Water network water diversion layout optimization method based on deep learning

Abstract

The invention discloses a water network water diversion layout optimization method based on deep learning. The method comprises the following steps: collecting and preprocessing multi-source environmental data of saline-alkali soil in an arid region; executing space-time fusion and feature extraction, constructing a feature incidence matrix, and generating a standardized multi-dimensional feature data set; the neural network model is cycled by using a coupling physical mechanism to obtain a predicted spatial-temporal distribution result; constructing a multi-objective optimization model; carrying out iterative solution on the multi-objective optimization model by adopting an improved non-dominated sorting genetic algorithm to obtain a target engineering scale parameter scheme and a target space layout scheme; performing error calculation to obtain an error evaluation result; and an updated coupling physical mechanism recurrent neural network model is formed, and a dynamic closed-loop process of prediction, optimization and correction of the water network water diversion layout is realized.

Description

Technical Field

[0001] This invention relates to the field of water resource allocation and water conservancy project planning, and in particular to a method for optimizing the layout of water diversion and regulation networks based on deep learning. Background Technology

[0002] The management of saline-alkali land in arid regions is one of the core issues for water resource utilization and sustainable agricultural development in Northwest my country. Affected by factors such as arid climate, scarce rainfall, high evaporation, and high groundwater salinity, the surface water and groundwater cycles in arid regions are extremely fragile, with widespread phenomena such as soil salinization, abnormal groundwater level fluctuations, and ecosystem degradation. To improve the regional water resource allocation pattern, reduce soil salinity, increase vegetation cover, and enhance ecosystem stability, water diversion and regulation projects have become a key engineering measure. However, water diversion and regulation projects involve coupled decisions on multiple factors such as canal system scale, pump station capacity, and spatial layout, and must simultaneously consider various constraints such as soil salinity reduction targets, groundwater depth safety targets, and ecological risk control targets, making project layout decisions extremely difficult.

[0003] Currently, traditional water diversion and transfer project planning methods mainly rely on static water balance analysis, empirical engineering layout methods, or rule-based water transfer scheme construction. These methods have significant limitations: First, traditional methods do not adequately consider the dynamic feedback patterns between environmental factors and water transfer projects, making it difficult to reflect the continuous spatiotemporal changes in water and salt transport processes. Second, their ability to integrate multi-source environmental data is limited, failing to fully utilize multi-dimensional spatiotemporal information such as remote sensing data, meteorological data, and soil and groundwater monitoring data. Third, multi-objective, multi-constraint engineering layout decisions often rely on human experience or simplified assumptions, failing to consider the complex trade-offs between multiple objectives such as investment costs, salinity control effectiveness, and ecological security. Fourth, traditional models fail to achieve a dynamic closed loop of "engineering operation—environmental change—model correction," making it difficult for planning results to adapt to changes in environmental conditions during actual operation. Summary of the Invention

[0004] One objective of this invention is to propose a deep learning-based method for optimizing the layout of water diversion projects in a water network. This invention fully utilizes multi-source environmental data processing technology, a recurrent neural network model coupled with physical mechanisms, and a multi-objective optimization algorithm. By making high-precision predictions of the water-salt transport process in arid saline-alkali lands and combining the prediction results with an engineering layout optimization model, it achieves intelligent optimization design of the scale and spatial layout of water diversion projects in a water network.

[0005] A water network water diversion layout optimization method based on deep learning according to an embodiment of the present invention includes the following steps:

[0006] Collect and preprocess multi-source environmental data of saline-alkali land in arid areas;

[0007] Spatiotemporal fusion and feature extraction are performed on the preprocessed multi-source environmental data to construct a feature correlation matrix and generate a standardized multidimensional feature dataset;

[0008] The standardized multidimensional feature dataset is input into the coupled physical mechanism recurrent neural network model to obtain the predicted spatiotemporal distribution results;

[0009] A multi-objective optimization model is constructed based on the predicted spatiotemporal distribution results;

[0010] An improved non-dominated sorting genetic algorithm is used to iteratively solve the multi-objective optimization model to obtain the target engineering scale parameter scheme and the target spatial layout scheme;

[0011] The water diversion project is implemented and operated in accordance with the target project scale parameter scheme and target spatial layout scheme. Real-time monitoring data is collected, and the error between the monitoring data and the predicted spatiotemporal distribution results is calculated to obtain the error assessment results.

[0012] When the error evaluation result meets the preset triggering condition, the parameters of the coupled physical mechanism recurrent neural network model are updated to form the updated coupled physical mechanism recurrent neural network model.

[0013] Optionally, the multi-source environmental data includes remote sensing image data, meteorological sequence data, soil profile data, hydrogeological data, and topographic data, and the preprocessing includes denoising, normalization, and missing value imputation.

[0014] Optionally, the generation of the standardized multidimensional feature dataset specifically includes:

[0015] Spatial interpolation is performed on the preprocessed multi-source environmental data to unify the spatial resolution of various types of data in the preprocessed multi-source environmental data to a preset spatial grid scale, forming a spatially unified multi-source environmental dataset.

[0016] Temporal interpolation is performed on the spatially unified multi-source environmental dataset to unify the temporal resolution of various types of data in the preprocessed multi-source environmental data to a preset time step, forming a spatiotemporally unified multi-source environmental dataset. The corresponding environmental feature vector is then extracted. The extraction process uses each spatial grid scale and each time step as an index to stitch together remote sensing image feature data, meteorological time series feature data, soil profile feature data, hydrogeological feature data, and topographic feature data acquired under a unified spatial grid scale and a unified time step, according to a preset feature arrangement order, to construct a fixed-dimensional environmental feature vector.

[0017] The Pearson correlation coefficient matrix is ​​calculated based on environmental feature vectors. The calculation process involves calculating the deviation of the mean of two environmental feature vectors of different dimensions along the time series, accumulating the product of the deviations at each time step to obtain the accumulated deviation product, calculating the accumulated sum of the squares of the deviations of each environmental feature vector, and performing a square root operation on the accumulated sum of the squares of the deviations. The accumulated sum of the deviation products is then multiplied by the accumulated sum of the squares of the deviations after the square root operation and normalized to obtain the Pearson correlation coefficient matrix.

[0018] The mutual information matrix is ​​calculated based on environmental feature vectors. The calculation process is to construct a joint probability distribution and a corresponding marginal probability distribution based on the values ​​of two environmental feature vectors in the time series. For all possible combinations of values, the logarithm of the ratio of the product of the joint probability and the corresponding marginal probability is calculated, and the product of the logarithm and the corresponding joint probability is accumulated to obtain the mutual information value between the two environmental feature vectors, thus forming a mutual information matrix.

[0019] The Pearson correlation coefficient matrix and the mutual information matrix are weighted and fused according to their corresponding feature dimensions to obtain a feature association matrix. Based on the feature association matrix, the environmental feature vectors are filtered to remove environmental feature components with feature association degrees lower than a preset threshold, and retain environmental feature components with association degrees that meet the conditions, thus forming a standardized multidimensional feature dataset.

[0020] Optionally, obtaining the predicted spatiotemporal distribution results specifically includes:

[0021] The standardized multidimensional feature dataset is used to construct an input sequence according to a preset time step order, and the environmental feature vector corresponding to each time step of the input sequence is input into the coupled physical mechanism recurrent neural network model.

[0022] Within the coupled physical mechanism recurrent neural network model, a hidden state update operation is performed on the environmental feature vector at each time step. The hidden state update process is to initialize the pre-trained model parameters to obtain the initialized hidden state, and then linearly transform the environmental feature vector of the current time step and the hidden state of the previous time step with the corresponding weight matrix, add them together, and superimpose the weighted value of the physical mechanism constraint term to obtain the hidden state of the current time step.

[0023] The hidden states generated at each time step are input into the output layer of the coupled physical mechanism recurrent neural network model, and the predicted values ​​of water salt concentration and groundwater level at each time step are output.

[0024] The predicted water and salt concentration values ​​for all time steps are combined according to the spatial grid scale and time step order, and the predicted groundwater level values ​​for all time steps are combined according to the spatial grid scale and time step order to form the predicted spatiotemporal distribution results. The predicted spatiotemporal distribution results include the distribution results of the predicted groundwater level and soil salinity values ​​at the spatial grid scale and time step scale within a future preset time range.

[0025] Optionally, the construction of the multi-objective optimization model specifically includes:

[0026] Based on the predicted values ​​of water and salt concentration and groundwater level in the predicted spatiotemporal distribution results, target items for engineering investment cost, soil salinity compliance deviation, and ecological risk indicators are constructed. The target item for engineering investment cost is constructed based on the canal system design flow and canal system length. The target item for soil salinity compliance deviation is constructed based on the deviation between the predicted soil salinity value and the target soil salinity threshold. The target item for ecological risk indicators is constructed based on the relationship between the predicted groundwater level and ecological risk.

[0027] The system sets constraints on groundwater depth, salinity concentration at the drainage end, and vegetation restoration. The groundwater depth constraint means that the predicted groundwater level is within the preset groundwater depth threshold. The salinity concentration at the drainage end constraint means that the salinity concentration at the drainage end is less than the preset maximum allowable salinity concentration. The vegetation restoration constraint means that the vegetation coverage reaches the preset restoration ratio.

[0028] The objective items are composed of the project investment cost, the soil salinity deviation, and the ecological risk index. The constraint conditions are composed of the groundwater depth limit, the drainage end salinity concentration limit, and the vegetation restoration constraint. This forms a multi-objective optimization model. The project investment cost objective item is constructed using the canal design flow, canal length, and pump station capacity as decision variables. These decision variables serve as input parameters for the multi-objective optimization model and participate in the multi-objective optimization solution together with the predicted water and salt concentration and the predicted groundwater level.

[0029] Optionally, obtaining the target project scale parameter scheme and the target spatial layout scheme specifically includes:

[0030] An improved non-dominated sorting genetic algorithm is used to iteratively solve the multi-objective optimization model. Based on the expression of canal system scale, pump station capacity and spatial layout in the multi-objective optimization model, the design flow of each canal system, the length of each canal system, the pump station capacity of each pump station and the plane coordinates of the engineering layout nodes are combined to construct a decision variable vector.

[0031] The decision variable vector is used as the representation of an individual. An initial population is randomly generated according to a preset population size and set as the first generation population. The decision variable vector of each individual in the first generation population is substituted into the multi-objective optimization model.

[0032] Non-dominated sorting is performed on all individuals in the first generation population according to the non-dominated relationship between the target item sets. A non-dominated level is assigned to each individual, and the crowding distance of each individual is calculated based on the adjacent distance of the individual in the target space.

[0033] An improved tournament selection operator is introduced based on non-dominance level and crowding distance to select a set of parent individuals. Crossover and mutation operations are then performed on the set of parent individuals to generate a set of offspring individuals. The process of selecting the set of parent individuals is as follows: in each selection, candidate individuals are randomly selected from the current population. The non-dominance level and crowding distance of the candidate individuals are compared in turn. Individuals with lower non-dominance levels are selected as the winners. When the non-dominance levels of the candidate individuals are the same, individuals with larger crowding distances are selected as the winners. Each winner is added to the set of parent individuals in turn until the set of parent individuals reaches a preset number.

[0034] The first generation population and the set of offspring individuals are merged to form a temporary population. Non-dominated sorting and crowding distance calculation are performed on the temporary population, and a preset number of individuals are selected from it as the next generation population based on the non-dominated level and crowding distance.

[0035] When the number of generations of the next generation is less than the preset maximum number of generations, the next generation is used as the new current population and the target item calculation, non-dominated sorting, crowding distance calculation, parent individual selection, crossover operation, mutation operation and population update operation are performed in sequence. When the number of generations of the next generation reaches the preset maximum number of generations, the final population is obtained and the non-dominated solution set is extracted as the target engineering scale parameter scheme and the target spatial layout scheme.

[0036] Optionally, obtaining the error evaluation result specifically includes:

[0037] During the implementation and operation of the water diversion project in accordance with the target project scale parameter scheme and target spatial layout scheme, monitoring data is collected through a preset monitoring frequency. The monitoring data includes groundwater level monitoring data, soil electrical conductivity monitoring data and vegetation index monitoring data.

[0038] The monitoring data is organized according to the spatial grid scale and time step to obtain the monitoring vector;

[0039] The predicted values ​​corresponding to the position and time step of the monitoring vector are extracted from the predicted spatiotemporal distribution results to form a predicted vector with the same spatial position and time step as the monitoring vector.

[0040] Based on the comparison between the monitoring vector and the prediction vector one by one, the difference value corresponding to each monitoring location and time step is calculated, and the difference value is used as the corresponding error vector. The calculation process is to subtract the value of each monitoring indicator in the corresponding monitoring vector from the value of the prediction indicator in the same location and time step in the prediction vector at each monitoring location and time step to obtain the difference value of each indicator.

[0041] The error vectors of all monitoring locations and time steps are summarized according to the preset error measurement rules to form the error assessment results.

[0042] Optionally, the formation of the updated coupled physical mechanism recurrent neural network model specifically includes:

[0043] The error evaluation results are input into the coupled physical mechanism recurrent neural network model to generate error information for adjusting the model parameters. The generation process involves passing the error evaluation results back and forth along the network structure from the output layer to the front layer in the coupled physical mechanism recurrent neural network model, calculating the error contribution corresponding to the input weight matrix, hidden weight matrix and bias term in each layer, and summing up the error contributions of each layer to form error information.

[0044] By using error information, parameter update operations are performed on the input weight matrix, hidden weight matrix, and bias term in the coupled physical mechanism recurrent neural network model to obtain the updated input weight matrix, updated hidden weight matrix, and updated bias term;

[0045] Based on the updated input weight matrix, the updated hidden weight matrix, and the updated bias term, the physical constraint coefficients in the coupled physical mechanism recurrent neural network model are adjusted to form the updated physical constraint coefficients.

[0046] The updated input weight matrix, hidden weight matrix, bias term, and updated physical constraint coefficients are written into the coupled physical mechanism recurrent neural network model to form the updated coupled physical mechanism recurrent neural network model.

[0047] The beneficial effects of this invention are:

[0048] This invention achieves high-precision prediction of water and salt transport processes in arid saline-alkali lands and scientific optimization of water diversion and regulation project layouts by constructing a spatiotemporal fusion system of multi-source environmental data, a recurrent neural network model coupled with physical mechanisms, and a multi-objective optimization model. Compared to traditional layout methods that rely on empirical judgment, static water balance, or simplified mathematical models, this invention can utilize multi-source data such as remote sensing data, meteorological data, and soil and groundwater monitoring data to extract key features reflecting regional water and salt evolution patterns, significantly improving the comprehensiveness and accuracy of data processing. Furthermore, this invention introduces physical mechanism constraints of water and salt transport into the deep learning model, enabling the model to maintain physical consistency while retaining the strong representational capabilities of deep learning, thereby effectively improving the credibility, stability, and robustness of the prediction results.

[0049] In terms of engineering optimization, this invention deeply couples the predicted spatiotemporal distribution results with a multi-objective optimization model, enabling a dynamic balance between canal system scale, pump station capacity, and spatial layout among engineering investment costs, soil salinity control effectiveness, and ecological risk control. An improved non-dominated sorting genetic algorithm is used to solve the constructed multi-objective optimization model, obtaining a richer and more reasonable set of non-dominated solutions in a complex decision space with high dimensions and multiple constraints. This provides water diversion layout schemes with selectivity, diversity, and engineering feasibility. This optimization method overcomes the limitations of traditional methods in simultaneously considering multiple engineering objectives, effectively improving the scientific nature and decision-making quality of engineering layout design.

[0050] Furthermore, this invention constructs an error assessment and model parameter update mechanism based on real-time monitoring data, enabling the prediction model to continuously adapt and adjust according to actual monitoring data during project operation, forming a dynamic closed-loop system of "monitoring-prediction-optimization-update". This mechanism not only improves the model's adaptability in long-term operation but also allows the project layout scheme to be dynamically optimized according to changes in natural conditions, thereby further ensuring the long-term effectiveness and ecological safety of saline-alkali land management projects in arid areas. In summary, this invention significantly outperforms existing technologies in terms of data fusion quality, model prediction accuracy, optimization decision-making ability, and operational adaptability, and has significant engineering application value and promotional significance. Attached Figure Description

[0051] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0052] Figure 1 This is an overall flowchart of a deep learning-based water diversion and regulation layout optimization method proposed in this invention.

[0053] Figure 2This is a schematic diagram illustrating the construction of a coupled physical mechanism recurrent neural network model for a deep learning-based water diversion and regulation layout optimization method proposed in this invention.

[0054] Figure 3 This is a schematic diagram of the structure of the improved non-dominated sorting genetic algorithm after training, which is a deep learning-based water network diversion and regulation layout optimization method proposed in this invention. Detailed Implementation

[0055] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0056] refer to Figures 1-3 A deep learning-based method for optimizing the layout of water diversion and regulation networks includes the following steps:

[0057] Collect and preprocess multi-source environmental data of saline-alkali land in arid areas;

[0058] Spatiotemporal fusion and feature extraction are performed on the preprocessed multi-source environmental data to construct a feature correlation matrix and generate a standardized multidimensional feature dataset;

[0059] The standardized multidimensional feature dataset is input into the coupled physical mechanism recurrent neural network model to obtain the predicted spatiotemporal distribution results;

[0060] A multi-objective optimization model is constructed based on the predicted spatiotemporal distribution results;

[0061] An improved non-dominated sorting genetic algorithm is used to iteratively solve the multi-objective optimization model to obtain the target engineering scale parameter scheme and the target spatial layout scheme;

[0062] The water diversion project is implemented and operated in accordance with the target project scale parameter scheme and target spatial layout scheme. Real-time monitoring data is collected, and the error between the monitoring data and the predicted spatiotemporal distribution results is calculated to obtain the error assessment results.

[0063] When the error evaluation result meets the preset triggering condition, the parameters of the coupled physical mechanism recurrent neural network model are updated to form the updated coupled physical mechanism recurrent neural network model.

[0064] In this embodiment, the multi-source environmental data includes remote sensing image data, meteorological sequence data, soil profile data, hydrogeological data, and topographic data. The preprocessing includes denoising, normalization, and missing value imputation.

[0065] In this embodiment, the generation of the standardized multidimensional feature dataset specifically includes:

[0066] Spatial interpolation is performed on the preprocessed multi-source environmental data to unify the spatial resolution of various types of data in the preprocessed multi-source environmental data to a preset spatial grid scale, forming a spatially unified multi-source environmental dataset.

[0067] Temporal interpolation is performed on the spatially unified multi-source environmental dataset to unify the temporal resolution of various types of data in the preprocessed multi-source environmental data to a preset time step, forming a spatiotemporally unified multi-source environmental dataset. The corresponding environmental feature vector is then extracted. The extraction process uses each spatial grid scale and each time step as an index to stitch together remote sensing image feature data, meteorological time series feature data, soil profile feature data, hydrogeological feature data, and topographic feature data acquired under a unified spatial grid scale and a unified time step, according to a preset feature arrangement order, to construct a fixed-dimensional environmental feature vector.

[0068] The Pearson correlation coefficient matrix is ​​calculated based on environmental feature vectors. The calculation process involves calculating the deviation of the mean of two environmental feature vectors of different dimensions along the time series, accumulating the product of the deviations at each time step to obtain the accumulated deviation product, calculating the accumulated sum of the squares of the deviations of each environmental feature vector, and performing a square root operation on the accumulated sum of the squares of the deviations. The accumulated sum of the deviation products is then multiplied by the accumulated sum of the squares of the deviations after the square root operation and normalized to obtain the Pearson correlation coefficient matrix.

[0069] The mutual information matrix is ​​calculated based on environmental feature vectors. The calculation process is to construct a joint probability distribution and a corresponding marginal probability distribution based on the values ​​of two environmental feature vectors in the time series. For all possible combinations of values, the logarithm of the ratio of the product of the joint probability and the corresponding marginal probability is calculated, and the product of the logarithm and the corresponding joint probability is accumulated to obtain the mutual information value between the two environmental feature vectors, thus forming a mutual information matrix.

[0070] The Pearson correlation coefficient matrix and the mutual information matrix are weighted and fused according to their corresponding feature dimensions to obtain a feature association matrix. Based on the feature association matrix, the environmental feature vectors are filtered to remove environmental feature components with feature association degrees lower than a preset threshold, and retain environmental feature components with association degrees that meet the conditions, thus forming a standardized multidimensional feature dataset.

[0071] In this embodiment, obtaining the predicted spatiotemporal distribution result specifically includes:

[0072] The standardized multidimensional feature dataset is used to construct an input sequence according to a preset time step order, and the environmental feature vector corresponding to each time step of the input sequence is input into the coupled physical mechanism recurrent neural network model.

[0073] Within the coupled physical mechanism recurrent neural network model, a hidden state update operation is performed on the environmental feature vector at each time step. The hidden state update process is to initialize the pre-trained model parameters to obtain the initialized hidden state, and then linearly transform the environmental feature vector of the current time step and the hidden state of the previous time step with the corresponding weight matrix, add them together, and superimpose the weighted value of the physical mechanism constraint term to obtain the hidden state of the current time step.

[0074] The hidden states generated at each time step are input into the output layer of the coupled physical mechanism recurrent neural network model, and the predicted values ​​of water salt concentration and groundwater level at each time step are output.

[0075] The predicted water and salt concentration values ​​for all time steps are combined according to the spatial grid scale and time step order, and the predicted groundwater level values ​​for all time steps are combined according to the spatial grid scale and time step order to form the predicted spatiotemporal distribution results. The predicted spatiotemporal distribution results include the distribution results of the predicted groundwater level and soil salinity values ​​at the spatial grid scale and time step scale within a future preset time range.

[0076] In this embodiment, the construction of the multi-objective optimization model specifically includes:

[0077] Based on the predicted values ​​of water and salt concentration and groundwater level in the predicted spatiotemporal distribution results, target items for engineering investment cost, soil salinity compliance deviation, and ecological risk indicators are constructed. The target item for engineering investment cost is constructed based on the canal system design flow and canal system length. The target item for soil salinity compliance deviation is constructed based on the deviation between the predicted soil salinity value and the target soil salinity threshold. The target item for ecological risk indicators is constructed based on the relationship between the predicted groundwater level and ecological risk.

[0078] The system sets constraints on groundwater depth, salinity concentration at the drainage end, and vegetation restoration. The groundwater depth constraint means that the predicted groundwater level is within the preset groundwater depth threshold. The salinity concentration at the drainage end constraint means that the salinity concentration at the drainage end is less than the preset maximum allowable salinity concentration. The vegetation restoration constraint means that the vegetation coverage reaches the preset restoration ratio.

[0079] The objective items are composed of the project investment cost, the soil salinity deviation, and the ecological risk index. The constraint conditions are composed of the groundwater depth limit, the drainage end salinity concentration limit, and the vegetation restoration constraint. This forms a multi-objective optimization model. The project investment cost objective item is constructed using the canal design flow, canal length, and pump station capacity as decision variables. These decision variables serve as input parameters for the multi-objective optimization model and participate in the multi-objective optimization solution together with the predicted water and salt concentration and the predicted groundwater level.

[0080] In this embodiment, obtaining the target project scale parameter scheme and the target spatial layout scheme specifically includes:

[0081] An improved non-dominated sorting genetic algorithm is used to iteratively solve the multi-objective optimization model. Based on the expression of canal system scale, pump station capacity and spatial layout in the multi-objective optimization model, the design flow of each canal system, the length of each canal system, the pump station capacity of each pump station and the plane coordinates of the engineering layout nodes are combined to construct a decision variable vector.

[0082] The decision variable vector is used as the representation of an individual. An initial population is randomly generated according to a preset population size and set as the first generation population. The decision variable vector of each individual in the first generation population is substituted into the multi-objective optimization model.

[0083] Non-dominated sorting is performed on all individuals in the first generation population according to the non-dominated relationship between the target item sets. A non-dominated level is assigned to each individual, and the crowding distance of each individual is calculated based on the adjacent distance of the individual in the target space.

[0084] An improved tournament selection operator is introduced based on non-dominance level and crowding distance to select a set of parent individuals. Crossover and mutation operations are then performed on the set of parent individuals to generate a set of offspring individuals. The process of selecting the set of parent individuals is as follows: in each selection, candidate individuals are randomly selected from the current population. The non-dominance level and crowding distance of the candidate individuals are compared in turn. Individuals with lower non-dominance levels are selected as the winners. When the non-dominance levels of the candidate individuals are the same, individuals with larger crowding distances are selected as the winners. Each winner is added to the set of parent individuals in turn until the set of parent individuals reaches a preset number.

[0085] The first generation population and the set of offspring individuals are merged to form a temporary population. Non-dominated sorting and crowding distance calculation are performed on the temporary population, and a preset number of individuals are selected from it as the next generation population based on the non-dominated level and crowding distance.

[0086] When the number of generations of the next generation is less than the preset maximum number of generations, the next generation is used as the new current population and the target item calculation, non-dominated sorting, crowding distance calculation, parent individual selection, crossover operation, mutation operation and population update operation are performed in sequence. When the number of generations of the next generation reaches the preset maximum number of generations, the final population is obtained and the non-dominated solution set is extracted as the target engineering scale parameter scheme and the target spatial layout scheme.

[0087] In this embodiment, obtaining the error evaluation result specifically includes:

[0088] During the implementation and operation of the water diversion project in accordance with the target project scale parameter scheme and target spatial layout scheme, monitoring data is collected through a preset monitoring frequency. The monitoring data includes groundwater level monitoring data, soil electrical conductivity monitoring data and vegetation index monitoring data.

[0089] The monitoring data is organized according to the spatial grid scale and time step to obtain the monitoring vector;

[0090] The predicted values ​​corresponding to the position and time step of the monitoring vector are extracted from the predicted spatiotemporal distribution results to form a predicted vector with the same spatial position and time step as the monitoring vector.

[0091] Based on the comparison between the monitoring vector and the prediction vector one by one, the difference value corresponding to each monitoring location and time step is calculated, and the difference value is used as the corresponding error vector. The calculation process is to subtract the value of each monitoring indicator in the corresponding monitoring vector from the value of the prediction indicator in the same location and time step in the prediction vector at each monitoring location and time step to obtain the difference value of each indicator.

[0092] The error vectors of all monitoring locations and time steps are summarized according to the preset error measurement rules to form the error assessment results.

[0093] In this embodiment, the formation of the updated coupled physical mechanism recurrent neural network model specifically includes:

[0094] The error evaluation results are input into the coupled physical mechanism recurrent neural network model to generate error information for adjusting the model parameters. The generation process involves passing the error evaluation results back and forth along the network structure from the output layer to the front layer in the coupled physical mechanism recurrent neural network model, calculating the error contribution corresponding to the input weight matrix, hidden weight matrix and bias term in each layer, and summing up the error contributions of each layer to form error information.

[0095] By using error information, parameter update operations are performed on the input weight matrix, hidden weight matrix, and bias term in the coupled physical mechanism recurrent neural network model to obtain the updated input weight matrix, updated hidden weight matrix, and updated bias term;

[0096] Based on the updated input weight matrix, the updated hidden weight matrix, and the updated bias term, the physical constraint coefficients in the coupled physical mechanism recurrent neural network model are adjusted to form the updated physical constraint coefficients.

[0097] The updated input weight matrix, hidden weight matrix, bias term, and updated physical constraint coefficients are written into the coupled physical mechanism recurrent neural network model to form the updated coupled physical mechanism recurrent neural network model.

[0098] Example 1:

[0099] This embodiment focuses on a typical plain river network irrigation area, characterized by its flat terrain, dense canal network, and coupled water diversion and drainage projects. In recent years, due to unstable water inflow, changes in irrigation regimes, and groundwater extraction, some areas have experienced a coexistence of groundwater level rise and soil salinity accumulation. During dry years, increased water diversion and replenishment are often implemented to ensure agricultural irrigation, leading to poor drainage in some low-lying areas and persistently high groundwater levels during the growing season. Conversely, during wet years, short-term concentrated water replenishment can cause shallow groundwater to rise, inducing salinization. The core challenge for irrigation area management is determining the spatial layout of water diversion channels, the design flow rate of the canal system, and the scheduling capacity of key nodes within a given budget constraint for canal and pumping station renovation. This aims to improve soil salinity compliance rates and maintain groundwater levels within an ecologically safe range over a future period, while simultaneously achieving an overall optimal balance between project investment costs and ecological risks. Traditional approaches often rely on empirical zoning and static planning parameters, which usually cannot depict the dynamic relationship of "multi-source environmental driving - water and salt spatiotemporal response - engineering layout feedback". This can easily lead to situations where engineering investment is large but the achievement of standards is unstable, or short-term salinity reduction leads to increased ecological risks.

[0100] In this scenario, multi-source environmental data of the irrigation area were first collected and unified to the same spatiotemporal scale. The spatial scale used a 500m×500m grid, covering the irrigation area and major drainage control sections. The temporal scale used a monthly time step, forming a continuous 36-month historical sequence for training and prediction of the next 12 months. Remote sensing image data was used to extract spectral features related to vegetation indices and surface salinization. Meteorological sequence data included precipitation, evapotranspiration, and temperature. Soil profile data included the texture and initial electrical conductivity of the 0–20cm and 20–40cm soil layers. Hydrogeological data included aquifer parameters and recharge conditions. Topographic data included elevation, slope, and depression index. After denoising, normalizing, and imputing missing values ​​for each data type, the multi-source features were concatenated in a preset order at each grid and each monthly time step to obtain a fixed-dimensional environmental feature vector. To avoid high-dimensional feature redundancy leading to model training instability, and to highlight driving factors strongly correlated with water and salt responses, Pearson correlation coefficient and mutual information matrices were calculated based on the environmental feature vector. The Pearson correlation coefficient matrix is ​​used to characterize the linear correlation structure, while the mutual information matrix is ​​used to capture nonlinear dependencies. The two are then weighted and fused along their corresponding dimensions to construct a feature association matrix. In this embodiment, the weight coefficients are determined to be 0.55 and 0.45 through cross-validation to ensure a relatively balanced contribution between linear and nonlinear structures. Based on the feature association matrix, feature filtering is performed on the environmental feature components, eliminating features with low correlation and high redundancy. Finally, the environmental feature dimension is compressed from 48 dimensions to 26 dimensions, resulting in a standardized multidimensional feature dataset. This process ensures that the input features retain interpretable environmental meaning while improving the sensitivity of subsequent prediction models to key driving factors.

[0101] In the deep learning prediction stage, a standardized multidimensional feature dataset is input into a recurrent neural network model coupled with physical mechanisms to predict the spatiotemporal distribution of groundwater level and soil salinity predictions for the next 12 months at a grid scale. The recurrent neural network model initializes its hidden state with a zero vector at the initial time step. At each time step, the current environmental feature vector and the hidden state from the previous time step are linearly transformed and summed, while a weighted value of the physical mechanism constraint term is added to obtain the current hidden state. In this embodiment, the physical mechanism constraint term is manifested as a constraint on the smoothness of groundwater level changes and a constraint on the directional consistency of salinity changes with evapotranspiration and recharge. Its weight coefficients are tuned through a validation set and can be adaptively fine-tuned during online updates. Through this coupling method, the model not only fits historical observation patterns but also suppresses short-term prediction jumps caused by noise or extreme weather, making the prediction results more consistent with the continuity of hydrogeological processes. The model output forms the predicted spatiotemporal distribution results, including the predicted groundwater level and soil salinity values ​​for each grid for the next 12 months, providing a quantitative basis for subsequent optimization.

[0102] During the engineering layout optimization phase, a multi-objective optimization model was constructed based on the predicted spatiotemporal distribution results. The engineering investment cost objective was constructed using the canal design flow rate, canal length, and pump station capacity as decision variables, reflecting the costs of canal excavation and lining, pump station expansion, and maintenance. The soil salinity compliance deviation objective was constructed based on the deviation between the predicted soil salinity value and the target soil salinity threshold, with the threshold set at 2.0 dS / m according to the local crop salt tolerance standard. The ecological risk indicator objective was constructed based on the relationship between the predicted groundwater level and ecological risk. When the groundwater depth is less than 1.2 m, the risk of salt return and waterlogging increases significantly; therefore, the risk indicator was constructed based on "the extent and duration of exceeding the ecological safety depth." Constraints included groundwater depth constraints, drainage end-point salinity concentration constraints, and vegetation restoration constraints. Among these, the vegetation restoration constraint used the increase in the remotely sensed vegetation index during the key growing season as one of the constraints on the engineering effect. An improved non-dominated sorting genetic algorithm is used to solve the multi-objective model. The decision variables employ a hybrid encoding of "engineering scale parameters and spatial layout coordinates." Non-dominated sorting and crowding levels are used to maintain solution set diversity. An improved tournament strategy is used in the selection phase, prioritizing individuals that meet constraints over those that violate them, and prioritizing the reduction of groundwater ecological risks when constraints conflict. To ensure the engineering feasibility of the spatial layout, feasibility repairs are performed for situations where channel discontinuity and overlapping pump station service areas occur after crossover mutation, allowing the output scheme to be directly used in engineering design. After iterative convergence, a set of Pareto optimal solutions is obtained. In this embodiment, a compromise solution that balances investment and compliance is selected as the recommended scheme.

[0103] After the project is put into operation, groundwater level monitoring wells and soil conductivity monitoring points are set up in typical areas of the irrigation district. At the same time, vegetation indices are obtained using monthly remote sensing to form monitoring vectors. The system extracts the predicted values ​​of the grid positions and time steps corresponding to the monitoring vectors from the predicted spatiotemporal distribution results, and compares them with the measured values ​​to obtain error assessment results. In this embodiment, the triggering conditions are set as follows: the monthly average RMSE of the groundwater level prediction is greater than 0.20m for two consecutive months or the MAE of soil conductivity converted to salinity is greater than 0.25dS / m for two consecutive months. After triggering, the parameters of the recurrent neural network model are updated through backpropagation, and the weight coefficients of the physical mechanism constraint terms are adjusted by small steps. The updated model is used for rolling prediction and optimization in the next cycle, thus forming a closed-loop operation mechanism of "prediction-optimization-monitoring-update".

[0104] To verify the beneficial effects, this embodiment selected three comparative strategies for periodic evaluation: one is a traditional static planning scheme, which only sets the canal flow and pumping station capacity based on the multi-year average inflow and empirical zoning; another is a recurrent neural network prediction without introducing physical mechanism constraints, combined with the same multi-objective optimization; and the third is an offline prediction optimization that introduces physical mechanism constraints but does not perform online triggered updates. Evaluation indicators include groundwater level prediction accuracy, soil salinity prediction accuracy, the proportion of compliant area, the number of months exceeding ecological risk limits, engineering investment, and unit cost of compliance. The results show that the proposed scheme outperforms the others in terms of prediction stability, compliance effect, and ecological risk control, and can significantly increase the compliant area and reduce the risk of salinization at similar investment levels.

[0105] Table 1 Comparison of Prediction Accuracy and Treatment Effect of Optimization Scheme for Water Diversion Layout in Irrigation Districts

[0106]

[0107] As shown in Table 1, traditional static planning lacks characterization of multi-source environmental drivers and temporal changes. The predicted RMSE for groundwater level reaches 0.34m, and the predicted MAE for soil salinity reaches 0.46dS / m. These large errors result in significant fluctuations in subsequent engineering parameters despite substantial investment. Introducing a recurrent neural network significantly improves prediction accuracy. However, without physical mechanism constraints, the model is more sensitive to fluctuations in extreme months, with the groundwater level RMSE still at 0.23m. This makes it prone to short-term "false drops" or "false rises" in low-lying areas. Adding physical mechanism constraints further reduces the groundwater level RMSE to 0.18m, indicating that physical constraints effectively suppress unreasonable jumps and make the predictions more consistent with the continuity of groundwater processes.

[0108] From the perspective of governance effectiveness, the proportion of areas meeting salinity standards is a key indicator for measuring whether the problem has been truly solved. Traditional static planning only achieves 61.8%, with many areas still remaining in a state of excessive salinity for extended periods. Predictive optimization without physical constraints can improve this to 72.4%, but localized recurrence of salinization still exists. Incorporating physical constraints without online updates further improves the rate to 76.9%, indicating that the introduction of physical constraints helps stabilize the salinity reduction path. After adopting the closed-loop scheme of this invention, the proportion of areas meeting standards increased to 84.6%, and the average deviation from standards decreased to 0.15 dS / m, demonstrating that at the same level of investment, this invention can more effectively "deploy engineering measures to truly sensitive spatial locations and time windows," achieving higher governance benefits.

[0109] From an ecological security perspective, shallow groundwater levels significantly increase the risk of salinization and waterlogging. Traditional static planning resulted in an ecological risk exceeding limit of 3.8 months, while the closed-loop scheme of this invention reduces this to 1.4. Furthermore, the number of times salinity concentration exceeded limits at the drainage end decreased from 7 to 2 times. This demonstrates that salinity control was achieved without sacrificing ecological balance; rather, through comprehensive optimization of groundwater depth constraints and risk targets, the project layout and scheduling are more aligned with ecological security boundaries. The improvement in vegetation index from +0.03 to +0.08 also indirectly confirms the combined effect of habitat improvement and enhanced crop growth conditions.

[0110] From an input-output perspective, the engineering investment cost of the closed-loop scheme of this invention is on the same order of magnitude as that of "physical constraint RNN + optimization," but the cost per unit compliant area is reduced to 13,200 yuan / percentage point. This indicates that the beneficial effects of this invention do not come from "higher investment," but rather from the technical chain of "more accurate prediction, more reasonable optimization, and adaptive updates." The table shows that the number of times an update is triggered is 3 times per year, indicating that online updates do not frequently disturb the system, but only perform necessary corrections when the error continuously exceeds the limit. This ensures stability and can cope with distribution drift caused by changes in inflow conditions and management strategies.

Claims

1. A method for optimizing the layout of water diversion and regulation systems in a water network based on deep learning, characterized in that, Includes the following steps: Collect and preprocess multi-source environmental data of saline-alkali land in arid areas; Spatiotemporal fusion and feature extraction are performed on the preprocessed multi-source environmental data to construct a feature correlation matrix and generate a standardized multidimensional feature dataset; The standardized multidimensional feature dataset is input into the coupled physical mechanism recurrent neural network model to obtain the predicted spatiotemporal distribution results; A multi-objective optimization model is constructed based on the predicted spatiotemporal distribution results; An improved non-dominated sorting genetic algorithm is used to iteratively solve the multi-objective optimization model to obtain the target engineering scale parameter scheme and the target spatial layout scheme; The water diversion project is implemented and operated in accordance with the target project scale parameter scheme and target spatial layout scheme. Real-time monitoring data is collected, and the error between the monitoring data and the predicted spatiotemporal distribution results is calculated to obtain the error assessment results. When the error evaluation result meets the preset triggering condition, the parameters of the coupled physical mechanism recurrent neural network model are updated to form the updated coupled physical mechanism recurrent neural network model.

2. The method for optimizing the layout of a water network based on deep learning according to claim 1, characterized in that, The multi-source environmental data includes remote sensing image data, meteorological sequence data, soil profile data, hydrogeological data, and topographic data. The preprocessing includes denoising, normalization, and missing value imputation.

3. The method for optimizing the layout of a water network based on deep learning according to claim 1, characterized in that, The generation of the standardized multidimensional feature dataset specifically includes: Spatial interpolation is performed on the preprocessed multi-source environmental data to unify the spatial resolution of various types of data in the preprocessed multi-source environmental data to a preset spatial grid scale, forming a spatially unified multi-source environmental dataset. Temporal interpolation is performed on the spatially unified multi-source environmental dataset to unify the temporal resolution of various types of data in the preprocessed multi-source environmental data to a preset time step, forming a spatiotemporally unified multi-source environmental dataset. The corresponding environmental feature vector is then extracted. The extraction process uses each spatial grid scale and each time step as an index to stitch together remote sensing image feature data, meteorological time series feature data, soil profile feature data, hydrogeological feature data, and topographic feature data acquired under a unified spatial grid scale and a unified time step, according to a preset feature arrangement order, to construct a fixed-dimensional environmental feature vector. The Pearson correlation coefficient matrix is ​​calculated based on environmental feature vectors. The calculation process involves calculating the deviation of the mean of two environmental feature vectors of different dimensions along the time series, accumulating the product of the deviations at each time step to obtain the accumulated deviation product, calculating the accumulated sum of the squares of the deviations of each environmental feature vector, and performing a square root operation on the accumulated sum of the squares of the deviations. The accumulated sum of the deviation products is then multiplied by the accumulated sum of the squares of the deviations after the square root operation and normalized to obtain the Pearson correlation coefficient matrix. The mutual information matrix is ​​calculated based on environmental feature vectors. The calculation process is to construct a joint probability distribution and a corresponding marginal probability distribution based on the values ​​of two environmental feature vectors in the time series. For all possible combinations of values, the logarithm of the ratio of the product of the joint probability and the corresponding marginal probability is calculated, and the product of the logarithm and the corresponding joint probability is accumulated to obtain the mutual information value between the two environmental feature vectors, thus forming a mutual information matrix. The Pearson correlation coefficient matrix and the mutual information matrix are weighted and fused according to their corresponding feature dimensions to obtain a feature association matrix. Based on the feature association matrix, the environmental feature vectors are filtered to remove environmental feature components with feature association degrees lower than a preset threshold, and retain environmental feature components with association degrees that meet the conditions, thus forming a standardized multidimensional feature dataset.

4. The water network diversion and regulation layout optimization method based on deep learning according to claim 1, characterized in that, The specific steps to obtain the predicted spatiotemporal distribution results include: The standardized multidimensional feature dataset is used to construct an input sequence according to a preset time step order, and the environmental feature vector corresponding to each time step of the input sequence is input into the coupled physical mechanism recurrent neural network model. Within the coupled physical mechanism recurrent neural network model, a hidden state update operation is performed on the environmental feature vector at each time step. The hidden state update process is to initialize the pre-trained model parameters to obtain the initialized hidden state, and then linearly transform the environmental feature vector of the current time step and the hidden state of the previous time step with the corresponding weight matrix, add them together, and superimpose the weighted value of the physical mechanism constraint term to obtain the hidden state of the current time step. The hidden states generated at each time step are input into the output layer of the coupled physical mechanism recurrent neural network model, and the predicted values ​​of water salt concentration and groundwater level at each time step are output. The predicted water and salt concentration values ​​for all time steps are combined according to the spatial grid scale and time step order, and the predicted groundwater level values ​​for all time steps are combined according to the spatial grid scale and time step order to form the predicted spatiotemporal distribution results. The predicted spatiotemporal distribution results include the distribution results of the predicted groundwater level and soil salinity values ​​at the spatial grid scale and time step scale within a future preset time range.

5. The method for optimizing the layout of a water network based on deep learning according to claim 1, characterized in that, The construction of the multi-objective optimization model specifically includes: Based on the predicted values ​​of water and salt concentration and groundwater level in the predicted spatiotemporal distribution results, target items for engineering investment cost, soil salinity compliance deviation, and ecological risk indicators are constructed. The target item for engineering investment cost is constructed based on the canal system design flow and canal system length. The target item for soil salinity compliance deviation is constructed based on the deviation between the predicted soil salinity value and the target soil salinity threshold. The target item for ecological risk indicators is constructed based on the relationship between the predicted groundwater level and ecological risk. The system sets constraints on groundwater depth, salinity concentration at the drainage end, and vegetation restoration. The groundwater depth constraint means that the predicted groundwater level is within the preset groundwater depth threshold. The salinity concentration at the drainage end constraint means that the salinity concentration at the drainage end is less than the preset maximum allowable salinity concentration. The vegetation restoration constraint means that the vegetation coverage reaches the preset restoration ratio. The objective items are composed of the project investment cost, the soil salinity deviation, and the ecological risk index. The constraint conditions are composed of the groundwater depth limit, the drainage end salinity concentration limit, and the vegetation restoration constraint. This forms a multi-objective optimization model. The project investment cost objective item is constructed using the canal design flow, canal length, and pump station capacity as decision variables. These decision variables serve as input parameters for the multi-objective optimization model and participate in the multi-objective optimization solution together with the predicted water and salt concentration and the predicted groundwater level.

6. The method for optimizing the layout of a water network based on deep learning according to claim 1, characterized in that, The specific methods for obtaining the target project scale parameter scheme and the target spatial layout scheme include: An improved non-dominated sorting genetic algorithm is used to iteratively solve the multi-objective optimization model. Based on the expression of canal system scale, pump station capacity and spatial layout in the multi-objective optimization model, the design flow of each canal system, the length of each canal system, the pump station capacity of each pump station and the plane coordinates of the engineering layout nodes are combined to construct a decision variable vector. The decision variable vector is used as the representation of an individual. An initial population is randomly generated according to a preset population size and set as the first generation population. The decision variable vector of each individual in the first generation population is substituted into the multi-objective optimization model. Non-dominated sorting is performed on all individuals in the first generation population according to the non-dominated relationship between the target item sets. A non-dominated level is assigned to each individual, and the crowding distance of each individual is calculated based on the adjacent distance of the individual in the target space. An improved tournament selection operator is introduced based on non-dominance level and crowding distance to select a set of parent individuals. Crossover and mutation operations are then performed on the set of parent individuals to generate a set of offspring individuals. The process of selecting the set of parent individuals is as follows: in each selection, candidate individuals are randomly selected from the current population. The non-dominance level and crowding distance of the candidate individuals are compared in turn. Individuals with lower non-dominance levels are selected as the winners. When the non-dominance levels of the candidate individuals are the same, individuals with larger crowding distances are selected as the winners. Each winner is added to the set of parent individuals in turn until the set of parent individuals reaches a preset number. The first generation population and the set of offspring individuals are merged to form a temporary population. Non-dominated sorting and crowding distance calculation are performed on the temporary population, and a preset number of individuals are selected from it as the next generation population based on the non-dominated level and crowding distance. When the number of generations of the next generation is less than the preset maximum number of generations, the next generation is used as the new current population and the target item calculation, non-dominated sorting, crowding distance calculation, parent individual selection, crossover operation, mutation operation and population update operation are performed in sequence. When the number of generations of the next generation reaches the preset maximum number of generations, the final population is obtained and the non-dominated solution set is extracted as the target engineering scale parameter scheme and the target spatial layout scheme.

7. The method for optimizing the layout of a water network based on deep learning according to claim 1, characterized in that, The error evaluation results are obtained specifically in the following ways: During the implementation and operation of the water diversion project in accordance with the target project scale parameter scheme and target spatial layout scheme, monitoring data is collected through a preset monitoring frequency. The monitoring data includes groundwater level monitoring data, soil electrical conductivity monitoring data and vegetation index monitoring data. The monitoring data is organized according to the spatial grid scale and time step to obtain the monitoring vector; The predicted values ​​corresponding to the position and time step of the monitoring vector are extracted from the predicted spatiotemporal distribution results to form a predicted vector with the same spatial position and time step as the monitoring vector. Based on the comparison between the monitoring vector and the prediction vector one by one, the difference value corresponding to each monitoring location and time step is calculated, and the difference value is used as the corresponding error vector. The calculation process is to subtract the value of each monitoring indicator in the corresponding monitoring vector from the value of the prediction indicator in the same location and time step in the prediction vector at each monitoring location and time step to obtain the difference value of each indicator. The error vectors of all monitoring locations and time steps are summarized according to the preset error measurement rules to form the error assessment results.

8. The method for optimizing the layout of a water network based on deep learning according to claim 1, characterized in that, The formation of the updated coupled physical mechanism recurrent neural network model specifically includes: The error evaluation results are input into the coupled physical mechanism recurrent neural network model to generate error information for adjusting the model parameters. The generation process involves passing the error evaluation results back and forth along the network structure from the output layer to the front layer in the coupled physical mechanism recurrent neural network model, calculating the error contribution corresponding to the input weight matrix, hidden weight matrix and bias term in each layer, and summing up the error contributions of each layer to form error information. By using error information, parameter update operations are performed on the input weight matrix, hidden weight matrix, and bias term in the coupled physical mechanism recurrent neural network model to obtain the updated input weight matrix, updated hidden weight matrix, and updated bias term. Based on the updated input weight matrix, the updated hidden weight matrix, and the updated bias term, the physical constraint coefficients in the coupled physical mechanism recurrent neural network model are adjusted to form the updated physical constraint coefficients. The updated input weight matrix, hidden weight matrix, bias term, and updated physical constraint coefficients are written into the coupled physical mechanism recurrent neural network model to form the updated coupled physical mechanism recurrent neural network model.

Images