Method and system for dynamic prediction and sustainable management of groundwater resources based on multi-source data fusion

Abstract

The application discloses to the technical field of groundwater resources, and particularly relates to a method and system for dynamic prediction and sustainable management of groundwater resources based on multi-source data fusion, which comprises the following steps: firstly, collecting multi-dimensional heterogeneous data related to the groundwater system, and outputting a complete preprocessed data set after quality evaluation, spatio-temporal alignment and missing value completion; then, performing feature extraction, causal modeling, multi-scale fusion and partition adaptation, and outputting a real-time updated fusion feature library; then, identifying driving factors and quantifying lag effects, constructing a physical and data dual-driven hybrid prediction model, dynamically calibrating the model, carrying out multi-scenario simulation deduction, quantifying prediction uncertainty and tracing optimization, and outputting a final prediction report and uncertainty optimization suggestions; based on the above, the application realizes synchronous online dynamic calibration of physical and data models, and effectively solves the problems of incomplete driving factor identification, model calibration lag and insufficient adaptability to extreme scenarios in the prior art.

Description

Technical Field

[0001] This invention relates to the field of groundwater resources technology, specifically to a method and system for dynamic prediction and sustainable management of groundwater resources based on multi-source data fusion. Background Technology

[0002] Groundwater resources are the most important water source for the plains of northern my country, the karst regions of southwestern China, and the arid regions of northwestern China, supporting approximately 60% of urban domestic water use and 40% of agricultural irrigation water nationwide. With the intensification of climate change and rapid economic and social development, problems such as over-extraction, pollution, and continuous decline in groundwater levels are becoming increasingly severe, posing a key bottleneck to regional sustainable development. Accurately understanding the spatiotemporal evolution patterns of groundwater resources and achieving dynamic prediction and refined management are major technical challenges that urgently need to be addressed in the field of water resources.

[0003] Significant progress has been made in groundwater prediction and management using existing technologies. For example, CN121599306A discloses a method for spatiotemporal dynamic monitoring and visualization analysis of groundwater levels, achieving multi-source data fusion, high-precision spatial interpolation, spatiotemporal dynamic visualization, and anomaly early warning; CN121279740A proposes a dynamic control method for groundwater levels based on digital twins, which realizes the simulation and risk quantification of groundwater systems by constructing a multi-field dynamically coupled digital twin model, and establishes a two-level closed-loop feedback mechanism; CN122022588A discloses a dynamic early warning method for groundwater ecological threshold classification that integrates multi-source indicators, and constructs a multi-scenario dynamic early warning mechanism; CN122113325A proposes a method for constructing and optimizing a unified groundwater monitoring network, which uses a machine learning model to realize the intelligent dynamic adjustment of the monitoring network.

[0004] However, the aforementioned existing technologies still have the following shortcomings: First, the prediction models of existing technologies are mostly based on a single physical field or a coupling of a few physical fields, lacking full-scale fusion and deep feature extraction of multi-source heterogeneous data of the groundwater system, resulting in incomplete identification of driving factors and inaccurate quantification of lag effects; Second, the calibration mechanisms of existing prediction models are mostly offline or periodically updated, making it difficult to achieve synchronous online dynamic calibration of physical models and data models, resulting in insufficient adaptability of the models under extreme hydrological scenarios; Third, although existing technologies have closed-loop feedback or dynamic control capabilities, there is a lack of deep integration between the prediction and management decision-making processes, and the uncertainty of the prediction results is not effectively transmitted to the management decision-making process, resulting in limited scientific validity and timeliness of management solutions.

[0005] Specifically, in the construction of existing prediction models, the identification of driving factors often relies solely on simple correlation analysis, failing to fully consider the lag effect of driving factors on groundwater level changes and their spatiotemporal heterogeneity. Model calibration is mostly based on batch training with historical data, lacking the ability to learn incrementally from new data. Multi-scenario simulation and uncertainty quantification are relatively independent, failing to form an organic whole of "prediction-uncertainty assessment-management decision-making." For example, although some regions have dynamic monitoring and early warning capabilities, in extreme scenarios where actual precipitation is significantly lower than predicted, the failure to calibrate the prediction model in a timely manner and the lack of sufficient uncertainty quantification lead to a lack of sufficient scientific basis for management decisions, making it impossible to reduce mining quotas in a timely manner, resulting in a greater-than-expected drop in groundwater levels. Summary of the Invention

[0006] To address the shortcomings of existing technologies in driver identification and lag effect quantification, model dynamic calibration, and deep integration of prediction and management, this invention provides the following technical solution:

[0007] A method for dynamic prediction and sustainable management of groundwater resources based on multi-source data fusion includes the following steps:

[0008] S1, Multi-source heterogeneous data acquisition and preprocessing: Comprehensive acquisition of multi-dimensional heterogeneous data related to the groundwater system, and output of a complete preprocessed dataset after quality assessment, spatiotemporal alignment and missing value completion;

[0009] S2, Spatiotemporal Multiscale Data Fusion Modeling: Performs feature extraction, causal modeling, multiscale fusion and partition adaptation on the complete preprocessed dataset, and builds and outputs a real-time updated fusion feature library;

[0010] S3, Coupled-driven dynamic prediction simulation:

[0011] S31, Driving Factor Identification: Based on a real-time updated fusion feature library, Pearson correlation coefficient and grey relational analysis methods are used to identify driving factors and quantify their contribution, outputting a set of key driving factors;

[0012] S32, Driving lag analysis: Based on the key driving factor set, cross wavelet transform and cross-correlation function are used to quantify lag duration and lag strength, and output a driving factor set with lag characteristics;

[0013] S33, Prediction Model Construction: Based on the set of driving factors with hysteresis features, construct a hybrid prediction model driven by both physical and data aspects, and output the initial groundwater level prediction model.

[0014] S34, Dynamic Model Calibration: Based on the initial groundwater level prediction model, incremental learning and data assimilation methods are used to achieve synchronous online updates of the physical model and the data model, and output the dynamically calibrated prediction model.

[0015] S35, Multi-scenario Simulation and Inference: Based on the dynamically calibrated prediction model, various typical scenarios are set to infer the spatiotemporal evolution trend of groundwater level and output a multi-scenario prediction result set;

[0016] S36, Quantification of Prediction Uncertainty: Based on the multi-scenario prediction result set, Monte Carlo simulation is used to quantify uncertainty and calculate confidence intervals, and a preliminary prediction report with confidence level is output.

[0017] S37, Uncertainty Source Tracing and Optimization: Based on the preliminary forecast report with confidence level, the variance decomposition method is used to decompose the total uncertainty into three sources: data, model parameters and scenarios, and quantify their contribution ratio. Uncertainty reduction measures are proposed, and the final forecast report and uncertainty optimization suggestions are output.

[0018] S4, Closed-loop Sustainable Management Decision-Making: Based on the final forecast report and uncertainty optimization suggestions, the groundwater carrying capacity is dynamically assessed, multi-objective optimization configuration is carried out, an executable management plan is generated, and the management plan is iteratively optimized through real-time monitoring and feedback.

[0019] As a preferred embodiment of the method for dynamic prediction and sustainable management of groundwater resources based on multi-source data fusion described in this invention, the specific steps of S1 are as follows:

[0020] S11, Multi-source data aggregation: Collects hydrogeological data, dynamic monitoring data, meteorological and environmental data, and socio-economic data, and outputs raw multi-source datasets;

[0021] S12, Data Quality Assessment: Based on the original multi-source dataset, outlier detection algorithms are used to identify outliers, remove erroneous data and mark missing data, and output a quality assessment report and the cleaned dataset;

[0022] S13, Spatiotemporal Alignment and Normalization: Based on the quality assessment report and the cleaned dataset, the data is unified to the same spatiotemporal coordinate system and the difference in units is eliminated, outputting a standardized dataset with spatiotemporal alignment.

[0023] S14, Intelligent missing value completion: Based on a standardized dataset with spatiotemporal alignment, a spatiotemporal generative adversarial network is used to complete missing data and output a complete preprocessed dataset.

[0024] As a preferred embodiment of the method for dynamic prediction and sustainable management of groundwater resources based on multi-source data fusion described in this invention, the specific steps of S2 are as follows:

[0025] S21, Feature extraction and encoding: Based on the complete preprocessed dataset, convolutional neural networks, long short-term memory networks and graph neural networks are used to extract spatial, temporal and topological features, and output a multi-source heterogeneous feature set;

[0026] S22, Hierarchical Causal Modeling: Based on multi-source heterogeneous feature sets, a convergent cross-mapping algorithm is used to identify multi-scale causal relationships, construct a hierarchical causal graph and remove false association features, and output a feature set with causal constraints;

[0027] S23, Multi-scale feature fusion: Based on the output feature set with causal constraints, a spatiotemporal attention mechanism is used to perform cross-scale fusion and output a fused feature vector;

[0028] S24, fusion effect verification: Based on the fusion feature vector, the mutual information method and redundancy analysis are used to remove redundant features and output the optimized core feature set;

[0029] S25, Zonal Feature Adaptation: Based on the optimized core feature set, combined with hydrogeological zoning, a feature importance ranking method is used to select a zonal-specific core feature subset and output the zonal core feature set.

[0030] S26, Feature Library Construction and Update: Based on the partitioned core feature set, construct a partitioned dynamic feature library and set an incremental update mechanism to output a fusion feature library that is updated in real time.

[0031] As a preferred embodiment of the method for dynamic prediction and sustainable management of groundwater resources based on multi-source data fusion described in this invention, the specific steps of step S4 are as follows:

[0032] S41, Dynamic assessment of bearing capacity: Based on the final prediction report and uncertainty optimization suggestions, combined with regional hydrogeological conditions and socio-economic development plans, the bearing capacity is dynamically assessed using the index system method, and overloaded areas and warning levels are identified, and a bearing capacity assessment report is output.

[0033] S42, Multi-objective optimization configuration: Based on the carrying capacity assessment report, with the objectives of groundwater level restoration, water supply security and minimization of economic losses, a multi-objective genetic algorithm is used to solve the optimal water resource allocation scheme and output a preliminary optimized configuration scheme.

[0034] S43, Management Plan Generation: Based on the preliminary optimized configuration plan, combined with the actual situation of the region and the demands of multiple stakeholders, the plan is broken down and implemented at the administrative region, industry and water user level, differentiated control measures are formulated, and an executable sustainable management plan is output.

[0035] S44, Feedback Iterative Optimization: Based on an executable sustainable management plan, real-time monitoring data after execution is collected and compared with predicted data to dynamically correct the predicted model parameters and management plan, and output the iteratively optimized management plan.

[0036] A dynamic prediction and sustainable management system for groundwater resources based on multi-source data fusion includes:

[0037] The multi-source heterogeneous data acquisition and preprocessing module comprehensively collects multi-dimensional heterogeneous data related to the groundwater system. After quality assessment, spatiotemporal alignment, and missing value completion, it outputs a complete preprocessed dataset.

[0038] The spatiotemporal multi-scale data fusion modeling module performs feature extraction, causal modeling, multi-scale fusion, and partition adaptation on the complete preprocessed dataset, and builds and outputs a fusion feature library that is updated in real time.

[0039] The coupled-driven dynamic prediction simulation module includes a driving factor identification unit, a driving lag analysis unit, a prediction model construction unit, a model dynamic calibration unit, a multi-scenario simulation and deduction unit, a prediction uncertainty quantification unit, and an uncertainty source tracing and optimization unit.

[0040] The driving factor identification unit, based on a real-time updated fusion feature library, uses Pearson correlation coefficient and grey relational analysis methods to identify driving factors and quantify their contribution, and outputs a set of key driving factors.

[0041] The driving lag analysis unit, based on the key driving factor set, uses cross wavelet transform and cross-correlation function to quantify lag duration and lag intensity, and outputs a driving factor set with lag characteristics.

[0042] The prediction model construction unit constructs a physical and data-driven hybrid prediction model based on a set of driving factors with hysteresis features, and outputs an initial groundwater level prediction model.

[0043] The model dynamic calibration unit, based on the initial groundwater level prediction model, uses incremental learning and data assimilation methods to achieve synchronous online updates of the physical model and the data model, and outputs the dynamically calibrated prediction model.

[0044] The multi-scenario simulation and deduction unit, based on the dynamically calibrated prediction model, sets up a variety of typical scenarios to deduce the spatiotemporal evolution trend of groundwater level and outputs a multi-scenario prediction result set.

[0045] The prediction uncertainty quantification unit, based on the multi-scenario prediction result set, uses Monte Carlo simulation to quantify uncertainty and calculate confidence intervals, and outputs a preliminary prediction report with confidence level.

[0046] The uncertainty source tracing and optimization unit, based on the preliminary prediction report with confidence level, uses the variance decomposition method to decompose the total uncertainty into three sources: data, model parameters, and scenarios, and quantifies their contribution ratios. It then proposes uncertainty reduction measures and outputs the final prediction report and uncertainty optimization suggestions.

[0047] The closed-loop sustainable management decision-making module dynamically assesses groundwater carrying capacity based on the final forecast report and uncertainty optimization suggestions, performs multi-objective optimization configuration, generates executable management plans, and iteratively optimizes the management plans through real-time monitoring and feedback.

[0048] As a preferred embodiment of the multi-source data fusion-based dynamic prediction and sustainable management system for groundwater resources described in this invention, the multi-source heterogeneous data acquisition and preprocessing module includes:

[0049] The multi-source data aggregation unit collects hydrogeological data, dynamic monitoring data, meteorological and environmental data, and socio-economic data, and outputs raw multi-source datasets;

[0050] The data quality assessment unit, based on the original multi-source dataset, uses an outlier detection algorithm to identify outliers, remove erroneous data and mark missing data, and outputs a quality assessment report and the cleaned dataset.

[0051] The spatiotemporal alignment and normalization unit, based on the quality assessment report and the cleaned dataset, unifies the data to the same spatiotemporal coordinate system and eliminates dimensional differences, outputting a spatiotemporally aligned standardized dataset.

[0052] The missing value intelligent completion unit, based on a spatiotemporally aligned standardized dataset, uses a spatiotemporal generative adversarial network to complete missing data and outputs a complete preprocessed dataset.

[0053] As a preferred embodiment of the multi-source data fusion-based dynamic prediction and sustainable management system for groundwater resources described in this invention, the spatiotemporal multi-scale data fusion modeling module includes:

[0054] The feature extraction encoding unit, based on the complete preprocessed dataset, uses convolutional neural networks, long short-term memory networks, and graph neural networks to extract spatial, temporal, and topological features, and outputs a multi-source heterogeneous feature set.

[0055] The hierarchical causal modeling unit, based on a multi-source heterogeneous feature set, uses a convergent cross-mapping algorithm to identify multi-scale causal relationships, constructs a hierarchical causal graph and removes false association features, and outputs a feature set with causal constraints.

[0056] The multi-scale feature fusion unit, based on the output feature set with causal constraints, uses a spatiotemporal attention mechanism to perform cross-scale fusion and outputs a fused feature vector.

[0057] The fusion effect verification unit, based on the fusion feature vector, uses mutual information and redundancy analysis to remove redundant features and outputs the optimized core feature set;

[0058] The partition feature adaptation unit, based on the optimized core feature set and combined with the hydrogeological partition, uses the feature importance ranking method to select a partition-specific core feature subset and outputs the partition core feature set.

[0059] The feature library construction and update unit constructs a partitioned dynamic feature library based on the partitioned core feature set and sets an incremental update mechanism to output a fusion feature library that is updated in real time.

[0060] As a preferred embodiment of the multi-source data fusion-based dynamic prediction and sustainable management system for groundwater resources described in this invention, the closed-loop sustainable management decision module includes:

[0061] The dynamic bearing capacity assessment unit, based on the final prediction report and uncertainty optimization suggestions, combined with regional hydrogeological conditions and socio-economic development plans, uses the index system method to dynamically assess the bearing capacity and identify overloaded areas and early warning levels, and outputs a bearing capacity assessment report.

[0062] The multi-objective optimization configuration unit, based on the carrying capacity assessment report, aims to restore the groundwater level, ensure water supply, and minimize economic losses. It uses a multi-objective genetic algorithm to solve for the optimal water resource allocation scheme and outputs a preliminary optimized configuration scheme.

[0063] The management solution generation unit, based on the preliminary optimized configuration plan, combined with the actual situation of the region and the demands of multiple stakeholders, decomposes and implements the plan to administrative regions, industries and water users, formulates differentiated control measures, and outputs an executable sustainable management plan.

[0064] The feedback iterative optimization unit, based on an executable sustainable management plan, collects monitoring data after execution in real time and compares and analyzes it with the predicted data, dynamically corrects the predicted model parameters and management plan, and outputs the iteratively optimized management plan.

[0065] Compared with existing technologies:

[0066] 1. By constructing a complete technical chain encompassing driving factor identification, lag analysis, dual-drive modeling, dynamic calibration, multi-scenario extrapolation, uncertainty quantification, and source tracing optimization, the predictive model's ability to characterize driving factors and their lag effects has been significantly improved. Synchronous online dynamic calibration of the physical model and the data model has been achieved, effectively solving the problems of incomplete driving factor identification, lagging model calibration, and insufficient adaptability to extreme scenarios in existing technologies.

[0067] 2. By deeply integrating high-precision dynamic prediction results with groundwater carrying capacity assessment and multi-objective optimization configuration, a differentiated management plan that can be implemented is generated, and a real-time monitoring feedback iteration mechanism is established. This invention realizes the effective transmission of the uncertainty of prediction results to management decisions, forming a closed-loop adaptive management of prediction, uncertainty assessment, management decision-making, and feedback correction. It overcomes the shortcomings of existing technologies where prediction and management links are relatively independent and uncertainty is not effectively transmitted to the decision-making process, and greatly improves the scientificity and timeliness of management decisions. Attached Figure Description

[0068] Figure 1 This is a schematic diagram of the overall framework of the present invention;

[0069] Figure 2 This is a schematic diagram of the framework of the multi-source heterogeneous data acquisition and preprocessing module of the present invention;

[0070] Figure 3 This is a schematic diagram of the spatiotemporal multi-scale data fusion modeling module framework of the present invention;

[0071] Figure 4 This is a schematic diagram of the coupled-driven dynamic prediction simulation module framework of the present invention;

[0072] Figure 5 This is a schematic diagram of the closed-loop sustainable management decision-making module framework of the present invention. Detailed Implementation

[0073] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0074] This invention provides a method and system for dynamic prediction and sustainable management of groundwater resources based on multi-source data fusion. Please refer to [link / reference]. Figures 1-5 This includes the following steps:

[0075] S1, Multi-source heterogeneous data acquisition and preprocessing: Comprehensive acquisition of multi-dimensional heterogeneous data related to the groundwater system, and output of a complete preprocessed dataset after quality assessment, spatiotemporal alignment and missing value completion;

[0076] S11, Multi-Source Data Aggregation: This involves collecting four core data categories: First, hydrogeological data, including aquifer thickness, permeability coefficient, and storage coefficient; second, dynamic monitoring data, including groundwater level, water temperature, water quality, and extraction volume; third, meteorological and environmental data, including precipitation, evaporation, temperature, and humidity; and fourth, socioeconomic data, including population, GDP, water consumption by various industries, and planting structure. The data originates from publicly available monitoring networks, satellite remote sensing products, statistical yearbooks, and field survey reports, outputting a raw multi-source dataset.

[0077] S12, Data Quality Assessment: Based on the original multi-source dataset, outlier detection algorithms are used to identify outliers. Specifically, the following steps are implemented: First, a threshold method is used to remove physically impossible data (such as negative water levels or rainfall exceeding regional extremes). Second, the standard deviation principle is used to identify statistical outliers. Finally, expert verification is used to manually confirm questionable data. Identified erroneous data is removed, and the location and duration of missing data are recorded. A quality assessment report and a cleaned dataset are then output.

[0078] S13, Spatiotemporal Alignment and Normalization: Based on the quality assessment report and the cleaned dataset, all data are unified to the same spatiotemporal coordinate system. The time dimension is unified to a daily scale, that is, monthly or yearly scale data are converted to daily scale data through interpolation or accumulation; the spatial dimension is unified to regular grid cells, with the grid size set according to the area and monitoring density, generally one kilometer by one kilometer. The min-max normalization method is used to map data of different dimensions to a unified interval, eliminating the impact of dimension differences on subsequent modeling, and outputting a spatiotemporally aligned normalized dataset.

[0079] S14, Intelligent Missing Value Imputation: Based on a spatiotemporally aligned standardized dataset, a spatiotemporal generative adversarial network (GAN) is used to impute missing data. This method comprises two neural networks: a generator and a discriminator. The generator uses the correlation between data from adjacent spatiotemporal points to generate candidate values ​​for missing values, while the discriminator judges the reasonableness of the candidate values. Through alternating training, the generator can generate imputed values ​​consistent with the true distribution. Cross-validation is used to verify the imputation accuracy; that is, some known data is randomly masked, and the difference between the imputed value and the true value is compared to ensure that the imputation accuracy meets the preset requirements, outputting a complete preprocessed dataset.

[0080] S2, Spatiotemporal Multiscale Data Fusion Modeling: Performs feature extraction, causal modeling, multiscale fusion and partition adaptation on the complete preprocessed dataset, and builds and outputs a real-time updated fusion feature library;

[0081] S21, Feature Extraction and Encoding: Based on the complete preprocessed dataset, three types of neural networks are used to extract features from different dimensions: Convolutional Neural Networks (CNNs) are used to extract spatial features, with their kernels sliding across the grid data to capture local spatial patterns; Long Short-Term Memory (LSTM) networks are used to extract temporal features, with their recurrent structure capable of remembering dependencies in long-term sequences; and Graph Neural Networks (GNNs) are used to extract topological relationships between hydrogeological units, constructing a graph structure with hydrogeological zones as nodes and groundwater flow relationships as edges. These three networks operate in parallel, outputting a multi-source heterogeneous feature set.

[0082] S22, Hierarchical Causal Modeling: Based on a multi-source heterogeneous feature set, a convergent cross-mapping algorithm is used to identify causal relationships between features at three scales: day, month, and year. Specifically, for any two feature sequences, their state spaces are reconstructed using a time-delay embedding method. The predictive power of one sequence's state space on the other is calculated. If the predictive power converges with increasing sample size, a causal relationship is determined. This method distinguishes between direct and indirect causal associations, constructs a multi-scale hierarchical causal graph, eliminates spurious association features, and outputs a feature set with causal constraints.

[0083] S23, Multi-scale Feature Fusion: Based on the output feature set with causal constraints, a spatiotemporal attention mechanism is used to adaptively allocate weights for different features. Specifically, an attention score is calculated for each feature, reflecting its importance to the current prediction task. During the fusion process, features at daily, monthly, and yearly scales are fused across scales according to hierarchical causal relationships. That is, features within the same scale are fused first, and then features at different scales are fused layer by layer through a hierarchical structure to capture the evolution patterns of the groundwater system at different time scales and output a fused feature vector.

[0084] S24, Fusion Effect Verification: Based on the fused feature vectors, the mutual information method and redundancy analysis are used to evaluate the fusion effect. The mutual information method is used to measure the correlation between features and the target variable (groundwater level), and redundancy analysis is used to measure the degree of duplication among features. Features that contribute little to the target variable or are highly redundant with other features are removed, and key features that contribute significantly to changes in groundwater level are retained, outputting the optimized core feature set.

[0085] The formula for calculating the mutual information is:

[0086]

[0087] in, For the first A fusion feature, The target variable is the groundwater level. For joint probability density, and This represents the marginal probability density.

[0088] The probability density estimation employs the kernel density estimation method.

[0089]

[0090] in, Gaussian kernel function , For bandwidth parameters, the Silverman rule is used to adaptively determine them:

[0091]

[0092] in, This represents the sample standard deviation.

[0093] The redundancy analysis uses mutual information between features as a metric:

[0094]

[0095] Set threshold and , retain satisfaction And for any preserved feature have Based on the characteristics, the optimized core feature set is output.

[0096] S25, Regional Feature Adaptation: Based on the optimized core feature set, and combined with regional hydrogeological zones (such as plains, karst areas, hilly areas, and desert areas), a feature importance ranking method trained on historical data (such as a random forest model) is used to select a unique subset of core features for each hydrogeological zone. Specifically, a random forest model is trained on different zones, the regional importance score of each feature is recorded, and the features with the highest importance scores are selected as the unique core features for that zone, outputting the regional core feature set.

[0097] S26, Feature Library Construction and Update: Based on the core feature set of each region, a dynamic feature library for regional groundwater resources is constructed. The feature library is stored in a database, and each record contains the feature name, its region, value, timestamp, and source information. An incremental update mechanism is set up so that newly collected preprocessed data is automatically integrated into the feature library of the corresponding region every day. That is, after new data is added, the feature values ​​within the current time window are recalculated and the old values ​​are replaced, outputting a real-time updated fused feature library.

[0098] S3, Coupled-driven dynamic prediction simulation: Based on the real-time updated fusion feature library, identify driving factors and quantify lag effects, construct a physical and data dual-driven hybrid prediction model and perform dynamic calibration, carry out multi-scenario simulation and deduction, quantify prediction uncertainty and trace the source optimization, and output the final prediction report and uncertainty optimization suggestions.

[0099] S31, Driving Factor Identification: Based on a real-time updated fusion feature library, Pearson correlation coefficient and grey relational analysis methods are used to identify the main driving factors affecting groundwater level changes. First, the Pearson correlation coefficient between each feature and the groundwater level is calculated, and features with larger absolute values ​​of correlation coefficients are selected. Second, grey relational analysis is used to calculate the geometric similarity between the feature sequence and the groundwater level sequence. The results of the two methods are combined to quantify the contribution of each factor and output the set of key driving factors in a ranked manner.

[0100] The formula for calculating the Pearson correlation coefficient is as follows:

[0101]

[0102] in, For the first Each feature at time The value of , For the groundwater level at time The observed values, and These are the means of the corresponding sequences. Where is the length of the time series. The calculated correlation coefficients are then subjected to a significance test using the t-test statistic:

[0103]

[0104] Will With degrees of freedom t-distribution critical value In comparison, if (significance level) If the value is positive, the feature is determined to be significantly correlated with the groundwater level; at the same time, the Benjamini-Hochberg method is used for multiple comparison correction to control the false discovery rate FDR < 0.1.

[0105] The calculation steps for the grey relational analysis are as follows:

[0106] (1) Determine the reference sequence (Groundwater level sequence) and comparison sequence (No. (feature sequences)

[0107] (2) Initialize the sequence:

[0108]

[0109] (3) Calculate the difference sequence:

[0110]

[0111] (4) Calculate the correlation coefficient:

[0112]

[0113] Among them, the resolution coefficient ,Pick ;

[0114] (5) Calculate the correlation degree:

[0115]

[0116] S32, Driving Lag Analysis: Based on a set of key driving factors, this study uses cross-wavelet transform and cross-correlation functions to quantify the lag duration and lag intensity of each driving factor in response to groundwater level changes. Cross-wavelet transform can simultaneously analyze the correlation in both the time and frequency domains, while the cross-correlation function determines the lag duration by calculating the correlation coefficient between two sequences at different time offsets. For each driving factor, its optimal lag days are determined, a driving factor sequence with lag parameters is constructed, and a set of driving factors with lag characteristics is output.

[0117] S33, Prediction Model Construction: Based on the set of driving factors with hysteresis features, construct a hybrid prediction model driven by both physical and data aspects, and output the initial groundwater level prediction model.

[0118] The hybrid model comprises two sub-models and a model integration layer:

[0119] First, a physical model is used, employing mature groundwater flow simulation software such as MODFLOW to simulate groundwater levels based on aquifer parameters and boundary conditions, and outputting a physical simulation water level sequence at the target monitoring point.

[0120] Second, the data-driven model uses a long short-term memory network, with a set of driving factors with lag features as input features. It learns the input-output mapping relationship based on historical data and outputs a data-driven predicted water level sequence.

[0121] Thirdly, the model integration layer employs an adaptive weighted integration strategy to fuse the prediction results of the physical model and the data-driven model. Specifically, using the historical prediction errors of the two models within the previous T time steps as input, a fully connected neural network dynamically calculates the integration weights, and the sum of the weights is normalized to 1. The physically simulated water level sequence and the data-driven predicted water level sequence are then weighted and summed to output the fused prediction result. The training objective of the fully connected neural network is to minimize the mean square error between the fused prediction result and the actual observed water level.

[0122] Output the initial groundwater level prediction model.

[0123] S34, Dynamic Model Calibration: Based on the initial groundwater level prediction model, incremental learning and data assimilation methods are used to achieve synchronous online updates of the physical model and the data model, and output the dynamically calibrated prediction model.

[0124] The synchronous online update includes a time-scale collaborative update mechanism:

[0125] (1) Online update of data-driven model: The weights of the long short-term memory network are fine-tuned daily using the latest groundwater monitoring data and the stochastic gradient descent method. The learning rate is set to one-tenth of the initial learning rate and the number of iterations is limited to one epoch, so that the model can adapt to the latest data distribution.

[0126] (2) Periodic updates of the physical model: Using the cumulative monitoring data within a week, the key parameters of the physical model (such as permeability coefficient and water storage coefficient) are corrected in batches by using the ensemble Kalman filter method.

[0127] The state equation, observation equation, and update formula for the ensemble Kalman filter are as follows:

[0128] Equations of state:

[0129]

[0130] in, The state vector consists of the logarithmic permeability coefficient and the logarithmic water storage coefficient. For the MODFLOW model operator, The model error is represented by the covariance matrix. .

[0131] Observation equation:

[0132]

[0133] in, This is the groundwater level observation vector. This is the observation matrix (which maps parameters to the observation space). The observation error is represented by the covariance matrix. .

[0134] Update steps:

[0135] (1) Forecast step: Generate Members of the set:

[0136]

[0137] (2) Calculate the mean and covariance of the forecast ensemble:

[0138]

[0139]

[0140] (3) Calculate the Kalman gain:

[0141]

[0142] (4) Analysis step update:

[0143]

[0144] Among them, the observation set Generate by adding perturbations to actual observations: , Number of members in a set Take 50-100.

[0145] (3) Dynamic adjustment of integration weights: After each data-driven model update or physical model rerun, the adaptive weights of the integration layer are recalculated based on the prediction error of the most recent T time steps, so as to realize the dynamic optimization of the fusion strategy.

[0146] Output the dynamically calibrated prediction model.

[0147] S35, Multi-Scenario Simulation and Extrapolation: Based on a dynamically calibrated prediction model, four typical scenarios are set: a baseline scenario (developing according to historical average trends), a drought scenario (precipitation decreases compared to the baseline), a wetland scenario (precipitation increases compared to the baseline), and an over-extraction remediation scenario (extraction is reduced according to the remediation plan). Driving factor data for different scenarios are input, and the driving factor data are adjusted accordingly based on the scenario settings. The model extrapolates the spatiotemporal evolution trend of groundwater levels over the next one to ten years, outputting a multi-scenario prediction result set.

[0148] S36, Quantification of Prediction Uncertainty: Based on a multi-scenario prediction result set, the Monte Carlo simulation method is used to quantify the uncertainties introduced by model parameters, input data, and scenario settings. Specifically, the uncertainties in the model parameters, input data, and scenario settings are randomly sampled to generate a large number (e.g., one thousand) of different parameter combinations; the prediction model is run on each parameter combination to obtain a set of prediction results; the distribution of all prediction results is statistically analyzed, confidence intervals are calculated, and a preliminary prediction report with confidence levels is output.

[0149] The specific implementation of the Monte Carlo simulation is as follows:

[0150] (1) Determine the uncertainties and their sampling distribution:

[0151] Model parameter uncertainty: physical model parameters (permeability coefficient) Water storage coefficient The weights follow a log-normal distribution, with the mean taken as the posterior mean of the Kalman filter set and the variance taken as the posterior variance. The data-driven model parameters (LSTM weights) are assumed to follow a Gaussian distribution, with the mean taken as the weights after training and the variance taken as an approximation of the inverse of the Hessian matrix of the validation set loss function during training.

[0152] Input data uncertainty: The driving factor data follow a normal distribution, with the mean taken from the observed value and the standard deviation taken as twice the observation error;

[0153] Scenario uncertainty: Scenario parameters (such as the rate of change in precipitation and the proportion of reduction in extraction) are uniformly distributed within the set range.

[0154] (2) Latin hypercube sampling: For each uncertainty factor, the Latin hypercube sampling method is used to generate... Group the samples to ensure uniform coverage of the entire distribution space;

[0155] (3) Model running and result statistics: Run the prediction model for each group of samples and obtain the results. Group prediction results ; Calculate the mean of the prediction results Standard deviation and 95% confidence interval ;

[0156] Output a preliminary forecast report with confidence level.

[0157] S37, Uncertainty Source Tracing and Optimization: Based on the preliminary forecast report with confidence levels, the total uncertainty is decomposed into three sources: data uncertainty, model parameter uncertainty, and scenario uncertainty using variance decomposition. Specifically, two sources are fixed, while only the third source is changed, and its resulting variance in the forecast results is calculated. The contribution of each source is quantified by comparing its proportion to the total variance. Based on the quantification results, targeted uncertainty reduction measures are proposed, such as supplementing monitoring data to reduce data uncertainty and improving parameter calibration methods to reduce model parameter uncertainty. The final forecast report and uncertainty optimization suggestions are then output.

[0158] The variance decomposition uses the Sobol first-order exponential method, and the total variance decomposition formula is as follows:

[0159]

[0160] in, For the predicted results, For the total variance, For the first The main effects variance of each factor The variance represents the interaction effect.

[0161] The formula for calculating the Sobol first-order sensitivity index is as follows:

[0162]

[0163] in, For the first One source of uncertainty ( Data uncertainty Model parameter uncertainty (Scenario uncertainty) Indicates except All other factors, For conditional expectation, The difference is for the other party.

[0164] The Monte Carlo estimate of the first-order exponent of Sobol is:

[0165]

[0166] in, and The model output is for two independent samples. To be The Middle Replace each factor with Mixed samples after corresponding values. Sample size. Use a range of 1000-2000 to ensure estimation accuracy.

[0167] according to Based on the quantitative results, targeted measures to reduce uncertainty are proposed: if (Data uncertainty) is the greatest; it is recommended to supplement monitoring data. (Model parameter uncertainty) is the greatest; it is recommended to improve the parameter calibration method or add physical constraints. (Situation uncertainty) is the greatest, and it is recommended to improve the scenario assumptions or increase the scenario coverage.

[0168] Output the final forecast report and uncertainty optimization suggestions.

[0169] S4, Closed-loop sustainable management decision-making: Based on the final forecast report and uncertainty optimization suggestions, the groundwater carrying capacity is dynamically assessed, multi-objective optimization configuration is carried out, an executable management plan is generated, and the management plan is iteratively optimized through real-time monitoring feedback;

[0170] S41, Dynamic Bearing Capacity Assessment: Based on the final prediction report and uncertainty optimization suggestions, combined with regional hydrogeological conditions and socio-economic development plans, an indicator system method is used to dynamically assess groundwater bearing capacity under different scenarios. An assessment system comprising multiple indicators is constructed, including water level depth, water level decline rate, extraction intensity, and water quality. Thresholds are set for each indicator, and the current and predicted values ​​are used to determine whether overload has occurred. Overloaded areas and warning levels (e.g., blue, yellow, red warnings) are identified, and a bearing capacity assessment report is output.

[0171] S42, Multi-Objective Optimization Allocation: Based on the carrying capacity assessment report, this method uses groundwater level restoration, water supply security, and minimizing economic losses as objective functions, and total extraction volume, industry water quotas, and ecological water demand as constraints. A multi-objective genetic algorithm is employed to solve for the optimal water resource allocation scheme. The multi-objective genetic algorithm generates a series of non-dominated solutions (Pareto front) by simulating the natural evolutionary process (selection, crossover, mutation), each solution representing a trade-off. Decision-makers can select the scheme that best meets their actual needs, outputting a preliminary optimized allocation scheme.

[0172] S43, Management Plan Generation: Based on the preliminary optimized allocation plan, and considering the actual situation of the region and the demands of multiple stakeholders, the plan is broken down and implemented at the administrative region, industry, and water user levels. Specific measures include: setting differentiated total extraction caps for different administrative regions; promoting water-saving irrigation technologies and providing subsidies to agricultural water users; increasing water prices and mandating water recycling for industrial water users; implementing water source replacement in water-scarce areas, using surface water or reclaimed water to replace groundwater; and implementing ecological water replenishment in ecologically sensitive areas. Differentiated extraction control, water-saving renovation, water source replacement, and ecological water replenishment measures are formulated, resulting in an executable sustainable management plan.

[0173] S44, Feedback Iterative Optimization: Based on an executable sustainable management plan, real-time groundwater monitoring data is collected after implementation. The actual data is compared and analyzed with the predicted data to dynamically correct the prediction model parameters and management plan, and output the iteratively optimized management plan.

[0174] The comparative analysis and dynamic correction include the following steps:

[0175] (1) Deviation calculation and statistical test: Calculate the absolute deviation between the actual water level and the predicted water level, and use the t test or Mann-Whitney U test based on the prediction error sequence of the most recent N time steps to determine whether the current deviation significantly exceeds the confidence interval of the historical prediction error (confidence level is 95%). If the deviation does not significantly exceed the confidence interval, it is judged as normal random fluctuation and no adjustment is triggered.

[0176] (2) Determining the source of deviation: If the deviation significantly exceeds the confidence interval, further determine the source of deviation:

[0177] If the deviation shows a continuous unidirectional shift (the actual water level is systematically lower or higher than the predicted water level for M consecutive time steps), and the prediction error directions of the physical model and the data-driven model are consistent, it is determined to be a systematic deviation of the model, triggering model parameter adjustment: updating the prediction model parameters through an online calibration mechanism, including fine-tuning of the network weights of the data-driven model and periodic batch correction of the physical model parameters.

[0178] If the deviation only occurs under specific scenario assumptions (such as during extreme drought or abundant water events), and the prediction accuracy is normal under other scenarios, it is determined to be a scenario assumption deviation, triggering a scenario library update: adding new extreme scenarios to the scenario library and re-conducting multi-scenario simulations;

[0179] If the deviation is concentrated at a few monitoring points and there are no synchronous anomalies at the surrounding monitoring points, it is determined to be a monitoring error or a local anomaly, triggering the data quality control process: the data of the monitoring point is checked for outliers, and if necessary, data from adjacent spatiotemporal points is used for interpolation and replacement, and the status of the monitoring point is marked for manual verification.

[0180] (3) Coordinated correction of management plan: Only when the deviation is statistically significant and the source is identified as systematic deviation of model or scenario assumption deviation, the mining quota or water replenishment plan in the management plan shall be corrected simultaneously; the correction range shall be determined based on the contribution ratio of scenario uncertainty in the uncertainty optimization suggestions. If scenario uncertainty is dominant, the scenario elasticity coefficient shall be adjusted first; if model parameter uncertainty is dominant, the safety margin of mining quota shall be tightened first.

[0181] (4) Iterative closed loop: The adjusted model parameters and management scheme are input into the next prediction cycle to form a closed-loop management mechanism of "prediction-execution-monitoring-statistical verification-source identification-correction".

[0182] Output the iteratively optimized management plan.

[0183] A dynamic prediction and sustainable management system for groundwater resources based on multi-source data fusion includes:

[0184] The multi-source heterogeneous data acquisition and preprocessing module comprehensively collects multi-dimensional heterogeneous data related to the groundwater system. After quality assessment, spatiotemporal alignment, and missing value completion, it outputs a complete preprocessed dataset.

[0185] The multi-source data aggregation unit collects four core data categories: the first category is hydrogeological data, including aquifer thickness, permeability coefficient, and storage coefficient; the second category is dynamic monitoring data, including groundwater level, water temperature, water quality, and extraction volume; the third category is meteorological and environmental data, including precipitation, evaporation, temperature, and humidity; and the fourth category is socioeconomic data, including population, GDP, water consumption by various industries, and planting structure. These data originate from publicly available monitoring networks, satellite remote sensing products, statistical yearbooks, and field survey reports, outputting a raw multi-source dataset.

[0186] The data quality assessment unit, based on the original multi-source dataset, uses an outlier detection algorithm to identify outliers. Specifically, it employs the following steps: first, a threshold method is used to remove physically impossible data (such as negative water levels or rainfall exceeding regional extremes); second, the standard deviation principle is used to identify statistical outliers; and finally, expert verification is used to manually confirm questionable data. Identified erroneous data is removed, and the location and duration of missing data are recorded. A quality assessment report and a cleaned dataset are then output.

[0187] Spatiotemporal alignment and normalization units, based on the quality assessment report and cleaned dataset, unify all data into the same spatiotemporal coordinate system. The time dimension is unified to a daily scale, i.e., monthly or yearly data are converted to daily scale data through interpolation or accumulation; the spatial dimension is unified to regular grid cells, with the grid size set according to the area and monitoring density, typically one kilometer by one kilometer. A min-max normalization method is used to map data of different dimensions to a unified interval, eliminating the impact of dimensional differences on subsequent modeling, and outputting a spatiotemporally aligned normalized dataset.

[0188] The intelligent missing value completion unit is based on a spatiotemporally aligned standardized dataset and employs a spatiotemporal generative adversarial network (GAN) to complete missing data. This method comprises two neural networks: a generator and a discriminator. The generator uses the correlation between data from adjacent spatiotemporal points to generate candidate values ​​for missing values, while the discriminator judges the reasonableness of the candidate values. Through alternating training, the generator can generate completed values ​​consistent with the true distribution. Cross-validation is used to verify the completion accuracy; that is, a portion of known data is randomly masked, and the difference between the completed values ​​and the true values ​​is compared to ensure that the completion accuracy meets the preset requirements, outputting a complete preprocessed dataset.

[0189] The spatiotemporal multi-scale data fusion modeling module performs feature extraction, causal modeling, multi-scale fusion, and partition adaptation on the complete preprocessed dataset, and builds and outputs a fusion feature library that is updated in real time.

[0190] The feature extraction and encoding unit, based on a complete preprocessed dataset, employs three types of neural networks to extract features across different dimensions: a convolutional neural network (CNN) for spatial features, where its kernels slide across the grid data to capture local spatial patterns; a long short-term memory (LSTM) network for temporal features, whose recurrent structure can remember dependencies in long-term sequences; and a graph neural network for extracting topological relationships between hydrogeological units, constructing a graph structure with hydrogeological zones as nodes and groundwater flow relationships as edges. These three networks operate in parallel, outputting a multi-source heterogeneous feature set.

[0191] The hierarchical causal modeling unit, based on a multi-source heterogeneous feature set, employs a convergent cross-mapping algorithm to identify causal relationships between features at three scales: day, month, and year. Specifically, for any two feature sequences, their state spaces are reconstructed using a time-delay embedding method. The predictive power of one sequence's state space on the other is calculated. If the predictive power converges with increasing sample size, a causal relationship is determined. This method distinguishes between direct and indirect causal associations, constructs a multi-scale hierarchical causal graph, eliminates spurious association features, and outputs a feature set with causal constraints.

[0192] The multi-scale feature fusion unit, based on an output feature set with causal constraints, adaptively allocates weights to different features using a spatiotemporal attention mechanism. Specifically, an attention score is calculated for each feature, reflecting its importance to the current prediction task. During the fusion process, features at daily, monthly, and yearly scales are fused across scales according to hierarchical causal relationships. That is, features within the same scale are fused first, and then features at different scales are fused layer by layer through a hierarchical structure to capture the evolution patterns of the groundwater system at different time scales and output a fused feature vector.

[0193] The fusion effect verification unit evaluates the fusion effect based on the fused feature vector, using mutual information and redundancy analysis. Mutual information measures the correlation between features and the target variable (groundwater level), while redundancy analysis measures the degree of overlap between features. Features with low contribution to the target variable or high redundancy with other features are removed, while key features with high contribution to groundwater level changes are retained, outputting the optimized core feature set.

[0194] The formula for calculating the mutual information is:

[0195]

[0196] in, For the first A fusion feature, The target variable is the groundwater level. For joint probability density, and This represents the marginal probability density.

[0197] The probability density estimation employs the kernel density estimation method.

[0198]

[0199] in, Gaussian kernel function , For bandwidth parameters, the Silverman rule is used to adaptively determine them:

[0200]

[0201] in, This represents the sample standard deviation.

[0202] The redundancy analysis uses mutual information between features as a metric:

[0203]

[0204] Set threshold and , retain satisfaction And for any preserved feature have Based on the characteristics, the optimized core feature set is output.

[0205] The regional feature adaptation unit, based on the optimized core feature set and combined with regional hydrogeological zones (such as plains, karst areas, hilly areas, and desert areas), uses a feature importance ranking method (such as a random forest model) trained on historical data to select a unique subset of core features for each hydrogeological zone. Specifically, a random forest model is trained on different zones, the regional importance score of each feature is recorded, and the features with the highest importance scores are selected as the unique core features for that zone, outputting the regional core feature set.

[0206] The feature library construction and update unit builds a dynamic feature library for regional groundwater resources based on the core feature set of each region. The feature library is stored in a database, with each record containing the feature name, its region, value, timestamp, and source information. An incremental update mechanism is set up to automatically integrate newly collected preprocessed data into the feature library of the corresponding region daily. That is, after new data is added, the feature values ​​within the current time window are recalculated, and the old values ​​are replaced, outputting a real-time updated fused feature library.

[0207] The coupled-driven dynamic prediction simulation module identifies driving factors and quantifies lag effects based on a real-time updated fusion feature library, constructs a physical and data-driven hybrid prediction model and performs dynamic calibration, conducts multi-scenario simulation and deduction, quantifies prediction uncertainty and traces the source of optimization, and outputs the final prediction report and uncertainty optimization suggestions.

[0208] The driving factor identification unit, based on a real-time updated fusion feature library, uses Pearson correlation coefficient and grey relational analysis methods to identify the main driving factors affecting groundwater level changes. First, the Pearson correlation coefficient between each feature and the groundwater level is calculated, and features with larger absolute values ​​of correlation coefficients are selected. Second, grey relational analysis is used to calculate the geometric similarity between the feature sequence and the groundwater level sequence. The results of both methods are combined to quantify the contribution of each factor and output a set of key driving factors in a ranked manner.

[0209] The formula for calculating the Pearson correlation coefficient is as follows:

[0210]

[0211] in, For the first Each feature at time The value of , For the groundwater level at time The observed values, and These are the means of the corresponding sequences. Where is the length of the time series. The calculated correlation coefficients are then subjected to a significance test using the t-test statistic:

[0212]

[0213] Will With degrees of freedom t-distribution critical value In comparison, if (significance level) If the value is positive, the feature is determined to be significantly correlated with the groundwater level; at the same time, the Benjamini-Hochberg method is used for multiple comparison correction to control the false discovery rate FDR < 0.1.

[0214] The calculation steps for the grey relational analysis are as follows:

[0215] (1) Determine the reference sequence (Groundwater level sequence) and comparison sequence (No. (feature sequences)

[0216] (2) Initialize the sequence:

[0217]

[0218] (3) Calculate the difference sequence:

[0219]

[0220] (4) Calculate the correlation coefficient:

[0221]

[0222] Among them, the resolution coefficient ,Pick ;

[0223] (5) Calculate the correlation degree:

[0224]

[0225] The lag analysis unit, based on a set of key driving factors, employs cross-wavelet transform and cross-correlation functions to quantify the lag duration and lag intensity of each driving factor in response to groundwater level changes. Cross-wavelet transform can simultaneously analyze correlations in both the time and frequency domains, while the cross-correlation function determines the lag duration by calculating the correlation coefficient between two sequences at different time offsets. For each driving factor, its optimal lag days are determined, a driving factor sequence with lag parameters is constructed, and a set of driving factors with lag characteristics is output.

[0226] The prediction model building unit constructs a hybrid prediction model driven by both physical and data aspects based on a set of driving factors with hysteresis features, and outputs an initial groundwater level prediction model.

[0227] The hybrid model comprises two sub-models and a model integration layer:

[0228] First, a physical model is used, employing mature groundwater flow simulation software such as MODFLOW to simulate groundwater levels based on aquifer parameters and boundary conditions, and outputting a physical simulation water level sequence at the target monitoring point.

[0229] Second, the data-driven model uses a long short-term memory network, with a set of driving factors with lag features as input features. It learns the input-output mapping relationship based on historical data and outputs a data-driven predicted water level sequence.

[0230] Thirdly, the model integration layer employs an adaptive weighted integration strategy to fuse the prediction results of the physical model and the data-driven model. Specifically, using the historical prediction errors of the two models within the previous T time steps as input, a fully connected neural network dynamically calculates the integration weights, and the sum of the weights is normalized to 1. The physically simulated water level sequence and the data-driven predicted water level sequence are then weighted and summed to output the fused prediction result. The training objective of the fully connected neural network is to minimize the mean square error between the fused prediction result and the actual observed water level.

[0231] Output the initial groundwater level prediction model.

[0232] The model dynamic calibration unit, based on the initial groundwater level prediction model, uses incremental learning and data assimilation methods to achieve synchronous online updates of the physical model and the data model, and outputs the dynamically calibrated prediction model.

[0233] The synchronous online update includes a time-scale collaborative update mechanism:

[0234] (1) Online update of data-driven model: The weights of the long short-term memory network are fine-tuned daily using the latest groundwater monitoring data and the stochastic gradient descent method. The learning rate is set to one-tenth of the initial learning rate and the number of iterations is limited to one epoch, so that the model can adapt to the latest data distribution.

[0235] (2) Periodic updates of the physical model: Using the cumulative monitoring data within a week, the key parameters of the physical model (such as permeability coefficient and water storage coefficient) are corrected in batches by using the ensemble Kalman filter method.

[0236] The state equation, observation equation, and update formula for the ensemble Kalman filter are as follows:

[0237] Equations of state:

[0238]

[0239] in, The state vector consists of the logarithmic permeability coefficient and the logarithmic water storage coefficient. For the MODFLOW model operator, The model error is represented by the covariance matrix. .

[0240] Observation equation:

[0241]

[0242] in, This is the groundwater level observation vector. This is the observation matrix (which maps parameters to the observation space). The observation error is represented by the covariance matrix. .

[0243] Update steps:

[0244] (1) Forecast step: Generate Each set member:

[0245]

[0246] (2) Calculate the mean and covariance of the forecast ensemble:

[0247]

[0248]

[0249] (3) Calculate the Kalman gain:

[0250]

[0251] (4) Analysis step update:

[0252]

[0253] Among them, the observation set Generate by adding perturbations to actual observations: , Number of members in a set Take 50-100.

[0254] (3) Dynamic adjustment of integration weights: After each data-driven model update or physical model rerun, the adaptive weights of the integration layer are recalculated based on the prediction error of the most recent T time steps, so as to realize the dynamic optimization of the fusion strategy.

[0255] Output the dynamically calibrated prediction model.

[0256] The multi-scenario simulation unit, based on a dynamically calibrated prediction model, sets four typical scenarios: a baseline scenario (developing according to historical average trends), a drought scenario (precipitation decreases compared to the baseline), a wet scenario (precipitation increases compared to the baseline), and an over-extraction remediation scenario (extraction is reduced according to the remediation plan). Driving factor data for different scenarios are input, and the driving factor data are adjusted accordingly based on the scenario settings. The model extrapolates the spatiotemporal evolution trend of groundwater levels over the next one to ten years, outputting a multi-scenario prediction result set.

[0257] The uncertainty quantification unit, based on a multi-scenario prediction result set, uses Monte Carlo simulation to quantify the uncertainties introduced by model parameters, input data, and scenario settings. Specifically, it involves randomly sampling uncertainties in model parameters, input data, and scenario settings to generate a large number (e.g., one thousand) of different parameter combinations; running the prediction model on each parameter combination to obtain a set of prediction results; statistically analyzing the distribution of all prediction results, calculating confidence intervals, and outputting a preliminary prediction report with confidence levels.

[0258] The specific implementation of the Monte Carlo simulation is as follows:

[0259] (1) Determine the uncertainties and their sampling distribution:

[0260] Model parameter uncertainty: physical model parameters (permeability coefficient) Water storage coefficient The weights follow a log-normal distribution, with the mean taken as the posterior mean of the Kalman filter set and the variance taken as the posterior variance. The data-driven model parameters (LSTM weights) are assumed to follow a Gaussian distribution, with the mean taken as the weights after training and the variance taken as an approximation of the inverse of the Hessian matrix of the validation set loss function during training.

[0261] Input data uncertainty: The driving factor data follow a normal distribution, with the mean taken from the observed value and the standard deviation taken as twice the observation error;

[0262] Scenario uncertainty: Scenario parameters (such as the rate of change in precipitation and the proportion of reduction in extraction) are uniformly distributed within the set range.

[0263] (2) Latin hypercube sampling: For each uncertainty factor, the Latin hypercube sampling method is used to generate... Group the samples to ensure uniform coverage of the entire distribution space;

[0264] (3) Model running and result statistics: Run the prediction model for each group of samples and obtain the results. Group prediction results ; Calculate the mean of the prediction results Standard deviation and 95% confidence interval ;

[0265] Output a preliminary prediction report with confidence level.

[0266] The uncertainty source tracing and optimization unit, based on the preliminary forecast report with confidence levels, uses variance decomposition to decompose the total uncertainty into three sources: data uncertainty, model parameter uncertainty, and scenario uncertainty. Specifically, it fixes two sources and changes only the third, calculating the resulting variance in the forecast results. By comparing the proportion of each source's variance to the total variance, the contribution of each source is quantified. Based on the quantification results, targeted uncertainty reduction measures are proposed, such as supplementing monitoring data to reduce data uncertainty and improving parameter calibration methods to reduce model parameter uncertainty. The final forecast report and uncertainty optimization suggestions are then output.

[0267] The variance decomposition uses the Sobol first-order exponential method, and the total variance decomposition formula is as follows:

[0268]

[0269] in, For the predicted results, For the total variance, For the first The main effects variance of each factor The variance represents the interaction effect.

[0270] The formula for calculating the Sobol first-order sensitivity index is as follows:

[0271]

[0272] in, For the first One source of uncertainty ( Data uncertainty Model parameter uncertainty (Scenario uncertainty) Indicates except All other factors, For conditional expectation, The difference is for the other party.

[0273] The Monte Carlo estimate of the first-order exponent of Sobol is:

[0274]

[0275] in, and The model output is for two independent samples. To be The Middle Replace each factor with Mixed samples after corresponding values. Sample size. Use a range of 1000-2000 to ensure estimation accuracy.

[0276] according to Based on the quantitative results, targeted measures to reduce uncertainty are proposed: if (Data uncertainty) is the greatest; it is recommended to supplement monitoring data. (Model parameter uncertainty) is the greatest; it is recommended to improve the parameter calibration method or add physical constraints. (Situation uncertainty) is the greatest, and it is recommended to improve the scenario assumptions or increase the scenario coverage.

[0277] Output the final forecast report and uncertainty optimization suggestions.

[0278] The closed-loop sustainable management decision-making module dynamically assesses groundwater carrying capacity based on the final forecast report and uncertainty optimization suggestions, performs multi-objective optimization configuration, generates executable management plans, and iteratively optimizes the management plans through real-time monitoring and feedback.

[0279] The dynamic carrying capacity assessment unit, based on the final prediction report and uncertainty optimization suggestions, and combined with regional hydrogeological conditions and socio-economic development plans, uses an indicator system method to dynamically assess groundwater carrying capacity under different scenarios. An assessment system comprising multiple indicators is constructed, including water level depth, water level decline rate, extraction intensity, and water quality. Thresholds are set for each indicator, and the current and predicted values ​​are used to determine whether overload has occurred. Overloaded areas and warning levels (e.g., blue, yellow, and red warnings) are identified, and a carrying capacity assessment report is output.

[0280] The multi-objective optimization allocation unit, based on the carrying capacity assessment report, uses groundwater level restoration, water supply security, and minimization of economic losses as objective functions, and total extraction volume, industry water quotas, and ecological water demand as constraints. It employs a multi-objective genetic algorithm to solve for the optimal water resource allocation scheme. The multi-objective genetic algorithm generates a series of non-dominated solutions (Pareto front) by simulating a natural evolutionary process (selection, crossover, and mutation), each solution representing a trade-off. Decision-makers can select the scheme that best meets their actual needs, outputting a preliminary optimized allocation scheme.

[0281] The management plan generation unit, based on the preliminary optimized configuration plan and considering the actual situation of the region and the demands of multiple stakeholders, breaks down and implements the plan to administrative regions, industries, and water users. Specific measures include: setting differentiated total extraction caps for different administrative regions; promoting water-saving irrigation technologies and providing subsidies to agricultural water users; increasing water prices and mandating water recycling for industrial water users; implementing water source replacement in water-scarce areas, using surface water or reclaimed water to replace groundwater; and implementing ecological water replenishment in ecologically sensitive areas. The unit formulates differentiated extraction control, water-saving renovation, water source replacement, and ecological water replenishment measures, and outputs an executable sustainable management plan.

[0282] The feedback iterative optimization unit, based on an executable sustainable management plan, collects groundwater monitoring data in real time after implementation, compares and analyzes the actual data with the predicted data, dynamically corrects the prediction model parameters and management plan, and outputs the iteratively optimized management plan.

[0283] The comparative analysis and dynamic correction include the following steps:

[0284] (1) Deviation calculation and statistical test: Calculate the absolute deviation between the actual water level and the predicted water level, and use the t test or Mann-Whitney U test based on the prediction error sequence of the most recent N time steps to determine whether the current deviation significantly exceeds the confidence interval of the historical prediction error (confidence level is 95%). If the deviation does not significantly exceed the confidence interval, it is judged as normal random fluctuation and no adjustment is triggered.

[0285] (2) Determining the source of deviation: If the deviation significantly exceeds the confidence interval, further determine the source of deviation:

[0286] If the deviation shows a continuous unidirectional shift (the actual water level is systematically lower or higher than the predicted water level for M consecutive time steps), and the prediction error directions of the physical model and the data-driven model are consistent, it is determined to be a systematic deviation of the model, triggering model parameter adjustment: updating the prediction model parameters through an online calibration mechanism, including fine-tuning of the network weights of the data-driven model and periodic batch correction of the physical model parameters.

[0287] If the deviation only occurs under specific scenario assumptions (such as during extreme drought or abundant water events), and the prediction accuracy is normal under other scenarios, it is determined to be a scenario assumption deviation, triggering a scenario library update: adding new extreme scenarios to the scenario library and re-conducting multi-scenario simulations;

[0288] If the deviation is concentrated at a few monitoring points and there are no synchronous anomalies at the surrounding monitoring points, it is determined to be a monitoring error or a local anomaly, triggering the data quality control process: the data of the monitoring point is checked for outliers, and if necessary, data from adjacent spatiotemporal points is used for interpolation and replacement, and the status of the monitoring point is marked for manual verification.

[0289] (3) Coordinated correction of management plan: Only when the deviation is statistically significant and the source is identified as systematic deviation of model or scenario assumption deviation, the mining quota or water replenishment plan in the management plan shall be corrected simultaneously; the correction range shall be determined based on the contribution ratio of scenario uncertainty in the uncertainty optimization suggestions. If scenario uncertainty is dominant, the scenario elasticity coefficient shall be adjusted first; if model parameter uncertainty is dominant, the safety margin of mining quota shall be tightened first.

[0290] (4) Iterative closed loop: The adjusted model parameters and management scheme are input into the next prediction cycle to form a closed-loop management mechanism of "prediction-execution-monitoring-statistical verification-source identification-correction".

[0291] Output the iteratively optimized management plan.

[0292] Although the present invention has been described above with reference to embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of the invention. In particular, as long as there is no structural conflict, the features in the disclosed embodiments can be combined with each other in any manner. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, the present invention is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for dynamic prediction and sustainable management of groundwater resources based on multi-source data fusion, characterized in that, Includes the following steps: S1, Multi-source heterogeneous data acquisition and preprocessing: Comprehensive acquisition of multi-dimensional heterogeneous data related to the groundwater system, and output of a complete preprocessed dataset after quality assessment, spatiotemporal alignment and missing value completion; S2, Spatiotemporal Multiscale Data Fusion Modeling: Performs feature extraction, causal modeling, multiscale fusion and partition adaptation on the complete preprocessed dataset, and builds and outputs a real-time updated fusion feature library; S3, Coupled-driven dynamic prediction simulation: S31, Driving Factor Identification: Based on a real-time updated fusion feature library, Pearson correlation coefficient and grey relational analysis methods are used to identify driving factors and quantify their contribution, outputting a set of key driving factors; S32, Driving lag analysis: Based on the key driving factor set, cross wavelet transform and cross-correlation function are used to quantify lag duration and lag strength, and output a driving factor set with lag characteristics; S33, Prediction Model Construction: Based on the set of driving factors with hysteresis features, construct a hybrid prediction model driven by both physical and data aspects, and output the initial groundwater level prediction model. S34, Dynamic Model Calibration: Based on the initial groundwater level prediction model, incremental learning and data assimilation methods are used to achieve synchronous online updates of the physical model and the data model, and output the dynamically calibrated prediction model. S35, Multi-scenario Simulation and Inference: Based on the dynamically calibrated prediction model, various typical scenarios are set to infer the spatiotemporal evolution trend of groundwater level and output a multi-scenario prediction result set; S36, Quantification of Prediction Uncertainty: Based on the multi-scenario prediction result set, Monte Carlo simulation is used to quantify uncertainty and calculate confidence intervals, and output a preliminary prediction report with confidence level; S37, Uncertainty Source Tracing and Optimization: Based on the preliminary forecast report with confidence level, the variance decomposition method is used to decompose the total uncertainty into three sources: data, model parameters and scenarios, and quantify their contribution ratio. Uncertainty reduction measures are proposed, and the final forecast report and uncertainty optimization suggestions are output. S4, Closed-loop Sustainable Management Decision-Making: Based on the final forecast report and uncertainty optimization suggestions, the groundwater carrying capacity is dynamically assessed, multi-objective optimization configuration is carried out, an executable management plan is generated, and the management plan is iteratively optimized through real-time monitoring and feedback.

2. The method for dynamic prediction and sustainable management of groundwater resources based on multi-source data fusion according to claim 1, characterized in that, The specific steps of S1 are as follows: S11, Multi-source data aggregation: Collects hydrogeological data, dynamic monitoring data, meteorological and environmental data, and socio-economic data, and outputs raw multi-source datasets; S12, Data Quality Assessment: Based on the original multi-source dataset, outlier detection algorithms are used to identify outliers, remove erroneous data and mark missing data, and output a quality assessment report and the cleaned dataset; S13, Spatiotemporal Alignment and Normalization: Based on the quality assessment report and the cleaned dataset, the data is unified to the same spatiotemporal coordinate system and the difference in units is eliminated, outputting a standardized dataset with spatiotemporal alignment. S14, Intelligent missing value completion: Based on a standardized dataset with spatiotemporal alignment, a spatiotemporal generative adversarial network is used to complete missing data and output a complete preprocessed dataset.

3. The method for dynamic prediction and sustainable management of groundwater resources based on multi-source data fusion according to claim 1, characterized in that, The specific steps of S2 are as follows: S21, Feature extraction and encoding: Based on the complete preprocessed dataset, convolutional neural networks, long short-term memory networks and graph neural networks are used to extract spatial, temporal and topological features, and output a multi-source heterogeneous feature set; S22, Hierarchical Causal Modeling: Based on multi-source heterogeneous feature sets, a convergent cross-mapping algorithm is used to identify multi-scale causal relationships, construct a hierarchical causal graph and remove false association features, and output a feature set with causal constraints; S23, Multi-scale feature fusion: Based on the output feature set with causal constraints, a spatiotemporal attention mechanism is used to perform cross-scale fusion and output a fused feature vector; S24, fusion effect verification: Based on the fusion feature vector, the mutual information method and redundancy analysis are used to remove redundant features and output the optimized core feature set; S25, Zonal Feature Adaptation: Based on the optimized core feature set, combined with hydrogeological zoning, a feature importance ranking method is used to select a zonal-specific core feature subset and output the zonal core feature set. S26, Feature Library Construction and Update: Based on the partitioned core feature set, construct a partitioned dynamic feature library and set an incremental update mechanism to output a fusion feature library that is updated in real time.

4. The method for dynamic prediction and sustainable management of groundwater resources based on multi-source data fusion according to claim 1, characterized in that, The specific steps of S4 are as follows: S41, Dynamic assessment of bearing capacity: Based on the final prediction report and uncertainty optimization suggestions, combined with regional hydrogeological conditions and socio-economic development plans, the bearing capacity is dynamically assessed using the index system method, and overloaded areas and warning levels are identified, and a bearing capacity assessment report is output. S42, Multi-objective optimization configuration: Based on the carrying capacity assessment report, with the objectives of groundwater level restoration, water supply security and minimization of economic losses, a multi-objective genetic algorithm is used to solve the optimal water resource allocation scheme and output a preliminary optimized configuration scheme. S43, Management Plan Generation: Based on the preliminary optimized configuration plan, combined with the actual situation of the region and the demands of multiple stakeholders, the plan is broken down and implemented at the administrative region, industry and water user level, differentiated control measures are formulated, and an executable sustainable management plan is output. S44, Feedback Iterative Optimization: Based on an executable sustainable management plan, real-time monitoring data after execution is collected and compared with predicted data to dynamically correct the predicted model parameters and management plan, and output the iteratively optimized management plan.

5. A dynamic prediction and sustainable management system for groundwater resources based on multi-source data fusion, characterized in that: include: The multi-source heterogeneous data acquisition and preprocessing module comprehensively collects multi-dimensional heterogeneous data related to the groundwater system. After quality assessment, spatiotemporal alignment, and missing value completion, it outputs a complete preprocessed dataset. The spatiotemporal multi-scale data fusion modeling module performs feature extraction, causal modeling, multi-scale fusion, and partition adaptation on the complete preprocessed dataset, and builds and outputs a fusion feature library that is updated in real time. The coupled-driven dynamic prediction simulation module includes a driving factor identification unit, a driving lag analysis unit, a prediction model construction unit, a model dynamic calibration unit, a multi-scenario simulation and deduction unit, a prediction uncertainty quantification unit, and an uncertainty source tracing and optimization unit. The driving factor identification unit, based on a real-time updated fusion feature library, uses Pearson correlation coefficient and grey relational analysis methods to identify driving factors and quantify their contribution, and outputs a set of key driving factors. The driving lag analysis unit, based on the key driving factor set, uses cross wavelet transform and cross-correlation function to quantify lag duration and lag intensity, and outputs a driving factor set with lag characteristics. The prediction model construction unit constructs a physical and data-driven hybrid prediction model based on a set of driving factors with hysteresis features, and outputs an initial groundwater level prediction model. The model dynamic calibration unit, based on the initial groundwater level prediction model, uses incremental learning and data assimilation methods to achieve synchronous online updates of the physical model and the data model, and outputs the dynamically calibrated prediction model. The multi-scenario simulation and deduction unit, based on the dynamically calibrated prediction model, sets up a variety of typical scenarios to deduce the spatiotemporal evolution trend of groundwater level and outputs a multi-scenario prediction result set. The prediction uncertainty quantification unit, based on the multi-scenario prediction result set, uses Monte Carlo simulation to quantify uncertainty and calculate confidence intervals, and outputs a preliminary prediction report with confidence level. The uncertainty source tracing and optimization unit, based on the preliminary prediction report with confidence level, uses the variance decomposition method to decompose the total uncertainty into three sources: data, model parameters, and scenarios, and quantifies their contribution ratios. It then proposes uncertainty reduction measures and outputs the final prediction report and uncertainty optimization suggestions. The closed-loop sustainable management decision-making module dynamically assesses groundwater carrying capacity based on the final forecast report and uncertainty optimization suggestions, performs multi-objective optimization configuration, generates executable management plans, and iteratively optimizes the management plans through real-time monitoring and feedback.

6. The groundwater resource dynamic prediction and sustainable management system based on multi-source data fusion according to claim 5, characterized in that, The multi-source heterogeneous data acquisition and preprocessing module includes: The multi-source data aggregation unit collects hydrogeological data, dynamic monitoring data, meteorological and environmental data, and socio-economic data, and outputs raw multi-source datasets; The data quality assessment unit, based on the original multi-source dataset, uses an outlier detection algorithm to identify outliers, remove erroneous data and mark missing data, and outputs a quality assessment report and the cleaned dataset. The spatiotemporal alignment and normalization unit, based on the quality assessment report and the cleaned dataset, unifies the data to the same spatiotemporal coordinate system and eliminates dimensional differences, outputting a spatiotemporally aligned standardized dataset. The missing value intelligent completion unit, based on a spatiotemporally aligned standardized dataset, uses a spatiotemporal generative adversarial network to complete missing data and outputs a complete preprocessed dataset.

7. The groundwater resource dynamic prediction and sustainable management system based on multi-source data fusion according to claim 5, characterized in that, The spatiotemporal multi-scale data fusion modeling module includes: The feature extraction encoding unit, based on the complete preprocessed dataset, uses convolutional neural networks, long short-term memory networks, and graph neural networks to extract spatial, temporal, and topological features, and outputs a multi-source heterogeneous feature set. The hierarchical causal modeling unit, based on a multi-source heterogeneous feature set, uses a convergent cross-mapping algorithm to identify multi-scale causal relationships, constructs a hierarchical causal graph and removes false association features, and outputs a feature set with causal constraints. The multi-scale feature fusion unit, based on the output feature set with causal constraints, uses a spatiotemporal attention mechanism to perform cross-scale fusion and outputs a fused feature vector. The fusion effect verification unit, based on the fusion feature vector, uses mutual information and redundancy analysis to remove redundant features and outputs the optimized core feature set; The partition feature adaptation unit, based on the optimized core feature set and combined with the hydrogeological partition, uses the feature importance ranking method to select a partition-specific core feature subset and outputs the partition core feature set. The feature library construction and update unit constructs a partitioned dynamic feature library based on the partitioned core feature set and sets an incremental update mechanism to output a fusion feature library that is updated in real time.

8. The groundwater resource dynamic prediction and sustainable management system based on multi-source data fusion according to claim 5, characterized in that, The closed-loop sustainable management decision-making module includes: The dynamic bearing capacity assessment unit, based on the final prediction report and uncertainty optimization suggestions, combined with regional hydrogeological conditions and socio-economic development plans, uses the index system method to dynamically assess the bearing capacity and identify overloaded areas and early warning levels, and outputs a bearing capacity assessment report. The multi-objective optimization configuration unit, based on the carrying capacity assessment report, aims to restore the groundwater level, ensure water supply, and minimize economic losses. It uses a multi-objective genetic algorithm to solve for the optimal water resource allocation scheme and outputs a preliminary optimized configuration scheme. The management solution generation unit, based on the preliminary optimized configuration plan, combined with the actual situation of the region and the demands of multiple stakeholders, decomposes and implements the plan to administrative regions, industries and water users, formulates differentiated control measures, and outputs an executable sustainable management plan; The feedback iterative optimization unit, based on an executable sustainable management plan, collects monitoring data after execution in real time and compares and analyzes it with the predicted data, dynamically corrects the prediction model parameters and management plan, and outputs the iteratively optimized management plan.

Images