Green infrastructure multi-objective co-optimization system

Abstract

The embodiment of the application provides a kind of green infrastructure multi-objective collaborative optimization system, including: graph neural network space topology construction module is configured to establish space adjacency graph based on input geographic space layer, and node feature is encoded and aggregated, obtain multiple space embedding feature vectors.Hydrology simulation and feedback evaluation module is configured to generate multidimensional simulation simulation dataset based on physical simulator.Distillation agent modeling module is configured to train GNN model based on multidimensional simulation simulation dataset and multiple space embedding feature vectors;And light weight GNN agent model is obtained.Lightweight. Multi-agent reinforcement learning optimization decision module is configured to calculate the optimal green infrastructure layout combination of node based on multiple space embedding feature vectors under the restriction of multi-objective constraint condition and facility layout upper limit.Hydrology simulation and feedback evaluation module is further configured to verify the optimal green infrastructure layout combination in physical level.

Description

Technical Field

[0001] This application relates to the field of smart water resource management and ecological restoration technology, and in particular to a multi-objective collaborative optimization system for green infrastructure. Background Technology

[0002] Riverbank ecological restoration is an important research direction in the field of environmental science and engineering. It aims to address water pollution, flooding and ecological degradation caused by urbanization by synergistically optimizing water purification, peak runoff reduction and biodiversity enhancement through green infrastructure (such as wetlands, vegetation buffer zones and rain gardens).

[0003] The coordinated optimization of green infrastructure layout is a multi-objective, multi-scale, and dynamically evolving system decision-making problem. Summary of the Invention

[0004] In view of this, the purpose of this application is to propose a multi-objective collaborative optimization system for green infrastructure.

[0005] Based on the above objectives, this application provides a multi-objective collaborative optimization system for green infrastructure, including: a graph neural network spatial topology construction module, a distillation agent modeling module, a hydrological simulation and feedback evaluation module, and a multi-agent reinforcement learning optimization decision-making module; The graph neural network spatial topology construction module is configured to build a spatial adjacency graph based on the input geospatial layer, and encode and aggregate the node features in the spatial adjacency graph to obtain multiple spatial embedding feature vectors; the multiple spatial embedding feature vectors correspond one-to-one with multiple nodes in the spatial adjacency graph; wherein, the nodes in the spatial adjacency graph correspond to sub-catchments, and the edges in the spatial adjacency graph represent the hydraulic connection relationship from the upstream sub-catchment to the downstream sub-catchment; The hydrological simulation and feedback assessment module is configured to generate multidimensional simulation datasets in batches based on a physical simulator, taking into account the upper limits of the deployment of various green infrastructure families and the mapping of land use change. The multidimensional simulation datasets are used for training the GNN model. Each set of data in the multidimensional simulation datasets includes a combination of green infrastructure layouts, as well as the corresponding peak runoff, total runoff, pollutant load and landscape diversity index. The distillation proxy modeling module is configured to train a GNN model based on the multidimensional simulation dataset and the multiple spatial embedding feature vectors; then, using a model distillation method, the GNN proxy model is lightweighted to obtain a lightweight GNN proxy model; wherein, the parameter size of the lightweight GNN proxy model is smaller than that of the GNN model; the GNN model is used to fit the mapping relationship between the layout of green infrastructure families and hydrological performance; the lightweight GNN proxy model can replace the physical simulator to provide performance or reward estimation during reinforcement learning; The multi-agent reinforcement learning optimization decision-making module is configured to, based on the multiple spatially embedded feature vectors, calculate the green infrastructure layout combination of the node through a reinforcement learning algorithm under multi-objective constraints and facility deployment limits, in order to determine the optimal green infrastructure layout combination of the node; wherein, a sub-catchment corresponds to a single agent; the reward estimation is determined based on a weighted combination of peak runoff reduction rate, pollutant removal rate, and landscape diversity enhancement predicted by the lightweight GNN surrogate model; The hydrological simulation and feedback evaluation module is also configured to perform physical-level verification of the optimal green infrastructure layout combination, so that the optimal green infrastructure layout combination has feasibility and hydrological effectiveness under actual rainfall conditions.

[0006] In some embodiments, training the GNN model based on the multidimensional simulation dataset and the plurality of spatial embedding feature vectors includes: The node features of the node's direct upstream nodes are aggregated using the first layer propagation in the GCN model to simulate the impact of the upstream nodes' LID facilities on the node. The hidden states of the direct upstream nodes of the node are aggregated using the second layer propagation in the GCN model to obtain the final hidden state, in order to capture the influence of indirect hydraulic paths on the node. Based on the final hidden state, the target hydrological and ecological indicators of the node under a given combination of green infrastructure deployment are predicted, and the reward value for reinforcement learning is calculated based on the target hydrological and ecological indicators, thus obtaining the GNN model.

[0007] In some embodiments, the influence of the upstream node's LID facility on the node is determined based on a first weight matrix, the hydraulic connection relationship with adjacent nodes, the hydrological feature vector of the direct upstream node, and a first bias vector.

[0008] In some embodiments, the influence of the indirect hydraulic path on the node is determined based on a second weight matrix, the hydraulic connection relationship with adjacent nodes, the hidden state of the direct upstream node, and a second bias vector.

[0009] In some embodiments, the reward value is determined based on the final hidden state, the third weight matrix, and the third bias vector; wherein the dimensions of the first weight matrix / second weight matrix / third weight matrix are set according to the node feature dimension and the hidden layer size to achieve a non-linear mapping from node features to hidden state and output metric.

[0010] In some embodiments, the green infrastructure includes multiple facility families; the multiple facility families have different ecological service functions; and the upper limit area of ​​the deployment of the multiple facility families is based on different deployment conditions.

[0011] In some embodiments, the maximum deployment area of ​​the facility family is determined based on at least one of the following: the area of ​​the sub-catchment, the distance from the river channel, the slope, and the road type.

[0012] In some embodiments, the multiple facility families include a first facility family, a second facility family, and a third facility family; The maximum deployment area of ​​the first facility family is determined based on the area of ​​the sub-catchment area and the distance correction factor corresponding to the distance from the river channel; The maximum deployment area of ​​the second facility family is determined based on the area of ​​the sub-catchment area, the correction parameter corresponding to the distance from the river channel, and the first slope correction factor; The maximum deployment area of ​​the third facility family is determined based on the road type correction factor and the second slope correction factor.

[0013] In some embodiments, the distance correction factor includes multiple distance correction factors; the values ​​of the multiple distance correction factors are different, and the multiple distance correction factors correspond one-to-one with multiple distances from the river channel; The first slope correction factor includes multiple first slope correction factors; the values ​​of the multiple first slope correction factors are different, and the multiple first slope correction factors correspond one-to-one with multiple first slopes; The road type correction factor includes multiple road type correction factors; the values ​​of the multiple road type correction factors are different, and the multiple road type correction factors correspond one-to-one with multiple road types; The second slope correction factor includes multiple second slope correction factors; the values ​​of the multiple second slope correction factors are different, and the multiple second slope correction factors correspond one-to-one with multiple second slopes.

[0014] In some embodiments, the average effective water depth of the first facility family is determined based on the effective reservoir capacity and the deployable area of ​​the first facility family; The second facility family has multiple outflow configuration types; The various facility families each have a corresponding mapping matrix; the mapping matrix is ​​used to characterize the land use category and changes in key land parameters within the facility family; the key land parameters include surface layer parameters, soil layer parameters, pollutant scour parameters, and other parameters. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in this application or related technologies, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart for multi-objective optimization coupled with SWMM for NSGA-II.

[0017] Figure 2 This is a system overall data flow diagram of an embodiment of this application.

[0018] Figure 3 This is a diagram of the GNN-based spatial topology and lightweight agent modeling framework in an embodiment of this application. Figure 4 This is a structural diagram of the MASAC module training process in an embodiment of this application.

[0019] Figure 5 This is a comparison diagram of the SWMM direct call and lightweight proxy process in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with specific embodiments and the accompanying drawings.

[0021] It should be noted that, unless otherwise defined, the technical or scientific terms used in the embodiments of this application should have the ordinary meaning understood by one of ordinary skill in the art to which this application pertains. The terms "first," "second," and similar terms used in the embodiments of this application do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Terms such as "comprising" or "including" mean that the element or object preceding the word encompasses the elements or objects listed after the word and their equivalents, without excluding other elements or objects. Terms such as "connected" or "linked" are not limited to physical or mechanical connections, but can include electrical connections, whether direct or indirect. Terms such as "upper," "lower," "left," and "right" are only used to indicate relative positional relationships; when the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0022] In recent years, with the acceleration of urbanization, the intensification of climate change, and the frequent occurrence of extreme rainfall events, the planning of green infrastructure in riparian zones has shifted from traditional grey infrastructure (such as dams and drainage networks) to nature-based solutions (NBS). NBS emphasizes managing water resources using natural processes, enhancing ecosystem resilience, and taking into account environmental, economic, and social benefits. As a core strategy for sustainable urban development, the layout optimization technology of green infrastructure (GI) has evolved from a single-function orientation to a multi-objective synergy.

[0023] Early GI (Geographic Infrastructure) layout optimization primarily focused on single functions, such as stormwater runoff control or water purification, emphasizing the supplementation of grey infrastructure. GI layout optimization methods relied on scenario analysis and manual enumeration, using hydrological models such as Storm Water Management Models (SWMMs) to simulate rainfall runoff and determine the scale and location of GIs. However, these methods had a single objective function, high computational cost, small search space, and lacked multi-dimensional evaluation capabilities.

[0024] With increasing urbanization and growing demand for ecological restoration, GI optimization has incorporated "economic efficiency" indicators, forming a dual-objective collaborative optimization model of "peak shaving and cost reduction." Multiple-Criteria Decision Analysis (MCDA) and Geographic Information System (GIS) are employed in GI layout optimization. The Non-Dominated Sorting Genetic Algorithm (NSGA-II) has begun to be applied to multi-objective optimization, generating Pareto fronts to balance cost and environmental benefits. However, these methods are primarily static optimizations, making them ill-suited to dynamic hydrological conditions, and their insufficient quantification of spatial synergy limits the overall effectiveness of GI.

[0025] In recent years, artificial intelligence (AI) technologies (such as machine learning and reinforcement learning) and advanced distributed hydrological modeling have driven the rapid development of GI optimization techniques. Some studies have attempted to use algorithms such as Deep Q-Learning (DQN) and Proximal Policy Optimization (PPO) to handle dynamic climate uncertainties, and introduced random forests and neural networks to predict water quality and runoff responses, thereby improving the accuracy of layout schemes. Numerous studies have employed non-dominated sorting genetic algorithms for multi-objective optimization, and have implemented targeted optimization and dynamic adjustment of objective weights through partitioning. These algorithms can improve the systematicity and adaptability of GI layout to some extent. However, the following key problems remain: the NSGA-II genetic algorithm is prone to getting trapped in local optima, and its computation time is excessively long (>30 hours) when optimizing tens of thousands of variables, making it difficult to support large-scale urban networks. Optimizing the layout of green infrastructure is a complex systems engineering project in urban planning and ecological construction. Its objectives typically encompass multiple dimensions, including ecological protection, flood control, climate regulation, landscape beautification, social equity, and economic feasibility. Significant conflicts exist among these dimensions; for example, ecological space protection often limits land development intensity. Therefore, multi-objective optimization algorithms (such as NSGA-II and MOEA / D (Multiobjective Evolutionary Algorithm Based On Decomposition)) are needed to find Pareto optimal solutions among different objectives. Furthermore, the layout of green infrastructure exhibits significant spatial dependence, nonlinear feedback characteristics related to ecological network connectivity and surface process interactions, requiring GI optimization to comprehensively consider the coupling of ecological processes. In addition, the operating environment of green infrastructure continuously changes with climate change, urban expansion, and socio-economic dynamics; therefore, GI layout needs to possess adaptability and robustness to maintain stable performance under uncertain environments. Thus, optimizing the layout of green infrastructure is a multi-objective, multi-scale, and dynamically evolving systems decision-making problem. To address challenges such as multi-objective conflicts, spatial coupling, and non-stationary environments, it is necessary to combine multi-objective evolutionary algorithms, spatial simulation models, and uncertainty optimization methods to achieve synergistic improvement in ecological, social, and economic benefits.

[0026] Some studies employ automated modeling and optimization techniques, particularly the mainstream Non-Dominated Sort Genetic Algorithm (NSGA-II). Its core idea is to simulate natural selection and genetic mechanisms to continuously approach the Pareto optimal front during the iterative process. Based on fast non-dominated sorting and crowding distance mechanisms, it can maintain the diversity and uniformity of solutions under multiple objectives. Although the coupling technique of NSGA-II and SWMM has achieved significant results in the optimization of green infrastructure layout, in urban-scale applications with a large number of catchment units, a complete simulation of one round of NSGA-II can take more than 24 hours, resulting in a severe computational bottleneck.

[0027] Some studies focus on phased optimization, including the NSGA-II + SWMM multi-objective optimization framework. This approach generates Pareto solutions using NSGA-II, simulates hydrological responses using SWMM, and selects the optimal solution using the Technique for Order Preference by Similarity to Ideal Solution (TOPSIS). Figure 1 As shown, this framework still suffers from several technical bottlenecks. These are primarily manifested in its static, single-cycle optimization strategy, which lacks responsiveness to dynamic rainfall patterns and spatial interactions. Furthermore, while the TOPSIS decision-making method can sort solution sets, it cannot reflect the spatial synergy or long-term evolution between facility layouts. In addition, frequent SWMM calls lead to low computational efficiency, making it difficult to adapt to complex decision-making needs at the regional scale or in multiple scenarios. This scheme still falls short in terms of system intelligence, timeliness, and ecological-spatial collaborative processing, urgently requiring the introduction of intelligent optimization mechanisms with greater generalization and structural awareness capabilities.

[0028] Therefore, current optimization of green infrastructure layout has significant shortcomings in terms of spatial synergy, dynamic adaptability, ecological integration, computational efficiency, and climate resilience. The SWMM model itself lacks the ability to express spatial dependencies; facility placement is only used as a static decision variable input, lacking the identification and response to upstream-downstream connections and spatial fragmentation issues. Furthermore, facility placement often exhibits a "downstream oversaturation, upstream blank" phenomenon, causing peak-shaving spatial mismatch and affecting overall system performance. Additionally, NSGA-II optimization is a single-round batch processing flow, lacking strategy evolution and environmental adaptability, generating only a solution set or a single solution, unable to provide a fast solution in the form of a strategy. TOPSIS only provides static solution set sorting, unable to handle multi-period inputs, complex strategy spaces, or phased updates. Moreover, traditional objective functions only consider peak runoff control and costs, lacking expression of landscape harmony and bioecological richness, and cannot support ecological planning objectives.

[0029] Based on this, such as Figure 2As shown, this application provides a collaborative optimization decision-making system for urban green infrastructure (GI) based on Graph Convolutional Network (GNN) and Multi-Agent Soft Actor-Critic (MASAC) algorithm to improve the intelligence level of current urban stormwater management and ecological landscape collaborative design. The system of this application, based on the deep coupling of Graph Convolutional Network (GNN), Multi-Agent Soft Actor-Critic (MASAC), and Solid State Simulation Model (SWMM), can provide collaborative optimization configuration of GI facility types, locations, and control parameters in complex urban water catchment spaces, ultimately generating an executable solution that can be embedded in urban planning platforms (such as ArcGIS). It can, to a certain extent, solve technical problems in current urban stormwater management and ecological resilience design, such as lack of overall layout coordination, poor adaptability to dynamic changes, narrow objectives, computational inefficiency, and weak strategy transferability. The system of this application can realize the scientific layout and dynamic collaborative control of green facilities, and has good adaptability and generalization ability. It can provide integrated intelligent decision support for facility regulation in complex urban hydrological scenarios. It also provides deployable smart tools for fields such as green stormwater management, ecological landscape design, and resilient urban planning.

[0030] The green infrastructure multi-objective collaborative optimization system provided in this application embodiment may include: a graph neural network spatial topology construction module, a GNN distillation agent modeling module, a hydrological simulation and feedback evaluation module, and a MASAC multi-agent reinforcement learning optimization decision module. The modules of the green infrastructure multi-objective collaborative optimization system in this application embodiment form a closed loop through data flow and feedback mechanisms, realizing a closed-loop optimization process from spatial understanding and policy generation to hydrological verification. These modules collaborate and divide labor within the closed loop: the hydrological simulation and feedback evaluation module, as the physical layer, provides physical constraints, basic hydrological structure, and dynamic verification functions, serving as the underlying engine of the entire system. The graph neural network spatial topology construction module, as the spatial cognition layer, reads the SWMM benchmark model file (including nodes, pipes, and catchment units) and establishes the sub-catchment topology map and adjacency matrix; the distillation agent modeling module, as the knowledge abstraction layer, uses SWMM simulation results (i.e., teacher signals) to train student models, forming a GNN agent network that can quickly predict hydrological performance; the multi-agent reinforcement learning optimization decision module, as the policy decision layer, calls a lightweight GNN agent model for rapid policy exploration and periodically calls SWMM for verification and relearning.

[0031] The graph neural network spatial topology construction module is configured to build a spatial adjacency graph based on the input geospatial layer, and encode and aggregate the node features in the spatial adjacency graph to obtain the spatial embedding features of each sub-catchment, which are used for subsequent agent modeling and reinforcement learning optimization. That is, multiple spatial embedding feature vectors are obtained; each of these multiple spatial embedding feature vectors corresponds one-to-one with multiple nodes in the spatial adjacency graph. The nodes in the spatial adjacency graph correspond to sub-catchments, and the edges in the spatial adjacency graph represent the hydraulic connections from upstream to downstream sub-catchments. This improves the global collaborative efficiency of local facility configuration and avoids the overall performance degradation caused by optimizing only a single sub-catchment independently.

[0032] Among them, such as Figure 5 As shown, the hydrological simulation and feedback assessment module is configured to generate a multidimensional simulation dataset based on a physical simulator, considering the upper limits of various green infrastructure families and land use change mapping; the multidimensional simulation dataset is used for training the GNN surrogate model. Each set of data in the multidimensional simulation dataset includes a combination of green infrastructure layouts, as well as the corresponding peak runoff, total runoff, pollutant load, and landscape diversity index.

[0033] Among them, such as Figure 3 As shown, the distillation proxy modeling module is configured to train a GNN model based on the multidimensional simulation dataset. The GNN model is then lightweighted using a model distillation method to obtain a lightweight GNN proxy model. The original GNN proxy model is a complete GNN proxy model used to fit the mapping relationship between facility layout and hydrological performance. The lightweight GNN proxy model has a smaller parameter scale than the original GNN proxy model. In reinforcement learning, the lightweight GNN proxy model can replace the physical simulator to provide performance or reward estimation. This significantly reduces the number of physical model calls, improves convergence speed, and reduces computational resource requirements.

[0034] Among them, such as Figure 4 As shown, the multi-agent reinforcement learning optimization decision-making module is configured to, based on the multiple spatially embedded feature vectors, calculate the optimal green infrastructure layout combination for each node using a reinforcement learning algorithm under multi-objective constraints. Each sub-catchment corresponds to a single agent. The reward estimate is determined by a weighted combination of peak runoff reduction rate, pollutant removal rate, and landscape diversity enhancement predicted by a lightweight GNN surrogate model. This enables the generation of intelligent configuration strategies that balance multiple objectives (runoff reduction, water purification, landscape diversity, etc.).

[0035] Among them, such as Figure 5As shown, the hydrological simulation and feedback evaluation module is also configured to perform physical-level verification of the optimal green infrastructure layout combination, so as to ensure the feasibility and hydrological effectiveness of the optimal green infrastructure layout combination under actual rainfall conditions. This achieves a balance between physical interpretability and computational efficiency, combining engineering reliability with cutting-edge technology.

[0036] refer to Figure 2 As shown, the green infrastructure multi-objective collaborative optimization system of this application uses SWMM as the underlying physical engine and combines data-driven and intelligent learning to achieve collaborative optimization of the layout of urban green infrastructure. The system first imports multi-source spatial and hydrological data, including digital elevation model, land use, catchment areas, pipeline nodes, and rainfall sequences, to construct an SWMM baseline model (.inp). The SWMM baseline model defines sub-catchment areas, pipelines, and boundary conditions, and provides a consistent hydrological structural framework for subsequent modules. Based on this, the system extracts pipeline node and pipeline information from the SWMM, establishes a sub-catchment topology map and adjacency matrix, and generates a high-dimensional spatial feature representation through the GNN spatial topology module. Subsequently, the SWMM module performs batch simulations of different facility and rainfall scenario configurations, outputting indicators such as peak flow, total runoff, and storage capacity to form a complete sample. The distillation proxy modeling module uses this to train a lightweight network, learning the mapping between facility layout and hydrological performance to achieve high-precision, low-cost performance prediction. During the optimization phase, the MASAC multi-agent reinforcement learning optimization decision module uses each sub-catchment as an agent, iterates policies based on the prediction results of the GNN surrogate model, and generates the optimal GI type and scale configuration. The green infrastructure multi-objective collaborative optimization system periodically calls SWMM for sampling verification during the iteration process and feeds the simulation results back to the surrogate model for re-distillation correction. Finally, the system outputs the policy under the optimal configuration and converts it into a software script that can be embedded in GIS, realizing spatial visualization and secondary simulation verification of the optimization results. The entire data flow of the green infrastructure multi-objective collaborative optimization system forms a closed loop of "physical simulation—spatial modeling—intelligent optimization—physical feedback," enabling efficient, interpretable, and transferable optimization of green infrastructure layout.

[0037] In some embodiments, the hydrological simulation and feedback evaluation module is configured to establish a complete rainfall-runoff simulation scenario based on the SWMM (StormWater Management Model) physical simulator, converting the optimization results into SWMM input files, including sub-catchment parameters, LIDLow-Impact Development, facility configuration, and drainage network connection structure. By inputting different design storm intensities or durations, the module simulates the storage and peak shaving processes of green infrastructure during rainfall events, calculating indicators such as peak flow, total runoff, peak occurrence time, and storage capacity to obtain a multidimensional simulation dataset. Typically, the simulation results in the multidimensional simulation dataset can be compared with the predictions of the GNN surrogate model to evaluate the accuracy of the surrogate model or trigger a redistillation process.

[0038] In some of these embodiments, see Figure 5 The hydrological simulation and feedback assessment module integrates land use layers into the traditional SWMM simulation framework and automatically calculates the ecological Shannon index to avoid the reliance on expert experience in landscape optimization inherent in traditional SWMM simulation frameworks. This module is also configured to automatically adapt traditional LID measures to various facility family structures with more clearly defined ecological significance (e.g., facility family G1-Facility family G3). Furthermore, the module maintains the traditional SWMM simulation framework's ability to simulate peak runoff and pollution loads and provides an open interface for interaction with the GNN-MASAC system.

[0039] In some embodiments, during the hydraulic runoff generation and collection calculation stage, the hydrological simulation and feedback assessment module divides the catchment area into a permeable zone A2, an impermeable sub-catchment area A3 without depression storage, and an impermeable sub-catchment area A1 containing depression storage. When the precipitation is greater than the depression storage capacity dp or greater than both the depression storage capacity and field capacity, runoff generation is calculated for each underlying surface type. Then, based on the pipe network and hydrodynamic equations, the runoff is collected at the outlet section and discharged into the urban river network and stormwater pipe network. When reflecting the surface runoff hydraulic process, the above-described catchment area often uses sub-catchments as the research object, analyzing regional changes based on the runoff evolution of each sub-catchment. The infiltration model used in the runoff generation calculation employs a modified Green-Ampt model with clear physical meaning, including soil saturated hydraulic conductivity. Ks The parameters include three aspects: (mm / hr), soil capillary head (mm), and initial soil moisture deficit (fraction). The soil saturated hydraulic conductivity will be calculated based on land use, soil type, and aggregate structure. Under stable rainfall conditions, when the cumulative infiltration F is less than the saturated infiltration... Fs hour, f = i The Green-Ampt model is modified as follows: When saturated: .in, iRainfall intensity (m / s); f The infiltration rate is (m / s). Permeation rate (m / s); K s Saturated hydraulic conductivity (m / s); Saturated capillary head (m); This indicates an initial water deficit.

[0040] In some embodiments, the hydrological simulation and feedback assessment module treats each sub-catchment as a nonlinear reservoir and the runoff as a flood process flowing into pipes and channels, approximating a one-dimensional wave motion process. The slope outflow Q is solved by simultaneously applying the Manning equations and the water balance equations. Key parameters include catchment area, catchment depression capacity, characteristic width of the catchment area, catchment slope, and Manning roughness coefficient. Pipeline runoff is generally calculated using wave motion without sideflows. The complete St. Vennat equations are used to solve for Q in both pipes and channels, with key parameters including cross-sectional area, Manning roughness coefficient of the pipe (or channel), and pipe length. Since riparian wetland ecosystems are river bypass wetlands, channels are typically generalized as open pipes or ditches to participate in hydrological simulations as a pipe network.

[0041] In some embodiments, the simultaneous equations for the overburden flow Q are: ; .in, V The volume of water collected is (m³). D The depth of the water accumulation (m); A The catchment area (m²) 2 ); Q The outflow rate is (m / s). W The characteristic width (m) can be represented by the ratio of the catchment area to the slope flow length. n is the Manning coefficient, representing the roughness. d p The depression depth (mm) represents the equivalent depth of rainfall that can be retained by surface micro-topography. When the cumulative rainfall depth does not exceed... d p No slope runoff occurs, exceeding d p Then it enters the slope flow Q The calculation. S This represents the slope of the catchment area. When permeable or impermeable areas are involved, the outflow of each area is calculated separately based on its proportion of the total area and then summed to obtain the final total flow. Additionally, based on the Hargeaves method, daily average temperature and geographic latitude are used to estimate the daily potential evaporation. Finally, the monthly evaporation rate is corrected using the pan coefficient (PAN). This method is suitable for global use without further calibration, and is particularly suitable for regions lacking detailed meteorological data.

[0042] In some embodiments, the SWMM pollutant load calculation module is divided into accumulation and scour kinetic models. The accumulation mechanism involves pollutants accumulating on the surface of sub-catchment areas during non-rainfall periods through dry deposition and traffic emissions. This accumulation is influenced by the maximum pollutant accumulation rate (highest sensitivity, measured in kg / ha), the accumulation rate constant, the half-saturation constant (used for the exponent / saturation function), and the number of preceding dry days. Nutrients are typically calculated using a saturation function. During rainfall, the scour process of surface pollutants by runoff is affected by runoff intensity (mm / hr). Key parameters include the scour coefficient R (sensitive to low rainfall intensity) and the scour exponent n (sensitive to high rainfall intensity). The pollutant transport process is implemented through a user-defined function, where C represents the pollutant concentration. HRT The hydraulic residence time for the corresponding facility or path, used to characterize concentration decay and transport effects, is expressed as: .in, C in The concentration of influent pollutants entering the node or facility unit, C out The concentration of pollutants in the effluent after treatment by the facility or transport along a hydraulic path.

[0043] In some embodiments, green infrastructure may include multiple families of facilities. These families of facilities have different ecosystem service functions, and their maximum deployment area is determined by different deployment conditions. This allows riparian areas to fulfill their functions as ecological corridors and recreational spaces, serving as transitional zones between water and land systems.

[0044] In some embodiments, the multiple facility families may include a first facility family, a second facility family, and a third facility family. Exemplarily, the first facility family may be G1, an offline ecological pond / wetland family; the second facility family may be G2, a shoreline strip infiltration family; and the third facility family, G3, may be a permeable pavement family. The maximum deployment area for each facility family is... A max It can be based on the area of ​​the sub-catchment area A sub Distance from the river channel d rive ,slope S And at least one of the road types must be determined. In this way, the ecological service functions of the riparian zone, such as runoff control, pollutant purification, and recreation, can be enhanced through the synergy of the three facility families.

[0045] In some embodiments, the maximum deployment area of ​​the first facility family is determined based on the area of ​​the sub-catchment and a distance correction factor corresponding to the distance from the river channel. The distance correction factor may include multiple distance correction factors; the values ​​of the multiple distance correction factors are different, and each of the multiple distance correction factors corresponds one-to-one with a multiple distances from the river channel.

[0046] In some embodiments, the maximum deployment area of ​​the first facility family can be based on a formula. Calculation. Among them, Based on river distance d The correction factor is specifically: when d When ≤50m, =1. When 50m < d When ≤100m, =0.75. When 100m < d When ≤200m, =0.5. When d When >200m, =0.

[0047] In some embodiments, the maximum deployment area of ​​the second facility family is determined based on the area of ​​the sub-catchment, a correction parameter corresponding to the distance from the river channel, and a first slope correction factor. The first slope correction factor includes multiple first slope correction factors; the values ​​of the multiple first slope correction factors are different, and the multiple first slope correction factors correspond one-to-one with multiple first slopes.

[0048] In some embodiments, the maximum deployment area of ​​the second facility family can be based on a formula. Calculation. Among them, IMP The impermeability of the catchment area represents the increased demand for nearshore interception due to the degree of hardening. This is the first slope correction factor. When the first slope... S When ≤8%, =1. When 8% < first slope S When ≤12%, =0.7. When the first slope S > 12%, =0.

[0049] In some embodiments, the maximum deployment area of ​​the third facility family is determined based on a road type correction factor and a second slope correction factor. The road type correction factor includes multiple road type correction factors; the values ​​of the multiple road type correction factors are different, and each of the multiple road type correction factors corresponds one-to-one with a multiple road type. The second slope correction factor includes multiple second slope correction factors; the values ​​of the multiple second slope correction factors are different, and each of the multiple second slope correction factors corresponds one-to-one with a multiple second slope.

[0050] In some embodiments, the maximum deployment area of ​​the third facility family can be based on a formula. Calculation. Among them, κ 3 is the default coefficient (usually 0.5), indicating that a maximum of 50% of the road can be permeable; This is a road type correction factor. =0.8 (sidewalk). =0.6 (branch road / park road). =0.3 (main road). This is the second slope correction factor. Second slope hour, =1.4% < Second Slope hour, =0.7. Second slope hour, =0.4.

[0051] In some embodiments, the method further includes determining whether a facility family meets the deployment conditions based on whether its deployment area overlaps with a red line (such as a building, a cultural relic protection area, etc.). If the deployment area of ​​a facility family overlaps with the red line, it does not meet the deployment conditions, the facility family is prohibited from deployment, the `forbid` value is set to 1, and it is excluded from the calculation.

[0052] In some embodiments, it may also include calculating the effective water depth range of multiple facility families (e.g. ). This can be understood as the permissible range (m) of free water depth for the operation of green infrastructure. Its minimum water depth is positively correlated with the target volume of hydrological regulation, and its upper limit is the ecological safety limit. For the first facility family, the average effective water depth can be determined based on the effective reservoir capacity and the deployable area of ​​the first facility family. The effective reservoir capacity can be determined using the formula... Calculation. Among them, Effective storage capacity, in m³. ∈(0,1] is the safety factor (e.g., it can be 0.5–0.8). Asub is the upper limit of the available deployment area for this facility family, in m². P D,TThe unit can be mm, and it can be the total rainfall for that duration / recurrence period. (t) as well as (t) The unit can be m³ / s, which can represent the design inflow / outflow rate, respectively. Mean effective water depth. . A use The actual layout area of ​​the calculated facility family includes the actual layout area of ​​the facility family obtained in a certain round of calculation; and the final actual layout area obtained after the calculation (if the actual land area ≤ A). max Let A use =A max Finally, the effective water depth range is determined, and the output is: .in, The permissible range of free water depth for the operation of green infrastructure ( m (its minimum water depth) The final water depth (i.e., the upper limit) is positively correlated with the target volume of hydrological regulation. These are limits set for ecological security. (During implementation...) like The site is too large and constrained by the landscape, so the footprint of the first facility family can be increased (the final footprint still needs to be located in A). max (within the range) to reduce the required water depth.

[0053] In some embodiments, for the second facility family, the shoreline strip infiltration / grass ditch is characterized by shallow burial to shallow water. Among these, It can be used to determine the effective water depth of trenches / infiltration. h surface This refers to the surface water storage depth. It primarily focuses on the surface water storage depth (surface retention in trenches / depressions). h surface For depths ∈ [0.2, 0.3] m, if it is a grassy ditch + infiltration bed, and the infiltration layer thickness is 0.3–0.6 m, it is not considered as free water depth. The final effective water depth output for the second facility family of grassy ditch + infiltration bed type can be: =[0.2,0.3].

[0054] In some embodiments, for the second facility family, due to the potential for significant tidal and backwater phenomena in the riparian zone, it is necessary to determine the outflow structure type. The outflow structure type can be set to multiple types, for example, including three. Specifically, if tidal / backwater is significant or a constant weir crest is required, the outflow structure type is 'weir'; if fine-tuning of the drainage time is required and water level fluctuations are large, the outflow structure type is 'orifice'; if it is an infiltration wetland with deep groundwater and requires maintaining saturation / slow drainage, the outflow structure type is 'orifice + underdrain' or 'underdrain'.

[0055] In some embodiments, the various facility families can each have a corresponding mapping matrix; the mapping matrix is ​​used to characterize the changes in land use categories and key land parameters within the facility family. Specifically, the land use category within the facility family can include the area of ​​the original land category occupied by the facility, and the proportional relationship between the area of ​​the new land category and the area of ​​the original land category. The LUCC area ledger can be updated iteratively. Land use categories can refer to the first-level classification standard of the SinoLC-1 national land cover map data, including major categories such as: forest, cultivated land, shrubland, grassland, road, building, water body, wetland, snow and ice, tundra, barren land, and sparse vegetation. The key land parameters can include surface layer parameters, soil layer parameters, pollutant erosion parameters, and other parameters. For example, the matrix for each facility family can satisfy... .in, A GI Let Δ be the area of ​​facility family g. A LU This represents the change in each category of LU. M g The meaning is the change matrix of land use categories in the layout of each facility family. The specific mapping relationship can be seen in Table 1, and the specific configuration of other key parameters can be seen in Table 2 below.

[0056] Table 1. Matrix of changes in land use categories in the layout of each facility family. M g

[0057] Note: This table is based on a newly deployed area of ​​1 m², with a recommended value of α = 0.8. No facilities are installed when the slope of the shoreline is ≥3%.

[0058] Table 2 Reference Values ​​for Key Parameters of Facility Families

[0059] The hydrological simulation and feedback assessment module of this application embodiment, by defining facility families and land use change matrices, can specifically adapt to the green infrastructure configuration of riparian regions. The hydrological simulation and feedback assessment module of this application embodiment can simultaneously and automatically calculate the Shannon diversity index. To enhance the quantitative effect of ecological landscape diversity. 。 in, pk For the first k The proportion of land area of ​​each type. Area k The area (m²) covered by land of type k within the sub-catchment area is derived from LU_by_SubID.xlsx. . H′ is based on p k The calculated measure of diversity (Shannon diversity index). H′ base H′ is calculated before GI / facility deployment (baseline land use ledger); H′ GI H′ is calculated after the GI / facility deployment (updated by the land use change matrix Mg). Additionally, the quantitative indicators for the ecosystem service function of the GI are the pollutant removal rate (PRR) and peak flow reduction rate (PFRR) of the total outflow before and after the GI deployment, reflecting both the pollutant removal rate (PRR) and the hydrological regulation level. . Where M is the total pollutant load (such as COD, TSS, TP, TN) simulated per unit time (such as a rainstorm process), with units such as mg / L×m³, or directly in mass units (kg). M base The total load calculated for the same rainfall process and the same outlet / node before GI deployment (baseline). M GI The total load calculated for the same rainfall process and the same outlet / node after GI deployment (GI scheme). Peak runoff at the target outlet / node of interest under the same rainfall event before GI deployment (baseline). The peak runoff at the target outlet / node of interest under the same rainfall event after GI deployment (GI scheme).

[0060] In some embodiments, the graph neural network spatial topology construction module can reflect the spatial topological structure of hydraulic, geographical, and ecological relationships within a catchment area, providing a high-dimensional spatial feature representation for subsequent strategy learning and collaborative optimization. The geospatial layer can include sub-catchment boundaries, flow direction information, topographic elevation, and infrastructure distribution. The spatial adjacency graph can be a directed spatial topology graph, such as G=(V,E). Node V represents each sub-catchment unit, and edge E represents the hydrological connectivity or geographical proximity between upstream and downstream. This can be understood as follows: in urban drainage systems, the construction of the spatial topology graph originates from the sub-catchment division and hydraulic connectivity analysis of the SWMM model: the riparian zone is subdivided into sub-catchments, each sub-catchment is connected to the downstream through an outlet, forming a directed confluence network, where upstream runoff and pollutants are driven downstream by gravity. This graph structure simulates hydrological processes, ensuring that the impact of upstream treatment on downstream can be quantified and propagated. A graph neural network is used to perform a "neighborhood aggregation" message passing mechanism on the graph structure, encoding the information of each node and its upstream and downstream nodes into the representation.

[0061] In some embodiments, nodes Represents a sub-catchment area, with edges This indicates that the outflow from sub-catchment j flows directly into the sub-catchment. This forms a hydraulic connection from upstream to downstream. Node characteristics may include node number, area, slope, impermeability, outflow outlet, GI configuration (e.g., including type, scale, and treatment area), and runoff Q. Edge characteristics may include the distance of the hydrological path connecting sub-sinks. The upstream-to-downstream transitive relationship is encoded as a directed acyclic graph (DAG), and the connectivity of sub-catchments is constructed as an adjacency matrix A.

[0062] In some embodiments, a side weight matrix W can be introduced based on the proportion of catchment volume between adjacent sub-catchments on a unit flow path. ]. Where Q represents the sub-sink runoff, in cubic meters per second (m³). 3 / s; This represents the length of the water flow path, in meters (m). After normalization, the weighted adjacency matrix for stable propagation can be obtained. . This allows for a more accurate depiction of the upstream's contribution to the downstream, avoiding distortion in the transmission process.

[0063] In some embodiments, node features can be encoded and aggregated using a graph convolutional network, such as a GCN (Graph Convolutional Network). Through message passing and weight updates between nodes, the resulting GNN model can capture the collaborative characteristics of each sub-catchment in terms of geographical location, hydrological response, and facility impact.

[0064] In some embodiments, training the GNN model based on the multidimensional simulation dataset and the plurality of spatial embedding feature vectors includes: The node features of the node's direct upstream nodes are aggregated using the first layer propagation in the GCN model to simulate the impact of the upstream nodes' LID facilities on the node.

[0065] The hidden states of the node's direct upstream nodes are aggregated using the second-layer propagation method in the GCN model to obtain the final hidden state, thereby capturing the influence of indirect hydraulic paths on the node. Indirect hydraulic paths can be, for example, upstream of upstream.

[0066] Based on the final hidden state, the target hydrological and ecological indicators of the node under a given combination of green infrastructure deployments are predicted, and the reward value for reinforcement learning is calculated based on the predicted indicators, thus obtaining the GNN model.

[0067] In some embodiments, the influence of the upstream node's LID facilities on the node is determined based on a first weight matrix, the hydraulic connectivity with adjacent nodes, the hydrological feature vector of the directly upstream node, and a first bias vector. Specifically, the influence of the upstream node's LID facilities on the node can be expressed as follows: .in W 1 represents the first weight matrix [8, 64]. h For Zihui District The integrated rainfall-runoff response after propagation through the first layer of GCN. b1 is

[64] , which is the bias vector and is retained as the base water volume. The activation function has the following mathematical expression: . For water flow connection strength, Let be the hydrological feature vector of the upstream sub-catchment area j.

[0068] In some embodiments, the influence of the indirect hydraulic path on the node is determined based on a second weight matrix, the hydraulic connection relationship with adjacent nodes, the hidden state of the direct upstream node, and a second bias vector.

[0069] In some embodiments, the total reward is determined based on the final hidden state, a third weight matrix, and a third bias vector. The dimension of the weight matrix is ​​set according to the node feature dimension and the hidden layer size to achieve a non-linear mapping from node features to the hidden state and output metric. The final hidden state is... The total reward for each node is... .in, hi (2) This is the final hidden state. R This represents the total reward under the current LID configuration. W 2 is the first weight matrix [8, 64]. W out The third weight matrix has a shape of [64, 1], representing the compression of the 64-dimensional global representation into a 1-dimensional output. b2 is

[64] , representing the bias vector. B3 is [1], representing the bias vector. Each element... Wout [k, 0] represents the contribution weight of the k-th feature to hydrological performance.

[0070] In some embodiments, to determine the parameters of the GNN surrogate model, the GNN surrogate model is trained using a training dataset generated by the hydrological simulation and feedback evaluation module. Typically, when generating the training dataset, 100-500 LID (i.e., GI) layout schemes (including LID type, location, and scale, etc.) can be randomly designed to simulate a 5-minute resolution rainfall sequence (2-10 year return period), and the total peak runoff reduction rate is output as... The GNN proxy model in this application embodiment uses mean squared error loss. 2 The optimizer is Adam (learning rate). Training for 200-300 rounds until the validation set is reached. R 2 >0.95. Initial weights W1, W2, W out and biases b1, b2 and b out Using Xavier initialization and updating via backpropagation, the prediction error is ensured to be less than 5%. This constitutes the initial construction of a complete GNN proxy model.

[0071] In some of these embodiments, see Figure 3As shown, the distillation surrogate modeling module is configured to lightweight the fully constructed GNN surrogate model. It introduces KL divergence (Kullback-Leibler (KL) divergence (also known as relative entropy and I divergence)) and MSE (Mean Squared Error) as a composite loss function. It utilizes soft-label transfer of complete GCN knowledge, inputs the same topological graph, and outputs approximate predicted values, resulting in a 1-2 layer lightweight GNN surrogate model. The GNN model is then trained based on a multi-dimensional simulation dataset. The composite loss function can be... . The mean squared error loss for the lightweight surrogate model and the true labels in SWMM. For lightweight proxy models and complete soft tags ( (output distribution) Divergence loss. .in, p all The output probability distribution of the complete GNN surrogate model. The output probability distribution of a lightweight GNN surrogate model. α The balancing parameter is 0.5.

[0072] In some embodiments, when training a lightweight GNN surrogate model based on a multidimensional simulation dataset, iterative training can be performed for 100 rounds with a learning rate of 0.0005 to keep the accuracy loss within 5%, monitor the KL divergence convergence curve to avoid over-distillation, and significantly reduce parameter and inference latency. The lightweight GNN is fine-tuned using different rainfall intensities (e.g., including rainfall intensities with a 1-10 year return period) and GI configuration variants to ensure generalization ability. Validation metrics can include mean squared error (MSE < 0.02) and correlation coefficient (R² > 0.95) to compare the prediction consistency between the lightweight GNN surrogate model and the SWMM. The lightweight GNN surrogate model is then embedded into MASAC to replace SWMM for state-reward feedback. During MASAC action updates, the merits of the GI configuration are quickly evaluated with approximate values, and multi-objective decision-making (e.g., water quality / water quantity / spatial distribution multi-objective weights) is supported, ultimately achieving accelerated optimization convergence (reducing the number of simulations by more than 60%).

[0073] In some of these embodiments, please refer to Figure 4MASAC's multi-agent reinforcement learning optimization decision-making module can automatically generate green infrastructure (GI) deployment schemes and optimal strategies for riparian areas, achieving systematic control of runoff, improvement of water quality, and enhancement of ecological synergy. MASAC treats the green infrastructure deployment problem as a multi-agent decision-making process, with each sub-catchment considered as an agent. Through local states, neighborhood information, and global rewards, it learns the optimal strategy for deploying GIs. Specifically, the MASAC multi-agent reinforcement learning optimization decision-making module maintains independent policy networks (Actors) and value networks (Critics) for each agent, and utilizes a framework of centralized training and distributed execution to address the non-independence problem among multiple agents. During training, the algorithm achieves stable convergence of policies and collaborative behavior learning in complex environments through soft updates to the target network and an experience replay mechanism.

[0074] In some embodiments, this application has adapted and technically coupled MASAC, embedding a phased reward mechanism and GNN into MASAC to optimize the dynamic adaptability of riparian LID layout, especially for dynamic adaptability in scenarios with high rainfall and complex terrain conditions. The phased reward mechanism dynamically adjusts reward weights based on rainfall intensity (e.g., light, moderate, and heavy rain) and terrain complexity (e.g., slope or connectivity), ensuring consistent optimization goals for LID layout across different scenarios.

[0075] In some embodiments, the reward R can be divided into three stages, primarily determined based on rainfall P (mm) and slope S (%). Specifically, for light rain (P<20), the reward R = 0.6 * water quality + 0.2 * runoff + 0.2 * Shannon index. For moderate rain (20≤P<50), the reward R = 0.4 * water quality + 0.4 * runoff + 0.2 * Shannon index. For heavy rain (P≥50), the reward R = 0.3 * water quality + 0.5 * runoff + 0.2 * Shannon index. Here, water quality represents the pollutant reduction rate, runoff represents the peak flow reduction rate, the Shannon index reflects biodiversity, and the weighting varies with rainfall to emphasize peak flow control.

[0076] In some embodiments, the weights can also be adjusted based on the regional average slope S2 (%). Specifically, if the regional average slope S2 > 5%, the runoff weight is increased by 0.1, and the water quality weight is decreased by 0.1, to prioritize flood reduction in steep slope scenarios. At each iteration, the agent can predict from the lightweight GNN surrogate model based on current rainfall and terrain data. Dynamically select stage weights and update rewards.

[0077] In some embodiments, a lightweight GNN agent model replaces SWMM in the MASAC embedding part, capturing the spatial synergistic effect between sub-sectors and improving the adaptability of MASAC. The modified MASAC algorithm can be understood as follows: First, each agent has its own Actor network π_θ(a|oi) (oi is the local observation), and a Critic network... Shared globally. Initialize parameters θ (Actor) and ψ (Critic) to random values, and the learning rate... =0.0003-0.001. MASAC optimization relies on real-time feedback of the environmental state s. The hydrological environment of riparian zones varies due to rainfall intensity and topographic complexity, and traditional single-state inputs are insufficient to cope with dynamic changes. Therefore, a lightweight GNN surrogate model is introduced to extract the spatial topological relationships and attributes of sub-catchment intervals from SWMM data to generate dynamic states s. The lightweight GNN surrogate model uses an adjacency matrix A[N, N] (where N is the number of sub-catchments, ...). A ij = 1 means Upstream connection j) and node features xi [8] (area, impermeability, GI attribute) simulate hydraulic conduction. After two layers of convolution and pooling in the GNN module, the total reward R is obtained as the new state. Input MASAC to ensure that the state corresponding to the action reflects the dynamic changes in rainfall (P, mm) and slope (S, %), laying the foundation for reward calculation and guiding GI layout adjustments.

[0078] Since each sub-catchment needs to independently select GI actions, but traditional MASAC ignores spatial cooperation, it is prone to getting trapped in local optima. Therefore, for each sub-catchment agent... Choose an action ai (GI type and area), and a joint action a = [a1, ...,an]. Each sub-sector agent is governed by its own Actor policy network. The action probabilities are adjusted based on the GNN output η. In the next stage, the action a is input to the Critic for evaluation, driving the optimization loop.

[0079] Critic Network The system receives the global state s and the joint action a of all agents, calculates the joint Q-value Q(s,a), guides the policy update of each agent, and shares the global value. Before convergence, the GNN agent module is called to return the new state s' and reward R, and the transition (s, a, R, s') is stored in the experience replay buffer D. Then, according to the Soft Actor-Critic objective function update strategy, a mini-batch of data (s, a, R, s') is randomly sampled from D, and the objective Q-value is calculated, where the parameters include a discount factor γ=0.99 and an entropy coefficient α=0.2; then the loss function is automatically calculated to complete the Critic update.

[0080] Then, the Actor, based on Critic feedback and the gradient ascent optimization strategy, softly updates the target network with a step size of 0.005 to maximize the Q-value and entropy. After each trial and error, the strategy is updated. This is done to increase the probability of selecting high-reward actions. Finally, a soft update is performed, repeating the steps from Critic update to soft update for over 200 iterations, checking the learning curve and mixing rate. The communication overhead ψ index is used until the average reward R > 0.8 and the fluctuation < 0.05. If convergence is not achieved, s (new rainfall) is updated, and the iteration is returned. Finally, the optimal joint action is obtained from the output. Extract the sub-collection ID, LID type, and area, and map them to the DEM latitude and longitude to generate... The optimal policy was determined by validating SWMM modeling using geographic information system software, confirming R² > 0.95. The policy was then exported to ONNX format and used in the script. Load to enable seamless integration with the ArcGIS Python environment.

[0081] The green infrastructure multi-objective collaborative optimization system provided in this application mainly includes topographic parameters, land use classification (LUCC), rainfall pattern scenario settings, existing GI deployment schemes, and hydraulic control facility operation parameters as data inputs. Through spatial preprocessing and standardization, node and edge features are generated for use by the GNN distillation agent modeling module in constructing the topology graph and the reinforcement learning module. Table 3 below shows the detailed system input data parameters.

[0082] Table 3 System Input Data Table

[0083] The system uses multi-source heterogeneous spatial data as input, combines graph neural networks to construct a spatial collaborative relationship map, and generates the optimal green infrastructure layout strategy through a multi-objective collaborative optimization algorithm. The output results include not only facility spatial layout files and optimization score matrices, but also automatically generated strategy script files. The scripts can be embedded in geographic information systems (such as QGIS and ArcGIS) for deployment simulation and secondary optimization.

[0084] Table 4 System Output Data Table

[0085] Referring to Table 4, the system outputs include, but are not limited to: the optimal combination of GI deployment locations and types, the collaboratively optimized hydraulic control strategy (pump station / gate control parameters), and system operation indicators (runoff reduction rate, pollution reduction rate, ecological Shannon index score, etc.) based on physical model / surrogate model simulations. All output results can be embedded into the GIS system in graphical form or provided to the urban design decision-making platform via API (Application Programming Interface), demonstrating good system portability and potential for secondary development.

[0086] In some embodiments, a report generation module may also be included, which is configured to generate a report by combining the optimal green infrastructure layout for analysis.

[0087] In some embodiments, a visualization module may also be included to visualize the report for easier viewing. This can improve user satisfaction to some extent.

[0088] The green infrastructure spatial optimization system in this application adopts a collaborative architecture of "intelligent decision-making module and physical simulation core". The hydrological simulation and feedback evaluation module serves as the underlying physical engine of the system, operating throughout all stages of system operation. It provides a consistent hydrological structure foundation, teacher data, and dynamic feedback for the graph neural network spatial topology construction module, the GNN distillation agent modeling module, and the MASAC multi-agent reinforcement learning optimization decision-making module. The four modules are organically linked through three main collaborative lines: a data flow collaborative line, a knowledge distillation line, and a physical feedback line.

[0089] The data flow coordination mechanism is primarily manifested in the system initialization phase. The hydrological simulation and feedback evaluation module first establishes a basic simulation file (.inp) based on the target watershed. This file can contain data such as catchment area distribution, node connections, pipeline topology, discharge outlets, and hydrological parameters. The graph neural network spatial topology construction module reads the node and pipeline structures from this basic model, generates a spatial graph structure G=(V,E), and simultaneously extracts the physical characteristics (area, slope, land use, etc.) of sub-catchments to form a node attribute matrix. Thus, the graph neural network spatial topology construction module achieves structural-level synchronization with the hydrological simulation and feedback evaluation module, ensuring consistency between spatial modeling and physical simulation.

[0090] The knowledge distillation line is primarily manifested in the hydrological simulation and feedback evaluation module's role as a complete model during the distillation surrogate model training phase. It generates a training sample set through batch simulations of different facility configurations, including multi-dimensional response indicators such as peak flow, total runoff, pollutant concentration, and Shannon diversity index. These simulated samples, after being formatted, are input into the GNN distillation surrogate modeling module to train the student model, enabling it to learn the mapping relationship between facility layout, spatial topology, and various indicators. This GNN distillation surrogate modeling module inherits the physical characteristics of the hydrological simulation and feedback evaluation module, but with significantly reduced computational complexity, achieving predictions in milliseconds. After training, the hydrological simulation and feedback evaluation module retains its role as an accurate simulation reference, providing a data source for redistillation and accuracy calibration for the GNN distillation surrogate modeling module.

[0091] The physical feedback loop is primarily manifested in the reinforcement learning optimization phase, where the MASAC optimizer invokes a GNN surrogate model to perform large-scale policy search and generates multiple rounds of candidate layouts based on the surrogate model's predictions. After every few iterations, the system automatically invokes SWMM to perform sampling simulations of some representative layout schemes to obtain real hydrological response data. These simulation results are fed back to the GNN distillation surrogate modeling module to correct lightweight model parameters (Re-Distillation) and optimize the weight calculation logic of the reward function. Once the optimization process is complete, the hydrological simulation and feedback evaluation module again assumes a verification role, performing a full simulation of the final optimal layout scheme and outputting an authoritative hydrological benefit assessment report.

[0092] The collaborative logic of the green infrastructure multi-objective collaborative optimization system provided in this application embodiment may include: (1) a basic model generation unit in the hydrological simulation and feedback evaluation module, used to provide the watershed basic structure and physical parameters; (2) a graph neural network spatial topology construction module, used to construct the graph structure and extract the node matrix; (3) a GNN distillation surrogate modeling module, used to perform surrogate modeling based on SWMM training data; (4) a MASAC multi-agent reinforcement learning optimization decision module, used for intelligent strategy optimization; and (5) an SWMM verification feedback unit in the hydrological simulation and feedback evaluation module, used to output physical simulation results and perform surrogate correction. Each module forms a dynamic closed loop through data flow, knowledge flow and feedback flow, realizing multi-level collaboration from physical simulation to intelligent optimization.

[0093] The method and execution steps of the green infrastructure multi-objective collaborative optimization system provided in this application embodiment may include: Step S10: Collect riparian surface environmental data, process missing or low-quality data, and construct the input for the basic model in the hydrological simulation and feedback evaluation module. Obtain topographic data in the form of Tin, DEM, and Terrien data for the study area, and perform hydrological analysis using ArcGIS, QGIS, or professional water treatment tools (flow direction-catchment accumulation-drainage point-watershed tools) to obtain slope and flow direction attributes. Based on the topographic slope and flow direction analysis, divide the area into multiple sub-catchment areas; if the storm drains of the drainage network are used as the center of stormwater collection, it is recommended to divide the sub-catchment areas using the Thiessen polygon method based on streets or the distribution of storm drains. The division result should be a continuous vector polygon file with watersheds as the boundary, serving as the boundary of the sub-catchment in SWMM 5.2 and above (hereinafter referred to as SWMM) models; assign a unique SubID number to each sub-catchment for subsequent data association and model calling.

[0094] Step S20: Obtain the upstream and downstream river channel data flowing through the sub-capital area, input the basic model in the hydrological simulation and feedback evaluation module, and establish the pipeline and manhole nodes and their attribute tables.

[0095] Step S30: Acquire rainfall data, calculate the 24-hour total rainfall for 1, 2, 5, and 10-year return periods using the local IDF (Intensity-duration-frequency) formula, and convert it into a standardized eccentric rainfall pattern with a 5-minute step size using the Chicago distribution parameter r. Name the patterns according to the return periods and input them into the base model in the hydrological simulation and feedback evaluation module. Simultaneously, extract the temperature sequence and corrected 3-hour rainfall from BCC-CSM2-MR under the SSP245 / 585 scenario, and reconstruct them to the same 24-hour total and 5-minute resolution using Huff quartile curves. Convert the climate file to .dat and input it into SWMM. Subsequently, perform uniform interpolation correction on both types of sequences to form a consistent time-P(mm) CSV file, add scenario labels to the file header, and input it into the SWMM model. Finally, when constructing the training set, the Chicago and future rain patterns are randomly selected at a 1:1 ratio and merged into rainset.npy for MASAC to call, realizing "dual-scenario" training of historical benchmark testing and future adaptive evaluation.

[0096] Step S40: Under the CGCS 2000 iso-area projection, UAV or high-resolution imagery acquired within the last 6 months with a resolution ≤2 m is used for machine learning / visual interpretation based on the first-level classification standard of the SinoLC-1 national land cover map data. This includes major categories: forest, cultivated land, shrubland, grassland, roads, buildings, water bodies, wetlands, snow and ice, tundra, and barren and sparse vegetation. A topologically correct LU_clean.shp is generated by eliminating <10 m² fragments. At least 200 validation points are then extracted for cross-validation to ensure an overall accuracy ≥ 85% and Kappa ≥ 0.80. If using the SinoLC-1 national land cover map data, a value table is exported, and an area table is created to obtain the land use cover table LU_by_SubID.xlsx for each sub-sink. If it is a vector, the data is written to the Shape_Area field and a Spatial Join is performed with the sub-sink water area layer to obtain LU_by_SubID.xlsx, which records the SubID, land use type, and area of ​​each category. Input LU_by_SubID.xlsx into the sub-sink land use [LANDUSES] control properties of SWMM. Then, based on the land use data, calculate the sub-sink parameters such as impermeability, Manning roughness, and depression depth (mm) for each sub-sink. Finally, calculate the Shannon index and archive the results along with the image metadata LU_meta.xml to provide high-precision and traceable land use input for subsequent MASAC-GNN optimization.

[0097] Step S50: In the SWMM, define pollutant attributes in the [POLLUTANTS] control, including name, unit, drought concentration, initial concentration, decay coefficient, etc. In the [BUILDUP] control, define the cumulative function for each pollutant and parameters such as maximum cumulative amount (kg / ha), number of half-saturation days (number of days to reach 50% maximum cumulative amount), and time index. Then, set the number of drought days before the model configuration in [OPTIONS], and define the washoff function for each pollutant and parameters such as event average concentration, washoff coefficient, and washoff index in the [WASHOFF] module.

[0098] Step S60: If there are LID facilities that can control runoff, such as bioretention ponds, vegetated swales, rainwater buckets, and infiltration channels in the sub-catchment area, the existing facility area and hydraulic control facility operation parameters (pump station flow rate, weir gate height) need to be configured in the LID control of SWMM. Step S70: In the common page of the displayed simulation options dialog box, select motion wave as the flow calculation method, set the infiltration method to Modified Green-Ampt, and use the SSURGO (Soil Survey Geographic Database) soil database or measured data to match the infiltration parameters.

[0099] Step S80: Obtain a record of a local rainfall event and observe the peak runoff Q and pollutant concentration at the downstream outlet or node of interest in the field. Input the rainfall sequence, run the model, record the predicted peak runoff Q and pollutant concentration at the downstream outlet or node of interest, and calculate the Nessler efficiency coefficient NSE. If NSE > 0.70, the model runoff calibration is qualified. If the total pollutant load deviation is less than 20%, the model pollutant parameters are qualified, and the optimization steps can continue. Otherwise, if the measured peak value is greater than the simulated peak value, adjust the scour function parameters to reduce the accumulation. If the peak runoff time is too early, fine-tune the Manning roughness coefficient by ±0.005. If the peak runoff is too large, increase the depression depth by ±2 (mm) or increase the soil permeability coefficient until the model calibration is qualified.

[0100] Step S90: Provide the sub-catchment area Asub, distance from the river channel d (m), and slope S (%) of the qualified model. The system will automatically calculate the upper limit of area Amax and the effective water depth range for each sub-catchment and each facility family according to the formula. The table includes parameters for outflow structure types, forbidden facility family markers, and land use change matrix Mg.

[0101] Step S100: Extract the sub-sump ID from the sub-sumption sections [SUBCATCHMENTS] in the SWMM baseline model, treat it as a graph node V, and add the area. Slope S, Impermeability IMP, Outflow Extract GI configuration from [LID_USAGE] The type of the generated proxy dictionary contains... This serves as a component of the node feature vector. Using the water system and flow direction data from step S1, the -Outlet field in SWMM is added as the downstream sub-sink ID, and the terminal sub-sink is set as the outlet node. Then, the upstream sub-sink IDs are traversed in reverse, and each sub-sink to downstream connection is defined as a directed edge E. The baseflow and sub-sinks of the actual rainfall event are obtained from [SUBAREA]. Hydrological path distance from the discharge point to the inflow point j Calculate the hydraulic connection weights , all After normalization, we get A directed graph G of a graph neural network with hydrological pathways as its structure is constructed.

[0102] Step S110: Using a graph neural network framework, such as GCN, define a propagation function based on neighbor aggregation. After two layers of propagation, obtain the embedding vector of each sub-sink unit node. h i (2) This vector represents the fusion of its upstream and downstream structural features and hydrological attributes within the entire watershed map. This implicit vector... As input to the agent model or policy network, it enables the optimization policy to have spatial coordination awareness.

[0103] Step S120: Using multiple sets of rainfall time series, differentiate the attributes of the LID of the existing sub-catchment areas, through... The interface uses a batch processing method to run a complete rainfall simulation once for each scheme and save the results as multiple training datasets. Each sample has a different combination of rainfall time series and features such as subslope, impermeability, facility family type, and facility treatment area. The output of the samples includes inlet and outlet pollutant reduction rate, peak runoff reduction rate, and Shannon index change rate, resulting in a tensor dataset (input state, facility deployment actions, and simulation output indicators).

[0104] Step S130: Based on the GNN interpretable neural network structure, perform regression modeling on the relation pairs in the tensor dataset, with at least two convolutional layers. Use the Adam optimizer algorithm for training iterations with a learning rate of 0.001, a batch size of 32, and 200-500 iterations. After convergence, calculate the mean squared error (MSE). When MSE < 0.05, a complete surrogate model is obtained. Save the trained teacher model weight parameters Θ and the output predicted physical vector y. The temperature function (temperature coefficient, usually taken as 4) is a smoothed probability distribution P of y, i.e., hard output and soft output, which is used as the input for the next distillation step.

[0105] Step S140: To lightweight the SWMM surrogate model, it is necessary to reduce the number of units and parameters of the full model, remove redundant attention mechanisms or skip layers, and retain only the basic graph convolution operator. During distillation, the loss function of the lightweight surrogate model is defined by two parts: the hard-labeled MSE term and the soft-labeled KL divergence term (used to measure the distribution difference between the output of the lightweight surrogate model and the output of the full surrogate). The total loss function is: ,in α=0.5~0.7, used to balance accuracy and knowledge transfer strength, the optimizer used is Adam, the learning rate is 0.005, the batch size is 32, and the iteration is 100. 20% of the dataset is used as the validation set, and the average error with the real SWMM output is calculated. When the average error is ≤2% and R²>0.95, the lightweight proxy model is considered to be substitutable.

[0106] Step S150: In the MASAC algorithm training loop, replace the SWMM simulation with a lightweight model; Step S160: Start the system and load the GNN output file (containing...) h i (2)

[64] Vector and adjacency matrix A [N,N], where N is the number of sub-sectors). In the configuration file " In the configuration, the number of agents N is set to the number of sub-sectors, the initial action ai is a random facility family type (1-3) and area (0-100 m²), and the Q-value network parameters are initialized to 0.

[0107] Step S170: The system calls the lightweight GNN model, input... h i Using AI, we obtain peak runoff reduction rate, pollutant reduction rate, and Shannon index changes to predict hydrological response. Based on a rainfall intensity of P=30 mm, weights w1=0.4 (water quality), w2=0.4 (runoff), and w3=0.2 (Shannon index) are set in the "reward_weights.txt" file. The reward is calculated as: R = 0.4·water quality + 0.4·runoff + 0.2·Shannon index. Finally, the (hi, ai, R, hi') tuple for the current cycle is stored in the experience pool for each Agent.

[0108] Step S180: The system periodically samples from the experience pool, and each agent i reads the facility family parameter table (containing Amax, ...) of its sub-sector. (forbid, etc.), according to the strategy Select AI. There is a 10% probability of random exploration, generating a new action file "actions_new.csv".

[0109] Step S190: The system loads the Q-value network and calculates Q(s, a). Target Q-value: = R +0.99·( -0.2·action_score), where action_score is the entropy value (0-0.5). Range [-1, 1]. Loss Adjust parameters 100 times to Save to " ".

[0110] Step S200: Update based on Goal = Q - 0.2 · action_score Save to "policy_updated". Soft update target network: new _target = 0.005 · current + 0.995 · old _target, update "target _network". ".

[0111] Step S210: Repeat steps S18-20 for 300 iterations. Check the R value in "reward_log.txt", the target R > 0.8, the fluctuation < 0.05, and generate "optimal_actions.csv".

[0112] Step S220: The user loads "optimal_actions.csv" (containing the optimization results of all sub-actions). Combine this with "params_table_i.csv" to validate the constraints.

[0113] Step S230: The system calls the lightweight model, input... and real time h i Output =0.88, save to " ".

[0114] Step S240: ... Turn to Example fields include "sub _id": 1, "lid _type": 1, "area": ​​75, etc., saved as " _ The user loads the file in ArcGIS.

[0115] The embodiments of this application can overcome multiple technical limitations of current green infrastructure layout optimization methods in terms of spatial perception, dynamic adaptation, multi-objective coordination, and ecological resilience expression. Specifically, the technical problems that can be solved include the following five dimensions: 1. Addressing the issue of fragmented spatial configuration and enhancing upstream and downstream collaborative layout capabilities: Existing GI optimization schemes fail to model the hydraulic and geographical relationships between sub-catchments, resulting in fragmented and uneven spatial deployment of facilities. This leads to non-cooperative phenomena such as "downstream stacking and upstream gaps," weakening the overall peak-shaving capacity and operational efficiency of the system. A graph neural network is introduced to construct a spatial topology map of the sub-catchment areas, and a spatial collaborative reward is designed. Through reinforcement learning, the collaborative and distributed deployment of GIs in space is guided, improving the collaborative control capabilities and structural rationality of the entire regional drainage system.

[0116] 2. Address the issue of static, single-execution optimization schemes and enhance the ability to dynamically adapt to policy transfer. Traditional NSGA-II schemes employ offline, single-cycle optimization, which does not support policy updates and real-time responses across multiple time periods, different rainfall patterns, or climate scenarios, and lacks a dynamic adjustment mechanism. We construct a multi-agent reinforcement learning strategy based on GNN and MASAC, enabling the model to perform policy adaptation and multi-round interactive optimization in future IPCC climate scenarios, achieving a shift from "static multi-objective optimality" to "dynamically transferable optimality."

[0117] 3. Address the lack of clear goals regarding biodiversity and social equity, and enhance the multifunctional integration capabilities of optimization models. Existing optimization objectives primarily focus on engineering indicators such as cost, runoff, and pollution reduction, failing to measure the ecological value of green facilities (e.g., biodiversity) and cultural values ​​related to human senses, such as landscape diversity. Introduce the Shannon index into the optimization objective function and reward mechanism to drive the system to generate GI schemes with diverse combinations of types, thereby balancing ecosystem resilience, landscape richness, and social sustainability.

[0118] 4. Addressing the computational time consumption issue in large-scale simulation processes and improving optimization efficiency at the regional scale: In the NSGA-II framework, each individual evaluation requires a complete call to the SWMM simulation. Especially in regional scale or multi-scenario simulations, the optimization process can take more than 24 hours, resulting in high computational resource consumption. This method was tested on three benchmark networks, with an average relative error of less than 5% and an observation ratio of only 5%. This indicates that GNN can serve as a surrogate model for SWMM. By leveraging the policy approximation capability learned by GNN, partial replacement of SWMM hydrological simulations is achieved, significantly reducing the frequency of SWMM calls during model iterations and lowering the computational load by more than 50%, thus adapting to the time constraints of practical engineering projects.

[0119] The multi-objective optimization strategy proposed in this application can theoretically reduce the number of model simulation calls by more than 60% within the same deployment cycle, significantly improving deployment efficiency and demonstrating high responsiveness for engineering applications. Preliminary tests show that this method can still stably output optimization strategies in areas with high land use heterogeneity and a large number of sub-catchments (>100), exhibiting good convergence speed and decision reliability. The upstream and downstream topological state inputs of sub-catchments constructed through a spatial graph neural network can effectively capture cross-regional hydraulic transmission relationships, providing "cooperative perception capabilities" for the MASAC strategy network, and is expected to improve deployment imbalance problems such as "downstream stacking and upstream gaps." The reward function is designed to consider three dimensions: pollution reduction, water volume shaving, and habitat heterogeneity, making it applicable to various ecological shoreline planning scenarios, such as urban riverside green corridors, industrial park stormwater systems, and ecological restoration of agricultural irrigation canals.

[0120] Those skilled in the art should understand that the discussion of any of the above embodiments is merely exemplary and is not intended to imply that the scope of this disclosure (including the claims) is limited to these examples; within the framework of this disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other variations of different aspects of the embodiments of this disclosure as described above, which are not provided in detail for the sake of brevity.

[0121] Although this disclosure has been described in conjunction with specific embodiments thereof, many substitutions, modifications and variations of these embodiments will be apparent to those skilled in the art from the foregoing description.

Claims

1. A multi-objective collaborative optimization system for green infrastructure, characterized in that, include: The system includes a graph neural network spatial topology construction module, a distillation agent modeling module, a hydrological simulation and feedback evaluation module, and a multi-agent reinforcement learning optimization decision-making module. The graph neural network spatial topology construction module is configured to build a spatial adjacency graph based on the input geospatial layer, and encode and aggregate the node features in the spatial adjacency graph to obtain multiple spatial embedding feature vectors; the multiple spatial embedding feature vectors correspond one-to-one with multiple nodes in the spatial adjacency graph; wherein, the nodes in the spatial adjacency graph correspond to sub-catchments, and the edges in the spatial adjacency graph represent the hydraulic connection relationship from the upstream sub-catchment to the downstream sub-catchment; The hydrological simulation and feedback assessment module is configured to generate a multidimensional simulation dataset based on a physical simulator, taking into account the upper limits of the deployment of various green infrastructure families and the mapping of land use change; the multidimensional simulation dataset is used for training the GNN model; each set of data in the multidimensional simulation dataset includes a combination of green infrastructure layouts, as well as the corresponding peak runoff, total runoff, pollutant load and landscape diversity index. The distillation proxy modeling module is configured to train a GNN model based on the multidimensional simulation dataset and the multiple spatial embedding feature vectors; to perform lightweight processing on the GNN model using a model distillation method to obtain a lightweight GNN proxy model; wherein the parameter size of the lightweight GNN proxy model is smaller than that of the GNN model; the GNN model is used to fit the mapping relationship between the layout of green infrastructure families and hydrological performance; the lightweight GNN proxy model can replace the physical simulator to provide performance or reward estimation during reinforcement learning; The multi-agent reinforcement learning optimization decision module is configured to, based on the multiple spatial embedding feature vectors, calculate the green infrastructure layout combination of the node through a reinforcement learning algorithm under multi-objective constraints and facility deployment limits, so as to determine the optimal green infrastructure layout combination of the node; wherein, a sub-catchment area corresponds to a single agent; The hydrological simulation and feedback evaluation module is also configured to perform physical-level verification of the optimal green infrastructure layout combination, so that the optimal green infrastructure layout combination has feasibility and hydrological effectiveness under actual rainfall conditions.

2. The green infrastructure multi-objective collaborative optimization system according to claim 1, characterized in that, The training of the GNN model based on the multidimensional simulation dataset and the multiple spatial embedding feature vectors includes: The node features of the node's direct upstream nodes are aggregated using the first layer propagation in the GCN model to simulate the impact of the upstream nodes' LID facilities on the node. The hidden states of the direct upstream nodes of the node are aggregated using the second layer propagation in the GCN model to obtain the final hidden state, in order to capture the influence of indirect hydraulic paths on the node. Based on the final hidden state, the target hydrological and ecological indicators of the node under a given combination of green infrastructure layouts are predicted, and the reward value for reinforcement learning is calculated based on the target hydrological and ecological indicators, thus obtaining the GNN model.

3. The green infrastructure multi-objective collaborative optimization system according to claim 2, characterized in that, The influence of the upstream node's LID facility on the node is determined based on the first weight matrix, the hydraulic connection relationship with adjacent nodes, the hydrological feature vector of the direct upstream node, and the first bias vector.

4. The green infrastructure multi-objective collaborative optimization system according to claim 2, characterized in that, The influence of the indirect hydraulic path on the node is determined based on the second weight matrix, the hydraulic connection relationship with adjacent nodes, the hidden state of the direct upstream node, and the second bias vector.

5. The green infrastructure multi-objective collaborative optimization system according to claim 4, characterized in that, The reward value is determined based on the final hidden state, the third weight matrix, and the third bias vector.

6. The green infrastructure multi-objective collaborative optimization system according to claim 1, characterized in that, The green infrastructure includes multiple facility families; the ecological service functions of the multiple facility families are different; and the upper limit area of ​​the deployment of the multiple facility families is based on different deployment conditions.

7. The green infrastructure multi-objective collaborative optimization system according to claim 6, characterized in that, The maximum deployment area of ​​the facility family is determined based on at least one of the following: the area of ​​the sub-catchment area, the distance from the river channel, the slope, and the road type.

8. The green infrastructure multi-objective collaborative optimization system according to claim 7, characterized in that, The multiple facility families include a first facility family, a second facility family, and a third facility family; The maximum deployment area of ​​the first facility family is determined based on the area of ​​the sub-catchment area and the distance correction factor corresponding to the distance from the river channel; The maximum deployment area of ​​the second facility family is determined based on the area of ​​the sub-catchment area, the correction parameter corresponding to the distance from the river channel, and the first slope correction factor; The maximum deployment area of ​​the third facility family is determined based on the road type correction factor and the second slope correction factor.

9. The green infrastructure multi-objective collaborative optimization system according to claim 8, characterized in that, The distance correction factor includes multiple distance correction factors; the values ​​of the multiple distance correction factors are different, and the multiple distance correction factors correspond one-to-one with multiple distances from the river channel; The first slope correction factor includes multiple first slope correction factors; The values ​​of the plurality of first slope correction factors are different, and the plurality of first slope correction factors correspond one-to-one with the plurality of first slopes; The road type correction factor includes multiple road type correction factors; the values ​​of the multiple road type correction factors are different, and the multiple road type correction factors correspond one-to-one with multiple road types; The second slope correction factor includes multiple second slope correction factors; The values ​​of the plurality of second slope correction factors are different, and the plurality of second slope correction factors correspond one-to-one with the plurality of second slopes.

10. The green infrastructure multi-objective collaborative optimization system according to claim 8, characterized in that, The average effective water depth of the first facility family is determined based on the effective reservoir capacity and the deployable area of ​​the first facility family; The second facility family has multiple outflow configuration types; The various facility families each have a corresponding mapping matrix; the mapping matrix is ​​used to characterize the land use category and changes in key land parameters within the facility family; the key land parameters include surface layer parameters, soil layer parameters, pollutant scour parameters, and other parameters.

Images