Multi-dimensional index associated natural resource intelligent aid decision-making generation method and system

Abstract

The invention relates to the technical field of natural resource intelligent decision making, and discloses a multi-dimensional index associated natural resource intelligent auxiliary decision making generation method and system, and the method comprises the steps: collecting multi-dimensional drainage basin state index data, and carrying out the heterogeneous recognition and normalization processing, and obtaining a feature vector set; generating an index association intensity matrix and embedded representation by using a dynamic graph neural network and comparative learning; constructing a decision knowledge graph based on scene clustering and graph embedding; forming a decision generation model and a resource allocation scheme by combining scene matching retrieval and knowledge priori fused reinforcement learning; outputting a comprehensive decision report by using digital twinning and multi-scenario simulation; an interactive interface and an interpretable report are provided by means of attention visualization and anti-factual reasoning; and dynamic decision updating and continuous model optimization are realized by adopting model predictive control and online learning. According to the invention, intelligent mining of index association and decision generation of knowledge guidance are realized, and scientificity and interpretability of resource configuration decision are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent decision-making technology for natural resources, and more specifically, to a method and system for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation. Background Technology

[0002] The current imbalance between supply and demand of natural resources is becoming increasingly prominent, and the allocation of water resources in river basins faces complex challenges such as the coupling of multiple dimensions of indicators, spatiotemporal unevenness, and high uncertainty. Traditional methods rely on empirical rules and simplified models, which are insufficient to characterize the dynamic relationships between multiple dimensions of indicators and have limited ability to cope with systemic risks. While existing data-driven methods can process large amounts of data, their interpretability is insufficient, which restricts their practical management applications.

[0003] Existing indicator correlation analysis methods mostly employ static correlation calculations, failing to capture the dynamic evolution of indicator correlation patterns over time and watershed conditions. Predictive accuracy declines when correlation patterns undergo abrupt changes under extreme climatic conditions. Existing decision generation methods lack a systematic organization and effective utilization of historical decision-making experience, failing to organically integrate domain knowledge with data-driven learning, resulting in insufficient rationality and feasibility of generated decision solutions in complex scenarios. Existing decision evaluation methods are mostly based on simplified effect prediction models, failing to fully consider the impact of multi-process coupling and uncertainties, making it difficult to accurately assess the robustness of decision solutions in actual implementation.

[0004] Therefore, there is an urgent need for an intelligent decision-making assistance method that can intelligently mine the dynamic correlation of indicators, integrate knowledge to guide decision-making, and support comprehensive simulation and interpretable analysis, so as to improve the scientificity, rationality and credibility of resource allocation and support the sustainable use and refined management of natural resources. Summary of the Invention

[0005] This invention provides a method and system for generating intelligent auxiliary decisions for natural resources based on multidimensional index correlation, which solves the technical problems in related technologies such as inaccurate mining of multidimensional index correlation, lack of knowledge guidance in decision generation, and poor interpretability of decision schemes.

[0006] This invention provides a method for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation, comprising the following steps: S1. Collect raw data of multi-dimensional watershed state indicators, and use heterogeneous data type identification and multi-level normalization processing methods to obtain a set of multi-dimensional indicator feature vectors. S2, obtain a set of multi-dimensional indicator feature vectors, and use dynamic graph neural network and contrastive learning to obtain the indicator correlation strength matrix and indicator embedding representation; S3 receives the indicator association strength matrix and indicator embedding representation, and uses scene clustering and graph embedding methods to obtain the decision knowledge graph; S4 receives the indicator association strength matrix, indicator embedding representation and decision knowledge graph, and adopts a reinforcement learning framework that combines scene matching retrieval and fusion of prior knowledge to obtain a decision generation model and resource allocation decision scheme; S5 receives the resource allocation decision plan, and uses digital twin technology and multi-scenario simulation methods to obtain a comprehensive decision report; S6 receives the indicator correlation strength matrix and indicator embedding representation, decision generation model and decision scheme, and comprehensive decision report. It uses attention visualization and counterfactual reasoning to obtain an interactive visualization interface and an interpretable report. S7 receives real-time monitoring data and effect feedback during the execution of the decision-making scheme, and uses model predictive control and online learning methods to obtain dynamically updated decision-making schemes and iteratively optimized model parameters.

[0007] In a preferred embodiment, S1 includes: Data on indicators from the hydrological and meteorological, resource supply and demand, ecological and environmental, socio-economic, and policy constraint dimensions of a multi-source monitoring network are collected. Data are then filled in using interpolation based on the missing rate, and outliers are identified using the standard deviation criterion. An indicator type identification model is constructed to classify indicators into numerical continuous indicators, numerical discrete indicators, and categorical indicators, and the maximum-minimum value normalization, standard deviation normalization, and one-hot coding methods are used to process them respectively. The initial weights of the indicators are determined based on information entropy values ​​and domain expert experience scores, and then dynamically adjusted using the coefficient of variation method. Extract time series features and spatial distribution features, and concatenate them to form a set of multidimensional indicator feature vectors.

[0008] In a preferred embodiment, S2 includes: Each indicator is treated as a node in the graph. Directed edges are established for indicator pairs with physical causal relationships based on domain knowledge. The Pearson correlation coefficient of the indicator time series is calculated. When the absolute value of the correlation coefficient is greater than the preset association threshold, an undirected edge is established. Design a multi-layer dynamic graph convolutional network, use an attention mechanism to calculate the aggregate weights of neighboring nodes, and achieve information propagation layer by layer by stacking multi-layer graph convolution operations; Gated cyclic units are used to model time series data, and the hidden state that integrates historical association patterns and current association patterns is output. Design an association mutation detection module to calculate the distance between the current time and the historical average association strength matrix, and trigger an association mutation flag when the distance exceeds a dynamic threshold. Construct positive and negative sample pairs for comparative learning training, and output the index association strength matrix and index embedding representation.

[0009] In a preferred embodiment, S3 includes: For each historical case, extract watershed status features, decision-making measures content, and effect evaluation data, and construct a scenario feature vector by combining the indicator correlation strength matrix and node embedding vector; Hierarchical clustering algorithm is used to perform cluster analysis on scene feature vectors, and historical cases are divided into several typical scene categories; Define decision-making scenario entities, indicator characteristic entities, decision-making behavior entities, and effect evaluation entities; design scenario inclusion relationships, scenario applicability relationships, behavior generation relationships, and indicator impact relationships; and transform historical cases into entities and relationships in the graph. We employ a translation-based embedding model for training and output a vector representation of the decision knowledge graph.

[0010] In a preferred embodiment, S4 includes: The current scene feature vector is input into the scene encoder to output the embedding vector. The historical scene with the highest cosine similarity is retrieved in the decision knowledge graph as the candidate matching scene, and the decision behavior entity associated with the candidate scene is retrieved as the reference strategy. Define the state space and action space for sequential decision-making, and adopt a hierarchical decision-making strategy; The reward function is defined to include rewards for water supply satisfaction, rewards for ecological protection compliance, rewards for economic benefits, and penalties for decision-making stability. We construct actor networks and critic networks, design a knowledge-guided attention module, and the final output strategy of the actor network is a weighted fusion of its own learning strategy and prior strategy according to the knowledge-guided weight. The algorithm is trained using a near-end strategy optimization algorithm, and outputs a decision generation model and a resource allocation decision scheme.

[0011] In a preferred embodiment, S5 includes: A digital twin model is constructed, which includes a hydrological sub-model, a hydraulic sub-model, a water quality sub-model, an ecological sub-model, and a socio-economic sub-model. An error correction model is constructed using a data-driven approach, and data assimilation is performed using an ensemble Kalman filter method. Input the resource allocation decision scheme into the digital twin model for coupled simulation; Multiple scenarios are set up to address uncertainties in weather, water demand, and unforeseen events. Sensitivity analysis is used to calculate the sensitivity index of uncertain factors on the decision-making effect, and decision parameters are adjusted to target weak links to generate a comprehensive decision report.

[0012] In a preferred embodiment, S6 includes: A force-directed layout algorithm is used to visualize the index correlation graph; The attention weight vectors of the actor network are extracted and plotted as a heatmap; Construct counterfactual scenarios for key decision variables, input the original and counterfactual scenarios into a digital twin model for simulation, and quantify the impact of adjusting decision variables; Generate natural language decision explanation text and build an interactive decision explanation system.

[0013] In a preferred embodiment, S7 includes: Real-time monitoring data is collected through a distributed monitoring network across the watershed. The deviation of key indicators is calculated by comparing real-time monitoring data with the expected state of the decision-making plan. When the deviation exceeds the preset warning threshold, an early warning is triggered. A rolling optimization method is adopted using model predictive control, which regenerates the optimization decision scheme based on the current real-time monitored watershed status and the latest forecast.

[0014] In a preferred embodiment, S7 further includes: The input conditions, parameters, and actual implementation effects of the decision-making schemes are used to create new decision-making cases and supplement the historical decision-making case database. Incremental updates to the decision knowledge graph based on an expanded case library; Online learning methods are used to fine-tune and update the parameters of the decision generation model.

[0015] In a preferred embodiment, a multi-dimensional index-based intelligent auxiliary decision-making generation system for natural resources is used to perform the steps in the aforementioned multi-dimensional index-based intelligent auxiliary decision-making generation method for natural resources, including: The data acquisition and preprocessing module is used to collect raw data of multi-dimensional watershed state indicators, and uses heterogeneous data type identification and multi-level normalization processing methods to obtain a set of multi-dimensional indicator feature vectors. The indicator association mining module is used to obtain a multi-dimensional indicator feature vector set. It uses dynamic graph neural network and contrastive learning to obtain the indicator association strength matrix and indicator embedding representation. The knowledge graph construction module receives the indicator association strength matrix and indicator embedding representation, and uses scene clustering and graph embedding methods to obtain the decision knowledge graph; The decision generation module receives the indicator association strength matrix, indicator embedding representation, and decision knowledge graph. It uses a reinforcement learning framework that combines scene matching retrieval and fusion of prior knowledge to obtain the decision generation model and resource allocation decision scheme. The simulation evaluation module receives resource allocation decision schemes and uses digital twin technology and multi-scenario simulation methods to obtain a comprehensive decision report. The interpretability analysis module receives the index correlation strength matrix and index embedding representation, decision generation model and decision scheme, and comprehensive decision report. It uses attention visualization and counterfactual reasoning to obtain an interactive visualization interface and interpretable report. The dynamic optimization module receives real-time monitoring data and effect feedback during the execution of the decision-making scheme. It uses model predictive control and online learning methods to obtain dynamically updated decision-making schemes and iteratively optimized model parameters.

[0016] The beneficial effects of this invention are as follows: By employing dynamic graph neural networks and contrastive learning methods to mine the correlations between multidimensional indicators, and by constructing an indicator correlation map and introducing a time-aware mechanism and a correlation mutation detection module, the dynamic evolution characteristics of indicator correlation patterns with time and watershed status can be accurately captured. Even under extreme climatic conditions, it can still effectively identify correlation pattern mutations, improving the accuracy and adaptability of indicator correlation analysis, providing reliable correlation information support for subsequent decision generation, and solving the problem that existing static correlation analysis methods cannot cope with dynamic changing scenarios. This paper adopts a reinforcement learning framework that combines scene matching retrieval with prior knowledge to generate resource allocation decision schemes. By constructing a decision knowledge graph to systematically organize historical decision-making experience, a knowledge-guided attention module is designed to achieve the organic integration of data-driven learning and domain knowledge. At the same time, digital twin technology is combined to conduct multi-scenario simulation evaluation and interpretability analysis such as attention visualization and counterfactual reasoning. This enables decision-makers to understand the generation logic and key influencing factors of decisions, improving the rationality, robustness and credibility of decision schemes. It solves the problems of lack of knowledge guidance and poor interpretability in decision generation of existing methods. Attached Figure Description

[0017] Figure 1 This is a flowchart of a method for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation, according to the present invention. Figure 2 This is a flowchart of a method for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation, according to the present invention. Detailed Implementation

[0018] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0019] At least one embodiment of the present invention discloses a method for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation, such as... Figures 1 to 2 As shown, it includes the following steps: S1. Collect raw data of multi-dimensional watershed state indicators, and use heterogeneous data type identification and multi-level normalization processing methods to obtain a set of multi-dimensional indicator feature vectors. S11, collect raw data of multi-dimensional watershed status indicators from the multi-source monitoring network; Acquire raw data of multi-dimensional watershed status indicators collected by a multi-source monitoring network. Collect hydrological and meteorological indicator data, including upstream inflow volume acquired in real time through hydrological station flow monitoring equipment, precipitation acquired through a network of rain gauges at meteorological stations, evaporation calculated using evaporation pans and meteorological models, and runoff acquired through monitoring at the watershed outlet section.

[0020] Quality control was performed on the collected raw data. For missing data, linear interpolation was used to complete the missing data when the missing rate was less than 5%, spatial interpolation based on adjacent sites was used when the missing rate was between 5% and 20%, and the data for that period was marked as unusable when the missing rate exceeded 20%. For outlier data, outliers were identified using the three-standard-deviation criterion. Manual verification was conducted to confirm whether they were genuine extreme events or measurement errors. Confirmed measurement error data was then removed or corrected.

[0021] Data on resource supply and demand dimensions are collected. These resource supply and demand dimensions include agricultural water demand calculated based on irrigated area and crop water demand quotas, industrial water demand calculated based on industrial output and water use efficiency statistics, domestic water demand calculated based on population size and water use quotas, and ecological water demand determined based on river ecological flow regulations and wetland maintenance needs.

[0022] Data on ecological and environmental indicators are collected, including river ecological flow obtained through key section flow monitoring, water pollution index calculated by collecting indicators such as chemical oxygen demand and ammonia nitrogen from water quality monitoring stations, and wetland area obtained through remote sensing image interpretation.

[0023] Data on socio-economic indicators are collected, including GDP obtained from statistical yearbooks, water use efficiency calculated from the economic value generated per unit of water consumption, and water supply guarantee rate obtained from statistics on the number of times historical water supply standards have been met. Data on policy constraints are also collected, including total water consumption control targets obtained from river basin management policy documents and water function zoning requirements obtained from environmental protection department regulations.

[0024] The data collection time granularity of the multi-dimensional indicators is on a daily scale, and the collection time span covers a complete hydrological year of 5 consecutive years in history, ensuring that the data includes various typical watershed states such as high water years, normal water years, low water years, and extremely low water years.

[0025] S12, classifies and identifies heterogeneous type indicator data and performs multi-level normalization processing; To address the heterogeneity of data types for indicators across different dimensions, an indicator type identification model is constructed to automatically classify the collected indicators. This model categorizes indicators into three types: numerical continuous indicators, numerical discrete indicators, and categorical indicators.

[0026] For data identified as numerical continuous indicators, the maximum-minimum normalization method is used to linearly map the data to a standard interval of 0 to 1, eliminating the influence of differences in the dimensions and numerical ranges of different indicators. For data identified as numerical discrete indicators, the standard deviation normalization method is used to convert the data into a standard normal distribution with a mean of 0 and a standard deviation of 1, eliminating the influence of the dimensions of data fluctuation.

[0027] For data identified as categorical indicators, a one-hot encoding method is used for transformation. Each category label is converted into a binary vector with a length equal to the total number of categories. The element corresponding to the category position is 1, and the rest are 0. For the one-hot encoded categorical indicators, their binary vectors are flattened and concatenated with the normalized values ​​of numerical indicators to form a unified numerical feature representation. This ensures that subsequent weight allocation and feature extraction can uniformly process all types of indicator data.

[0028] S13, Design an adaptive weight allocation strategy to dynamically determine the indicator weights; Given the significant differences in the contribution of various indicators to the decision-making objectives of watershed resource allocation, an adaptive weight allocation strategy is designed to dynamically determine the indicator weights. The information entropy value of each indicator is calculated to characterize its uncertainty and discriminative ability; a smaller information entropy value indicates higher certainty in the indicator's value and a greater contribution to the discriminative decision-making process.

[0029] By combining information entropy values ​​and the experience scores of domain experts on the importance of each indicator, a linear weighting method is used to determine the initial weights of the indicators. The information entropy contribution weight accounts for 0.6 and the expert experience weight accounts for 0.4, which takes into account the advantages of data-driven and knowledge-driven approaches.

[0030] The initial weights are dynamically adjusted using the coefficient of variation method. The coefficient of variation of each indicator is calculated within the current decision-making period. The coefficient of variation is defined as the ratio of the standard deviation of the indicator to its mean. The larger the coefficient of variation, the stronger the volatility of the indicator and the more significant its impact on the current decision. The initial weights are then adjusted based on the coefficient of variation, assigning higher weights to indicators with high volatility and lower weights to indicators with low volatility.

[0031] S14, extract spatiotemporal features and construct a set of multidimensional indicator feature vectors with unified representation; Feature enhancement processing is performed on the normalized and weighted multidimensional indicator data to extract richer indicator representations. Time series features of the indicators are extracted, and the mean of each indicator over the past 30-day time window is calculated to reflect the central trend of the indicator, the variance to reflect the degree of volatility of the indicator, the rate of change to reflect the growth trend of the indicator, and the periodicity is identified by Fourier transform to identify the dominant periodic component and reflect the seasonality of the indicator.

[0032] The spatial distribution characteristics of the indicators are extracted, and the gradient differences of the corresponding indicators in the three functional zones of upstream, midstream and downstream are calculated to reflect the spatial heterogeneity of the watershed. The spatial clustering of the indicators in each functional zone is calculated, and the spatial autocorrelation coefficient is used to measure the similarity of the indicators in adjacent regions to reflect the spatial dependence.

[0033] The original normalized index, time series features, and spatial distribution features are concatenated column by column to form a high-dimensional feature vector for each index. The feature vector comprehensively describes the numerical, temporal evolution, and spatial distribution characteristics of the index.

[0034] For the daily-scale data of the past 5 years, after the above processing, a unified set of multidimensional indicator feature vectors containing all monitoring times is obtained. Assuming that there are N total monitoring indicators, T time steps of historical data, the original normalized feature dimension of each indicator is 1, the time series feature dimension is 4 including the periodicity of the mean, variance, and rate of change, and the spatial distribution feature dimension is 2 including gradient differences and spatial clustering, then the feature vector dimension of each indicator is 7, and the dimension of the entire feature vector set is T multiplied by N multiplied by 7.

[0035] S2, obtain a set of multi-dimensional indicator feature vectors, and use dynamic graph neural network and contrastive learning to obtain the indicator correlation strength matrix and indicator embedding representation; S21, Construct the topological structure of the multidimensional indicator correlation graph; The system receives the unified representation of the multidimensional indicator feature vector set output from step 1 and trains a dynamic graph neural network model using five years of historical data to learn the relationships between indicators. First, it constructs the topological structure of the multidimensional indicator correlation graph to explicitly model the relationships between indicators.

[0036] Each indicator involved in the watershed resource system is treated as a node in the graph. These nodes include hydrological and meteorological nodes such as precipitation and runoff nodes; resource supply and demand nodes such as agricultural water demand and industrial water demand nodes; ecological and environmental nodes such as ecological flow and water pollution index nodes; and socio-economic nodes such as GDP and water use efficiency nodes. The total number of nodes equals the total number of monitoring indicators. The features of each node are initialized as the feature vector of the corresponding indicator obtained in step 1.

[0037] Based on knowledge of watershed resources and statistical analysis of data, the edge connection relationship between nodes is initialized. For index pairs with clear physical causal relationships, directed edges are established. For example, if precipitation affects runoff, a directed edge is established from the precipitation node to the runoff node. If runoff affects agricultural water supply, a corresponding directed edge is established.

[0038] For indicator pairs that may influence each other, the Pearson correlation coefficient of the time series of the two indicators is calculated. When the absolute value of the correlation coefficient is greater than the threshold of 0.3, the two indicators are considered to be significantly correlated, and an undirected edge is established. The absolute value of the correlation coefficient is used as the initial weight of the edge; the larger the absolute value, the stronger the correlation. For indicator pairs that have not been found to be correlated by domain knowledge and data analysis, no edge connection is established initially. Implicit correlations may be discovered later through graph neural network learning.

[0039] S22, Design a multi-layer dynamic graph convolutional network to capture the correlation of metrics; A multi-layer dynamic graph convolutional network was designed to capture the direct correlations and multi-hop indirect propagation relationships between metrics. The specific structural parameters of the graph convolutional network were set as follows: node feature dimension of 64, number of graph convolutional layers of 3, and hidden feature dimensions of 128, 64, and 32 dimensions for each layer, respectively. The ReLU function was used as the activation function.

[0040] For any node i in the graph, its feature update at layer l is achieved by aggregating the features of neighboring nodes. Specifically, the feature at layer l of node i is obtained by aggregating the feature at layer l (1) of node i itself and the feature at layer l (1) of all neighboring nodes connected to node i by edges. The aggregation contribution of neighboring node j to node i is determined by the edge weights and the feature vectors of neighboring nodes.

[0041] An attention mechanism is used to calculate the aggregate weight of neighboring nodes. The method for calculating the attention weight is to concatenate the features of node i and node j and input them into a single-layer feedforward neural network. The feedforward network outputs a scalar attention score. The attention score is multiplied by the initial edge weight to obtain the comprehensive weight. The comprehensive weight of all neighboring nodes of node i is normalized to obtain the final aggregate weight of each neighboring node. The larger the aggregate weight, the more important the influence of the neighboring node on node i.

[0042] The aggregated feature of node i is calculated by weighting the features of all neighboring nodes according to the final aggregate weight. The l-1 layer features of node i and the aggregated features are concatenated and then input into a nonlinear activation function to obtain the l-th layer updated features of node i.

[0043] By stacking three layers of graph convolution operations, information is propagated layer by layer in the graph. The first layer captures the first-order association between directly adjacent indicators, the second layer captures the second-order indirect association between nodes, and the third layer captures the transmission relationship over a longer distance, so that the final feature of each node integrates the association information of its neighborhood range of multiple hops.

[0044] S23 introduces a time-aware mechanism and an associated mutation detection module; A time-aware mechanism is introduced to capture the dynamic evolution of indicator correlation patterns over time. Multidimensional indicator data from T consecutive historical time steps are input into the model in chronological order, with T set to 30 time steps to capture time dependencies at the monthly scale. For each time step t, an indicator correlation graph is constructed, with node features being the indicator feature vector at time t.

[0045] A gated recurrent unit (GRU) is used to model the time series data. The hidden state dimension of the GRU is set to 64 dimensions, and the hidden state of the GRU stores the index association pattern information of historical time points. At time t, the node features output by the graph convolutional network at time t and the hidden state of the GRU at time t minus 1 are used as inputs. The GRU controls how much historical information to retain by updating the gate and how much irrelevant information to forget by resetting the gate, and outputs a new hidden state at time t. The new hidden state integrates the historical association pattern and the current association pattern.

[0046] Through cyclical connections over time, the model learns how the strength of the correlation between indicators changes with water availability, such as the increased competition between agricultural water demand and ecological flow during the dry season and the weakening of this competition during the wet season.

[0047] To address the potential for abrupt changes in index correlation patterns under extreme climatic conditions, a correlation abrupt change detection module is designed. The Frobenius norm distance between the index correlation strength matrix at the current time t and the historical average correlation strength matrix over the past 30 time steps is calculated; this distance measures the overall degree of difference between the two matrices.

[0048] The dynamic threshold is set to the historical distance mean plus twice the historical distance standard deviation. When the current distance exceeds the dynamic threshold, the association mutation flag is triggered. The system records the time of the mutation and the association pattern after the mutation. The data before and after the mutation time are studied in detail to extract the characteristics and patterns of the association pattern mutation for early warning and response to future extreme events.

[0049] S24, a contrastive learning method is used to train a dynamic graph neural network and output the correlation analysis results; The parameters of a dynamic graph neural network are optimized using a contrastive learning method to improve the accuracy of association mining. Five years of historical data are divided into training, validation, and test sets in a 7:2:1 ratio. The training set is used for model parameter learning, the validation set for hyperparameter tuning and early stopping detection, and the test set for final performance evaluation.

[0050] Positive and negative sample pairs are constructed for comparative learning training. Positive sample pairs are defined as indicator pairs that exhibit a strong correlation in actual watershed operation. For example, precipitation and runoff are identified as a strongly correlated positive sample pair based on domain expert knowledge, and indicator pairs with an absolute correlation coefficient greater than 0.7 are identified as positive sample pairs based on historical data statistics. Negative sample pairs are defined as indicator pairs with a weak or no correlation. Negative sample pairs are constructed by randomly sampling indicators from different dimensions that have no causal relationship, such as socio-economic indicators and hydro-meteorological indicators that have no direct connection.

[0051] The optimization goal of contrastive learning is to maximize the cosine similarity of positive sample pairs in the embedding space while minimizing the cosine similarity of negative sample pairs. The two indicator node features of the positive sample pairs are input into the graph neural network to obtain the embedding vectors, and the cosine similarity of the two embedding vectors is calculated. The same operation is performed on the negative sample pairs.

[0052] The loss function is designed as a contrastive loss. For positive sample pairs, a penalty is imposed when the cosine similarity is less than the set high similarity threshold of 0.8, which encourages the model to increase the similarity of positive sample pairs. For negative sample pairs, a penalty is imposed when the cosine similarity is greater than the set low similarity threshold of 0.2, which encourages the model to decrease the similarity of negative sample pairs.

[0053] The Adam optimizer is used to optimize the contrastive loss function, with the learning rate set to 0.001, the batch size set to 32, and the maximum number of training epochs set to 1000. When the validation set loss does not decrease for 20 consecutive epochs, an early stopping mechanism is triggered to stop training and update the parameters of the graph neural network. After training, the model can accurately identify the real correlation between indicators.

[0054] After the model is trained, the input is multidimensional indicator data from the current decision time and the previous T consecutive time steps. The trained dynamic graph neural network outputs a multidimensional indicator correlation graph at the current time. The edge weights in the correlation graph represent the correlation strength between indicators. The weight values ​​are between 0 and 1, with values ​​closer to 1 indicating a stronger correlation and values ​​closer to 0 indicating a weaker correlation.

[0055] The adjacency matrix form of the association graph is extracted to obtain the index association strength matrix. The rows and columns of the matrix correspond to each index, and the matrix element values ​​are the association strength of the corresponding index pair. At the same time, the final embedding vector of each index node after multi-layer graph convolution and temporal modeling is output. The embedding vector integrates the index's own features, neighborhood association information, and temporal evolution pattern, and serves as the index embedding representation of the integrated association information for subsequent decision generation.

[0056] The indicator association strength matrix and indicator embedding representation at the current moment are used as the output of step 2 and passed to steps 3 and 4. For each historical case in the historical decision case database, the indicator data at the corresponding moment is also input into the trained model, and the indicator association strength matrix and node embedding vector at that historical moment are extracted as the association pattern features of the historical case for subsequent knowledge graph construction.

[0057] S3 receives the indicator association strength matrix and indicator embedding representation, and uses scene clustering and graph embedding methods to obtain the decision knowledge graph; S31, Extract the state characteristics and decision-making measures of historical decision-making cases; The system receives the index correlation strength matrix and node embedding vector corresponding to the historical cases output in step 2, as well as the historical decision case database pre-built by the system, and systematically extracts decision rules and experiential knowledge. The historical decision case database records the decision-making process and results of watershed resource allocation over the past few years. Each historical case includes watershed status information at the time of decision-making, the historical decision-making scheme formulated and implemented at that time, and the actual effect evaluation data after the implementation of the historical scheme.

[0058] For each historical case, the watershed state characteristics at the decision-making time are extracted. These watershed state characteristics include water inflow characteristics, which are quantified by the ratio of upstream water inflow to the multi-year average to represent the state of abundance or scarcity; water demand characteristics, which are quantified by the ratio of water demand to available water in each functional area to represent the degree of supply-demand contradiction; and ecological status characteristics, which are quantified by the ratio of ecological flow at key sections to a specified threshold to represent ecological pressure.

[0059] Extract the decision-making measures taken in historical cases. The decision-making measures include water allocation plans for each functional area, recording the water volume allocated to the upstream water conservation area, the midstream agricultural irrigation area, the downstream industrial water use area, and the ecological protection area, as well as the proportion of the total available water volume. The scheduling sequence records the release plan of the allocated water volume in the time dimension.

[0060] Extract the actual effect evaluation data after the implementation of historical decision-making schemes. The effect evaluation data includes the water supply guarantee rate statistics, the percentage of days when the actual water supply in each functional area meets the water demand, the ecological flow satisfaction statistics, the compliance rate of the flow of key sections meeting the ecological flow requirements, and the economic benefit calculation, which includes the sum of agricultural output value, industrial output value, and ecological service value brought about by resource allocation.

[0061] S32, a hierarchical clustering method is used to classify typical decision-making scenarios into categories; The design decision-making scenario clustering method divides massive historical cases into different typical scenario categories to facilitate knowledge organization. For each historical case, the indicator association strength matrix and node embedding vector extracted in step 2 are used as indicator association pattern features. Combined with watershed state features, a scenario feature vector is constructed. The scenario feature vector includes key state indicators such as the indicator association strength matrix flattened into a vector, watershed water flow abundance / dampness index, supply and demand contradiction index, and ecological pressure index.

[0062] Hierarchical clustering algorithm is used to perform cluster analysis on the scene feature vectors of all historical cases. Hierarchical clustering first treats each case as an independent cluster, and then iteratively merges the clusters with the most similar features. The similarity is measured by Euclidean distance. The clustering process continues until the number of clusters reaches a preset value or the distance between clusters is greater than a threshold.

[0063] Based on the actual needs of watershed resource management, eight scenario categories are pre-defined, including typical scenarios such as routine configuration in high-water years, routine configuration in normal-water years, emergency configuration in low-water years, and extreme response in exceptionally low-water years. Historical cases are divided into several typical scenario categories through cluster analysis, and the categories are named according to the main distinguishing factors of the scenario characteristics.

[0064] For each scenario category, the most frequently occurring decision-making action pattern in that category is counted as the typical decision-making strategy for that scenario. The average and variance of the decision-making effect evaluation are calculated, and the successful experiences and lessons learned in that scenario are identified. The key decision-making points corresponding to the successful experiences are extracted.

[0065] S33, Construct the ontology model and entity relationships of the decision knowledge graph; An ontology model for a decision knowledge graph is constructed to achieve a structured representation of decision knowledge. An entity type system for the knowledge graph is designed, defining decision scenario entity types. Each scenario category corresponds to one scenario entity, and the attributes of a scenario entity include scenario name, scenario description, and scenario occurrence frequency. An indicator feature entity type is also defined, with each key watershed status indicator corresponding to one indicator entity. The attributes of an indicator entity include indicator name, indicator value range, and typical value of the indicator in the scenario.

[0066] Define decision-making behavior entity types: Each decision measure corresponds to one behavior entity. The attributes of the behavior entity include measure name, measure parameters, and applicable conditions of the measure. Define effect evaluation entity types: Each evaluation dimension corresponds to one effect entity. The attributes of the effect entity include evaluation indicator name, expected target value, and actual achieved value.

[0067] Design a relation type system for the knowledge graph, define the scenario inclusion relation connecting decision scenario entities and indicator feature entities, define the scenario applicability relation connecting decision scenario entities and decision behavior entities, define the behavior generation relation connecting decision behavior entities and effect evaluation entities, and define the indicator influence relation connecting indicator feature entities and decision behavior entities.

[0068] For each decision case in the historical case database, it is transformed into entities and relationships in the knowledge graph according to the knowledge graph ontology model. The decision scenario category to which the case belongs is identified, and corresponding scenario entities are created or associated in the knowledge graph. Key indicator features of the case are extracted, indicator feature entities are created in the knowledge graph, and containment edges are established between scenario entities and indicator feature entities.

[0069] Extract the decision-making measures taken in the case study, create a decision-making behavior entity in the knowledge graph, establish applicable relationship edges from the scenario entity to the decision-making behavior entity, and establish influence relationship edges from the indicator feature entity to the decision-making behavior entity. Extract the effect evaluation results of the case study, create an effect evaluation entity in the knowledge graph, and establish generation relationship edges from the decision-making behavior entity to the effect evaluation entity.

[0070] By transforming all historical cases, a large-scale decision knowledge graph containing thousands of entity nodes and tens of thousands of relation edges was constructed. This knowledge graph systematically organizes the decision knowledge of watershed resource allocation.

[0071] S34 uses the TransE model to transform the knowledge graph into a vector representation; Knowledge graph embedding technology is employed to transform decision knowledge graphs into vector representations to support efficient knowledge retrieval and reasoning. The translation-based embedding model TransE is used. For each triple in the knowledge graph, consisting of a head entity, a relation, and a tail entity, the head entity, relation, and tail entity are embedded as vectors. The embedding vector dimension is set to 128 dimensions, and the optimization objective is to make the head entity vector plus the relation vector approximately equal to the tail entity vector.

[0072] A training sample set is constructed, where positive samples are real-world triples from the knowledge graph, and negative samples are generated by randomly replacing the head or tail entities in positive triples. The ratio of negative to positive samples is set to 5:1. The loss function is designed as marginal loss, with a marginal margin of 1.0. The vector matching degree is measured by calculating the Euclidean distance between the head entity vector plus the relation vector and the tail entity vector; the smaller the distance, the higher the matching degree.

[0073] The Adam optimizer is used to optimize the loss function for learning entity and relation embedding vectors. The learning rate is set to 0.01, the batch size is set to 128, and the number of training epochs is set to 500. After training, each decision scenario entity, indicator feature entity, decision behavior entity, and effect evaluation entity corresponds to a 128-dimensional vector representation. In the vector space, similar decision scenarios have close vector positions, and related decision behaviors have close vector positions, which facilitates the rapid retrieval of related knowledge through vector similarity.

[0074] The trained knowledge graph vector representation is used as the output of step 3 and passed to step 4 for knowledge retrieval and guidance during decision generation.

[0075] S4 receives the indicator association strength matrix, indicator embedding representation and decision knowledge graph, and adopts a reinforcement learning framework that combines scene matching retrieval and fusion of prior knowledge to obtain a decision generation model and resource allocation decision scheme; S41: Perform intelligent matching between the current scenario and historical cases; The system receives the indicator correlation strength matrix and indicator embedding representation corresponding to the current watershed state output in step 2, and the decision knowledge graph and its vector representation output in step 3, and performs intelligent matching between the current decision scenario and historical reference cases. It extracts the scenario features of the current watershed state, including the indicator correlation strength matrix at the current moment, the current water inflow abundance / scarcity index, the supply-demand imbalance index, and the ecological pressure index, and concatenates these features to form the current scenario feature vector.

[0076] The current scene feature vector is input into the scene encoder, which is a multi-layer feedforward neural network. The network structure is such that the input layer dimension is equal to the scene feature vector dimension, the hidden layer contains two layers with 256 dimensions and 128 dimensions respectively, and the output layer dimension is 128 dimensions, which is consistent with the knowledge graph embedding dimension. The activation function is the ReLU function, and the output is a low-dimensional embedding vector of the current scene.

[0077] Retrieve the historical scene most similar to the current scene embedding vector from the scene entity embedding vector set of the decision knowledge graph, and calculate the cosine similarity between the current scene vector and each historical scene entity vector. The cosine similarity value is between -1 and 1, and the closer the value is to 1, the more similar the scenes are.

[0078] The top 5 historical scenarios with the highest cosine similarity are selected as candidate matching scenarios. For each candidate scenario, the decision-making entities associated with the scenario are retrieved through relational queries of the knowledge graph. The content of decision-making measures and effect evaluation results corresponding to these decision-making entities are extracted. The retrieved decision-making measures are used as reference strategies for the current decision, and the effect evaluation results are used as prior knowledge of the strategy effect.

[0079] S42: Model the sequential decision-making process and define the state-action space; The current watershed resource allocation decision generation problem is modeled as a sequential decision-making process and solved using a reinforcement learning framework. The state space of the sequential decision-making process is defined as follows: the state includes the current state of the watershed's multi-dimensional indicators (normalized values ​​of each indicator); the indicator association pattern features extracted in step 2 (association strength matrix and node embedding vector); the historical decision sequence recording the allocation decisions taken at several past moments; and the knowledge matching degree representing the similarity between the current state and the retrieved reference cases.

[0080] The action space for sequential decision-making is defined, with actions being the water allocation schemes and timing plans for each functional area. To avoid action space combinatorial explosion, a hierarchical decision-making strategy is adopted. First, the overall allocation mode is selected from several preset typical modes, such as the balanced allocation mode, the agriculture priority mode, and the ecological priority mode. Then, under the constraints of the selected mode, the water allocation for each functional area is fine-tuned. The fine-tuning range is limited to ±20% of the mode baseline value and discretized into 5 levels, so that the actual action space size is controlled within a reasonable range.

[0081] The reward function for sequential decision-making is defined to comprehensively evaluate the quality of decision-making. The reward function includes a water supply satisfaction reward item, which calculates the ratio of actual water supply to water demand in each functional area; an ecological protection compliance reward item, which calculates the ratio of ecological flow at key sections to the prescribed threshold; an economic benefit reward item, which calculates the comprehensive economic value of agriculture, industry, and ecology brought about by resource allocation; and a decision stability penalty item, which calculates the change in the current decision compared to the previous decision.

[0082] The total reward is a weighted sum of all reward items. The weights are determined based on the priority of the decision-making objectives. Under normal circumstances, the weights of each objective are balanced, while in extreme cases, the weight of safeguard objectives such as achieving ecological flow targets is increased.

[0083] S43: Constructing an actor critic network architecture that integrates prior knowledge; The agent network architecture for reinforcement learning is constructed using the actor-critic method. The actor network is responsible for policy generation, taking the encoded vector of the current state as input and outputting the probability distribution of each action selection in the action space. The actor network consists of multiple layers of long short-term memory networks and fully connected layers. The long short-term memory network encodes the temporal dependencies of the state sequence, and the fully connected layers output action probabilities. A flexible maximum function is used to transform the output into a probability distribution.

[0084] The critic network is responsible for value assessment. Its input is a state encoding vector, and its output is a value function estimate of that state. The critic network is a multi-layer fully connected network, and the last layer outputs a single scalar as the state value.

[0085] By integrating decision-making knowledge graphs into the actor network, the reference decision-making strategies retrieved during the scene matching process are transformed into prior policy distributions. Decision-making measures that frequently appear in the reference cases are assigned higher probabilities in the prior distribution, and the prior probabilities of decision-making measures with good effect evaluations are further increased.

[0086] The design includes a knowledge-guided attention module with two parallel attention branches. The first branch calculates the semantic matching degree between the current state features and the prior policy distribution and outputs the semantic matching weight. The second branch calculates the numerical similarity between the current state and the reference case state to obtain the numerical matching weight. The matching weights of the two branches are fused through a gating mechanism to obtain the final knowledge-guided weight.

[0087] The final output strategy of the actor network is a weighted fusion of its own learning strategy and prior strategy according to the knowledge guidance weight. The fusion formula is: the probability of the final strategy equals the knowledge guidance weight multiplied by the prior strategy probability plus 1 minus the knowledge guidance weight multiplied by the learning strategy probability, thus realizing the organic combination of data-driven learning and knowledge guidance.

[0088] S44: The decision generation model is trained using a proximal strategy optimization algorithm; A near-end policy optimization algorithm is used to train a reinforcement learning model to generate the optimal decision policy. A simulation environment for watershed resource management is constructed, which is based on hydrological, hydraulic, ecological, and economic models and can simulate the watershed state transition and effect output under a given decision action.

[0089] The training process consists of multiple iterations, each including a data collection phase and a policy update phase. During the data collection phase, the agent interacts with the simulated environment based on the current actor network policy. Starting from the initial watershed state, the actor network outputs an action probability distribution, and specific actions are sampled. The environment then transitions to the next state based on the action and state transition function, calculating reward feedback. This process repeats until the end of the round. The end-of-round condition is reaching the decision time domain endpoint of 365 days or achieving the cumulative reward target.

[0090] The system collects a sequence of four-tuples representing states, actions, rewards, and next states from multiple rounds, with 10 rounds of data collected in each iteration to form an empirical dataset. During the policy update phase, the parameters of the actor network and the critic network are updated based on the collected empirical data. The critic network's update objective is to minimize the value function prediction error, while the actor network's update employs the policy gradient method.

[0091] Proximal policy optimization avoids training instability by limiting the policy update magnitude. The limiting method is policy pruning; pruning occurs when the probability ratio of the new and old policies exceeds 1 plus or minus 0.2. The Adam optimizer is used for parameter updates, with a learning rate of 0.0003 and a training batch size of 64.

[0092] After thousands of iterations of training, training was stopped when the average reward stopped increasing after 50 consecutive iterations. The actor network learned a strategy for making optimal resource allocation decisions under different watershed conditions.

[0093] After the model training is completed, the current actual watershed state is input into the trained actor network. The actor network outputs the action probability distribution, selects the action with the highest probability as the generated decision scheme, decodes the action encoding to obtain the specific water allocation value and allocation time sequence plan of each functional area, and constitutes a complete knowledge-guided resource allocation decision scheme.

[0094] The decision-making scheme clearly stipulates the water supply volume for each period in the future decision-making time domain for the upstream water conservation area, the midstream agricultural irrigation area, the downstream industrial water use area, and the ecological protection area, as well as the timing arrangement of water release, providing the basin management department with a decision-making basis that can be directly implemented.

[0095] S5 receives the resource allocation decision plan, and uses digital twin technology and multi-scenario simulation methods to obtain a comprehensive decision report; S51: Constructing a digital twin model of the watershed resource system; The resource allocation decision scheme generated by the knowledge-guided approach output from step 4 is received, and a digital twin model of the watershed resource system is constructed to support accurate simulation and evaluation of the decision scheme. The digital twin model is constructed using a hybrid modeling method combining physical mechanism modeling and data-driven modeling. The digital twin model comprises five core components: a hydrological sub-model, a hydraulic sub-model, a water quality sub-model, an ecological sub-model, and a socio-economic sub-model. These sub-models are coupled together to form a complete watershed resource system simulation platform.

[0096] A hydrological sub-model was constructed to simulate the precipitation-runoff generation process in the watershed. The sub-model adopted a distributed hydrological model framework, dividing the watershed into several computational units, each with an area of ​​1 square kilometer. The sub-model includes a rainfall interception module using the Horton interception model, with the maximum interception capacity determined based on vegetation type: 5 mm / h for forest, 2 mm / h for grassland, and 1 mm / h for farmland; an infiltration module using the Greenampt infiltration model, with saturated hydraulic conductivity determined based on soil texture: 50 mm / h for sandy soil, 10 mm / h for loam, and 2 mm / h for clay; an evapotranspiration module using the Penman-Montis formula, comprehensively considering meteorological factors such as net radiation, temperature, and wind speed; a runoff generation module employing a combined mechanism of infiltration-excess runoff and storage-saturation runoff; and a confluence module using the kinematic wave equation to simulate slope confluence and the Muskingan method to simulate river confluence. The parameters of the hydrological sub-model were calibrated using five years of historical hydrological observation data and optimized using a genetic algorithm to ensure that the Nash efficiency coefficient (NSE) is greater than 0.75.

[0097] A hydraulic sub-model was constructed to simulate the river flow process. This sub-model describes unsteady flow in an open channel based on the Saint-Venant equations, uses a continuity equation to describe flow variation along the channel, and employs dynamic equations to describe the spatiotemporal evolution of water depth and velocity. The model is solved using the finite difference method, with a spatial step of 1 km and a time step of 1 hour. The Manning formula is used for river roughness, with the roughness coefficient determined based on the riverbed material: 0.035 for gravel, 0.025 for sand, and 0.015 for concrete lining. The hydraulic sub-model outputs the water level and flow processes at various river cross-sections, providing hydrodynamic conditions for the water quality and ecological sub-models.

[0098] A water proton model was constructed to simulate the migration and transformation of pollutants in water bodies. Based on the convection-diffusion reaction equation, the model established migration and transformation equations for two key pollutants: chemical oxygen demand (COD) and ammonia nitrogen (NH3-N). The COD degradation rate coefficient was set at 0.1 per day, and the ammonia nitrogen nitrification rate coefficient at 0.2 per day. The water proton model outputs the pollutant concentration evolution process at various cross-sections of the water body.

[0099] An ecological sub-model was constructed to simulate the ecosystem's response to changes in hydrology and water quality. This sub-model includes a fish habitat suitability model using the Habitat Suitability Index (HSI) method, where HSI is the geometric mean of flow suitability, depth suitability, flow velocity suitability, and water quality suitability; a wetland vegetation growth model simulating vegetation cover changes based on water level and inundation duration, with an optimal inundation depth of 20 to 50 centimeters; and an ecosystem service function assessment model quantifying the value of ecosystem services such as water conservation and biodiversity maintenance. The ecological sub-model outputs indicators such as ecological flow satisfaction, habitat quality index, wetland vegetation cover, and ecosystem service value.

[0100] A socioeconomic sub-model is constructed to simulate the economic benefits and social impacts of resource allocation. This sub-model includes an agricultural production function using the Cobb-Douglas production function, with the output elasticity coefficient of irrigation water supply set to 0.4; an industrial production function, where industrial output significantly decreases when the water supply guarantee rate falls below 0.9, and the output loss rate equals 1 minus the square of the water supply guarantee rate; and a residents' welfare function, where residents' welfare significantly decreases when the adequacy of domestic water supply falls below 0.95. The socioeconomic sub-model outputs indicators such as agricultural output, industrial output, residents' welfare index, and overall economic benefits.

[0101] A data-driven approach was employed to correct systematic errors in the physical mechanism model to improve simulation accuracy. A gradient boosting decision tree (GBDT) model was constructed to construct the error correction model, with parameters set to 100 trees, a maximum depth of 5 layers, and a learning rate of 0.1. After error correction, the simulation accuracy of the digital twin model was significantly improved. On the historical data validation set, the relative error of the simulation of key indicators was controlled within 5%, the NSE of runoff simulation reached 0.85, the relative error of water quality concentration simulation was less than 10%, and the correlation coefficient of ecological indicator simulation was greater than 0.8.

[0102] Real-time monitoring and data assimilation techniques are used to keep the state of the digital twin model synchronized with that of the physical entity. An ensemble Kalman filter (EnKF) method is employed, with 50 ensemble members generated by perturbing the model parameters and initial conditions. When new monitoring data arrives, the Kalman gain is calculated based on the ensemble covariance and the observation error covariance, and the states of the ensemble members are adjusted. The adjusted ensemble mean is used as the optimal state estimate. Through data assimilation, the digital twin model can reflect the true state of the physical watershed in real time, and the state estimation error after data assimilation is reduced by more than 40% compared to the model prediction error.

[0103] S52: Perform full lifecycle simulation of decision-making schemes; The resource allocation decision scheme generated in step 4 is input into the digital twin model constructed in step 51 for full lifecycle simulation. The simulation time domain is set to the next year (365 days), with a time step of 1 day. The spatial scope covers the entire watershed, including the upstream water conservation area, the midstream agricultural irrigation area, the downstream industrial water use area, and the ecological protection area. Water supply boundary conditions for each functional area are set according to the decision scheme. At each time step, the digital twin model performs coupled simulation according to the following process.

[0104] First, the hydrological sub-model is run to simulate the precipitation and runoff generation process at this time step. Meteorological data such as precipitation, temperature, and wind speed at this time step are input. The hydrological sub-model calculates the interception, infiltration, evapotranspiration, and runoff of each calculation unit in the watershed. The runoff of each unit is then summarized to obtain the inflow at the watershed outlet section.

[0105] Then, the hydraulic sub-model is run to simulate the river flow. The inflow output from the hydrological sub-model is used as the upstream boundary condition. The water intake boundary along the river is set according to the water intake of each functional area specified in the decision-making scheme. The hydraulic sub-model solves the Saint-Venant equations to calculate the water level and flow process at each section of the river. Water is supplied from the river and reservoir to each functional area according to the water allocation value specified in the decision-making scheme, and the river flow and reservoir capacity status are updated.

[0106] Next, the water proton model is run to simulate the migration and transformation of pollutants. The flow rate and water level output by the hydraulic sub-model are used as hydrodynamic conditions. The pollutant emission rate at this time step is input, and the water proton model solves the convection-diffusion reaction equation to calculate the COD and ammonia nitrogen concentrations at each section of the water body.

[0107] Then, the ecological sub-model is run to simulate the ecosystem response. The flow and water depth output by the hydraulic sub-model and the water quality concentration output by the water quality sub-model are used as inputs. The ecological sub-model calculates the ecological flow satisfaction of key sections, assesses the fish habitat suitability index, simulates changes in wetland vegetation cover, and quantifies the value of ecosystem services.

[0108] Finally, the socio-economic sub-model is run to simulate the efficiency of resource allocation. Using the actual water supply of each functional area as input, the socio-economic sub-model calculates agricultural output and output value, assesses the water supply guarantee rate and output of industry, and calculates the adequacy of residential water supply and welfare index.

[0109] The simulation outputs detailed state variables for each spatiotemporal node, including the daily flow process lines of each section of the watershed, the daily actual water supply and water demand satisfaction rate of each functional zone, the daily ecological flow compliance status of key sections, the daily COD and ammonia nitrogen concentrations of water quality monitoring points, and the accumulated agricultural output value, industrial output value, and ecosystem service value.

[0110] Statistical analysis was performed on the key indicators of the simulation output. The water supply guarantee rate during the implementation of the decision-making scheme was calculated as the percentage of days in which the water supply met the water demand out of the total 365 days. The ecological flow compliance rate was calculated as the percentage of days in which the ecological flow met the prescribed requirements out of the total days. The annual total economic benefit was calculated as the sum of agricultural output value, processing industry output value, and ecological service value. The water quality compliance rate was calculated as the percentage of days in which the concentration at the water quality monitoring point met the standard.

[0111] Compared with traditional effect evaluation based on simplified models, digital twin simulation can predict the effect of decision-making more comprehensively and accurately, fully considering the multi-process coupling effects of hydrology, water quality, ecology and economy, and significantly improving the credibility and accuracy of simulation results.

[0112] S53: Set up multiple uncertainty scenarios to conduct robustness assessment; To address meteorological uncertainties, the Monte Carlo method was used to generate 1000 sets of stochastic simulations of future precipitation. Each set contains a daily precipitation sequence for the next year, covering a range of precipitation variations from exceptionally low to exceptionally high. Each set of precipitation simulations was then input into a digital twin model to simulate the effectiveness of the decision-making scheme. To address water demand uncertainties, scenarios were set up including a 5% increase in agricultural water demand, a 10% increase in industrial water demand, a combined scenario of both increases, and a scenario of water demand changes due to industrial restructuring. Digital twin simulations were run under each water demand scenario. To address unforeseen event uncertainties, a water pollution accident scenario was set up to simulate a chemical leak at a midstream factory causing continuous water pollution in a river section for 10 days. A water intake facility failure scenario was set up to simulate a water intake being shut down for 15 days due to equipment failure, requiring the activation of a backup water source. The emergency response capability of the decision-making scheme was evaluated. For each of the above uncertainty scenarios, the digital twin model was run to simulate the entire implementation process of the decision-making scheme, and the performance of key indicators was recorded. We statistically analyzed indicators such as water supply guarantee rate, ecological flow compliance rate, and economic benefits under all scenarios, and calculated the mean, standard deviation, minimum value, and 5th percentile to reflect average performance, stability, worst-case scenario, and extreme risk level, respectively.

[0113] The Sobol sensitivity analysis method was used to quantify the sensitivity of decision-making schemes to various uncertainties. Uncertainties were parameterized, and Latin hypercube sampling was employed to generate parameter sample points within the value range of each factor. These sample points were uniformly distributed in the parameter space to ensure coverage of various combinations. For each sample point, a digital twin simulation was run to obtain the decision performance index value. Based on the input parameters and output indicators of all sample points, the first-order sensitivity index and the total sensitivity index of each uncertainty factor on the decision performance were calculated. The first-order sensitivity index measures the contribution of the factor's individual change to the output variance, while the total sensitivity index measures the total contribution of the factor and its interactions with other factors to the output variance. Sensitivity index values ​​range from 0 to 1; a higher value indicates a more sensitive decision-making scheme to that factor. Highly sensitive factors and decision-making periods severely affected by them were identified. For example, the analysis revealed that the sensitivity index of the decision-making scheme to precipitation deviations reached 0.7 during the dry season, indicating that the water supply guarantee capacity during the dry season is very vulnerable to precipitation fluctuations and requires significant improvement.

[0114] S54: Design an adaptive optimization strategy based on the evaluation results; To address the weaknesses identified in sensitivity analysis, decision parameters are adjusted to enhance resilience against disturbances. For periods of insufficient water allocation margin, a strategy of increasing water reserves is adopted. During the high-water season before the arrival of the weak period, reservoir storage is appropriately increased. The amount of water to be stored is determined using an optimization algorithm. The goal is to minimize the risk of insufficient water supply during the weak period while avoiding excessive storage that could affect the current water supply.

[0115] For periods with significant supply-demand imbalances, a dynamic adjustment strategy for allocation ratios will be adopted. During the dry season, priority will be given to ensuring domestic water and ecological base flow, agricultural water will be supplied at 80% of the required level, and industrial water will be supplied in stages according to the urgency of demand.

[0116] For highly sensitive critical decision-making nodes, multiple alternative schemes are set up and switching trigger conditions are established. When the water intake flow rate is lower than the threshold or the water quality exceeds the standard in real time, the switching conditions are automatically triggered to activate the backup scheme.

[0117] In response to sudden pollution incidents, emergency dispatch plans are set up, including increasing the discharge upstream of the pollution source to dilute the concentration of pollutants, suspending water intake in the polluted river section and using clean water sources instead, and activating emergency water source protection in downstream functional areas.

[0118] The optimized decision-making scheme was re-input into the digital twin model for simulation verification to evaluate the optimization effect. Multiple scenario simulations, identical to those before optimization, were run to calculate the performance indicators of the optimized scheme under each scenario, comparing changes in key indicators such as water supply guarantee rate, ecological flow compliance rate, economic benefits, and sensitivity index before and after optimization.

[0119] If the robustness indicators of the optimized scheme are significantly improved, such as the minimum water supply guarantee rate increasing from 85% to 92%, the standard deviation of the ecological flow compliance rate decreasing from 0.15 to 0.08, and the maximum sensitivity index decreasing from 0.7 to 0.4, then the optimization is considered successful, and the scheme is output as the final decision.

[0120] If the optimization results are unsatisfactory, analyze the reasons and adjust the optimization strategy, then optimize and verify again to form an iterative closed loop. Set the iteration termination condition as follows: all key indicators meet the set threshold in all scenarios, or the number of iterations reaches the upper limit of 10.

[0121] After iterative optimization, a comprehensive decision report is generated, which includes multi-scenario simulation results, robustness evaluation indicators, sensitivity analysis results, and optimization adjustment suggestions. The decision report provides decision-makers with a comprehensive view of the expected effects, potential risks, and countermeasures of the decision-making plan, and provides a scientific basis for decision implementation.

[0122] S6 receives the indicator correlation strength matrix and indicator embedding representation, decision generation model and decision scheme, and comprehensive decision report. It uses attention visualization and counterfactual reasoning to obtain an interactive visualization interface and an interpretable report. S61: Visualize the weight distribution of multi-dimensional indicators; The system receives the current-time indicator association strength matrix and indicator embedding representation from step 2, the trained decision generation model and knowledge-guided resource allocation decision scheme from step 4, and the comprehensive decision report including robustness evaluation indicators and sensitivity analysis results from step 5. It extracts the attention weights and association strength information within the model to generate decision interpretability analysis. A force-directed layout algorithm is used to visualize the indicator association graph. The algorithm attracts nodes with strong associations and repels nodes with weak associations. In the visualization, node size represents the importance of the indicator using degree centrality. Node color represents the dimension to which the indicator belongs: blue for hydrological and meteorological categories, green for resource supply and demand categories, yellow for ecological and environmental categories, and red for socio-economic categories. Edge thickness represents the association strength: thick lines for values ​​greater than 0.7, medium-thick lines for values ​​between 0.4 and 0.7, and thin lines or no lines for values ​​less than 0.4. The generated indicator association network graph visually demonstrates which indicators in the watershed resource system have strong associations.

[0123] S62: Extract the attention weights of the decision generation model; The attention weight vector calculated by the actor network when generating the current decision is extracted. Each element of the weight vector corresponds to an attention score for an input feature; a higher score indicates a greater influence of that feature on the decision. The attention weights are plotted as a heatmap, with rows corresponding to time steps in the decision-making time domain and columns corresponding to input features. The color depth of the cells indicates the magnitude of the attention weight, with dark red representing high attention and light yellow representing low attention. The heatmap can identify which key indicators the model focuses on during the decision-making process. For example, during dry season decisions, the model's attention weight for water inflow, ecological flow, and agricultural water demand is significantly higher than other indicators, indicating that the model recognizes these as key constraints for decision-making during dry seasons.

[0124] S63: Use counterfactual reasoning to quantify the impact of decision variables; For key decision variables in the decision schemes generated in step 4, such as the allocated water volume for the midstream agricultural irrigation area, a counterfactual scenario analysis is constructed to analyze their impact. In the original scheme, the allocated water volume for the agricultural irrigation area is a baseline value. Counterfactual scenario 1 increases this allocated water volume by 10%, and counterfactual scenario 2 decreases it by 10%, while keeping the allocated water volumes for other functional areas and other decision parameters unchanged. The original scheme and the two counterfactual scenarios are input into the digital twin model simulation in step 5 to calculate key indicators such as water supply guarantee rate, ecological flow compliance rate, and economic benefits for the three schemes. The differences in indicators between the original scheme and the counterfactual scenarios are compared to quantify the impact of adjusting decision variables. Through counterfactual comparison, the multi-objective trade-offs of decision variables are revealed. For example, increasing agricultural water supply is beneficial to agricultural production and economic benefits but will crowd out ecological water use; reducing agricultural water supply is beneficial to ecological protection but will harm agricultural benefits. The baseline value selected in the original scheme achieves a balance among multiple objectives. Counterfactual analysis is performed on all key decision variables, including the allocated water volume for each functional area, reservoir scheduling rules, and emergency activation thresholds, to identify the key variables with the greatest impact on the decision-making effect.

[0125] S64: Generate natural language decision explanation texts and interactive systems; Generate natural language decision explanation text for decision-makers. The explanation text is divided into four parts: scenario description, decision considerations, key variable description, and risk warning. The scenario description section describes the main characteristics of the current watershed status based on the indicator data from step 1 and the correlation analysis results from step 2.

[0126] The decision-making considerations section describes the main basis for the decision based on the knowledge graph retrieval in step 3 and the decision generation process in step 4. The key variable description section describes the rationale for setting key decision variables and the multi-objective trade-offs based on the counterfactual analysis results.

[0127] The risk warning section, based on the robustness assessment in step 5, describes the potential risks and countermeasures of the decision-making plan. The four sections are then integrated to form a complete decision explanation report, with a length of 800 to 1200 words, using concise and accurate language to avoid technical jargon and ensure the decision-maker's comprehension.

[0128] Develop an interactive decision interpretation system to support decision-makers' in-depth exploration needs. Create a web-based visualization interface; the main interface displays an indicator correlation network diagram, an attention heatmap, a decision scheme parameter table, and effect prediction indicator cards.

[0129] It provides interactive functionality, allowing decision-makers to click on any indicator node in the network diagram to view detailed information, including the current value, historical trend curve, ranking of correlation strength with other indicators, and explanation of the indicator's role in decision-making.

[0130] Decision-makers can click on any decision variable in the decision-making scheme to view the counterfactual analysis results, including the impact curves of variable adjustments on each target indicator, sensitivity coefficients, and explanations of the setting basis. Decision-makers can adjust parameters of interest to conduct hypothetical scenario simulations, and the system will call the surrogate model in real time to predict the effect of the adjusted decision and update the display.

[0131] The system provides a decision-making option comparison function, allowing decision-makers to load multiple candidate options. The system then displays radar charts and indicator comparison tables showing the performance of each option under different scenarios. It also provides a decision interpretation report export function, allowing decision-makers to export complete reports, including visual charts and text descriptions, as document format.

[0132] Interactive systems enhance the transparency and credibility of decision-making, enabling decision-makers not only to know what the system recommends, but also to understand why the decision was made that way, thus giving them the ability to intervene and adjust the decision.

[0133] S7 receives real-time monitoring data and effect feedback during the execution of the decision-making scheme, and uses model predictive control and online learning methods to obtain dynamically updated decision-making schemes and iteratively optimized model parameters. S71: Collect real-time monitoring data and identify deviations; The system receives real-time monitoring data and effect feedback from the resource allocation decision scheme output in step 4 during actual implementation, constructing a closed-loop feedback mechanism for decision implementation to achieve dynamic optimization of the decision scheme. During the actual implementation of the decision scheme, real-time data is continuously collected through a distributed monitoring network across the watershed. This real-time monitoring data includes flow and water level monitoring at each river section, actual water intake monitoring at each intake point, actual water supply and consumption monitoring in each functional area, ecological flow and water quality monitoring at key sections, and real-time precipitation and evaporation monitoring from meteorological stations. The monitoring data is transmitted to the data center of the decision support system via the Internet of Things (IoT), with a data collection frequency of hourly to ensure timely capture of changes in the watershed status. The real-time monitoring data undergoes quality checks to identify data quality issues such as sensor malfunctions, communication interruptions, and outliers. Abnormal data is repaired using time-series interpolation or spatial interpolation methods from nearby stations to ensure data integrity and reliability. The real-time monitoring data is compared with the expected state of the decision scheme to identify deviations.

[0134] S72: Rolling optimization is performed using model predictive control; When formulating the decision-making plan, it is based on the predicted future watershed state trajectory. As time goes by, the actual watershed state monitored in real time is compared with the predicted state, and the deviation of key indicators is calculated, such as the percentage deviation between the actual water inflow and the predicted water inflow, the percentage deviation between the actual water demand and the predicted water demand, and the deviation between the actual value and the target value of ecological flow.

[0135] Set deviation warning thresholds. When a certain indicator deviates beyond the threshold, an warning is triggered. For example, a deviation of more than 20% in water inflow triggers a hydrological warning, and a deviation of more than 10% in ecological flow triggers an ecological warning. Analyze the causes of the deviations, distinguishing between those caused by changes in uncertain factors such as abnormal precipitation, and those caused by unreasonable design of the decision-making scheme itself.

[0136] For deviations caused by uncertainty, the adaptive adjustment strategies and alternative solutions preset in step 5 are activated, such as supplementing water supply with reserve water or switching to a backup water source. For deviations caused by design flaws in the decision-making scheme, the decision-making scheme needs to be corrected and optimized.

[0137] The decision-making parameters are dynamically adjusted based on real-time state deviations to adapt to changes in actual conditions. A model predictive control method is used for rolling optimization. At each decision adjustment moment, the current real-time monitored watershed state is used as the initial condition, the inflow forecast for the future period is updated based on the latest meteorological forecast, and the future water demand forecast is updated based on the real-time water demand report. The updated state and forecasts are input into the decision generation model to regenerate the optimized decision-making scheme for the future period.

[0138] The regenerated plan incorporates the latest information and is more realistic than the original plan. Comparing the new plan with the original plan, if the new plan's expected effect is significantly better than the original plan and the adjustment range is within acceptable limits, then the new plan will be adopted to replace the original plan, and the water supply scheduling instructions issued to each functional area will be updated.

[0139] If drastic adjustments to the new plan may affect execution stability, a gradual approach will be adopted, transitioning to the new plan over multiple time periods to avoid drastic changes. The rolling optimization cycle is determined based on the watershed management needs; daily optimization is possible during the flood season when water conditions change rapidly, while weekly optimization is possible during the relatively stable dry season, balancing decision adaptability and execution stability. Feedback data on the effectiveness of the decision-making plan is collected for continuous model learning and iterative optimization.

[0140] S73: Collect feedback data on the collection results for model iteration and optimization; After a period of implementation, the actual performance indicators are statistically analyzed, including the actual water supply guarantee rate of each functional area, the actual ecological flow compliance rate of key sections, the actual economic benefits brought by resource allocation, and the timeliness and stability of decision implementation. The input conditions, parameters, and actual implementation effects of the decision scheme are used to construct new decision cases, which are then added to the historical decision case database. These new cases contain the latest decision experience and effect verification. Based on the expanded case library, the decision knowledge graph from step 3 is incrementally updated, transforming new cases into entities and relationships within the knowledge graph. If a new case reveals a new type of decision scenario, a new scenario entity is created in the knowledge graph. If a new case verifies that a certain decision measure is effective in a certain scenario, the recommendation weight of that decision measure is increased. The decision generation model from step 4 is incrementally updated using an online learning method. New cases are used as training samples, and actual implementation effects are used as labels to fine-tune the model parameters. Fine-tuning uses a small learning rate to avoid destroying already learned knowledge. Through incremental learning, the model continuously absorbs new experience and improves its decision-making capabilities. Regularly evaluate the performance of the updated model and assess decision quality metrics on independent test sets to ensure that model performance continues to improve rather than overfitting new data.

[0141] Through long-term closed-loop feedback and iterative optimization, the decision support system has continuously improved its decision quality and enhanced its adaptability to watershed characteristics, gradually forming a customized intelligent decision-making capability for the watershed, providing strong technical support for the scientific management and sustainable utilization of watershed resources.

[0142] In one embodiment of the present invention, an application example is provided, focusing on the comprehensive allocation of water resources in a large, trans-regional river basin. The basin includes an upstream water conservation area of ​​5,000 square kilometers, a midstream agricultural irrigation area of ​​3,000 square kilometers planted with wheat and corn, a downstream industrial water use area comprising several manufacturing parks, and an important wetland ecological protection area of ​​800 square kilometers. The basin has an average annual precipitation of 600 mm, an average annual runoff of 2 billion cubic meters, and an average annual water demand of 1.8 billion cubic meters for all functional areas. Of this, agricultural water demand accounts for 55%, industrial water demand for 30%, domestic water demand for 10%, and ecological water demand for 5%. The overall water supply and demand in the basin is balanced, but the spatial and temporal distribution is uneven.

[0143] Examples of the obtained application data are shown in Tables 1 and 2.

[0144] Table 1: Examples of Multidimensional Indicator Monitoring Data for Typical Time Periods in Watersheds

[0145] Table 2: Examples of Water Allocation Results Generated by Intelligent Decision-Making Schemes

[0146] The decision-making scheme generated using the method of this invention, during the spring irrigation season, appropriately reduces industrial water use and increases agricultural allocation to meet the peak water demand of agriculture, while ensuring ecological base flow. The total allocated water volume of 520 million cubic meters is slightly higher than the inflow of 420 million cubic meters. By supplementing with reservoir reserves, a good water supply guarantee rate of 0.91 is achieved. During the summer flood season, the abundant inflow of 850 million cubic meters of water is fully utilized, and the water demand of all functional areas is adequately met. The total allocation is 550 million cubic meters, and ecological flow is increased to protect wetland ecology, achieving a water supply guarantee rate of 0.98. During the autumn normal water season, the inflow is 380 million cubic meters, and the total allocation is 420 million cubic meters. Through reasonable scheduling, a water supply guarantee rate of 0.95 is achieved. During the winter dry season, the inflow is 210 million cubic meters, and the total allocation is 300 million cubic meters. Through reservoir regulation and priority allocation, a water supply guarantee rate of 0.88 is achieved. Facing severe water shortages during the dry season of an extremely dry year, with an inflow of only 120 million cubic meters, the decision-making plan prioritized ensuring the essential water needs of domestic use, significantly reduced agricultural irrigation water use, moderately reduced industrial water use, and maintained the minimum ecological flow standard, allocating a total of 240 million cubic meters. Through reasonable emergency allocation strategies and reservoir regulation, a water supply guarantee rate of 0.75 was still achieved under extreme conditions, averting a water supply crisis. The implementation of the decision-making plan effectively guaranteed the water needs of various functional zones in the basin, maintained important ecological functions, promoted the sustainable use of water resources in the basin, and verified the effectiveness and practicality of the method of this invention.

[0147] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation, characterized in that, Includes the following steps: S1. Collect raw data of multi-dimensional watershed state indicators, and use heterogeneous data type identification and multi-level normalization processing methods to obtain a set of multi-dimensional indicator feature vectors. S2, obtain a set of multi-dimensional indicator feature vectors, and use dynamic graph neural network and contrastive learning to obtain the indicator correlation strength matrix and indicator embedding representation; S3 receives the indicator association strength matrix and indicator embedding representation, and uses scene clustering and graph embedding methods to obtain the decision knowledge graph; S4 receives the indicator association strength matrix, indicator embedding representation and decision knowledge graph, and adopts a reinforcement learning framework that combines scene matching retrieval and fusion of prior knowledge to obtain a decision generation model and resource allocation decision scheme; S5 receives the resource allocation decision plan, and uses digital twin technology and multi-scenario simulation methods to obtain a comprehensive decision report; S6 receives the indicator correlation strength matrix and indicator embedding representation, decision generation model and decision scheme, and comprehensive decision report. It uses attention visualization and counterfactual reasoning to obtain an interactive visualization interface and an interpretable report. S7 receives real-time monitoring data and effect feedback during the execution of the decision-making scheme, and uses model predictive control and online learning methods to obtain dynamically updated decision-making schemes and iteratively optimized model parameters.

2. The method for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation according to claim 1, characterized in that, S1 includes: Data on indicators from the hydrological and meteorological, resource supply and demand, ecological and environmental, socio-economic, and policy constraint dimensions of a multi-source monitoring network are collected. Data are then filled in using interpolation based on the missing rate, and outliers are identified using the standard deviation criterion. An indicator type identification model is constructed to classify indicators into numerical continuous indicators, numerical discrete indicators, and categorical indicators, and the maximum-minimum value normalization, standard deviation normalization, and one-hot coding methods are used to process them respectively. The initial weights of the indicators are determined based on information entropy values ​​and domain expert experience scores, and then dynamically adjusted using the coefficient of variation method. Extract time series features and spatial distribution features, and concatenate them to form a set of multidimensional indicator feature vectors.

3. The method for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation according to claim 1, characterized in that, S2 includes: Each indicator is treated as a node in the graph. Directed edges are established for indicator pairs with physical causal relationships based on domain knowledge. The Pearson correlation coefficient of the indicator time series is calculated. When the absolute value of the correlation coefficient is greater than the preset association threshold, an undirected edge is established. Design a multi-layer dynamic graph convolutional network, use an attention mechanism to calculate the aggregate weights of neighboring nodes, and achieve information propagation layer by layer by stacking multi-layer graph convolution operations; Gated cyclic units are used to model time series data, and the hidden state that integrates historical association patterns and current association patterns is output. Design an association mutation detection module to calculate the distance between the current time and the historical average association strength matrix, and trigger an association mutation flag when the distance exceeds a dynamic threshold. Construct positive and negative sample pairs for comparative learning training, and output the index association strength matrix and index embedding representation.

4. The method for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation according to claim 1, characterized in that, S3 includes: For each historical case, extract watershed status features, decision-making measures content, and effect evaluation data, and construct a scenario feature vector by combining the indicator correlation strength matrix and node embedding vector; Hierarchical clustering algorithm is used to perform cluster analysis on scene feature vectors, and historical cases are divided into several typical scene categories; Define decision-making scenario entities, indicator characteristic entities, decision-making behavior entities, and effect evaluation entities; design scenario inclusion relationships, scenario applicability relationships, behavior generation relationships, and indicator impact relationships; and transform historical cases into entities and relationships in the graph. We employ a translation-based embedding model for training and output a vector representation of the decision knowledge graph.

5. The method for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation according to claim 1, characterized in that, S4 includes: The current scene feature vector is input into the scene encoder to output the embedding vector. The historical scene with the highest cosine similarity is retrieved in the decision knowledge graph as the candidate matching scene, and the decision behavior entity associated with the candidate scene is retrieved as the reference strategy. Define the state space and action space for sequential decision-making, and adopt a hierarchical decision-making strategy; The reward function is defined to include rewards for water supply satisfaction, rewards for ecological protection compliance, rewards for economic benefits, and penalties for decision-making stability. We construct actor networks and critic networks, design a knowledge-guided attention module, and the final output strategy of the actor network is a weighted fusion of its own learning strategy and prior strategy according to the knowledge-guided weight. The algorithm is trained using a near-end strategy optimization algorithm, and outputs a decision generation model and a resource allocation decision scheme.

6. The method for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation according to claim 1, characterized in that, S5 includes: A digital twin model is constructed, which includes a hydrological sub-model, a hydraulic sub-model, a water quality sub-model, an ecological sub-model, and a socio-economic sub-model. An error correction model is constructed using a data-driven approach, and data assimilation is performed using an ensemble Kalman filter method. Input the resource allocation decision scheme into the digital twin model for coupled simulation; Multiple scenarios are set up to address uncertainties in weather, water demand, and unforeseen events. Sensitivity analysis is used to calculate the sensitivity index of uncertain factors on the decision-making effect, and decision parameters are adjusted to target weak links to generate a comprehensive decision report.

7. The method for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation according to claim 1, characterized in that, S6 includes: A force-directed layout algorithm is used to visualize the index correlation graph; The attention weight vectors of the actor network are extracted and plotted as a heatmap; Construct counterfactual scenarios for key decision variables, input the original and counterfactual scenarios into a digital twin model for simulation, and quantify the impact of adjusting decision variables; Generate natural language decision explanation text and build an interactive decision explanation system.

8. The method for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation according to claim 1, characterized in that, S7 includes: Real-time monitoring data is collected through a distributed monitoring network across the watershed. The deviation of key indicators is calculated by comparing real-time monitoring data with the expected state of the decision-making plan. When the deviation exceeds the preset warning threshold, an early warning is triggered. A rolling optimization method is adopted using model predictive control, which regenerates the optimization decision scheme based on the current real-time monitored watershed status and the latest forecast.

9. The method for generating intelligent auxiliary decisions for natural resources based on multi-dimensional index correlation according to claim 8, characterized in that, The S7 also includes: The input conditions, parameters, and actual implementation effects of the decision-making schemes are used to create new decision-making cases and supplement the historical decision-making case database. Incremental updates to the decision knowledge graph based on an expanded case library; Online learning methods are used to fine-tune and update the parameters of the decision generation model.

10. A multi-dimensional index-based intelligent auxiliary decision-making generation system for natural resources, used to execute the steps in the multi-dimensional index-based intelligent auxiliary decision-making generation method for natural resources as described in any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing module is used to collect raw data of multi-dimensional watershed state indicators, and uses heterogeneous data type identification and multi-level normalization processing methods to obtain a set of multi-dimensional indicator feature vectors. The indicator association mining module is used to obtain a multi-dimensional indicator feature vector set. It uses dynamic graph neural network and contrastive learning to obtain the indicator association strength matrix and indicator embedding representation. The knowledge graph construction module receives the indicator association strength matrix and indicator embedding representation, and uses scene clustering and graph embedding methods to obtain the decision knowledge graph; The decision generation module receives the indicator association strength matrix, indicator embedding representation, and decision knowledge graph. It uses a reinforcement learning framework that combines scene matching retrieval and fusion of prior knowledge to obtain the decision generation model and resource allocation decision scheme. The simulation evaluation module receives resource allocation decision schemes and uses digital twin technology and multi-scenario simulation methods to obtain a comprehensive decision report. The interpretability analysis module receives the index correlation strength matrix and index embedding representation, decision generation model and decision scheme, and comprehensive decision report. It uses attention visualization and counterfactual reasoning to obtain an interactive visualization interface and interpretable report. The dynamic optimization module receives real-time monitoring data and effect feedback during the execution of the decision-making scheme. It uses model predictive control and online learning methods to obtain dynamically updated decision-making schemes and iteratively optimized model parameters.

Images