An irrigation water resource optimal scheduling method combining deep learning prediction and double-layer sub-planning

Abstract

The application discloses a kind of irrigation water resource optimal scheduling methods combining deep learning prediction and double-layer fractional programming, belong to agricultural water resource optimal allocation and irrigation engineering technical field.Collect historical hydrology, meteorology, crop and engineering data of irrigation area, and construct LSTM and improved iTransformer coupling model, predict future daily scale water supply and water demand of irrigation area;Based on the prediction result, a double-layer fractional programming model is constructed, the upper layer is used to optimize the upper channel water distribution decision, the lower layer is used to optimize the lower channel water distribution decision, and the water balance, crop water requirement, channel flow, field water balance, drainage and non-negative constraint conditions are set;NSGA-II and Dinkelbach algorithm are used for nested solution, and TOPSIS method is used to select and recommend scheme from Pareto non-inferior solution set.The method can effectively improve the prediction accuracy of water supply and demand of irrigation area, and truly reflect the characteristics of layered water distribution decision.

Description

Technical Field

[0001] This invention belongs to the field of agricultural water resource optimization and irrigation engineering technology, and particularly relates to a method for optimizing the scheduling of irrigation district water resources by combining deep learning prediction and bi-level fractional programming. Background Technology

[0002] With the continued advancement of climate change and the demand for high-quality agricultural development, irrigation districts face multiple challenges in water resource allocation, including increased uncertainty in water supply, intensified fluctuations in crop water demand, and insufficient water transmission and distribution efficiency. As an important spatial unit for agricultural water allocation, the optimal allocation of water resources in irrigation districts is not only related to crop yield and economic benefits, but also to the efficiency of canal system operation, water conservation levels, and the sustainable development of regional agriculture.

[0003] Existing methods for water resource allocation in irrigation districts mainly include empirical statistical analysis, mechanistic simulation, and mathematical optimization. Empirical statistical methods are simple to calculate, but they are insufficient in characterizing the significant nonlinearity, dynamics, and multi-factor coupling relationships in the water supply and demand process of irrigation districts. Mechanistic models can reflect crop water consumption, field water balance, and canal water conveyance processes, but they often have many parameters, are complex to model, and have high requirements for the integrity of basic data and regional adaptability. Although traditional optimization models can be used for water allocation decisions, most of them are based on deterministic input conditions and are difficult to reflect the uncertainty in future changes in water supply and demand.

[0004] In recent years, deep learning technology has demonstrated strong advantages in the field of complex time series forecasting. In particular, Long Short-Term Memory (LSTM) neural networks have shown good performance in extracting time-series dependency information, while Transformer-type models excel in long-sequence modeling and multivariate correlation feature extraction. Introducing deep learning methods into irrigation district water supply and demand forecasting can help improve the accuracy of future water supply and demand predictions, thereby enhancing the reliability of optimized scheduling inputs.

[0005] On the other hand, the water allocation process in irrigation districts typically exhibits a clear hierarchical decision-making characteristic. Upper-level canals are responsible for total volume control and backbone distribution, while lower-level canals are responsible for the detailed allocation of water to the fields. The objectives at different levels are not entirely consistent: the upper level focuses more on the overall economic benefits of the system and the macro-allocation of water resources, while the lower level pays more attention to local water allocation efficiency, water loss, and crop irrigation satisfaction. Existing research mostly uses single-level optimization models, which are insufficient to accurately reflect the hierarchical decision-making mechanism in irrigation district management. Furthermore, the system's economic benefits and water loss control often have an input-output or benefit-loss ratio relationship, making fractional programming a more suitable approach for the actual management needs of irrigation districts. However, bi-level fractional programming problems typically exhibit nonlinearity, non-convexity, and multi-constraint coupling characteristics, making them difficult to solve directly using traditional analytical methods.

[0006] Therefore, there is an urgent need to propose a water resource optimization scheduling method for irrigation districts that can organically combine deep learning-based water supply and demand forecasting, bi-level fractional programming optimization for water allocation, and efficient solution methods under complex constraints. This method can address issues such as insufficient accuracy in water supply and demand forecasting in irrigation districts under the background of climate change, lack of hierarchical characteristics in water allocation decision-making, and difficulty in coordinating economic benefits with water transfer losses. Summary of the Invention

[0007] To address the shortcomings of existing technologies, this invention provides a method for optimizing water resource allocation in irrigation districts by combining deep learning prediction with bi-level fractional programming. This method constructs a coupled prediction model of LSTM and an improved iTransformer to predict daily water supply and demand in the irrigation district. Based on this, a bi-level fractional programming model is constructed and solved using an intelligent optimization algorithm, thereby obtaining an optimal water allocation scheme that adapts to changes in water supply and demand and climate uncertainty.

[0008] To achieve the above objectives, this invention provides a method for optimal water resource scheduling in irrigation districts that combines deep learning prediction with bi-level fractional programming, comprising the following steps:

[0009] S1: Basic data collection and preprocessing.

[0010] Historical hydrological, meteorological, crop, engineering and management data of irrigation districts were collected and studied, and data cleaning, missing data completion, normalization and time series reconstruction were performed to form an input dataset for water supply and demand forecasting and optimal scheduling.

[0011] S2: Prediction of water supply and demand in the irrigation area.

[0012] Based on historical input sequences, a coupled LSTM and improved iTransformer model is constructed to perform multi-step prediction of future daily water supply and demand, and generate the water supply and demand boundary conditions required for the optimized scheduling model.

[0013] S3: Construction of a two-layer fractional planning model for irrigation districts.

[0014] Based on the tiered water allocation management structure of irrigation areas, a two-level fractional programming model is constructed, which includes an upper-level decision-making model and a lower-level decision-making model. The upper-level model aims to maximize the economic benefits of the system, while the lower-level model aims to maximize the irrigation water productivity.

[0015] S4: Constraint Construction.

[0016] The bi-level fractional programming model introduces constraints on available water volume, channel water balance, crop water requirement, channel water conveyance capacity, field water balance, water conveyance time, drainage, and non-negativity.

[0017] S5: Model Solving.

[0018] The Non-Dominated Sorting Genetic Algorithm II (NSGA-II) is used to perform outer-layer optimization search on the upper-level multi-objective problem, and the Dinkelbach algorithm is used to transform the fractional programming objective into a parameterized non-fractional subproblem, which is then nested to solve the lower-level decision problem. If necessary, the genetic algorithm or other heuristic algorithms are combined to obtain the solutions to the subproblems.

[0019] S6: Recommended solution preferred.

[0020] The obtained Pareto non-dominated solution set is comprehensively evaluated using the TOPSIS method to output a recommended water distribution scheme for the irrigation district, including the water flow rate, water distribution duration, field irrigation water depth and corresponding system performance indicators of each level of canal at each time period.

[0021] Preferably, the basic data collection and preprocessing in step S1 includes: historical runoff, river flow, water level, precipitation, reference crop evapotranspiration, average temperature, maximum temperature, minimum temperature, relative humidity, average wind speed, sunshine duration, air pressure, channel design flow, channel length, channel seepage prevention parameters, channel water loss parameters, irrigated area, water requirement indicators for each growth stage of rice, initial water layer depth in the field, soil parameters, and irrigation system parameters.

[0022] Preferably, the prediction of irrigation area water supply and demand in step S2 includes: first, using LSTM to extract local dynamic features from the original multivariate time series to obtain latent state features; then, inputting the latent state features into the improved iTransformer module to learn the global correlation and long-term dependency features between variables; and finally, outputting the predicted values ​​of water supply and demand for the next H time steps through a fully connected layer.

[0023] Preferably, the construction of the two-level fractional programming model for the irrigation district in step S3 includes: an upper-level fractional programming model optimized to maximize the system's economic benefits per unit water loss; and a lower-level fractional programming model optimized to improve irrigation water productivity. Decision variables include the water distribution flow rate, water distribution duration, and field irrigation water depth of the upper and lower level channels.

[0024] The upper-level function is: ,

[0025] In the formula, The net economic benefit of the system corresponding to the water loss per unit volume of water conveyance and distribution in the canal system, CNY / m 3 . The price of rice in the study area is CNY / kg. i and j represent the upstream and downstream channels, respectively. For rice yield, kg / hm 2 . The cost of planting per unit area includes land contracting, seeds, fertilizers, pesticides, and tillage costs, in CNY / hm². 2 . To control the irrigated area through canals, hm 2 . To account for the flow loss during water transmission and distribution in upstream and downstream channels, m 3 / s. , The values ​​are the water delivery and distribution times for the upper and lower level channels, respectively, in seconds. This represents the water supply status of the channel (a 0 / 1 variable). When it equals 1, it indicates that water is being supplied to the lower-level channel; when it equals 0, no water is being supplied.

[0026] ,

[0027] In the formula, These are the soil permeability coefficients of the upper and lower channel beds, respectively. , These represent the water diversion flow rates at the headworks of the upper and lower channels, respectively, in m. 3 / s. , These are the soil permeability indices of the upper and lower level canal beds, respectively. , The lengths of the upper and lower level channels are respectively, in km.

[0028] The Stewart model is used to describe the dynamic response of crop yield to irrigation amount, and its expression is as follows:

[0029] ,

[0030] In the formula, For maximum rice yield, kg / hm 2 n represents the total number of rice growth stages. , where n is the crop water sensitivity index during the growth period. The actual evapotranspiration of the crop during the n-year growth period, in mm, can be obtained by adding the actual daily evapotranspiration of the crop. The maximum evapotranspiration during the growing season n, in mm, can be calculated using the crop coefficient method.

[0031] ,

[0032] In the formula, The actual evapotranspiration of the crop on day d of the growing season is expressed in mm. This refers to the number of days in the reproductive period.

[0033] ,

[0034] In the formula, This is the crop coefficient. To reference crop evapotranspiration (mm), the Penman-Monteith formula was used for calculation:

[0035] ,

[0036] In the formula, The gradient of the saturated vapor pressure temperature curve is given in kPa / ℃. Net radiation, MJ / m 2 ・d. Geothermal flux, MJ / m 2 ・d. The constant for humidity calculation is kPa / ℃. The average temperature at an altitude of 1.5~2.5m is expressed in °C. The wind speed at a height of 2m is in m / s. Hereinafter, denoted as saturated vapor pressure and actual vapor pressure at temperature T, respectively, in kPa.

[0037] The lower-level function is: ,

[0038] In the formula, Water productivity is the total crop yield per unit of net irrigation water, expressed in kg / m³. 3 . The amount of irrigation water used in the field on day d of the growing season is in mm.

[0039] Preferably, the field water balance constraint in step S4 includes the balance relationship between changes in irrigation amount, effective rainfall, evapotranspiration, deep infiltration, drainage and field surface water storage.

[0040] The drainage constraints are as follows: drainage is allowed when the field water depth is greater than the maximum water storage depth after rainfall during a certain period; and the drainage volume is zero when the field water depth is less than or equal to the maximum water storage depth after rainfall during that period.

[0041] Specifically, the constraint conditions include: available water quantity constraints, channel water balance constraints, channel water conveyance capacity constraints, irrigation time constraints, field water balance constraints, drainage constraints, and non-negativity constraints.

[0042] The available water constraint is that the water diverted from the headworks of the upstream canal must not exceed the total available water volume, and the actual net irrigation water received by the crops must be greater than the crop's water requirement. This constraint is expressed as:

[0043] ,

[0044] In the formula, For the total available water volume, m 3 . The water requirement for crops is expressed in mm.

[0045] The channel water balance constraint is that the inflow of each section of the main canal equals the inflow of the next section, the current section's losses, and the sum of the water distribution from all lower-level channels connected to this section. This constraint is expressed as:

[0046] ,

[0047] For the terminal section I of the main canal (without a sub-level main canal), the constraint is expressed as:

[0048] ,

[0049] The total water supply of the lower-level canals during the irrigation cycle is equal to the sum of the daily irrigation water depths in the fields during that cycle. This constraint is expressed as:

[0050] ,

[0051] In the formula, The efficiency of water utilization in the canal system.

[0052] The water conveyance capacity constraint of the channels is that the water flow rate of all levels of the channels must not exceed the channel design flow rate at any time. This constraint is expressed as:

[0053] ,

[0054] In the formula, , Minimum design traffic for upper and lower level channels, m 3 / s. , Maximum design traffic for upper and lower level channels, m 3 / s.

[0055] The irrigation time constraint stipulates that for any subordinate canal, the start and end times of water distribution should fall within the maximum permissible canal system water transport and distribution cycle. This constraint is expressed as:

[0056] ,

[0057] In the formula, h represents the maximum permissible water transport and distribution cycle in the canal system. , These are the start and end times of water distribution in the lower-level channels, respectively, in seconds. This refers to the set of lower-level channels connected to the upper-level channel segment i.

[0058] The field water balance constraint is that changes in water depth are jointly determined by water inputs (rainfall, irrigation) and water outputs (evapotranspiration, drainage, deep seepage). Initial conditions The initial water level after transplanting was set. The growing season irrigation quota constraint controls the total irrigation water consumption within the quota allocated to the irrigation district, reflecting the rigid constraint of water resources. This constraint is expressed as:

[0059] ,

[0060] In the formula, The field water depth on day d of the growing season, in mm. The initial field water depth after rice transplanting is measured in mm. The rainfall on day d of the growing season is in mm. The amount of irrigation water used in the field on day d of the growing season is in mm. The field drainage depth is measured in mm on day d of the growing season. The value represents deep seepage in the field, expressed in mm / d.

[0061] The drainage constraint stipulates that drainage can occur if the field water depth on a given day is greater than the maximum water storage depth after that day's rainfall. Conversely, drainage cannot occur if the field water depth on a given day is less than or equal to the maximum water storage depth after that day's rainfall. This constraint is expressed as:

[0062] ,

[0063] In the formula, The maximum water storage depth of the field surface water layer is measured in mm.

[0064] The nonnegativity constraint means that the decision variables in the model should satisfy the real-world physical meaning of nonnegativity, i.e., all variables are nonnegative. This constraint is expressed as:

[0065] ,

[0066] Preferably, the solution in step S5 includes using the Non-Dominated Sorting Genetic Algorithm II (NSGA-II) to perform outer-layer optimization search on the upper-level multi-objective problem, and using the Dinkelbach algorithm to transform the fractional programming objective into a parameterized non-fractional subproblem, and nesting the solution to the lower-level decision problem. If necessary, the solution to the subproblem can be obtained by combining the genetic algorithm or other heuristic algorithms.

[0067] Preferably, the recommended scheme in step S6 includes comprehensively evaluating the obtained Pareto solution set using the TOPSIS method to output a recommended water allocation scheme for the irrigation district, including the water flow rate, water allocation duration, and field irrigation water depth of each level of canal at each time period. The TOPSIS method comprehensively evaluates the candidate water allocation schemes based on one or more indicators among system economic benefits, total water loss, water productivity, and water allocation time.

[0068] Compared with the prior art, the present invention has the following beneficial effects:

[0069] 1. Accuracy of water supply and demand forecasting;

[0070] This invention combines LSTM with an improved iTransformer, taking into account both local temporal feature extraction and global multivariate correlation modeling, which can effectively improve the prediction accuracy and stability of irrigation district water supply and demand.

[0071] 2. It effectively reflects the tiered water allocation decision-making mechanism in irrigation areas;

[0072] By constructing a two-level split planning model, the scheduling decisions of upper-level and lower-level channels can be distinguished, which can better reflect the hierarchical management characteristics in the actual operation of irrigation districts.

[0073] 3. Synergistic optimization of economic benefits and water transfer losses;

[0074] This invention uses fractional programming to express the efficiency relationship between system benefits and water conveyance losses, making the optimization results better reflect the performance of unit water resource utilization.

[0075] 4. Adapting to climate change and uncertainties in water supply and demand;

[0076] This invention uses deep learning prediction results as optimization input, which can introduce uncertainties such as future water inflow fluctuations and water demand changes into the scheduling process, thereby improving the foresight and adaptability of the water distribution scheme.

[0077] 5. It has strong engineering application value;

[0078] The results obtained by this invention can directly output scheduling information such as flow rate of each channel, water delivery duration and irrigation water depth in the field, providing technical support for refined management of irrigation districts and water-saving irrigation decisions. Attached Figure Description

[0079] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0080] like Figure 1 As shown, a method for optimal water resource scheduling in irrigation districts that combines deep learning prediction with bi-level fractional programming includes the following steps:

[0081] S1: Basic Data Acquisition and Preprocessing

[0082] A rice-irrigated area was selected as the research subject, and historical data from many years were collected, including:

[0083] (1) Hydrological data: runoff, river inflow, water level, etc.;

[0084] (2) Meteorological data: precipitation, average temperature, maximum / minimum temperature, relative humidity, wind speed, sunshine duration, air pressure, reference evapotranspiration;

[0085] (3) Engineering data: length of main canal, branch canal and distribution canal, design flow rate of canal, seepage prevention parameters, water conveyance loss parameters;

[0086] (4) Farmland data: rice planting area, irrigation system at each growth stage, initial water depth in the field, and soil physicochemical parameters;

[0087] (5) Economic data: crop prices, irrigation costs, water conveyance management costs, etc.

[0088] The original data is processed by imputing missing values, removing outliers, standardizing the time scale, and normalizing to form the model input dataset.

[0089] S2: Irrigation District Water Supply and Demand Forecast

[0090] Suppose that at time t, the state of the irrigation system is composed of an input feature sequence consisting of N input variables from the past L time steps. The model uses data from the past L time steps to predict the water supply and demand for the next H time steps. The water supply can be represented by irrigation runoff or inflow, while the water demand can be calculated using reference evapotranspiration and crop coefficients, or directly represented by crop water demand.

[0091] In practice:

[0092] (1) Feed the multivariate input sequence into the LSTM module to extract local time-dependent features;

[0093] (2) Improve the iTransformer module by inputting the hidden state output of LSTM to extract global correlation features and long-term dependencies between variables;

[0094] (3) The prediction results of future water supply and demand are output through the fully connected layer.

[0095] S3: Construct a two-level fractional programming model

[0096] Based on the hierarchical structure of water distribution in the irrigation district, water distribution decisions are divided into upper and lower levels.

[0097] 1. Upper-level model

[0098] The objective of the upper-level model can be expressed as maximizing the system's economic benefits and optimizing the unit loss revenue, that is, increasing the net revenue corresponding to the unit water transfer loss based on considering crop revenue, water supply costs, and system operating costs.

[0099] 2. Lower-level model

[0100] The goal of the lower-level model is to maximize irrigation water productivity and improve irrigation efficiency while meeting the water requirements of crops during their critical growth stages.

[0101] S4: Constraint Construction

[0102] The constraints include, but are not limited to:

[0103] (1) Water balance constraint

[0104] The water intake of the upper-level channels should be balanced with the water distribution of each lower-level channel and the water loss of the channels.

[0105] (2) Crop water requirement constraints

[0106] Irrigation water volume at each growth stage should meet the upper and lower limits of the rice's normal growth requirements for field water layer and soil moisture.

[0107] (3) Channel traffic constraints

[0108] The actual water flow rate of each channel should not be less than the minimum design flow rate and not greater than the maximum design flow rate.

[0109] (4) Field water balance constraints

[0110] The change in field surface water volume over a certain period should satisfy the balance between field surface water storage, irrigation volume, effective rainfall, evapotranspiration, seepage and drainage in the previous period.

[0111] (5) Drainage constraints

[0112] Drainage is permitted when the water depth in the field is greater than the maximum water storage depth after rainfall on a given day; otherwise, the drainage volume is zero.

[0113] (6) Nonnegativity constraint

[0114] All channel water distribution flow, irrigation water volume, drainage volume, and related decision variables should be non-negative.

[0115] S5: Model Solving

[0116] Because the constructed model has characteristics such as double-layer nesting, fractional objectives, and coupling of nonlinear constraints, the following solution strategy is adopted:

[0117] (1) The outer layer uses NSGA-II to perform multi-objective optimization to maximize the economic benefits of the system, minimize the water loss of the canal system, and maximize the crop water productivity, generating a Pareto non-dominated solution set;

[0118] (2) For the given outer decision variables, the inner layer transforms the fractional programming problem into a parameterized non-fractional optimization problem using the Dinkelbach algorithm;

[0119] (3) In each parameter iteration, a genetic algorithm is used to search for the feasible optimal solution of the subproblem;

[0120] (4) When the Dinkelbach iteration satisfies the convergence condition, output the corresponding inner optimal solution;

[0121] (5) Repeat the evolution of the outer population until the maximum number of generations is reached or the Pareto front is stable.

[0122] S6: Recommended Solution Optimization

[0123] The obtained Pareto non-dominated solution set is used to construct a candidate scheme set, and the TOPSIS method is used for comprehensive evaluation to obtain the recommended water allocation scheme.

[0124] The final output includes:

[0125] (1) Water diversion flow rate of the main canal at different times;

[0126] (2) Water flow rate and duration of branch canals and distribution canals at different times;

[0127] (3) Irrigation water depth in the field at different growth stages in each irrigation unit;

[0128] (4) Performance indicators such as total economic benefits of the system, total water loss and water productivity.

Claims

1. A method for optimal water resource scheduling in irrigation districts that combines deep learning prediction with bi-level fractional programming, characterized in that, Includes the following steps: S1: Collect historical hydrological, meteorological, crop, engineering and management data of the irrigation area and perform preprocessing; S2: Construct a coupled LSTM and improved iTransformer model to predict the future daily water supply and demand in the irrigation district; S3: Construct a two-level fractional programming model for the irrigation district based on the prediction results. The upper-level model is used to optimize the water allocation decision of the upper-level canal, and the lower-level model is used to optimize the water allocation decision of the lower-level canal. S4: In the bi-level fractional programming model, set water balance constraints, crop water requirement constraints, channel flow constraints, field water balance constraints, drainage constraints, and non-negativity constraints. S5: The bilevel fractional programming model is solved using NSGA-II and Dinkelbach algorithms to obtain the Pareto non-dominated solution set; S6: The TOPSIS method is used to optimize the Pareto non-dominated solution set and output the optimal or near-optimal water distribution scheme for the irrigation area.

2. The irrigation district water resource optimization scheduling method combining deep learning prediction and bi-level fractional programming as described in claim 1, characterized in that: The historical hydrological, meteorological, crop, engineering and management data mentioned in step S1 include one or more of the following: historical runoff, precipitation, reference evapotranspiration, average temperature, maximum temperature, minimum temperature, relative humidity, average wind speed, sunshine duration, air pressure, water level, channel design flow, channel length, channel water loss parameters, rice planting area, soil parameters and irrigation system parameters.

3. The irrigation district water resource optimization scheduling method combining deep learning prediction and bi-level fractional programming as described in claim 1, characterized in that: The LSTM and improved iTransformer coupled model mentioned in step S2 includes: LSTM is used to extract local dynamic features of water supply and demand time series. We utilize an improved iTransformer to extract global associations and long-term dependency features among multiple variables; The fully connected layer is used to output the predicted values ​​of water supply and demand for the next H time steps.

4. The irrigation district water resource optimization scheduling method combining deep learning prediction and bi-level fractional programming as described in claim 1, characterized in that: The upper-level model in step S3 aims to maximize the economic benefits of the system, while the lower-level model aims to maximize the irrigation water productivity.

5. The irrigation district water resource optimization scheduling method combining deep learning prediction and bi-level fractional programming as described in claim 1, characterized in that: The objective function of the bi-level fractional programming model described in step S3 is expressed in fractional form to represent the relationship between system benefits, water transfer losses and resource utilization efficiency.

6. The irrigation district water resource optimization scheduling method combining deep learning prediction and bi-level fractional programming as described in claim 1, characterized in that: The field water balance constraints mentioned in step S4 include the balance relationship between changes in irrigation volume, effective rainfall, evapotranspiration, deep infiltration, drainage volume, and field surface water storage.

7. The irrigation district water resource optimization scheduling method combining deep learning prediction and bi-level fractional programming as described in claim 1, characterized in that: The drainage constraint in step S4 is as follows: drainage is allowed when the field water depth is greater than the maximum water storage depth after rainfall during a certain period; and the drainage volume is zero when the field water depth is less than or equal to the maximum water storage depth after rainfall during that period.

8. The irrigation district water resource optimization scheduling method combining deep learning prediction and bi-level fractional programming as described in claim 1, characterized in that: The NSGA-II mentioned in step S5 is used for upper-level multi-objective optimization search, and the Dinkelbach algorithm is used to transform the fractional programming problem into a parameterized non-fractional problem and solve iteratively.

9. The irrigation district water resource optimization scheduling method combining deep learning prediction and bi-level fractional programming as described in claim 1, characterized in that: The TOPSIS method described in step S6 comprehensively evaluates candidate water distribution schemes based on one or more of the following indicators: system economic benefits, total water transfer loss, water productivity, and water distribution time.

Images