Intelligent settlement water adaptability evaluation method adopting artificial intelligence machine learning

Abstract

The invention relates to the technical field of artificial intelligence machine learning, and discloses a settlement water adaptability intelligent evaluation method and system adopting artificial intelligence machine learning. The method comprises the following steps: constructing a time sequence dynamic heterogeneous graph sequence representing a settlement system; performing space-time alignment and attribute occurrence on the multi-source heterogeneous data; performing spatial neighborhood aggregation and time state evolution by using the space-time diagram neural network model, and extracting high-dimensional space-time features; a multi-dimensional water adaptability evaluation index vector is generated through an evaluation decoder, and water adaptability trend prediction in a future scene is supported. The system comprises a dynamic heterogeneous graph construction module, a multi-source data processing module, a spatial-temporal feature coding module, a water adaptability evaluation decoding module and an evaluation and simulation application module. According to the method, dynamic, fine and multi-dimensional evaluation and prospective prediction of the settlement water adaptability are realized, and the model precision and the decision support capability are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence and machine learning, and specifically relates to an intelligent evaluation method for settlement water suitability using artificial intelligence and machine learning. Background Technology

[0002] With the intensification of global climate change and the acceleration of urbanization, the interaction between settlements and the water environment is becoming increasingly complex. The scientific, dynamic, and precise evaluation of settlement water suitability has become a core issue in urban and rural planning, disaster prevention, and sustainable development. Water suitability not only involves the efficiency of water resource utilization in settlements but also encompasses their resilience and adaptability in the face of extreme hydrological events. Traditional evaluation methods often rely on expert experience or static indicator systems, making it difficult to characterize the coupling mechanism between the internal spatial structure of a settlement system and external hydrological processes, and also failing to respond to the dynamic evolution characteristics at multiple scales and dimensions.

[0003] Artificial intelligence and machine learning technologies have been introduced into settlement water suitability research in recent years, attempting to improve the objectivity and predictive ability of assessments through data-driven approaches. Typical methods include convolutional neural network classification based on remote sensing imagery and recurrent neural network modeling based on time series data.

[0004] However, these models generally treat settlements as regular grids or independent sample points, neglecting their inherent spatial topological relationships—such as the connectivity between building clusters, the network structure formed by roads and rivers, and the crucial role of water flow paths in risk propagation. This neglect of spatial heterogeneity and structural information makes the models vulnerable when migrating across regions or dealing with unseen scenarios.

[0005] Most existing machine learning models are statically trained and lack the ability to learn from changes themselves. When faced with new climate patterns, sudden water events, or untrained settlement patterns, the models need to collect a large amount of labeled data and train from scratch, which is time-consuming, labor-intensive, and difficult to meet the timeliness requirements of emergency decision-making.

[0006] Furthermore, existing methods fail to effectively integrate spatiotemporal evolution patterns with prior knowledge of graph structures, and cannot quickly adjust parameters to adapt to dynamic environments under limited new data conditions. Therefore, there is an urgent need for an intelligent evaluation method that can simultaneously model the spatial topological dependencies of settlements and possess rapid adaptability, in order to achieve a high-precision, highly generalizable, and transferable dynamic assessment of settlement water suitability. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide an intelligent evaluation method for settlement water suitability using artificial intelligence and machine learning, so as to overcome the shortcomings of the existing evaluation model, which is static and cannot effectively represent the spatial topology of settlement and its dynamic evolution law, resulting in insufficient generalization ability in future scenario prediction and response to sudden water events.

[0008] To address the aforementioned technical problems, this invention provides an intelligent evaluation method for settlement water suitability using artificial intelligence machine learning. This method constructs a dynamic heterogeneous graph of settlements, mapping the multi-source heterogeneous elements within a settlement, such as hydrology, geography, and socio-economic factors, and their interactions, into nodes and edges of the graph.

[0009] Then, a spatiotemporal graph neural network model is adopted to aggregate the neighborhood information of nodes in the spatial dimension to capture the influence of topological structure, and to learn the evolution law of graph structure and node attributes in the temporal dimension. Finally, by decoding the learned high-dimensional spatiotemporal features, a multi-dimensional, dynamic and refined evaluation and prediction result of the water suitability of settlements is generated.

[0010] According to an aspect of the present invention, a method for intelligent evaluation of settlement water suitability using artificial intelligence machine learning is provided, comprising the following steps:

[0011] A time-series dynamic heterogeneous graph sequence representing a settlement system is constructed. The dynamic heterogeneous graph sequence consists of a series of static heterogeneous graphs collected at a preset time step. Each static heterogeneous graph includes a set of nodes and a set of edges.

[0012] Acquire and process multi-source heterogeneous data associated with the settlement system, and perform spatiotemporal alignment and attribute assignment between the multi-source heterogeneous data and the nodes and edges in the dynamic heterogeneous graph sequence to generate a time-series dynamic heterogeneous graph sequence with attributes.

[0013] The attributed temporal dynamic heterogeneous graph sequence is input into a pre-trained spatiotemporal graph neural network model to perform deep representation learning of spatiotemporal features. The spatiotemporal graph neural network model sequentially performs spatial feature aggregation and temporal feature update on the heterogeneous graph at each time step, thereby outputting a high-dimensional temporal feature embedding vector for each node at each time step.

[0014] The high-dimensional temporal feature embedding vector is input into the evaluation decoder, which aggregates and performs nonlinear transformation on the feature embedding vector to generate a multi-dimensional water suitability evaluation index vector of the settlement system at the corresponding time step.

[0015] Based on the multi-dimensional water suitability evaluation index vector, the water suitability status of the settlement is assessed, and combined with preset future scenario parameters, the future water suitability evolution trend of the settlement is simulated and predicted.

[0016] As one embodiment of the present invention, the construction of the time-series dynamic heterogeneity graph sequence characterizing the settlement system specifically includes:

[0017] The spatial region of the settlement is divided into basic geographic cell networks, and each geographic cell network is defined as a basic node.

[0018] Based on functional attributes, key facilities within the settlement are identified as specific types of functional nodes, including water source nodes, water treatment plant nodes, pumping station nodes, flood storage area nodes, and important infrastructure nodes.

[0019] Define the connection relationships between nodes to form an edge set, which includes: geographical adjacency edges determined based on geographical adjacency; hydrological connectivity edges determined based on the physical connectivity of the water system network; pipeline connectivity edges determined based on the physical connectivity of the municipal pipeline network; and transportation connectivity edges determined based on the physical connectivity of the transportation road network.

[0020] As one embodiment of the present invention, the step of acquiring and processing multi-source heterogeneous data associated with the settlement system and assigning attributes to the data specifically includes:

[0021] Data on land use type, vegetation cover, and impervious surface ratio are obtained through remote sensing image interpretation and stored in the static attributes of basic nodes;

[0022] The elevation and slope information of each basic node is extracted from the digital elevation model data and stored in the static attributes of the basic node.

[0023] Data on water system distribution, pipeline layout, and transportation network are obtained through a geographic information system and stored in the static attributes of the corresponding hydrological connectivity edges, pipeline connectivity edges, and transportation connectivity edges. The static attributes include river width, pipeline diameter, and road grade.

[0024] Rainfall, temperature, and evaporation data are acquired in real time through a network of meteorological monitoring stations and stored in the dynamic attributes of basic nodes;

[0025] The data on river water level, flow rate, and water quality parameters are acquired in real time through the hydrological monitoring station network and stored in the dynamic attributes of the hydrological connectivity edges.

[0026] Population density and gross economic output data are obtained through socio-economic statistics and then stored in the dynamic attributes of basic nodes.

[0027] All acquired data are processed by timestamp alignment, spatial coordinate registration, data normalization, and missing value imputation to form a unified spatiotemporal attribute matrix, which corresponds to the node attribute matrix and edge attribute matrix in the dynamic heterogeneous graph sequence, respectively.

[0028] In one embodiment of the present invention, the spatiotemporal graph neural network model includes a spatial feature aggregation module and a temporal feature evolution module, and the deep representation learning of the spatiotemporal features specifically includes:

[0029] At each time step, for each target node in the graph, the spatial feature aggregation module is activated first.

[0030] The spatial feature aggregation module adopts a graph attention network mechanism to adaptively calculate attention weight coefficients for different types of neighboring nodes of the target node. The attention weight coefficients characterize the importance of different neighboring nodes to the state of the target node.

[0031] Based on the attention weight coefficients, the feature vectors of all neighboring nodes are weighted and summed to generate an intermediate feature vector that aggregates the spatial topology information at the current time.

[0032] The intermediate feature vector and the hidden state vector of the target node at the previous time step are input together into the time feature evolution module;

[0033] The time feature evolution module adopts a gated loop unit structure, which controls the degree of retention of historical information and the degree of reception of current information through its internal update gate and reset gate, and updates the hidden state vector of the target node.

[0034] Spatial feature aggregation and temporal feature update are sequentially performed at each time step in the dynamic heterogeneous graph sequence until the entire sequence is processed, ultimately obtaining a high-dimensional temporal feature embedding vector for each node at all time steps.

[0035] As one embodiment of the present invention, the evaluation decoder performs aggregation and nonlinear transformation on the feature embedding vector, specifically including:

[0036] The feature embedding vectors of all nodes within the same time step are aggregated using a readout function to generate a graph-level global feature vector. The readout function is to perform mean pooling or max pooling operations on all node vectors.

[0037] The graph-level global feature vector is input into a multilayer perceptron network, which is composed of multiple fully connected layers and nonlinear activation function layers stacked alternately.

[0038] The output layer of the multilayer perceptron network is the multidimensional water suitability evaluation index vector. Each dimension of the vector corresponds to a specific water suitability sub-index, which includes the flood control and drainage capacity index, the water resource security capacity index, the water environment health index, and the water ecosystem service value index.

[0039] As one embodiment of the present invention, the simulation prediction based on the multi-dimensional water suitability evaluation index vector specifically includes:

[0040] Construct a set of parameters for future climate scenarios, which includes design rainfall pattern data for different future return periods;

[0041] Construct a set of urban development scenario parameters, which includes graph node and graph edge information of future planned new impervious area, population growth distribution, and newly built water conservancy facilities;

[0042] The set of future climate scenario parameters and the set of urban development scenario parameters are used as inputs to the spatiotemporal graph neural network model. Through forward propagation of the model, a multi-dimensional water suitability evaluation index vector sequence under the future scenario is generated, thereby realizing quantitative prediction of the evolution trend of settlement water suitability.

[0043] According to another aspect of the present invention, a smart evaluation system for settlement water suitability employing artificial intelligence machine learning is provided, comprising:

[0044] The dynamic heterogeneous graph construction module is used to abstract the physical space and functional elements of a settlement system and construct a temporal dynamic heterogeneous graph sequence composed of nodes and edges to represent the spatiotemporal topological relationships of various elements within the settlement.

[0045] The multi-source data processing module is used to acquire data from multiple sources such as remote sensing, meteorology, hydrology, geographic information and socio-economic data, and perform spatiotemporal alignment, data cleaning and attribute assignment, and assign static and dynamic attributes to the nodes and edges of the dynamic heterogeneous graph.

[0046] The spatiotemporal feature encoding module integrates a spatiotemporal graph neural network model to receive dynamic heterogeneous graph sequences with attributes and perform spatial dimension neighborhood information aggregation and temporal dimension state evolution learning step by step to extract high-dimensional spatiotemporal dynamic features of the settlement system.

[0047] The water suitability evaluation decoding module is used to receive the high-dimensional features output by the spatiotemporal feature encoding module, and decode them into a multi-dimensional water suitability evaluation index vector that characterizes the overall state of the settlement through global information aggregation and nonlinear mapping.

[0048] The evaluation and simulation application module is used to quantitatively evaluate the current state based on the evaluation index vector, and supports input of future scenario parameters to drive the entire model to perform forward extrapolation in order to predict the future evolution trajectory of settlement water suitability under different development and climate scenarios.

[0049] As one embodiment of the present invention, the dynamic heterogeneous graph construction module specifically includes a spatial grid division unit, a functional node identification unit, and a connection relationship generation unit; the spatial grid division unit is used to discretize settlement areas into basic geographic unit nodes; the functional node identification unit is used to extract key facilities as functional nodes from geographic information data according to preset rules; the connection relationship generation unit is used to generate various types of edges connecting different nodes based on geographic adjacency, water system connectivity, pipeline layout, and transportation network data.

[0050] As one embodiment of the present invention, the spatiotemporal graph neural network model inside the spatiotemporal feature encoding module is composed of a graph attention network layer and a gated recurrent unit layer stacked together. The graph attention network layer is responsible for weighted aggregation of spatial heterogeneous information by calculating the attention coefficients between nodes within a single time step. The gated recurrent unit layer is responsible for performing temporal modeling on the aggregated features output by the graph attention network layer along the time axis to capture the long-term dependencies and evolutionary laws of the system state.

[0051] As one embodiment of the present invention, the water suitability evaluation decoding module specifically includes a global pooling unit and a multi-layer feedforward network unit; the global pooling unit is used to integrate the feature vectors of all nodes in the graph into a single graph-level representation vector; the multi-layer feedforward network unit is used to perform multi-layer nonlinear transformation on the graph-level representation vector and finally output a water suitability evaluation vector with fixed dimensions, each component of which corresponds to a specific water suitability evaluation sub-index.

[0052] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0053] I. This invention constructs a dynamic heterogeneous graph, which provides a unified mathematical representation of the complex spatial topology and multi-source heterogeneous information of settlements. This breaks through the limitation of traditional static models that cannot describe the network correlation between elements, and realizes the structured and dynamic modeling of settlement systems.

[0054] Second, this invention employs a spatiotemporal graph neural network model, which can simultaneously capture the deep-seated patterns of settlement water suitability evolution from both spatial and temporal dimensions. It effectively learns complex spatiotemporal coupling effects such as the propagation process of rainfall events in water systems and the impact of urban expansion on surface runoff, significantly improving the accuracy and generalization ability of the evaluation model.

[0055] Third, the multi-dimensional water suitability evaluation index system provided by this invention, compared with a single comprehensive score, can more comprehensively and meticulously reveal the specific advantages and disadvantages of settlement water suitability, and provide more targeted decision support for urban planning and water resource management.

[0056] Fourth, this invention has the ability to simulate and predict future scenarios. By adjusting the input parameters, it can quantitatively assess the long-term impact of different climate change scenarios or urban development planning schemes on the water suitability of settlements, providing a scientific tool for formulating forward-looking and adaptive urban development strategies. Attached Figure Description

[0057] Figure 1 This is a schematic diagram of the overall technical solution architecture of the intelligent evaluation method for settlement water suitability using artificial intelligence machine learning proposed in this invention.

[0058] Figure 2 This is a schematic diagram of the core principle framework of the spatiotemporal graph neural network model in this invention;

[0059] Figure 3 This is a logical flowchart of the construction of dynamic heterogeneous graphs and the assignment of attributes of multi-source heterogeneous data in this invention.

[0060] Figure 4 This is a logical framework diagram of the sequential processing of spatial feature aggregation and temporal feature evolution in this invention;

[0061] Figure 5 This is a flowchart illustrating the logical process of decoding water suitability evaluation and generating multi-dimensional indicators in this invention.

[0062] Figure 6 This is a schematic diagram of the interaction relationship and data flow between the settlement system and the evaluation and simulation prediction driven by future scenario parameters in this invention. Detailed Implementation

[0063] Please refer to Figures 1 to 6 This invention provides an intelligent evaluation method for settlement water suitability using artificial intelligence and machine learning. Its core lies in constructing a temporal dynamic heterogeneous graph sequence representing the settlement system, fusing multi-source heterogeneous data, and utilizing a spatiotemporal graph neural network model to deeply model the spatiotemporal evolution of settlement water suitability. Ultimately, this achieves a refined assessment of the current state and quantitative prediction of future scenarios. The following will describe each step in detail according to the steps S1 to S5 explicitly specified in the invention.

[0064] The method includes the following steps:

[0065] S1, Construct a sequence of temporal dynamic heterogeneous graphs representing the settlement system;

[0066] S2, acquire and process multi-source heterogeneous data associated with the settlement system, and perform spatiotemporal alignment and attribute assignment between the multi-source heterogeneous data and the nodes and edges in the dynamic heterogeneous graph sequence to generate a time-series dynamic heterogeneous graph sequence with attributes.

[0067] S3, the time-series dynamic heterogeneous graph sequence with attributes is input into a pre-trained spatiotemporal graph neural network model to perform deep representation learning of spatiotemporal features.

[0068] S4, the high-dimensional temporal feature embedding vector is input into the evaluation decoder to generate a multi-dimensional water suitability evaluation index vector of the settlement system at the corresponding time step;

[0069] S5. Based on the multi-dimensional water suitability evaluation index vector, the water suitability status of the settlement is assessed, and combined with the preset future scenario parameters, the future water suitability evolution trend of the settlement is simulated and predicted.

[0070] In step S1, a time-series dynamic heterogeneous graph sequence representing the settlement system is constructed. This sequence consists of a series of static heterogeneous graphs collected at a preset time step. Each static heterogeneous graph includes a set of nodes and a set of edges.

[0071] Specifically, the target settlement spatial area is first divided into basic geographic cell grids. The grid resolution is determined according to the settlement scale, typically 100×100 meters or 500×500 meters. Each geographic cell grid is defined as a basic node, used to carry comprehensive attribute information of that spatial location.

[0072] Secondly, based on functional attributes, key facilities within the settlement are identified from the geographic information system database and abstracted into specific types of functional nodes. These functional nodes include water source nodes, water treatment plant nodes, pumping station nodes, flood storage area nodes, and important infrastructure nodes. These nodes have semantic type identifiers that are different from basic nodes and will be assigned different message passing rules in subsequent graph neural network processing.

[0073] Next, define the connections between nodes to form an edge set, which includes four types:

[0074] The first type is the geographical adjacency edge, which is established by determining whether two basic nodes share a boundary or are within a preset neighborhood radius. It is used to characterize the spatial continuity of surface runoff diffusion and pollutant migration.

[0075] The second type is the hydrological connectivity edge, which is established based on the flow direction and connectivity of natural or artificial water systems such as rivers and ditches, connecting upstream and downstream nodes, and is used to depict the water flow path and flood propagation process;

[0076] The third type is the pipeline connection edge, which is established based on the physical topology of the municipal water supply and drainage pipeline network, connecting pumping stations, water plants and user nodes, and is used to simulate water supply security and sewage discharge capacity.

[0077] The fourth category is traffic connectivity edges, which are established based on the road network's grade and traffic capacity, connecting adjacent basic or functional nodes to reflect the accessibility of emergency evacuation and material allocation.

[0078] All edges are labeled with directionality and type, collectively forming a heterogeneous graph structure that supports multi-relational reasoning. The time step of this dynamic heterogeneous graph sequence is set according to the data update frequency, typically daily, weekly, or monthly, ensuring that it can capture the dual dynamic characteristics of the settlement system under both short-term extreme events and long-term development trends.

[0079] In step S2, multi-source heterogeneous data associated with the settlement system is acquired and processed, and the data is spatiotemporally aligned and attribute-assigned to generate a time-series dynamic heterogeneous graph sequence with attributes.

[0080] In practice, land use type, vegetation cover, and impervious surface ratio data are first obtained through remote sensing image interpretation. This data is then processed by supervised classification algorithms such as support vector machines or random forests, mapped to corresponding base nodes according to their spatial location, and stored as static attributes in the node attribute matrix. Secondly, the elevation and maximum slope values ​​of the center point of each base node are extracted from digital elevation model data. Elevation is used to calculate runoff time and water accumulation risk, while slope is used to assess surface runoff velocity; both are also assigned as static attributes to the base nodes.

[0081] Vector data of water system distribution, pipeline layout, and transportation network are obtained through a geographic information system. Structural parameters such as river width, pipeline diameter, and road grade are extracted from these data and stored in the static attribute fields of hydrological connectivity edges, pipeline connectivity edges, and transportation connectivity edges, respectively.

[0082] Rainfall, temperature, and evaporation data are acquired in real time through a network of meteorological monitoring stations. These data are recorded by timestamp and distributed to each basic node using inverse distance weighted interpolation as their dynamic attributes. When there are no direct observations within a certain time step, missing values ​​are filled using Kriging interpolation or the historical average for the same period.

[0083] The river water level, flow rate, and water quality parameters are acquired in real time through the hydrological monitoring station network. These data are associated with the corresponding hydrological connectivity edges according to the location of the monitoring sections, and serve as the dynamic attributes of the edges at the corresponding time steps.

[0084] Population density and gross economic output data are obtained through socioeconomic statistics. These annual or quarterly statistics are then converted into gridded data and stored in the dynamic attributes of the basic nodes through spatialization methods such as area weighting or nighttime light index correction.

[0085] All the above data must undergo a unified data preprocessing procedure before being fed into the model:

[0086] First, timestamp alignment is performed, resampling data from different sampling frequencies to a unified time step; second, spatial coordinate registration is performed to ensure all data are projected onto the same geographic coordinate system; third, data normalization is performed, using minimum-maximum scaling or... The standardization method maps each attribute value to 0 to 1 or a mean of 0. The interval;

[0087] Finally, missing value imputation is performed. For continuous variables, linear interpolation or cubic spline interpolation is used, and for categorical variables, mode imputation is used. After this processing, two structured spatiotemporal attribute matrices are formed: a node attribute matrix and a node attribute matrix. Where N is the total number of nodes. The attribute dimension for each node; the edge attribute matrix. Where M is the total number of edges. These represent the attribute dimensions of each edge. These two matrices relate to the topology of the dynamic heterogeneous graph. Together they constitute a complete sequence of temporally dynamic heterogeneous graphs with attributes. .

[0088] In step S3, the attributed temporal dynamic heterogeneous graph sequence is input into a pre-trained spatiotemporal graph neural network model for deep representation learning of spatiotemporal features. This model consists of a spatial feature aggregation module and a temporal feature evolution module sequentially, performing deep representation learning for each time step. Process sequentially. At each time step... For each target node in the graph The spatial feature aggregation module is started first. This module uses a graph attention network mechanism to target nodes. Different types of neighbor node sets

[0089] (k represents the neighbor type, such as geographical adjacency, hydrological connectivity, etc.), calculate the attention weight coefficients respectively. The specific calculation process is as follows: First, the target node... Current features with neighboring nodes Features splice and connect with the corresponding edges Attributes The features are fused to form a joint feature vector; then passed through a shared single-layer feedforward neural network. The joint features are subjected to a nonlinear transformation and the LeakyReLU activation function is applied. Finally, the attention weights are obtained by normalizing them over all similar neighbors using the softmax function. ;

[0090] in The feature transformation matrix is ​​a learnable feature. This represents a vector concatenation operation. Based on the calculated attention weight coefficients, the feature vectors of all neighboring nodes are weighted and summed to generate an intermediate feature vector that aggregates the spatial topology information at the current time. :

[0091] ;

[0092] in To ensure differentiated processing of information regarding heterogeneous relationships, specific transformation matrices are used for different types of neighbors. Subsequently, the intermediate feature vectors... With the target node The hidden state vector of the previous time step The data is input together into the time feature evolution module. This module adopts a gated loop unit structure, which contains an update gate. With Reset Door These controls the degree to which historical information is retained and the degree to which current information is received, respectively. The calculation formulas for the update gate and reset gate are as follows:

[0093] ;

[0094] ;

[0095] in It is the sigmoid activation function. and This represents the learnable parameter matrix. Candidate hidden states.

[0096] The aggregation is determined by the historical state modulated after the reset gate and the current aggregation features:

[0097] ;

[0098] Finally, the hidden state vector of the target node The update gate performs a weighted fusion of historical states and candidate states:

[0099] ;

[0100] Where ⊙ denotes element-wise multiplication. The spatial feature aggregation and temporal feature update operations are sequentially repeated for each time step in the dynamic heterogeneous graph sequence until the entire sequence has been processed for T time steps, ultimately yielding the result for each node. High-dimensional temporal feature embedding vector sequence at all time steps The embedding vector dimension is typically set to 128 or 256, which is sufficient to characterize the complex dynamic behavior of settlement systems under the multi-field coupling of electro-thermal-social fields.

[0101] In step S4, the high-dimensional temporal feature embedding vector is input into the evaluation decoder to generate a multi-dimensional water suitability evaluation index vector for the settlement system at the corresponding time step. The evaluation decoder first uses a readout function to evaluate the data at the same time step. Feature embedding vectors of all nodes Perform aggregation to generate a graph-level global feature vector. .

[0102] The readout function employs mean pooling, i.e. This operation is permutation invariant, ensuring that the output is independent of the node order. Subsequently, the global feature vector at the graph level... The input is fed into a multilayer perceptron network, which consists of three fully connected layers stacked alternately with ReLU activation functions. The output dimensions of each layer are 512, 256, and 4, respectively.

[0103] The last layer has no activation function and directly outputs a four-dimensional vector. This vector is the multi-dimensional water suitability evaluation index vector.

[0104] in, The higher the value of the flood control and drainage capacity index, the lower the risk of flooding in the settlement during rainstorm events. The corresponding water resource security capacity index reflects the stability of water supply and the adequacy of emergency backup water sources; The corresponding water environment health index represents the balance between water pollution load and self-purification capacity. The corresponding aquatic ecosystem service value index quantifies the regulatory and cultural service functions provided by ecological elements such as wetlands and riparian zones. Each dimension of this evaluation index vector has been trained under supervised training on independently labeled datasets to ensure that its physical meaning is clear and highly interpretable.

[0105] In step S5, based on the multi-dimensional water suitability evaluation index vector, the water suitability status of the settlement is assessed, and combined with preset future scenario parameters, the future water suitability evolution trend of the settlement is simulated and predicted. During the status assessment phase, the current time step is... Evaluation index vector The water suitability level is determined by comparing the results with the preset grading thresholds. For example, each sub-indicator is divided into four levels: excellent, good, medium, and poor. A visual heat map is generated to show the areas with shortcomings in the water quality.

[0106] In the simulation and prediction phase, the first step is to construct a set of parameters for future climate scenarios. This set of parameters includes data on design rainstorm patterns under different return periods, such as 24-hour rainfall process lines for 10-year, 50-year, and 100-year return periods. These data are generated using hydrological frequency analysis methods and converted into dynamic attributes that are stored in the basic nodes.

[0107] Secondly, a set of urban development scenario parameters is constructed, which includes a distribution map of newly added impervious areas in future plans, a spatial distribution map of population growth, and information on the location and type of newly built water conservancy facilities. The newly added impervious surface and population data are used to update the static and dynamic attributes of the basic nodes, while the newly built water conservancy facilities are added to the dynamic heterogeneous graph structure as new functional nodes and connecting edges.

[0108] Using the aforementioned future scenario parameters as initial conditions, corresponding parts of the original data are replaced to form a modified temporal dynamic heterogeneous graph sequence. Subsequently, this sequence is input into a pre-trained spatiotemporal graph neural network model and an evaluation decoder. Through model forward propagation, a multi-dimensional water suitability evaluation index vector sequence under the future scenario is generated. ,in For the predicted duration, typical values ​​are 5 years, 10 years, or 20 years.

[0109] This sequence clearly demonstrates the evolution trajectory of various sub-indicators of settlement water suitability under specific climate and urban development patterns, providing decision-makers with quantitative evidence to assess the long-term impact of different planning schemes and optimize adaptation strategies.

[0110] The intelligent evaluation system for settlement water suitability using artificial intelligence and machine learning, as another aspect of this invention, includes a dynamic heterogeneous graph construction module, a multi-source data processing module, a spatiotemporal feature encoding module, a water suitability evaluation decoding module, and an evaluation and simulation application module. The dynamic heterogeneous graph construction module is responsible for abstracting the physical space of the settlement into a mathematical graph structure. Internally, it includes spatial grid partitioning units, functional node identification units, and connection relationship generation units, which respectively perform tasks such as basic node generation, key facility extraction, and multi-type edge construction.

[0111] The multi-source data processing module integrates a remote sensing interpretation engine, a geographic information processing unit, a meteorological and hydrological data interface, and a socio-economic data spatialization tool to complete the entire process of data acquisition, cleaning, alignment, and attribute assignment. The spatiotemporal feature encoding module incorporates the aforementioned spatiotemporal graph neural network model and achieves end-to-end learning of high-dimensional spatiotemporal dynamic features through the collaborative work of a graph attention mechanism and a gated recurrent unit.

[0112] The water suitability assessment decoding module consists of a global pooling unit and a multi-layer feedforward network unit. The former compresses node-level features into a graph-level representation, while the latter maps this representation to a multi-dimensional evaluation index space. The assessment and simulation application module provides a human-computer interaction interface, supporting current state visualization, historical trend backtracking, and future scenario projection. Internally, it includes a scenario parameter configurator and a result analyzer, ensuring that the model output can directly serve urban planning and emergency management practices. The entire system is deployed on a high-performance computing server cluster, supporting parallel processing of large-scale graph data and meeting the real-time evaluation needs of city-level settlements.

Claims

1. A smart evaluation method for settlement water suitability using artificial intelligence and machine learning, characterized in that: include: A time-series dynamic heterogeneous graph sequence representing a settlement system is constructed. The dynamic heterogeneous graph sequence consists of a series of static heterogeneous graphs collected at a preset time step. Each static heterogeneous graph includes a set of nodes and a set of edges. Acquire and process multi-source heterogeneous data associated with the settlement system, and perform spatiotemporal alignment and attribute assignment between the multi-source heterogeneous data and the nodes and edges in the dynamic heterogeneous graph sequence to generate a time-series dynamic heterogeneous graph sequence with attributes. The sequence of attributed temporal dynamic heterogeneous graphs is input into a pre-trained spatiotemporal graph neural network model to perform deep representation learning of spatiotemporal features. The spatiotemporal graph neural network model sequentially performs spatial feature aggregation and temporal feature update on the heterogeneous graph at each time step, thereby outputting a high-dimensional temporal feature embedding vector for each node at each time step. The high-dimensional temporal feature embedding vector is input into the evaluation decoder, which aggregates and performs nonlinear transformation on the feature embedding vector to generate a multi-dimensional water suitability evaluation index vector of the settlement system at the corresponding time step. Based on the multi-dimensional water suitability evaluation index vector, the water suitability status of the settlement is assessed, and combined with preset future scenario parameters, the future water suitability evolution trend of the settlement is simulated and predicted.

2. The intelligent evaluation method for settlement water suitability using artificial intelligence machine learning as described in claim 1, characterized in that, Constructing a time-series dynamic heterogeneity graph sequence characterizing the settlement system, including: The spatial region of the settlement is divided into basic geographic cell networks, and each geographic cell network is defined as a basic node. Based on functional attributes, key facilities within the settlement are identified as specific types of functional nodes, including water source nodes, water treatment plant nodes, pumping station nodes, flood storage area nodes, and important infrastructure nodes. Define the connection relationships between nodes to form an edge set, the edge set including: Geographical adjacency edges determined based on geographic adjacency relationships; Hydrological connectivity edges determined based on the physical connection relationships of the water system network; Pipeline connectivity edges determined based on the physical connections of municipal pipeline networks; and traffic connectivity edges determined based on the physical connections of traffic road networks.

3. The intelligent evaluation method for settlement water suitability using artificial intelligence machine learning according to claim 2, characterized in that, Acquire and process multi-source heterogeneous data associated with the settlement system, and assign attributes to the data, including: Data on land use type, vegetation cover, and impervious surface ratio are obtained through remote sensing image interpretation and stored in the static attributes of basic nodes; The elevation and slope information of each basic node is extracted from the digital elevation model data and stored in the static attributes of the basic node. Data on water system distribution, pipeline layout, and transportation network are obtained through a geographic information system and stored in the static attributes of the corresponding hydrological connectivity edges, pipeline connectivity edges, and transportation connectivity edges. The static attributes include river width, pipeline diameter, and road grade. Rainfall, temperature, and evaporation data are acquired in real time through a network of meteorological monitoring stations and stored in the dynamic attributes of basic nodes; The data on river water level, flow rate, and water quality parameters are acquired in real time through the hydrological monitoring station network and stored in the dynamic attributes of the hydrological connectivity edges. Population density and gross economic output data are obtained through socio-economic statistics and then stored in the dynamic attributes of basic nodes. All acquired data are processed by timestamp alignment, spatial coordinate registration, data normalization, and missing value imputation to form a unified spatiotemporal attribute matrix, which corresponds to the node attribute matrix and edge attribute matrix in the dynamic heterogeneous graph sequence, respectively.

4. The intelligent evaluation method for settlement water suitability using artificial intelligence machine learning according to claim 3, characterized in that, The attributed temporal dynamic heterogeneous graph sequence is input into a pre-trained spatiotemporal graph neural network model to perform deep representation learning of spatiotemporal features, including: At each time step, for each target node in the graph, the spatial feature aggregation module is activated; The spatial feature aggregation module adopts a graph attention network mechanism to adaptively calculate attention weight coefficients for different types of neighboring nodes of the target node. The attention weight coefficients characterize the importance of different neighboring nodes to the state of the target node. Based on the attention weight coefficients, the feature vectors of all neighboring nodes are weighted and summed to generate an intermediate feature vector that aggregates the spatial topology information at the current time. The intermediate feature vector and the hidden state vector of the target node at the previous time step are input together into the time feature evolution module; The time feature evolution module adopts a gated loop unit structure, which controls the degree of retention of historical information and the degree of reception of current information through its internal update gate and reset gate, and updates the hidden state vector of the target node. Spatial feature aggregation and temporal feature update are sequentially performed at each time step in the dynamic heterogeneous graph sequence until the entire sequence is processed, ultimately obtaining a high-dimensional temporal feature embedding vector for each node at all time steps.

5. The intelligent evaluation method for settlement water suitability using artificial intelligence machine learning according to claim 4, characterized in that, The spatial feature aggregation module employs a graph attention network mechanism, including: The current features of the target node are concatenated with the features of its neighboring nodes, and then fused with the attributes of the corresponding edges to form a joint feature vector. The joint feature is nonlinearly transformed using a shared single-layer feedforward neural network, and the LeakyReLU activation function is applied. The attention weight coefficients are obtained by normalizing over all similar neighbors using the softmax function.

6. The intelligent evaluation method for settlement water suitability using artificial intelligence machine learning according to claim 5, characterized in that, The time feature evolution module adopts a gated loop unit structure, including: The update gate and reset gate are calculated to control the degree to which historical information is retained and the degree to which current information is received, respectively. Based on the historical state modulated by the reset gate and the current aggregated features, the candidate hidden state is calculated; The hidden state vector of the target node is obtained by weighted fusion of the historical state and the candidate state through the update gate.

7. The intelligent evaluation method for settlement water suitability using artificial intelligence machine learning according to claim 6, characterized in that, The high-dimensional temporal features are embedded into a vector and input into the evaluation decoder to generate a multi-dimensional water suitability evaluation index vector for the settlement system at the corresponding time step, including: The feature embedding vectors of all nodes within the same time step are aggregated using a readout function to generate a graph-level global feature vector. The readout function is to perform mean pooling or max pooling operations on all node vectors. The graph-level global feature vector is input into a multilayer perceptron network, which is composed of multiple fully connected layers and nonlinear activation function layers stacked alternately. The output layer of the multilayer perceptron network is the multidimensional water suitability evaluation index vector. Each dimension of the vector corresponds to a specific water suitability sub-index, which includes the flood control and drainage capacity index, the water resource security capacity index, the water environment health index, and the water ecosystem service value index.

8. The intelligent evaluation method for settlement water suitability using artificial intelligence machine learning according to claim 7, characterized in that, Simulation and prediction based on the aforementioned multi-dimensional water suitability evaluation index vector includes: Construct a set of parameters for future climate scenarios, which includes design rainfall pattern data for different future return periods; Construct a set of urban development scenario parameters, which includes graph node and graph edge information of future planned new impervious area, population growth distribution, and newly built water conservancy facilities; The set of future climate scenario parameters and the set of urban development scenario parameters are used as inputs to the spatiotemporal graph neural network model. Through forward propagation of the model, a multi-dimensional water suitability evaluation index vector sequence under the future scenario is generated, thereby realizing quantitative prediction of the evolution trend of settlement water suitability.

9. The intelligent evaluation method for settlement water suitability using artificial intelligence machine learning according to claim 8, characterized in that, The design rainstorm pattern data in the future climate scenario parameter set is generated using hydrological frequency analysis and converted into dynamic attributes, which are then stored in the basic nodes.

10. The intelligent evaluation method for settlement water suitability using artificial intelligence machine learning according to claim 9, characterized in that, The newly added impermeable surface and population data in the urban development scenario parameter set are used to update the static and dynamic attributes of the basic nodes, while newly built water conservancy facilities are added to the dynamic heterogeneous graph structure as new functional nodes and connecting edges.

Images