A water resources climate impact assessment system and method

Abstract

This invention relates to a water resources climate impact assessment system and method, belonging to the field of electronic digital data processing technology. It constructs a physically-enhanced spatiotemporal graph neural network model, which learns spatial correlations and temporal dynamics through graph convolutional layers and long short-term memory (LSTM) network layers, respectively. A multi-factor collaborative attention mechanism is employed to analyze the complex interactions between different variables. Specifically, a differentiable physical constraint layer embeds domain physical equations, such as the water balance equation, as regularization terms into the network training process to ensure that the model output conforms to physical laws. This invention also incorporates incremental computation technology to achieve real-time interactive simulation and scenario extrapolation of large-scale graph structure systems. This invention overcomes the shortcomings of traditional physical models' insufficient adaptability and the physical inconsistencies of purely data-driven models, providing a general spatiotemporal prediction framework that balances physical consistency and high data-driven accuracy. It can be widely applied in fields such as water resources management, weather forecasting, and environmental monitoring.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, specifically to a water resources climate impact assessment system and method. Background Technology

[0002] As climate change increasingly impacts water resource systems, accurately assessing water resource responses to climate stress has become a core requirement for water resource management and adaptation planning. Existing technologies have made various attempts in this area, but significant limitations remain.

[0003] Chinese invention patent CN119830791B (A Precise Assessment System and Method for Water Resources in Inland River Basins) proposes to achieve accurate calculation of water resources by deploying a multimodal sensor network and establishing a implicit dynamic model of water exchange. Although this method has high accuracy in total water resource accounting, it still relies on traditional hydrodynamic mechanism models, making it difficult to effectively integrate multi-source heterogeneous data. Furthermore, it lacks the ability to deeply explore the complex nonlinear relationship between climate factors and water resources, and the assessment process exhibits static characteristics, failing to support dynamic scenario simulation and real-time interactive analysis.

[0004] Chinese invention patent CN119720590B (A Method and System for Optimizing Water Resource Allocation) predicts runoff sequences based on the SWAT model and establishes an optimization allocation model with multiple constraints. While this scheme performs well in water resource allocation under given conditions, its overall architecture relies on the traditional paradigm of "physical model prediction + mathematical programming optimization." This is limited by the physical model's ability to simulate complex climate scenarios and cannot achieve rapid assessment and adjustment based on real-time data, making it difficult to meet the dynamic adaptive management needs of water resource systems under the background of climate change.

[0005] Chinese invention patent CN111797493B (Method and System for Assessing the Water Resource Effects of Forest Planting with Biophysical Significance) assesses the hydrological effects of vegetation change by constructing a three-layer forest surface energy-land water balance model. Although this technology considers biophysical processes, its model structure is fixed and highly specialized, making it difficult to transfer to broader water resource and climate impact assessment scenarios. Furthermore, it lacks real-time simulation capabilities that allow for user interaction.

[0006] In summary, existing technologies generally suffer from three common shortcomings: first, the models lack flexibility, making it difficult to adapt to changing climate environments and complex data characteristics; second, the assessment process is static, lacking the ability to dynamically respond to uncertain scenarios; and third, the systems are highly closed, failing to provide real-time interactive and visual analysis functions for decision-making. Therefore, there is an urgent need for a new method and system for assessing the impact of water resources on climate that can deeply integrate physical mechanisms and data intelligence, supporting dynamic simulation and real-time interaction. Summary of the Invention

[0007] To address the problems of existing technologies, this invention provides a water resources climate impact assessment system and method.

[0008] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution: In a first aspect, a water resources climate impact assessment system includes: a data acquisition module configured to acquire first-type spatiotemporal sequence data and second-type spatiotemporal sequence data from multiple heterogeneous data sources through an application programming interface;

[0009] The data fusion and preprocessing module, connected to the data acquisition module, is configured to perform quality checks and format standardization on the first and second types of data, and to perform scale alignment based on a unified spatiotemporal grid benchmark to generate a fused dataset.

[0010] The intelligent analysis engine module is connected to the data fusion and preprocessing module. The intelligent analysis engine module loads and runs a pre-trained physical information-enhanced spatiotemporal graph neural network model.

[0011] Among them, the spatiotemporal graph neural network model with physical information enhancement is calculated based on a graph structure, which uses geospatial entities as nodes and connections that reflect their spatial relationships as edges.

[0012] The physically-enhanced spatiotemporal graph neural network model includes graph convolutional layers, long short-term memory (LSTM) network layers, and attention mechanism layers;

[0013] The output of the physically augmented spatiotemporal graph neural network model is connected to a differentiable physical constraint layer. This layer is used to substitute the intermediate prediction results of the physically augmented spatiotemporal graph neural network model into the predefined domain physical constraint equations for calculation. The calculation results are then added as regularization terms to the loss function of the physically augmented spatiotemporal graph neural network model to ensure that the output of the physically augmented spatiotemporal graph neural network model conforms to physical laws.

[0014] The output of the physically-enhanced spatiotemporal graph neural network model is the predicted value of the target variable;

[0015] The interactive output module, connected to the intelligent analysis engine module, is configured to output a visual representation of the predicted value of the target variable and provides user interface controls to receive scenario adjustment parameters. By maintaining a persistent communication connection with the front-end interface, it triggers the intelligent analysis engine module to perform incremental calculations and dynamically updates the visual representation in real time.

[0016] Preferably, the domain physical constraint equation upon which the differentiable physical constraint layer is based is the watershed water balance equation, the expression of which is:

[0017]

[0018] in, For precipitation, Evaporation rate For runoff, This represents the change in water storage variables;

[0019] The differentiable physical constraint layer is configured to: calculate the difference between the two sides of the water balance equation, and then... Norm as a physical constraint loss term ,Right now:

[0020] .

[0021] Preferably, the attention mechanism layer is a multi-factor spatiotemporal collaborative attention layer, specifically configured to perform the following operations:

[0022] The cross-attention score is calculated between different variables in the first and second categories of data. The cross-attention score is calculated using a query-key-value mechanism, where the query vector... and key vector Feature representations from different variables;

[0023] Based on the upstream and downstream location relationships of geospatial entities within a watershed, the spatial attention score is modulated. Specifically, when calculating attention, downstream nodes are assigned a weight based on the flow path distance, derived from information from upstream nodes. Enhancer ,in , These are learnable parameters or preset constants;

[0024] The time attention score is modulated based on the season in which the time step occurs, specifically by setting different learnable biases for the wet season and the dry season. and .

[0025] Preferably, the physically-enhanced spatiotemporal graph neural network model is obtained through the following steps:

[0026] S1: Prepare the training dataset ,in As input features, This refers to the true water resource status data, which has been corrected using a traditional hydrological model.

[0027] S2: Construct a network architecture for implementing a water resources climate impact assessment system;

[0028] S3: Define the composite loss function Its expression is:

[0029]

[0030] in, Predictions from a spatiotemporal graph neural network model that enhances physical information with truth value Mean squared error loss between them The physical constraint loss is calculated from the differentiable physical constraint layer. The hyperparameter used to balance the two losses has a range of values. The optimal value is determined by grid search on the validation set;

[0031] S4: Minimize the composite loss function using the error backpropagation algorithm. All parameters in the spatiotemporal graph neural network model that enhances physical information Joint optimization was carried out.

[0032] Preferably, when performing scale alignment, the data fusion and preprocessing module is specifically used to: unify spatial resolution through an inverse distance weighted spatial interpolation algorithm and unify temporal frequency through a linear resampling method;

[0033] The data fusion and preprocessing module also includes an intelligent repair unit, which is configured to impute missing values ​​in the data using a machine learning algorithm based on random forest regression.

[0034] The interactive output module is further configured to visualize the weight information calculated by the attention mechanism layer in the form of a chart to indicate the degree of influence of different input variables on the predicted value of the target variable.

[0035] Preferably, the persistent communication connection is a full-duplex communication connection based on the WebSocket protocol.

[0036] Secondly, a method for assessing the climate impact of water resources includes the following steps:

[0037] S10: Obtain first-type and second-type spatiotemporal sequence data from multiple heterogeneous data sources through the data interface of the computing system;

[0038] S20: Perform quality checks, format standardization, and spatiotemporal scale alignment on the first and second categories of data to generate a fused dataset;

[0039] S30: Input the fused dataset into a physically-enhanced spatiotemporal graph neural network model for computation to obtain the predicted value of the target variable;

[0040] S31: Among them, the spatiotemporal graph neural network model with physical information enhancement learns spatial correlation through graph convolution operations, learns temporal dynamics through long short-term memory networks, and assigns weights to input variables through an attention mechanism;

[0041] S40: In the training process of the spatiotemporal graph neural network model with enhanced physical information, hydrological physical constraints are introduced through a differentiable physical constraint layer, and physical laws are added as regularization terms to the loss function.

[0042] S50: Outputs a visualization of the predicted values ​​of the target variable;

[0043] S60: Receives scenario parameters input by the user, triggers a spatiotemporal graph neural network model with physical information enhancement to perform incremental forward propagation calculations through persistent connections, and dynamically updates the visualization results within a second-level delay.

[0044] Preferably, the hydrophysical constraint is the watershed water balance equation. .

[0045] Preferably, the step of assigning weights to input variables through an attention mechanism specifically includes: calculating cross-attention between different variables and modulating attention scores based on watershed location and season.

[0046] Preferably, the training loss function of the physically-enhanced spatiotemporal graph neural network model is a weighted sum of data fitting loss and physical constraint loss, and its composite loss function is: Among them, the two losses are used to balance the two losses. The range of hyperparameter values ​​is .

[0047] The beneficial effects of this invention are as follows:

[0048] 1. This invention deeply embeds hydrological physical mechanisms into a spatiotemporal graph neural network in the form of differentiable constraints, and adopts a new intelligent evaluation paradigm that deeply integrates physical laws and data-driven approaches. This not only breaks through the limitations of traditional physical models in characterizing complex nonlinear processes, but also overcomes the defects of pure data-driven models that may violate the laws of nature, thus achieving a unified improvement in scientific rigor, accuracy, and adaptability.

[0049] 2. The system integrates innovative incremental computing mechanisms and multi-factor collaborative attention mechanisms, successfully achieving a leap from static evaluation to dynamic inference. This not only ensures the real-time interactive efficiency of large-scale spatial simulation, but also makes the model decision-making process transparent and credible through visualized attribution analysis, significantly enhancing the practical value and operability of the technology in actual management.

[0050] 3. This technology system constructs a complete solution from data fusion and intelligent analysis to interactive decision-making. Its generalized architecture design is not only applicable to watershed water resource management, but can also be extended to multiple fields such as agricultural water conservation, ecological protection, and urban water security, providing advanced and reliable technical support for comprehensively improving water resource security capabilities under the background of climate change. Attached Figure Description

[0051] Figure 1 This is a schematic diagram of the system framework of the present invention.

[0052] Figure 2 This is a schematic diagram of the spatiotemporal graph neural network model architecture for physical information enhancement according to the present invention.

[0053] Figure 3 This is a schematic diagram of the calculation process of the multi-factor spatiotemporal collaborative attention mechanism of the present invention.

[0054] Figure 4 This is a schematic diagram of the model training and incremental calculation process of the present invention. Detailed Implementation

[0055] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0056] like Figures 1 to 4 This paper presents a water resources climate impact assessment system and method.

[0057] Part 1: Physically Augmented Spatiotemporal Graph Neural Networks. (Combined with...) Figure 1 The architecture, forward propagation process, and key components are as follows.

[0058] The model is mainly composed of four interconnected components: a graph convolutional network module, a temporal modeling module, a multi-factor spatiotemporal collaborative attention module, and a differentiable physical constraint layer. The data flows through these modules sequentially, ultimately outputting the predicted value of the target variable. Simultaneously, throughout the training process, the differentiable physical constraint layer is responsible for calculating the physical consistency loss to guide the optimization of model parameters.

[0059] The specific implementation of the differentiable physical constraint layer is to embed hydrophysical laws into the neural network in a differentiable form, ensuring that the model's output is not only accurate in terms of data, but also physically reasonable.

[0060] Forward propagation process: This layer itself does not change the dimensions of the data; its core function is to calculate the physical constraint loss. In the forward propagation, it receives the predicted output from the model backbone. For example, the model predicts precipitation. Evaporation Runoff and changes in water storage .

[0061] Physical loss calculation: This layer is based on the water balance equation. Calculate the loss. Specifically, the calculation formula is as follows:

[0062]

[0063] in, The number of samples in a training batch. express The square of the norm. This loss measures the degree to which the model's predictions violate the law of water balance.

[0064] Backpropagation and gradient flow: Since all operations (subtraction, squaring) are differentiable, the loss during backpropagation... Relative to input parameters , , , gradient , , , These gradients can be calculated precisely. Through the backpropagation algorithm, these gradients force the model backbone to adjust its parameters during the optimization process, so that the final prediction gradually tends to satisfy the physical laws.

[0065] Implementation example: In a Python environment, using the PyTorch framework, this layer can be implemented as a custom `torch.autograd.Function` or directly encapsulated in an `nn.Module`. Its forward propagation function performs the loss calculation described above, while backpropagation is handled automatically by the framework.

[0066] The specific implementation of the multi-factor spatiotemporal collaborative attention layer.

[0067] This module is used to finely quantify the dynamic interactions between different climate and hydrological variables, and its algorithm flow is as follows.

[0068] Input and Initialization: The input to this layer includes the feature matrices of all spatiotemporal nodes from the temporal modeling module. ,in For the number of nodes, For time step, This is the feature dimension. Features are grouped according to the source of the variable (such as temperature, precipitation, humidity).

[0069] Cross-attention calculation:

[0070] For any two distinct sets of variables and Query vectors are generated through linear transformations respectively. and key vector .

[0071] in and This is a learnable weight matrix.

[0072] Calculate the original cross-attention score ,in It is the dimension of the key vector.

[0073] Time-based air conditioning control mechanism:

[0074] Spatial modulation (based on watershed location): Define a modulation matrix based on flow path distance. For nodes and ,like lie in Upstream, ,in The distance of the flow path. , As a learnable parameter, the initial value can be set to =0.5, =0.1. Otherwise, =1. The modulated attention score is ,in This indicates element-wise multiplication.

[0075] Time modulation (based on season): based on time steps A learnable bias vector is introduced based on the season (high water season or low water season). The bias term is added to the modulated attention score, and then Softmax normalization is performed again to obtain the final attention weights.

[0076] Output: Value vector adjusted using the final attention weights Perform a weighted summation to obtain the variables. Subject to variable The new features after the influence are represented, among which This is a learnable weight matrix. This process is repeated for all variable pairs, and the results are fused to output an enhanced feature with synergistic effect information.

[0077] The complete process of model training.

[0078] The following pseudocode outlines the complete training steps of the model, demonstrating how physical constraints are integrated into the optimization process:

[0079] text

[0080] / / Initialize the model and optimizer, and define the composite loss function

[0081] model = PhysicallyInformedSTGNN(...)

[0082] optimizer = torch.optim.Adam(model.parameters(), lr=1e-3)

[0083] lambda_physics = 1.0 / / Determine the value in the range [0.01, 10] using a grid search

[0084] for epoch in range(total_epochs):

[0085] for batch_data, batch_labels in dataloader:

[0086] optimizer.zero_grad()

[0087] / / Forward propagation

[0088] predictions, physics_loss = model(batch_data)

[0089] / / Calculate the data fitting loss

[0090] data_loss = F.mse_loss(predictions, batch_labels)

[0091] / / Calculate composite loss

[0092] total_loss = data_loss + lambda_physics physics_loss

[0093] / / Backpropagation and parameter update

[0094] total_loss.backward()

[0095] optimizer.step()

[0096] Through the above specific implementation, the high expressive power of data-driven approaches is combined with the reliability constraints of physical mechanisms, thus solving the accuracy and consistency problems faced by traditional methods in complex scenarios.

[0097] The second part is the construction of the training dataset.

[0098] Build a high-quality training dataset that can be used for supervised learning. .

[0099] Data source and region selection:

[0100] Example Area: The target watershed area has a dense river system, complex climate, and prominent challenges in water resource management, making it representative.

[0101] First type of data (spatiotemporal sequence data):

[0102] Meteorological data: including monthly precipitation Surface temperature Sun's downward shortwave radiation The spatial resolution is 0.5 degrees, and the time range is 1982-2022.

[0103] Remote sensing data: Monthly leaf area index (LAI) and land surface albedo were obtained using satellite remote sensing products. Etc., spatial resolution 500 meters.

[0104] Second type of data (spatiotemporal sequence data):

[0105] Hydrological data: Collects monthly runoff data released by the water resources department. Water storage capacity of major reservoirs Observational data, etc.

[0106] True value label data: To obtain reliable training targets, monthly runoff depths calibrated and validated by the SWAT distributed hydrological model were used as the true values ​​predicted by the model. The SWAT model was calibrated using data from 1991 to 2011, and validated using data from 2012 to 2018 to ensure its simulation accuracy (Nash-Sutcliffe efficiency coefficient > 0.75).

[0107] Sample construction and preprocessing:

[0108] Spatiotemporal gridding: All data are uniformly resampled to a uniform latitude and longitude grid (e.g., 0.1° x 0.1°) and monthly time step.

[0109] Graph Structure Construction: A watershed topology is generated based on a digital elevation model, treating each grid cell as a node in the graph. Edges between nodes are determined by the watershed flow direction, with upstream nodes connected to their immediate downstream counterparts.

[0110] Sample splitting: Data from 1982-2010 was used as the training set, data from 2011-2018 as the validation set, and data from 2019-2022 as the test set. Ultimately, approximately [number missing] samples were constructed. =50,000 valid spatiotemporal node samples.

[0111] Hyperparameter determination and model training process.

[0112] Successful training of the model depends on the careful selection of key hyperparameters. The specific training steps are as follows:

[0113] Model initialization and optimizer settings:

[0114] The model weights are initialized using a uniform Xavier distribution.

[0115] Using the Adam optimizer, with its initial learning rate Set as Momentum parameters , The default values ​​of 0.9 and 0.999 are used respectively.

[0116] Physical constraint weights Determination:

[0117] This is a core hyperparameter. Its optimal value is determined on the validation set using a grid search.

[0118] Search scope: Traverse the set {0.01,0.1,0.5,1,5,10}.

[0119] Predicted values ​​of the target variable: The Nash efficiency coefficient on the validation set is used as the evaluation standard. The higher the NSE, the better the model predicts the runoff process and the "true value".

[0120] Determined Result: In this embodiment, when When = 1, the validation set NSE reaches its maximum (0.86), so it is determined as the final training hyperparameter.

[0121] Training loops and early stop mechanisms:

[0122] Batch size: Set to 32.

[0123] Number of training rounds: The maximum number of training rounds is 300.

[0124] Early stopping strategy: Monitor the validation set loss. If the validation set loss no longer decreases after 20 consecutive training epochs, terminate training early and restore the model parameters to the level with the lowest validation set loss. This effectively prevents overfitting.

[0125] The following pseudocode clearly summarizes the core training loop:

[0126] Python

[0127] # Hyperparameters

[0128] lambda_physics = 1.0 # Determine via grid search

[0129] learning_rate = 1e-3

[0130] batch_size = 32

[0131] # Initialization

[0132] model = PhysicallyInformedSTGNN(...)

[0133] optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)

[0134] for epoch in range(max_epochs):

[0135] for batch_data, batch_labels in train_dataloader:

[0136] optimizer.zero_grad()

[0137] # Forward Propagation

[0138] predictions, physics_loss = model(batch_data)

[0139] # Calculate loss

[0140] data_loss = F.mse_loss(predictions, batch_labels)

[0141] total_loss = data_loss + lambda_physics physics_loss

[0142] # Backpropagation and Optimization

[0143] total_loss.backward()

[0144] optimizer.step()

[0145] # Evaluate on the validation set and determine if premature stopping is necessary.

[0146] val_loss = evaluate_on_validation_set(model, val_dataloader)

[0147] if early_stopping_criteria_met(val_loss):

[0148] break

[0149] Training environment

[0150] Hardware: Training was conducted on a server equipped with two NVIDIA A100 40GB GPUs.

[0151] Software: Implemented based on Python 3.8 and PyTorch 1.12.1 frameworks, and relies on the PyTorch Geometric library for graph neural network operations.

[0152] Training duration: With the data scale in this embodiment, the model typically converges in about 150 rounds, with a total training time of approximately 12 hours.

[0153] The third part employs an incremental calculation mechanism based on local updates of the graph structure. This mechanism works as follows: when a user modifies the scenario parameters of a node, the system automatically identifies the affected area and recalculates only that subgraph, rather than processing the entire watershed network.

[0154] Identification of affected nodes and subgraph construction

[0155] Initial state and cache:

[0156] After completing a full graph calculation, the system maintains the complete state of the entire watershed graph in memory, including the latest feature representation of each node. And the predicted value of the target variable output by the model.

[0157] At the same time, the system persistently stores the topological adjacency table of the flow domain. Parameters of graph neural network models .

[0158] Determining the propagation domain:

[0159] When a user modifies a target node through the interactive interface When the parameters are changed (for example, increasing the precipitation by 20%), the system immediately initiates the incremental calculation process.

[0160] Scope of impact: The system performs a directed breadth-first search based on the pre-stored watershed topology. The search starts from the target node. Initially, the traversal is performed only along the direction of water flow (i.e., downstream).

[0161] Subgraph extraction: Extract all subgraphs within the preset maximum influence depth. (For example, The set of downstream nodes visited within a 5-level river channel is denoted as Let the subgraph formed by these nodes and the edges between them be denoted as .

[0162] Local forward propagation calculation.

[0163] Computational context preparation:

[0164] For subgraph Each node in the dataset has its initial node features derived from the modified new data (for...). (or old features in the cache (for affected upstream and downstream nodes)).

[0165] To ensure the accuracy of the calculation, the computation of each node in the subgraph depends on the information of its neighboring nodes. Therefore, when computing the subgraph, it is necessary to introduce its first-order neighbor boundary nodes. The characteristics of these boundary nodes are read from the global cache and are considered fixed in this calculation.

[0166] Efficient local forward propagation:

[0167] Only subgraph The boundary nodes are input into the trained spatiotemporal graph neural network to perform a local forward propagation.

[0168] This process leverages the local connectivity property of graph neural networks. In the graph convolutional layers of the model, during message passing, the update of each node depends only on its direct neighbors. Therefore, for subgraphs... For internal nodes, the computation results are entirely new; for boundary nodes, they only act as information providers and their own state is not updated.

[0169] The mathematical process of local forward propagation is the same as that of full graph computation, but the computational cost is significantly reduced. Its output is a subgraph. Predicted values ​​of the target variable after updates at all nodes .

[0170] Status updates and visualization refresh.

[0171] Global state synchronization: This involves synchronizing the calculated new metrics. Update the global state cache, overwriting the old values ​​of the affected nodes.

[0172] Communication and Rendering: The system immediately pushes updated node data to the front end via a persistent WebSocket connection. After receiving the data, the front-end visualization engine only redraws the areas on the map that have changed.

[0173] Implementation examples and performance guarantees.

[0174] The following pseudocode summarizes the core process of incremental computation:

[0175] Python

[0176] def incremental_update(target_node_vt, new_parameters, global_cache):

[0177] # 1. Identify the affected subgraph

[0178] affected_nodes = bfs_downstream(vt, global_cache['topology'],max_depth=L)

[0179] subgraph_nodes = affected_nodes ∪ get_one_hop_neighbors(affected_nodes) # Contains boundaries

[0180] subgraph = extract_subgraph(global_cache, subgraph_nodes)

[0181] # 2. Update target node parameters

[0182] subgraph.node_features[vt] = new_parameters

[0183] # 3. Local forward propagation

[0184] with torch.no_grad(): # Do not calculate gradients, speeding up inference

[0185] new_predictions = model(subgraph)

[0186] # 4. Update global state

[0187] global_cache['predictions'][affected_nodes]= new_predictions[affected_nodes]

[0188] #5. Push Updates

[0189] websocket.send(serialize(affected_nodes, new_predictions[affected_nodes]))

[0190] Performance Guarantee: Due to the locality of the watershed topology, the number of nodes affected by a single parameter modification is far less than the total number of nodes in the network. In the embodiments of this invention, for a target watershed graph containing more than 10,000 nodes, the average response time of a single incremental calculation can be controlled within 500 milliseconds, achieving an interactive experience with "second-level latency".

[0191] Through the above mechanism, while ensuring the accuracy of the evaluation results, the real-time interactive capability of the large-scale spatiotemporal simulation system was realized, which significantly improved the system's practicality and user experience.

[0192] Example: Application of water resources and climate impact assessment in target flow areas

[0193] 1. Data preparation and preprocessing.

[0194] 1.1 Study area and data sources.

[0195] Research period: January 2000 - December 2020 (monthly data);

[0196] Training set: Data from 2000 to 2015;

[0197] Validation set: Data from 2016 to 2018;

[0198] Test set: Data from 2019-2020;

[0199] Data source details:

[0200] Meteorological data: CRUNCEP v7 dataset, including precipitation ( ), temperature ( ), solar radiation ( ), relative humidity ).

[0201] Remote sensing data: MODIS products, including leaf area index (LAI), land surface temperature (LST), and land surface albedo.

[0202] Hydrological data: collected runoff observation data and reservoir storage data.

[0203] Topographic data: SRTM 90m DEM data was used to extract watershed drainage systems.

[0204] 1.2 Data preprocessing process.

[0205] All data were uniformly resampled to a 0.1°×0.1° spatial grid.

[0206] The time scale is unified to monthly data.

[0207] The topology of the watershed was extracted based on the DEM, and a graph structure (12,345 nodes in total) was constructed.

[0208] The missing data was imputed using random forest regression.

[0209] 2. Model training and parameter setting.

[0210] 2.1 Model architecture parameters.

[0211] Python

[0212] # Key parameters of the model

[0213] graph_conv_layers = [128, 128] # Two-layer graph convolution, each layer is 128-dimensional

[0214] lstm_hidden_size = 128 # LSTM hidden layer dimension

[0215] attention_heads = 8 # Multi-head attention mechanism

[0216] physics_weight = 1.0 # Physics constraint weights (determined via grid search)

[0217] 2.2 Training process.

[0218] Optimizer: ;

[0219] Batch size: 32;

[0220] Training rounds: 250 rounds (actually stopped at 218 rounds);

[0221] Training hardware: 2×NVIDIA A100 40GB GPUs;

[0222] Total training time: 14.5 hours;

[0223] Training loss convergence status:

[0224] Initial total loss: 8.76;

[0225] Final total loss: 1.23;

[0226] Data loss (MSE): 0.89;

[0227] Physical constraint loss: 0.34;

[0228] 3. Performance evaluation results.

[0229] 3.1 Prediction accuracy index.

[0230] Table 1. Comparison of prediction accuracy of different models on the test set:

[0231]

[0232] 3.2 Physical consistency assessment.

[0233] Statistics on water balance closure error (all samples in the test set):

[0234] Mean absolute error (MAE): 1.3 mm / month;

[0235] Root mean square error (RMSE): 2.1 mm / month;

[0236] Maximum error: 8.7 mm / month (occurred during a typhoon in July 2019);

[0237] Error distribution: 95% of the samples have a closure error within ±5mm / month;

[0238] 3.3 Computational efficiency test.

[0239] Training phase:

[0240] Full model training convergence time: 14.5 hours;

[0241] Single round training time: approximately 3.5 minutes;

[0242] Reasoning stage:

[0243] Full graph reasoning time (12,345 nodes): 18.2 seconds;

[0244] Incremental computing performance:

[0245] Average response time for single-node modification: 0.42 seconds;

[0246] The average response time for simultaneous modification of 10 nodes is 1.35 seconds.

[0247] Maximum test (modification of upstream critical nodes): affected 856 nodes, response time 2.8 seconds.

[0248] 4. Explainability Case Study: The Spring Drought Event of 2020.

[0249] 4.1 Case Background.

[0250] From March to May 2020, the target area suffered a severe spring drought, with average precipitation 45% less than normal, and the water levels of many reservoirs were close to dead water levels.

[0251] 4.2 Attention weight analysis.

[0252] Analyzing the drought-stricken month of April 2020, the model assigns attention weights to the main driving factors as follows:

[0253] Table 2. Attention weight analysis of the main driving factors of the spring drought event in 2020:

[0254]

[0255] 4.3 Comparison between model predictions and actual results.

[0256] Predicted water shortage index: 0.68 (high risk);

[0257] Prediction accuracy: The risk of severe water shortage was predicted one month in advance (NSE=0.87).

[0258] 5. Comparative experimental analysis.

[0259] 5.1 Ablation experiment.

[0260] Table 3. Comparison of ablation test results:

[0261]

[0262] 5.2 Advantages compared with traditional methods.

[0263] Prediction accuracy: approximately 58% higher than the traditional SWAT model (RMSE reduced from 12.5 to 5.2);

[0264] Physical rationality: The water balance error is controlled within an acceptable range;

[0265] Practicality: Incremental computation enables near real-time interaction with a response time of less than 3 seconds;

[0266] Interpretability: The attention mechanism clearly reveals the main driving factors.

[0267] 6. System application effect.

[0268] In practical water resource management decision support, this system demonstrates the following value:

[0269] Scenario simulation: It can quickly assess the impact of water resources under different climate scenarios;

[0270] Risk warning: A high-risk warning for the 2020 drought was issued one month in advance;

[0271] Decision support: Providing quantitative basis for water allocation in target watersheds;

[0272] Public services: Disseminating water resource information to the public through a visual interface;

[0273] Conclusion: This embodiment fully demonstrates that the technical solution described in this invention significantly outperforms existing technologies in terms of prediction accuracy, physical consistency, computational efficiency, and interpretability, possessing outstanding technological advancements and practical value. All performance indicators have been rigorously verified through quantitative testing, and the results are reliable and reproducible.

[0274] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A water resources climate impact assessment system, characterized in that, include: The data acquisition module is configured to acquire first-type and second-type spatiotemporal sequence data from multiple heterogeneous data sources via an application programming interface. The data fusion and preprocessing module, connected to the data acquisition module, is configured to perform quality checks and format standardization on the first and second types of data, and perform scale alignment based on a unified spatiotemporal grid benchmark to generate a fused dataset. The intelligent analysis engine module is connected to the data fusion and preprocessing module. The intelligent analysis engine module loads and runs a pre-trained physical information-enhanced spatiotemporal graph neural network model. The physical information-enhanced spatiotemporal graph neural network model is calculated based on a graph structure, which uses geospatial entities as nodes and connections reflecting their spatial relationships as edges. The physically enhanced spatiotemporal graph neural network model includes graph convolutional layers, long short-term memory network layers, and attention mechanism layers. The output of the physically enhanced spatiotemporal graph neural network model is connected to a differentiable physical constraint layer, which is used to substitute the intermediate prediction results of the physically enhanced spatiotemporal graph neural network model into the predefined domain physical constraint equation for calculation, and add the calculation result as a regularization term to the loss function of the physically enhanced spatiotemporal graph neural network model to ensure that the output of the physically enhanced spatiotemporal graph neural network model conforms to physical laws. The output of the physically enhanced spatiotemporal graph neural network model is the predicted value of the target variable; An interactive output module, connected to the intelligent analysis engine module, is configured to output a visual representation of the predicted value of the target variable and provide user interface controls to receive scenario adjustment parameters. By maintaining a persistent communication connection with the front-end interface, it triggers the intelligent analysis engine module to perform incremental calculations and dynamically updates the visual representation in real time. The domain physical constraint equation upon which the differentiable physical constraint layer is based is the watershed water balance equation, and its expression is: in, For precipitation, Evaporation rate For runoff, This represents the change in water storage variables; The differentiable physical constraint layer is configured to: calculate the difference between the two sides of the water balance equation, and then... Norm as a physical constraint loss term ,Right now: 。 2. The water resources climate impact assessment system according to claim 1, characterized in that: The attention mechanism layer is a multi-factor spatiotemporal collaborative attention layer, specifically configured to perform the following operations: Calculate the cross-attention score between different variables in the first and second data categories. This cross-attention score is calculated using a query-key-value mechanism, where the query vector... and key vector Feature representations from different variables; Based on the upstream and downstream location relationships of geospatial entities within a watershed, the spatial attention score is modulated. Specifically, when calculating attention, downstream nodes are assigned a weight based on the flow path distance, derived from information from upstream nodes. Enhancer ,in , These are learnable parameters or preset constants; The time attention score is modulated based on the season in which the time step occurs, specifically by setting different learnable biases for the wet season and the dry season. and .

3. The water resources climate impact assessment system according to claim 1, characterized in that: The physically enhanced spatiotemporal graph neural network model is obtained through the following steps: S1: Prepare the training dataset ,in As input features, This refers to the true water resource status data, which has been corrected using a traditional hydrological model. S2: Construct a network architecture for a water resources climate impact assessment system; S3: Define the composite loss function Its expression is: in, Predictions from a spatiotemporal graph neural network model that enhances physical information with truth value Mean squared error loss between them The physical constraint loss calculated for the differentiable physical constraint layer. The hyperparameter used to balance the two losses has a range of values. The optimal value is determined by grid search on the validation set; S4: Minimize the composite loss function using the error backpropagation algorithm. All parameters in the spatiotemporal graph neural network model that enhances physical information Joint optimization was carried out.

4. The water resources climate impact assessment system according to claim 1, characterized in that: When performing scale alignment operations, the data fusion and preprocessing module is specifically used to: unify spatial resolution through an inverse distance weighted spatial interpolation algorithm and unify time frequency through a linear resampling method. The data fusion and preprocessing module also includes an intelligent repair unit, configured to: impute missing values ​​in the data using a machine learning algorithm based on random forest regression; The interactive output module is further configured to visualize the weight information calculated by the attention mechanism layer in the form of a chart to indicate the degree of influence of different input variables on the predicted value of the target variable.

5. A water resources climate impact assessment system according to claim 1, characterized in that: The persistent communication connection is a full-duplex communication connection based on the WebSocket protocol.

6. A method for assessing the climate impact of water resources, characterized in that, To implement a water resources climate impact assessment system according to any one of claims 1-5, the system comprises the following steps: S10: Obtain first-type and second-type spatiotemporal sequence data from multiple heterogeneous data sources through the data interface of the computing system; S20: Perform quality checks, format standardization, and spatiotemporal scale alignment on the first and second types of data to generate a fused dataset; S30: Input the fused dataset into a physically-enhanced spatiotemporal graph neural network model for calculation to obtain the predicted value of the target variable; S31: Wherein, the physical information-enhanced spatiotemporal graph neural network model learns spatial correlation through graph convolution operations, learns temporal dynamics through long short-term memory networks, and assigns weights to input variables through an attention mechanism; S40: In the training process of the spatiotemporal graph neural network model with enhanced physical information, hydrological physical constraints are introduced through a differentiable physical constraint layer, and physical laws are added as regularization terms to the loss function. S50: Output the visualization results of the predicted values ​​of the target variable; S60: Receive the scenario parameters input by the user, trigger the spatiotemporal graph neural network model with physical information enhancement to perform incremental forward propagation calculation through persistent connection, and dynamically update the visualization results within a second-level delay.

7. The water resources climate impact assessment method according to claim 6, characterized in that: The step of assigning weights to input variables through an attention mechanism specifically includes: calculating the cross-attention between different variables and modulating the attention score based on the watershed location and season.

Images