual operation, weather forecasts have significant uncertainties, and purely data-driven simulation models often ignore the physical logic of farmland water balance, causing the simulated state within the forecast period to easily deviate from physical reality. Summary of the Invention
[0004] The purpose of this invention is to overcome the shortcomings of existing irrigation decision-making methods, such as insufficient foresight, difficulty in handling weather forecast errors, and lack of physical constraints in the simulation process, and to provide an intelligent irrigation decision-making method with a forward-looking perspective. This method establishes a rolling optimization mechanism for irrigation decisions with risk perception capabilities by integrating multi-step weather forecasting, Bayesian probability-based decision risk correction, and deep learning state extrapolation technology under physical constraints. It can provide timely decision support for the dynamic simulation and coordinated control of the entire process of supply, transmission, distribution, irrigation, and drainage in the construction of digital twin irrigation districts, improve the scientific and systematic nature of irrigation water management in large and medium-sized irrigation districts, and ensure the rational allocation and efficient utilization of water resources.
[0005] To achieve the above objectives, the present invention adopts the following technical solution: A smart irrigation decision-making method with predictive capabilities includes the following steps: Step 1. Obtain the future irrigation area n Daily weather forecast sequence for the day, current water depth in the field h1. And crop growth period parameters. The daily weather forecast sequence shall include at least the weather forecast data such as maximum temperature, minimum temperature, and weather type forecast; the crop growth period parameters shall include at least the crop coefficient and the suitable irrigation mode for the crop (lower limit of irrigation trigger). h min and the maximum allowable water storage after rain H p ).
[0006] Step 2. Slice the daily weather forecast sequence into segments according to time sequence. k A series of consecutive forecast subsequences, in which k = n - L +1, the length of each subsequence is L sky, L The preset forecast period length; Step 3. Set the iteration variable t Its initial value is 1; Step 4. Place the first t One forecast subsequence and current field surface water depth h t The input is fed into the irrigation decision model, and the output is the first... t Irrigation decision-making actions a t The irrigation decision model is a data-driven intelligent decision model, and its state space input includes at least the current water depth. h t and the length is L The forecast subsequence for the day; Step 5. Use a Bayesian probability model to assess the uncertainty risk of rainfall during the forecast period and adjust preliminary irrigation decisions accordingly. a t Make adjustments to obtain the final irrigation execution volume. I t ; Step 6. Using a physically constrained Long Short-Term Memory (LSTM) network model, input the irrigation execution amount. I t Current water depth h t Based on weather forecast data, the prediction and calculation of the first... t +1 day initial field forecast water depth h t+1 ; Step 7. Set the iteration variable t Add 1, and update the current forecast subsequence to the 1st digit. t +1 subsequence, update the current field surface water depth to h t+1 ; Step 8. Determine t Is it less than or equal to? k If yes, return to step 4; otherwise, output the irrigation execution sequence for the entire forecast period. I 1, I 2, …, I k ].
[0007] Furthermore, the irrigation decision model employs a reinforcement learning algorithm as its machine learning method, and the design of its reward function comprehensively considers the rainfall utilization efficiency and crop water stress degree after the irrigation decision is implemented.
[0008] Furthermore, the reinforcement learning algorithm employs an offline learning algorithm based on value function approximation, fitting the state-action value function through a neural network. Q The neural network adopts a dual-network structure, namely, it includes a master network. Q Network and a goal Q Network, where the target Q The network regularly from the main Q Network update parameters to eliminate Q Value estimation bias; the selection of irrigation decision actions is based on the main Q Network output Q The value is used to select the optimal irrigation decision action through a greedy strategy. a t .
[0009] Furthermore, the irrigation execution volume I t The correction process is as follows: 5.1 Constructing the probability density function of historical forecast errors f ( E ): Collecting historical daily rainfall forecasts R fore,hist Compared with the measured rainfall R real,hist Calculate the forecast error E = R real,hist - R fore,hist For different forecast rainfall levels, the error is fitted using a Gaussian distribution function. E probability distribution:
[0010] In the formula, E This represents the forecast error value; μ This represents the average historical forecast error. σ This represents the standard deviation of historical forecast errors.
[0011] 5.2 Based on forecasted rainfall R fore Calculate the actual rainfall during the forecast period using Bayes' theorem. R real The posterior probability distribution:
[0012] In the formula, P ( R real | R fore The known forecast rainfall is... R fore At that time, the actual rainfall was R real The probability distribution; P ( R fore | R real ) is the likelihood function, derived from the historical forecast error probability density function. f ( E This indicates that the actual rainfall was [determined / confirmed]. R real Time forecast is R fore The probability of; P ( R real ) represents the prior probability, which is the probability of rainfall based on historical statistics.
[0013] 5.3 Calculate the risk index of post-irrigation rain based on the posterior probability distribution RI :
[0014] In the formula, RI This is a risk indicator with a value range of [0, 1], representing the probability of drainage caused by rain after irrigation. C filed The remaining water storage capacity in the field is calculated using the following formula: C filed = H p -( h t + a t - AND c,t - P t ),in H p This is the upper limit for water storage allowed after rain. h t For the first t Initial water layer deptha t This is a preliminary recommendation for irrigation volume. AND c,t To forecast crop water requirements; P t To predict the amount of seepage in the field.
[0015] 5.4 Calculate the final irrigation execution volume I t :
[0016] In the formula, I t The amount of irrigation to be implemented in the final decision; the The preset risk correction coefficient is used to adjust the sensitivity of the decision to rainfall risk. The higher the coefficient, the stronger the tendency to utilize rainfall and the more conservative the decision, that is, the more inclined to "irrigate less and wait for rain".
[0017] Furthermore, the process of predicting water depth using a physically constrained LSTM model is as follows: 6.1 The revised irrigation execution volume I t Weather subsequence characteristics, crop coefficient K c,t and current water depth h t As the input vector of the LSTM model X t ; 6.2 Introducing physical constraints based on the principle of water balance S phy : S phy = h t + R t + I t - D t - AND c,t - P t In the formula, R t For the first t Forecast rainfall for the day; D t For forecasting drainage volume; AND c,t To forecast crop water requirements; P t To predict field seepage; 6.3 Constructing an LSTM model under physical constraints to calculate the first... t+ 1-day initial field forecast water depth h t+1 The specific formula is as follows: h t+1 =ReLU((1- ω )· f LSTM ( X t )+ ω · S phy )) In the formula, f LSTM This is the residual correction term for the output of the Long Short-Term Memory network; ω The physical constraint weights are assigned coefficients, with values ranging from [0,1]; ReLU is the non-negative constraint operator, mathematically expressed as max(0, x This ensures that the water depth is not negative.
[0018] Furthermore, the predicted crop water requirement AND c,t Obtained through the following means: according to the... t The highest temperature on the first day of the forecast subsequence T max,t and lowest temperature T min,t The Hargreaves-Samani model was used to calculate the evapotranspiration of a reference crop. AND 0,t Combined with the crop coefficient in the crop growth period parameters mentioned above K c,t ,get AND c,t = K c,t · AND 0,t .
[0019] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of the above-described method.
[0020] Meanwhile, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.
[0021] This invention establishes a rolling optimization mechanism for irrigation decisions with risk perception capabilities by introducing multi-step weather forecasting, decision risk correction based on Bayesian probability, and deep learning state inference technology under physical constraints. It can provide accurate decision inputs with high robustness and timeliness for the simulation of the entire process of water supply, transmission, distribution, irrigation and drainage in digital twin irrigation districts, significantly improving the predictability, scientificity and systematicness of irrigation water management in large and medium-sized irrigation districts in the face of variable weather environments, and providing solid technical support for the efficient allocation of water resources and agricultural water conservation.
[0022] The role and effect of invention This invention proposes an intelligent irrigation decision-making method with a forecast period by introducing multi-step weather forecasting, Bayesian risk assessment, and a physically constrained rolling decision-making mechanism. The method first obtains the environmental parameters required for the irrigation area in the current decision-making cycle. After processing these parameters, a data-driven intelligent decision-making model generates a preliminary irrigation strategy. Then, a Bayesian probability model assesses the forecast period risk and corrects it to obtain the final irrigation execution amount. A Long Short-Term Memory (LSTM) network model under physical constraints predicts and calculates the initial forecast water depth for day t+1, and updates the weather forecast data subsequence to the t+1th subsequence. Subsequently, based on the updated environmental parameters, the irrigation decision for day t+1 of the forecast period is obtained. If the current decision is for the last forecast period, the irrigation execution amount sequence for that period is output; otherwise, the next day's irrigation decision is formulated, continuing until the desired irrigation decision for the last forecast period is obtained.
[0023] This invention effectively addresses the irrigation scheduling risks posed by weather uncertainties, solves the physical distortion problem that may occur in state extrapolation using purely data-driven models, and realizes a transformation from traditional passive response irrigation to proactive, predictive, and precise risk control management. By providing refined irrigation plans for the next few days, it offers core decision support with high physical consistency and robustness for the coordinated control of the entire process of supply, transmission, distribution, irrigation, and drainage in digital twin irrigation districts, significantly improving the predictability, systematicness, and precision of water resource management in irrigation districts. The implementation of this method helps to fully realize the potential for agricultural water conservation, greatly improve rainfall utilization, and reduce ineffective irrigation and drainage, providing an effective technical approach for achieving optimal allocation and efficient utilization of water resources. It is of great significance for promoting the intelligent upgrading and sustainable development of modern irrigation districts. Attached Figure Description
[0024] Figure 1 This is a flowchart of the present invention.
[0025] Figure 2 This is an operation flowchart of an embodiment of the present invention. Detailed Implementation
[0026] The present invention will now be further described with reference to the accompanying drawings.
[0027] like Figure 2 As shown, the intelligent irrigation decision-making method with a forecast period of 1 to 4 days provided in this embodiment includes the following steps: S110. Obtain the irrigation area for the next 10 days ( n Daily weather forecast data (=10) including maximum temperature, minimum temperature, and rainfall forecasts; simultaneously, crop growth period parameters, including crop coefficient K, are obtained. c Suitable water depths for each crop growth stage (lower limit of irrigation trigger) h min Permissible water storage depth after rain H p ); and the initial field water depth on day 1 was measured. h 1.
[0028] S120. Divide the 10-day weather forecast sequence into 4 consecutive subsequences in chronological order. k =10-7+1), each subsequence is 7 days long, namely: forecast data for days 1-7, 2-8, 3-9, and 4-10.
[0029] S130. Combine the first subsequence (forecast data for days 1-7) with the current water depth. h 1. Input the common data into the pre-trained irrigation decision model, and output the irrigation decision action for day 1. a 1.
[0030] The irrigation decision model is constructed based on a deep reinforcement learning algorithm, characterized in that: The state space of the model includes the current field water depth and the weather forecast sequence data for the next 7 days. The model's action space is discretely designed, including five irrigation levels: {0, 15, 30, 45, 60} mm. The model employs a dual-network structure, updating parameters by minimizing temporal difference error; its loss function is: L ( θ ) = Ε[( r + γQ ( s ', argmax Q ( s ′, a ′; θ ); θ − ) - Q ( s , a ; θ )) 2] in, θ and θ − These represent the parameters of the main network and the target network, respectively. γ Discount factor; The parameters of the target network are periodically updated synchronously from the main network, and the update formula is: θ − ← τθ + (1 − τ ) θ − in, τ This is a soft update coefficient; The reward function of the model is designed as follows: R t = α ⋅ RU t + β ⋅ WSI t in, RU t Indicates the first t Rainfall utilization rate per day reflects the proportion of actual rainfall utilized to the total rainfall. WSI t This represents the water stress index calculated based on the Jenson water production function model; α and β This is the weighting coefficient, which can be set to 1.
[0031] S140. Use a Bayesian probability model to assess the uncertainty risk of rainfall during the forecast period and to influence preliminary irrigation decisions. a t Make adjustments to obtain the final irrigation execution volume. I t The correction process is characterized by the following: 1) Constructing the error distribution: Collect historical daily forecast and measured rainfall data for the region, and calculate the forecast error. E The error was fitted using a Gaussian distribution function. E From the probability distribution, we obtain the probability density function of historical forecast errors. f ( E ); 2) Posterior probability derivation: based on the forecast rainfall for day 1 R fore The Bayesian formula was used to calculate the actual rainfall during the forecast period. R real posterior probability distribution P ( R real| R fore ); 3) Risk assessment: Calculate the post-irrigation rain risk index for the decision-making process on day 1 based on the posterior probability distribution. RI The calculation formula is as follows:
[0032] In the formula, RI This is a risk indicator with a value range of [0, 1], representing the probability of drainage caused by rain after irrigation. C filed The remaining water storage capacity in the field is calculated using the following formula: C filed = H p -( h t + a t - AND c,t - P t ),in H p This is the upper limit for water storage allowed after rain. h t For the first t Initial water layer depth a t This is a preliminary recommendation for irrigation volume. AND c,t To forecast crop water requirements; P t To predict field seepage; 4) Decision Adjustment: The final irrigation execution amount is calculated using the risk adjustment operator. I 1= a 1·(1- the · RI ),in the This is a preset risk correction factor.
[0033] S150. Using a physically constrained Long Short-Term Memory (LSTM) network to predict and calculate the initial field forecast water depth for the second day. h 2. Its characteristic is that its prediction process is as follows: 1) Input Construction: This will include the corrected irrigation execution volume. I t Weather subsequence characteristics, crop coefficient K c,t and current water depth h t As the input vector of the LSTM model X t ; 2) Physical constraints: Introduce physical constraint terms based on the principle of water balance. S phy = h 1+ R 1+ I 1- D 1- AND c,1 - P 1, among which, R 1. D 1. AND c,1 and P 1. The forecast includes the predicted rainfall, predicted drainage, predicted crop water requirement, and predicted field seepage for the first day; wherein, the... AND c,1 The calculation process is as follows: First, the Hargreaves-Samani model was used to calculate the evapotranspiration of the reference crop. AND 0,HS : AND 0,HS = C · R a · ( T max - T min ) E · [( T max + T min ) / 2 + 17.8] Where parameters C and E Calibration was performed using the least squares method based on historical Penman-Monteith model results, and the recommended values are as follows: C = 0.0023, E = 0.5; R a It is solar radiation outside the Earth; T max and T min These are the predicted maximum and minimum temperatures, respectively; then, combined with the crop coefficient... K c Calculate and forecast crop water requirements AND c,1 .
[0034] 3) Hybrid prediction: Constructing a physically constrained LSTM model to calculate... h 2. The specific formula is as follows: h 2 = ReLU((1- ω )· f LSTM ( X 1)+ ω · S phy )) in, f LSTM This is the residual correction term for the output of the Long Short-Term Memory network; ω The coefficients are assigned to the physical constraint weights; ReLU is a non-negative constraint operator that ensures the water depth is not negative.
[0035] S160. Change the iteration variable t Add 1, and update the current forecast subsequence to the second subsequence (days 2-8), and update the current field surface water depth to... h 2; Return to the execution steps S130 to S150 and output the final irrigation execution amount for day 2. I 2; S170. Judgment t Is it less than or equal to the total number of subsequences? k Due to the current situation t =2 < 4, return to step S130, and continue iterating to generate irrigation decisions for day 3 and day 4.
[0036] S180. After completing the decision-making process with a 4-day forecast period, output the final sequence of actual irrigation execution amounts. I 1, I 2, I 3, I 4).
[0037] This embodiment establishes a robust and physically consistent irrigation decision optimization method by deeply coupling deep reinforcement learning, Bayesian risk assessment correction, and rolling forecasting using a Long Short-Term Memory (LSTM) network under physical constraints. The discretized action space design ensures both the engineering practicality of the decision and reduces algorithmic complexity; the calibration of the Hargreaves-Samani model parameters ensures the accuracy of the reference evapotranspiration calculation; the reward function, based on the Jenson water production function model, maximizes irrigation benefits by comprehensively balancing rainfall use efficiency and crop water stress; the Bayesian risk correction mechanism quantifies the uncertainty risk of weather forecasts in real time by modeling the distribution of historical forecast errors and deriving posterior probabilities, utilizing risk indicators... RI The dynamic adjustment of the decision by the correction operator effectively avoids resource waste caused by "rain after irrigation" and ensures the robustness of the decision under variable weather conditions. The LSTM model under physical constraints adopts a dual-drive mode of "data feature extraction + physical water balance", using LSTM to capture the nonlinear temporal characteristics of farmland water changes, while introducing...S phy The physical baseline value and the ReLU nonnegative constraint operator are rigidly verified to ensure that the simulation of water layer depth evolution within the forecast period conforms to the physical constraints. By obtaining the irrigation decision sequence in the short term, this method provides accurate decision input data for the simulation of the entire process of water supply, transmission, distribution, irrigation and drainage in digital twin irrigation districts. It significantly improves the scientificity, systematicness and predictability of irrigation water process management in large and medium-sized irrigation districts, and provides effective technical support for achieving refined allocation and efficient utilization of water resources.
Claims
1. A smart irrigation decision-making method with predictive forecasting, characterized in that, Includes the following steps: Step 1. Obtain the future irrigation area n Daily weather forecast sequence for the day, current water depth in the field h 1. And crop growth period parameters; Step 2. Slice the daily weather forecast sequence into segments according to time sequence. k A series of consecutive forecast subsequences, in which k = n - L +1, the length of each subsequence is L sky, L The preset forecast period length; Step 3. Set the iteration variable t Its initial value is 1; Step 4. Place the first t One forecast subsequence and current field surface water depth h t The input is fed into the irrigation decision model, and the output is the first... t Irrigation decision-making actions a t The irrigation decision model is a data-driven intelligent decision model. Step 5. Use a Bayesian probability model to assess the uncertainty risk of rainfall during the forecast period and adjust preliminary irrigation decisions accordingly. a t Make adjustments to obtain the final irrigation execution volume. I t ; Step 6. Using a physically constrained Long Short-Term Memory (LSTM) network model, input the irrigation execution amount. I t Current water depth h t Based on weather forecast data, the prediction and calculation of the first... t +1 day initial field forecast water depth h t+1 ; Step 7. Set the iteration variable t Add 1, and update the current forecast subsequence to the 1st digit. t +1 subsequence, update the current field surface water depth to h t+1 ; Step 8. Determine t Is it less than or equal to? k If so, return to step 4; If not, output the irrigation execution sequence for the entire forecast period. I 1, I 2, …, I k ].
2. The intelligent irrigation decision-making method with predictive timeframe according to claim 1, characterized in that: in, In step 1, the daily weather forecast sequence must include at least the weather forecast data such as maximum temperature, minimum temperature, and weather type forecast; the crop growth period parameters must include at least the crop coefficient and the suitable irrigation pattern for the crop (lower limit of irrigation trigger). h min and the maximum allowable water storage after rain H p ).
3. The intelligent irrigation decision-making method with predictive timeframe according to claim 1, characterized in that: in, The irrigation decision model in step 4 is a model trained using machine learning methods, and its state space input includes at least the current water depth. h t and the length is L The forecast subsequence for each day; the design of its reward function comprehensively considers the rainfall utilization efficiency and crop water stress degree after the implementation of irrigation decisions.
4. The intelligent irrigation decision-making method with predictive timeframe as described in claim 3, Its features are: The reinforcement learning algorithm employs an offline learning algorithm based on value function approximation, which fits the state-action value function through a neural network. Q The neural network adopts a dual-network structure, namely, it includes a master network. Q Network and a goal Q Network, where the target Q The network regularly from the main Q Network update parameters to eliminate Q Value estimation bias; the selection of irrigation decision actions is based on the main Q Network output Q The value is used to select the optimal irrigation decision action through a greedy strategy. a t .
5. The intelligent irrigation decision-making method with predictive timeframe according to claim 1, characterized in that: in, In step 5, irrigation execution volume I t The correction process is as follows: 5.1 Constructing the probability density function of historical forecast errors f ( E ): Collecting historical daily rainfall forecasts R fore,hist Compared with the measured rainfall R real,hist Calculate the forecast error E = R real,hist - R fore,hist For different forecast rainfall levels, the error is fitted using a Gaussian distribution function. E probability distribution: In the formula, E This represents the forecast error value; μ This represents the average historical forecast error. σ This represents the standard deviation of historical forecast errors; 5.2 Based on forecasted rainfall R fore Calculate the actual rainfall during the forecast period using Bayes' theorem. R real The posterior probability distribution: In the formula, P ( R real | R fore The known forecast rainfall is... R fore At that time, the actual rainfall was R real The probability distribution; P ( R fore | R real ) is the likelihood function, derived from the historical forecast error probability density function. f ( E This indicates that the actual rainfall was [determined / confirmed]. R real Time forecast is R fore The probability of; P ( R real ) represents the prior probability, which is the probability of rainfall based on historical statistics; 5.3 Calculate the risk index of post-irrigation rain based on the posterior probability distribution RI : In the formula, RI This is a risk indicator with a value range of [0, 1], representing the probability of drainage caused by rain after irrigation. C filed The remaining water storage capacity in the field is calculated using the following formula: C filed = H p -( h t + a t - ET c,t - P t ),in H p This is the upper limit for water storage allowed after rain. h t For the first t Initial water layer depth a t This is a preliminary recommendation for irrigation volume. ET c,t To forecast crop water requirements; P t To predict field seepage; 5.4 Calculate the final irrigation execution volume I t : In the formula, I t The amount of irrigation to be implemented in the final decision; η The preset risk correction coefficient is used to adjust the sensitivity of decision-making to rainfall risk. The higher the coefficient, the stronger the tendency to utilize rainfall and the more conservative the decision-making, that is, the more inclined to "irrigate less and wait for rain".
6. The intelligent irrigation decision-making method with predictive timeframe according to claim 1, characterized in that: in, In step 6, the process of predicting water depth using a physically constrained LSTM model is as follows: 6.1 The revised irrigation execution volume I t Weather subsequence characteristics, crop coefficient K c,t and current water depth h t As the input vector of the LSTM model X t ; 6.2 Introducing physical constraints based on the principle of water balance S phy : S phy = h t + R t + I t - D t - ET c,t - P t In the formula, R t For the first t Forecast rainfall for the day; D t For forecasting drainage volume; ET c,t To forecast crop water requirements; P t To predict field seepage; 6.3 Constructing an LSTM model under physical constraints to calculate the first... t+ 1-day initial field forecast water depth h t+1 The specific formula is as follows: h t+1 =ReLU((1- ω )· f LSTM ( X t )+ ω · S phy )) In the formula, f LSTM This is the residual correction term for the output of the Long Short-Term Memory network; ω The physical constraint weights are assigned coefficients, with values ranging from [0,1]; ReLU is the non-negative constraint operator, mathematically expressed as max(0, x This ensures that the water depth is not negative.
7. The intelligent irrigation decision-making method with predictive timeframe according to claim 1, characterized in that: in, In step 5, forecast crop water requirements. ET c,t Obtained through the following means: according to the... t The highest temperature on the first day of the forecast subsequence T max,t and lowest temperature T min,t The Hargreaves-Samani model was used to calculate the evapotranspiration of a reference crop. ET 0,t Combined with the crop coefficient in the crop growth period parameters mentioned above K c,t ,get ET c,t = K c,t · ET 0,t .
8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method as described in any one of claims 1 to 6.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by a processor, it implements the steps of the method as described in any one of claims 1 to 7.