Method and system for predicting and tracing concentration of endocrine disrupter in lake and reservoir

Abstract

The invention discloses a lake and reservoir endocrine disrupter concentration prediction and tracing method and system, and relates to the technical field of environment monitoring and pollution control, the method comprises the following steps: collecting and fusing multi-source heterogeneous monitoring data of a basin-lake and reservoir system, and constructing a standardized spatio-temporal data set; based on a hydrodynamic force and pollutant migration and transformation physical mechanism, key spatio-temporal features are extracted from the fusion data, and a high-dimensional feature tensor is constructed; constructing and training a deep neural network model fusing causal discovery and physical constraint, and predicting spatial and temporal distribution of the concentration of the endocrine disrupter in the lake and reservoir by using a feature tensor; analyzing the dynamic contribution degree of different pollution sources to the concentration of the endocrine disrupter in the target area to realize pollution source tracing; and carrying out integrated visualization on a concentration prediction result and pollution source contribution degree information, and generating graded early warning signals and targeted treatment decision suggestions. According to the method, the accuracy and robustness of lake and reservoir endocrine disrupter concentration prediction can be remarkably improved, and pollution traceability is realized.

Description

Technical Field

[0001] This invention relates to the field of environmental monitoring and pollution control technology, and in particular to a method and system for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs. Background Technology

[0002] Water environment safety is a core issue in national ecological civilization construction and public health protection. Lakes, reservoirs, and other enclosed or semi-enclosed water bodies serve as important drinking water sources and ecological barriers; their water quality directly impacts regional sustainable development capabilities. Endocrine disruptors (EDCs) are exogenous chemical substances that interfere with the function of biological endocrine systems. Typical pollutants include bisphenol A (BPA), nonylphenol, 17α-ethynylestradiol, and diethylhexyl phthalate (DET), which are widely derived from domestic sewage discharge, chemical and pharmaceutical wastewater, agricultural non-point source runoff, and urban surface runoff. Even at extremely low concentrations, these substances can cause reproductive toxicity, developmental abnormalities, and population decline in aquatic organisms, and can also accumulate through the food chain, threatening human health. They have been included in the global priority list of pollutants. Therefore, accurately predicting the dynamic concentration of EDCs in lakes and reservoirs and efficiently tracing pollution sources has become a key technological bottleneck that urgently needs to be overcome in the intelligent water environment monitoring and risk prevention system.

[0003] Among them, the technology for predicting the concentration of endocrine disruptors in lakes and reservoirs and apportioning pollution sources focuses on integrating hydrological processes, human activity disturbances, and pollutant migration and transformation mechanisms to construct an intelligent analysis framework with spatiotemporal evolution simulation capabilities and causal attribution logic. This technical direction aims to overcome the limitations of traditional water quality models in terms of complex nonlinear responses, multi-scale driving coupling, and lack of interpretability. By integrating external input signals at the watershed scale with the hydrodynamic-water quality interaction processes within lakes and reservoirs, it establishes a full-chain response relationship from "source" to "sink," thereby supporting forward-looking early warning and targeted governance decisions.

[0004] Existing technologies for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs suffer from systemic defects: First, mainstream prediction methods often employ multiple linear regression, support vector machines, or black-box deep neural networks. While these methods can achieve high fitting accuracy on specific datasets, they lack clear physical mechanisms and causal logic within the models, failing to distinguish between correlation and causation. This results in a significant decrease in the robustness of predictions under varying operating conditions or extreme events, and makes it difficult to support pollution liability determination. Second, existing monitoring systems are limited by cost and operational capabilities, with sampling frequencies typically weekly or monthly and sparse spatial distribution, making it impossible to capture endocrine disruptors. The problems include: sudden concentration fluctuations and spatial heterogeneous distribution characteristics on short timescales; fragmented modeling of the watershed-lake system, with a lack of unified data representation and coupling mechanisms between upstream agricultural non-point source loads, industrial point source emissions, urban runoff, and processes such as water exchange, sedimentation degradation, and photolysis within the lakes and reservoirs, making it difficult for the model to reflect the complete pathway of cross-scale material transport; and finally, pollution source apportionment relies heavily on receptor models or empirical emission inventories, failing to deeply couple with dynamic hydrological and meteorological conditions and pollutant migration processes, resulting in a lack of spatiotemporal dynamism and physical consistency in the source tracing results, making it difficult to support precise enforcement and source reduction. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs, in order to solve the problems of existing water quality prediction models lacking causal interpretability, having low monitoring frequency, limited spatial resolution, and insufficient cross-scale data fusion.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: On the one hand, a system for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs, comprising the following components: The multi-source heterogeneous data acquisition and fusion module is used to acquire multi-dimensional monitoring data at the watershed scale and within lakes and reservoirs, and to perform data standardization and spatiotemporal alignment processing. The physical mechanism-guided spatiotemporal feature extraction module is used to extract physically meaningful spatiotemporal features from the fused data and construct feature tensors. The causal inference-driven concentration prediction module is used to build and train a deep learning model that integrates causal constraints based on extracted spatiotemporal features, so as to realize the dynamic prediction of the concentration of endocrine disruptors in lakes and reservoirs. The Dynamic Pollution Source Analysis and Contribution Measurement Module is used to analyze the spatiotemporal dynamic contribution of different pollution sources to the concentration of secretory interfering substances in the target area based on the internal mechanism and input characteristics of the prediction model. The visualization early warning and decision support module is used to visualize the prediction results and traceability information, and generate risk early warning and control suggestion reports; To achieve the above objectives, the present invention provides the following technical solution: On the other hand, a method for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs, the specific steps of which are as follows: Step S110: Collect and fuse multi-source heterogeneous monitoring data of the watershed-lake-reservoir system to construct a standardized spatiotemporal dataset; Step S120: Based on the physical mechanisms of hydrodynamics and pollutant migration and transformation, key spatiotemporal features are extracted from the fused data to construct a high-dimensional feature tensor; Step S130: Construct and train a deep neural network model that integrates causal discovery and physical constraints, and use feature tensors to predict the spatiotemporal distribution of endocrine disruptors concentration in lakes and reservoirs. Step S140: Based on the trained prediction model, analyze the dynamic contribution of different pollution sources to the concentration of secretory interfering substances in the target area, and realize pollution source tracing. Step S150: Integrate and visualize the concentration prediction results with the pollution source contribution information, and generate graded early warning signals and targeted governance decision-making suggestions; Preferably, the multi-source heterogeneous data acquisition and fusion module specifically includes a hydrological and meteorological monitoring unit, a water quality parameter monitoring unit, a human activity intensity monitoring unit, and a remote sensing inversion data acquisition unit. The hydrological and meteorological monitoring unit is responsible for collecting data on precipitation, evaporation, runoff, water temperature, wind speed, and wind direction, with a sampling frequency of no less than once per hour. The water quality parameter monitoring unit is responsible for collecting laboratory analysis concentration values ​​of water body pH, dissolved oxygen, turbidity, chlorophyll a, and target endocrine disruptors, with on-site sensor monitoring frequency of once every two hours and laboratory sampling frequency of once per week. The human activity intensity monitoring unit acquires data on industrial wastewater discharge, urban domestic sewage discharge, agricultural fertilizer application, and livestock and poultry breeding volume by accessing regional statistical yearbooks, pollutant discharge permit databases, and land use data, with an update frequency of once per month. The remote sensing inversion data acquisition unit processes multispectral and hyperspectral satellite imagery to invert spatial distribution maps of suspended solids concentration, colored dissolved organic matter concentration, and water temperature, with a spatial resolution better than 30 meters and a temporal resolution of no less than once every three days. The data fusion submodule uses a collaborative correction algorithm based on spatiotemporal kriging interpolation to unify the spatiotemporal scale and fill in missing values ​​of the above multi-source data, generating a fused dataset with a consistent spatiotemporal grid.

[0007] Furthermore, the spatiotemporal feature extraction module guided by the physical mechanism employs a multi-channel feature construction strategy. This module first defines a set of fundamental physical features based on the hydrodynamic equation and the pollutant convection-diffusion-reaction equation, including water age distribution, pollutant transport paths, sedimentation rate, and photolysis rate. Water age distribution is obtained by solving the conservative mass transport equation of the lake / reservoir and is used to characterize water retention time. Pollutant transport paths are simulated using the Lagrange particle tracking algorithm, calculating the trajectory and time of particles from various potential pollution sources to a specified grid within the lake / reservoir. Sedimentation rate is calculated based on Stokes' law and particulate matter adsorption characteristics, while photolysis rate is determined based on the molar absorptivity of the target endocrine disruptor and the water body's light intensity profile. Secondly, this module extracts statistical and deep learning features from the fused dataset. Statistical features include the mean, variance, skewness, and kurtosis of each monitoring indicator within the time window, as well as the local Moran's index within the spatial neighborhood. Deep learning features are extracted from the spatial distribution map of water quality parameters retrieved from remote sensing using a pre-trained convolutional autoencoder. The encoder part of this autoencoder outputs a 256-dimensional compressed feature vector. Finally, all features are concatenated and organized into a four-dimensional tensor, with dimensions representing the time step, the number of spatial grid rows, the number of spatial grid columns, and the number of feature channels.

[0008] Furthermore, the core of the causal inference-driven concentration prediction module is a hybrid architecture integrating a graph neural network and a long short-term memory network. This architecture includes a graph structure learning layer for automatically discovering causal dependencies between variables from high-dimensional feature tensors. The graph structure learning layer employs an attention-based mechanism to calculate the causal strength score between each pair of feature variables; edges with scores exceeding a preset threshold of 0.05 are retained, forming a dynamic causal graph. This causal graph serves as the adjacency matrix for subsequent graph convolution operations, ensuring information propagates among causally related features. The long short-term memory network layer is used to capture the long-term temporal dependencies of the concentration sequence, with its hidden layer dimension set to 128. A physical consistency loss term is introduced during model training, penalizing the deviation between the predicted concentration and the concentration change derived from the mass conservation equation. The physical consistency loss function is defined as: in, Let physical consistency loss function be used. Let be the predicted concentration at time t. For the velocity field, Where is the diffusion coefficient. The overall degradation rate constant, This is a hyperparameter with a value of 0.1. The total loss function is a weighted sum of the prediction mean square error and the physical consistency loss.

[0009] Preferably, the dynamic pollution source analysis and contribution measurement module implements a hybrid source tracing method based on integral gradient and structural equation modeling. This module first uses the integral gradient algorithm to calculate the gradient integral of the prediction model output with respect to each input feature (corresponding to different pollution source types and spatial locations), obtaining the initial contribution score of each feature to the final predicted concentration. The integration path is a straight line, with 50 steps. Subsequently, the module constructs a structural equation model, using the initial contribution score as the observed variable and pollution source emission intensity, hydrological and meteorological conditions, and geographical distance as latent variables, establishing a linear structural relationship between the latent variables and the observed variables. The parameters of the structural equation model are solved using the maximum likelihood estimation method; finally, the contribution of pollution source i to monitoring point j at time t is calculated. Calculated using the following formula: in, Let be the integral gradient value of the feature corresponding to pollution source i at time t, i.e. Let be the integral gradient value of the feature corresponding to pollution source k at point j and time t. For the path coefficients of the latent variable i of pollution source in the structural equation model, represents the path coefficients of the latent variable k for pollution source in the structural equation model, where N is the total number of pollution sources. This contribution result exhibits spatiotemporal dynamics and can be accurate to hourly and 100-meter grid scales.

[0010] Furthermore, the visualization-based early warning and decision support module integrates geographic information systems and dynamic dashboard technology. This module overlays concentration prediction results onto a high-precision electronic map in the form of a heatmap, with color mapping ranging from blue (concentration below 1 ng / L) to red (concentration above 100 ng / L). Pollution source contribution results are displayed as a stacked area map and flow arrows, dynamically reflecting changes in the proportion of different source contributions and pollutant migration paths. The early warning submodule automatically triggers different levels of early warning signals based on preset concentration thresholds and pushes them to designated management personnel via SMS and email. The decision support submodule has a built-in rule base containing recommendations for governance measures based on different pollution source types and contribution levels. For example, when agricultural non-point source pollution exceeds 60%, it is recommended to initiate ecological ditch construction and fertilizer reduction plans; when industrial point source pollution suddenly increases and exceeds 40%, it is recommended to initiate on-site verification of key polluting enterprises.

[0011] Compared with the prior art, the beneficial technical effects of the present invention are as follows: This invention significantly improves the accuracy and robustness of predicting the concentration of endocrine disruptors in lakes and reservoirs by integrating feature extraction guided by physical mechanisms with a deep learning model driven by causal inference. The model achieves a Nash efficiency coefficient of over 0.85 on the test set, which is more than 20% higher than that of traditional black-box models, and reduces the prediction error by more than 35% under extreme rainfall events.

[0012] This invention employs a dynamic pollution source analysis method based on integral gradient and structural equation modeling, which enables accurate, quantitative, and spatiotemporal dynamic source tracing of pollution contribution. The spatial resolution of the source tracing results reaches 100 meters, and the temporal resolution reaches 1 hour, providing a reliable basis for precise law enforcement and source reduction.

[0013] This invention constructs a complete technology system from multi-source data fusion, feature extraction, concentration prediction to pollution source tracing and decision support. It effectively overcomes the problems of fragmented modeling of watershed-lake-reservoir systems, poor model interpretability, and static source tracing results in existing technologies, and supports the forward-looking early warning and targeted governance of water environment risks. Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the overall technical architecture of a method and system for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs proposed in this invention. Figure 2 This is a schematic diagram of the dynamic hybrid framework of the concentration prediction module driven by causal inference and the dynamic pollution source analysis module in this invention; Figure 3 This is a schematic diagram illustrating the specific steps of a method for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs proposed in this invention. Figure 4 This is a schematic diagram of the core principle framework of the visualization early warning and decision support module of this invention. Detailed Implementation

[0015] The features and exemplary embodiments of various aspects of the present invention will now be described in detail. To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely intended to explain the present invention and not to limit the present invention. For those skilled in the art, the present invention can be practiced without some of these specific details. The following description of the embodiments is merely to provide a better understanding of the present invention by illustrating examples of the invention.

[0016] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0017] In the embodiments of the present invention, the same reference numerals denote the same components, and for the sake of brevity, detailed descriptions of the same components are omitted in different embodiments. It should be understood that the thickness, length, width, and other dimensions of various components in the embodiments of the present invention shown in the accompanying drawings, as well as the overall thickness, length, width, and other dimensions of the integrated device, are merely illustrative and should not constitute any limitation on the present invention; the term "multiple" in the present invention refers to two or more (including two).

[0018] Example 1 In eutrophic areas of large, shallow lakes, a multi-source heterogeneous data acquisition and fusion module conducts comprehensive monitoring and data integration at the watershed scale and within the lake / reservoir environment. The hydro-meteorological monitoring unit continuously collects data on precipitation, evaporation, runoff, water temperature, wind speed, and wind direction using water level gauges, current meters, and weather stations deployed in major inflow rivers and at the lake's center. All sensors maintain a sampling frequency of at least once per hour to ensure the complete dynamic process of rainfall-runoff events is captured. The water quality parameter monitoring unit employs a combination of in-situ sensor networks and laboratory sampling, deployed at 10 representative cross-sections of the lake to monitor basic parameters such as pH, dissolved oxygen, turbidity, and chlorophyll a in real time. Multi-parameter water quality sensors automatically upload data every 2 hours, while a laboratory sampling vessel collects water samples weekly and sends them to a certified laboratory. Liquid chromatography-mass spectrometry (LC-MS) is used to analyze the concentration of target endocrine disruptors, with a detection limit of 0.1 nanograms per liter. The Human Activity Intensity Monitoring Unit, by connecting to the provincial environmental statistics database and the territorial spatial planning platform, acquires data on wastewater discharge from key industrial enterprises, effluent from urban sewage treatment plants, agricultural non-point source fertilizer application, and pollution discharge from large-scale livestock and poultry farms within the basin. All socio-economic data are updated monthly to ensure the timeliness of the pollution source inventory. The Remote Sensing Inversion Data Acquisition Unit processes multispectral images from Sentinel-2 and Landsat-8, using semi-analytical algorithms to invert the spatial distribution of suspended solids concentration, colored dissolved organic matter concentration, and surface water temperature. The product achieves a spatial resolution of 20 meters, a revisit period of 3 days, and complete coverage of the entire lake area. The data fusion submodule adopts a collaborative correction algorithm based on spatiotemporal kriging interpolation. First, data from different sources are uniformly projected onto the same geographic coordinate system to establish a regular spatial grid of 100m×100m. Then, with a time step of 1 hour, spatiotemporal scale matching is performed on the multi-source data. The optimal interpolation parameters are determined by calculating the variogram function. Missing values ​​in sparse areas of monitoring stations are scientifically filled in, and finally a fused dataset with a consistent spatiotemporal grid is generated, with data integrity reaching over 98%.

[0019] The physical mechanism-guided spatiotemporal feature extraction module implements a multi-channel feature construction strategy to extract spatiotemporal features with clear physical meaning from the fused dataset. This module first calculates four types of fundamental physical features based on the three-dimensional hydrodynamic equations and the pollutant convection-diffusion-reaction equations. Water age distribution features are obtained by solving the conservative mass transport equations of the lake / reservoir. The computational domain is discretized using the finite volume method, and the residence time of the water from the inlet to that location is calculated in each grid cell, forming a water age distribution map with a time scale of 72 hours. Pollutant transport path features are simulated using the Lagrange particle tracking algorithm. Virtual particles are released at the main inflow river mouths and potential pollution source locations, and the fourth-order Runge-Kutta method is used to solve for the particle trajectories, recording the transmission time and path length of each particle to each grid within the lake. Settling rate features are calculated based on Stokes' law, combining on-site collected suspended particulate matter size distribution data with water temperature and viscosity parameters to determine the settling velocity of particles of different sizes. The photolysis rate characteristics are determined based on the molar absorptivity of the target endocrine disruptors and the water body illumination intensity profile. The photolysis reaction rate constants at different depths are calculated by measuring the light attenuation coefficients at different water depths. Next, this module extracts statistical and deep learning features from the fused dataset. Statistical feature calculations include the mean, variance, skewness, and kurtosis of each monitoring indicator within a 24-hour time window, as well as the local Moran's index within a 3×3 spatial neighborhood, used to quantify spatial autocorrelation. Deep learning features are extracted from the spatial distribution map of water quality parameters retrieved from remote sensing using a pre-trained convolutional autoencoder. This autoencoder employs an encoder-decoder structure, with the encoder containing 5 convolutional layers and 3 fully connected layers, ultimately outputting a 256-dimensional compressed feature vector representing the spatial distribution pattern of water quality parameters. After standardization, all features are concatenated and organized into a four-dimensional tensor, with dimensions corresponding to 720 time steps, 150 spatial grid rows, 200 spatial grid columns, and 85 feature channels, forming a complete feature tensor for use by downstream models.

[0020] The causal inference-driven concentration prediction module employs a hybrid architecture combining graph neural networks and long short-term memory networks to process feature tensors. This module first automatically discovers causal dependencies between variables through a graph structure learning layer, and then uses a multi-head attention mechanism to calculate the causal strength score between each pair of feature variables. For each feature channel in the feature tensor, its attention score with all other channels is calculated, and after normalization using the softmax function, a causal strength matrix is ​​obtained. A threshold of 0.05 is set, retaining only edges with strength scores exceeding this threshold, forming a dynamic causal graph. This causal graph serves as the adjacency matrix for subsequent graph convolutional layers, ensuring information propagates among causally related features and avoiding interference from spurious correlations. The long short-term memory network layer consists of two stacked layers with a hidden layer dimension of 128 and a dropout rate of 0.2, used to capture the long-term temporal dependencies of the concentration sequence. During model training, a physical consistency loss term is introduced, which penalizes the deviation between the predicted concentration and the concentration change derived from the mass conservation equation. The physical consistency loss function is defined as the mean square error between the predicted concentration time derivative and the calculated value from the convection-diffusion-reaction equation. The convection term is calculated based on the measured velocity field, the diffusion term uses the anisotropic diffusion tensor, and the reaction term comprehensively considers hydrolysis, photolysis, and biodegradation processes. The total loss function is the weighted sum of the prediction mean square error and the physical consistency loss, with the weight coefficient λ determined to be 0.1 through grid search. The model is trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, a training period of 500 epochs, and an early stopping patience value of 50 epochs.

[0021] The dynamic pollution source analysis and contribution measurement module implements a hybrid source tracing method based on integral gradient and structural equation modeling. This module first uses the integral gradient algorithm to calculate the gradient integral of the prediction model output relative to each input feature, obtaining the initial contribution score of each feature to the final predicted concentration. The integration path adopts a straight line from the baseline to the current input, with the baseline selected as the historical minimum concentration, and the integration step count set to 50 steps to ensure the numerical stability of the gradient integral. Subsequently, the module constructs a structural equation model, using the initial contribution score as the observed variable and pollution source emission intensity, hydrological and meteorological conditions, and geographical distance as latent variables, establishing a linear structural relationship between the latent and observed variables. The structural equation model consists of two parts: a measurement model and a structural model. The measurement model defines the relationship between observed and latent variables, while the structural model defines the causal relationship between latent variables. Model parameters are solved using the maximum likelihood estimation method, and the Bollen-Stine bootstrap method is used to evaluate the model's goodness of fit, requiring a comparison fit index greater than 0.9 and an approximate root mean square error less than 0.08. Finally, the contribution of pollution source i to monitoring point j at time t is calculated by normalizing the product of the integral gradient value and the path coefficient of the structural equation model, ensuring that the sum of the contributions of all pollution sources is 100%. This contribution result has spatiotemporal dynamics, with a temporal resolution of 1 hour and a spatial resolution of 100-meter grid scale, which can accurately reflect the dynamic impact process of different pollution sources in rainfall-runoff events.

[0022] The causal inference-driven concentration prediction module and the dynamic pollution source analysis and contribution quantification module are coupled and cooperate to build a dynamic hybrid framework. Specifically, their coupling is reflected in the following two levels: Data flow coupling, where the prediction result provides input for parsing, such as... Figure 2 As shown, the output of the "causal inference-driven concentration prediction module"—that is, the future concentration prediction result—is the core driver of the entire system. The startup and operation of the "dynamic pollution source analysis module" take this prediction result as its input and target. The core task of the analysis module is to answer: "What characteristics (pollution sources) led the prediction module to arrive at such a specific concentration prediction value?" Without the prediction result, the analysis loses its object and target.

[0023] Causal structural coupling, meaning a shared and unified causal worldview. "Causal discovery" in the prediction module: Figure 2The "graph structure learning layer" in the algorithm is the causal discovery engine of the entire system. It automatically learns from the data and outputs a dynamic causal graph. This graph defines which variables have directed causal dependencies, i.e., edges with a causal strength score exceeding a preset threshold of 0.05. The "causal constraints" in the parsing module: the generated dynamic causal graph is used as the adjacency matrix for subsequent graph convolution operations, ensuring that the information propagation of the prediction model is constrained by causal relationships. More importantly, this causal graph also implicitly constrains the attribution range of the parsing module. When the integral gradient algorithm calculates contributions, it performs backpropagation on this established causal network structure. Only variables identified as causally related by the causal graph can be assigned significant contributions. This avoids incorrectly attributing high contributions to variables that are only statistically correlated but have no causal relationship.

[0024] In summary, the causal inference-driven concentration prediction module and the dynamic pollution source analysis and contribution quantification module are tightly coupled through two key links: "forward prediction output" and "shared causal graph." Together, they transform a prediction problem into an interpretable and traceable causal analysis system. The prediction module ensures the accuracy of the results, while the analysis module, based on this, endows the results with causal credibility, enabling the entire system not only to "predict the future" but also to "understand the reasons," thus providing a solid technical foundation for precise environmental management.

[0025] See Figure 4The visualization-based early warning and decision support module integrates geographic information systems and dynamic dashboard technology, transforming analysis results into intuitive visualizations. This module overlays concentration prediction results as a heatmap onto a high-precision electronic map, using a continuous color map from blue to red to represent the concentration gradient. Areas with concentrations below 1 nanogram per liter are displayed in blue, while areas with concentrations above 100 nanograms per liter are displayed in red. Intermediate concentrations are determined by linear interpolation. Pollution source contribution results are displayed as a stacked area map and flow arrows. The stacked area map shows the change in the contribution ratio of different pollution sources over time, and the flow arrows visualize the migration paths of pollutants based on particle tracking results. The early warning submodule automatically triggers tiered early warnings based on preset concentration thresholds. The initial warning threshold is set at 10 nanograms per liter, and the severe warning threshold is set at 50 nanograms per liter. A yellow warning signal is triggered when the predicted concentration exceeds 10 nanograms per liter, and a red severe warning signal is triggered when it exceeds 50 nanograms per liter. Early warning information is pushed to environmental monitoring center staff in real time via SMS gateway and email server. The decision support submodule includes a built-in rule base of 128 governance rules, with rule conditions set based on a combination of pollution source type, contribution level, and hydrological and meteorological conditions. When agricultural non-point source pollution exceeds 60%, the system automatically generates suggestions for ecological ditch construction and fertilizer reduction plans, including suitable plant configuration schemes and fertilizer application reduction ratios. When industrial point source pollution suddenly increases and exceeds 40%, the system generates on-site verification plans for key polluting enterprises, clearly defining the list of enterprises to be verified, key areas to be verified, and sampling point layout plans. All decision recommendations are verified through an expert knowledge base to ensure the feasibility and effectiveness of the measures.

[0026] Example 2 In the scenario of protecting drinking water sources in deep-water reservoirs, the multi-source heterogeneous data acquisition and fusion module adjusts monitoring strategies to address the unique hydrological characteristics of stratified water bodies. The hydro-meteorological monitoring unit deploys a vertical profile monitoring system in the reservoir area, using a temperature, salinity, and depth profiler to collect vertical distribution data of water temperature, conductivity, and dissolved oxygen every 6 hours, achieving a depth resolution of 0.5 meters and accurately capturing the thermocline. The water quality parameter monitoring unit adds high-frequency monitoring buoys near the main tributary inlets and water intakes in the reservoir area, increasing the monitoring frequency to once per hour, focusing on the risk of endocrine disruptors release due to hypoxia in the bottom water during stratification periods. The human activity intensity monitoring unit adds a tourism activity intensity monitoring indicator, using mobile phone signaling data to calculate the number of tourists around the reservoir and combining it with GPS trajectory data from cruise ships to estimate the tourism pollution load. The remote sensing inversion data acquisition unit introduces synthetic aperture radar data to monitor the distribution of floating debris and the extent of algal blooms, compensating for the monitoring blind spots of optical remote sensing during cloudy and rainy weather. The data fusion submodule uses a three-dimensional spatiotemporal kriging interpolation algorithm to add vertical layers on the basis of the horizontal grid, and establishes a three-dimensional grid system of 100m×100m×2m to more accurately characterize the vertical heterogeneity of deep-water reservoirs.

[0027] The physical mechanism-guided spatiotemporal feature extraction module optimizes physical feature calculations for stratified water bodies. Water age distribution features consider the effect of the thermocline hindering vertical mixing, calculating the water age distribution of the surface and bottom layers separately. Pollutant transport path features add a vertical migration component to simulate the accumulation and release of pollutants at the stratification interface. Sedimentation rate features consider the impact of water temperature stratification on particulate matter sedimentation, using measured vertical temperature gradients to correct the water viscosity parameter in the Stokes equation. When calculating the light intensity at each depth layer, the photolysis rate feature employs a more accurate light attenuation model, considering the influence of chlorophyll and suspended solids concentrations on light penetration. Statistical feature calculations add a vertical gradient index to quantify the concentration differences of various water quality parameters above and below the thermocline. During deep learning feature extraction, the training data incorporates vertical distribution patterns of water quality parameters specific to stratified water bodies, enhancing feature representation capabilities.

[0028] The causal inference-driven concentration prediction module optimizes the model structure for deep-water reservoirs. Vertical connections are added to the graph structure learning layer to capture causal relationships between different water depths. An attention mechanism is introduced into the long short-term memory network layer to enhance feature extraction at key time points of thermocline formation and ablation. A vertical diffusion term is added to the physical consistency loss function to more accurately describe the pollutant migration process in stratified water bodies. The model training data includes monitoring records for a complete hydrological year, ensuring the model grasps the concentration variation patterns during different hydrological stages such as stratification and mixing periods.

[0029] The dynamic pollution source apportionment and contribution quantification module adds vertical dimension analysis to the contribution calculation. The integral gradient algorithm calculates the contribution of surface and bottom layer input features to the predicted concentration separately. The structural equation model incorporates water stability as a latent variable to quantify the impact of stratification intensity on the vertical distribution of pollution source contributions. The final contribution results provide pollution source apportionment at different water depths, accurately identifying the relative contributions of bottom layer endogenous releases and surface layer exogenous inputs.

[0030] The visualization-based early warning and decision support module adjusts early warning thresholds and decision rules to suit the characteristics of drinking water sources. Early warning thresholds are adjusted according to drinking water source quality standards, with a warning threshold set at 5 nmol / L and a severe warning threshold at 20 nmol / L. The decision support rule base adds specific governance measures for water sources. When the contribution of bottom-level endogenous pollution exceeds 30%, it is recommended to initiate bottom-level aeration or sediment passivation projects; when the contribution of tourism pollution exceeds 25%, it is recommended to implement tourist flow control and ecological bank protection construction schemes. The visualization display adds a vertical profile view, simultaneously showing the distribution of concentration and pollution source contribution along water depth, providing three-dimensional decision support for stratified water source management.

[0031] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Therefore, all equivalent changes made in accordance with the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs, characterized in that, The method includes the following steps: Step S110: Collect and fuse multi-source heterogeneous monitoring data of the watershed-lake-reservoir system to construct a standardized spatiotemporal dataset; Step S120: Based on the physical mechanisms of hydrodynamics and pollutant migration and transformation, key spatiotemporal features are extracted from the fused data to construct a high-dimensional feature tensor; Step S130: Construct and train a deep neural network model that integrates causal discovery and physical constraints, and use feature tensors to predict the spatiotemporal distribution of endocrine disruptors concentration in lakes and reservoirs. Step S140: Based on the trained prediction model, analyze the dynamic contribution of different pollution sources to the concentration of secretory interfering substances in the target area. Step S150: Integrate and visualize the concentration prediction results with the pollution source contribution information, and generate graded early warning signals and targeted governance decision-making suggestions.

2. The method for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs according to claim 1, characterized in that, In step S110, the multi-source heterogeneous monitoring data includes hydrological and meteorological data, water quality parameter data, human activity intensity data, and remote sensing inversion data. The data fusion adopts a collaborative correction algorithm based on spatiotemporal kriging interpolation to unify the spatiotemporal scale and fill in missing values ​​of the multi-source heterogeneous monitoring data, generating a fused dataset with a consistent spatiotemporal grid.

3. The method for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs according to claim 1, characterized in that, In step S130, the deep neural network model is a hybrid architecture that integrates graph neural networks and long short-term memory networks; a physical consistency loss term is introduced during model training.

4. A system for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs, applied to the method for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs as described in any one of claims 1-3, characterized in that, The system includes the following components: The multi-source heterogeneous data acquisition and fusion module is used to acquire multi-dimensional monitoring data at the watershed scale and within lakes and reservoirs, and to standardize and align the data in time and space. The multi-source heterogeneous data acquisition and fusion module specifically includes a hydrological and meteorological monitoring unit, a water quality parameter monitoring unit, a human activity intensity monitoring unit, and a remote sensing inversion data acquisition unit. The physical mechanism-guided spatiotemporal feature extraction module is used to extract physically meaningful spatiotemporal features from the fused data and construct feature tensors. The spatiotemporal feature extraction module guided by the physical mechanism adopts a multi-channel feature construction strategy, and defines basic physical features based on the hydrodynamic equation and the pollutant convection-diffusion-reaction equation, including water age distribution, pollutant transport path, sedimentation rate and photolysis rate. The causal inference-driven concentration prediction module is used to build and train a deep learning model that integrates causal constraints based on extracted spatiotemporal features, so as to realize the dynamic prediction of the concentration of endocrine disruptors in lakes and reservoirs. The core of the concentration prediction module driven by causal inference is a hybrid architecture that integrates graph neural networks and long short-term memory networks, including a graph structure learning layer for automatically discovering causal dependencies between variables from high-dimensional feature tensors. The Dynamic Pollution Source Analysis and Contribution Measurement Module is used to analyze the spatiotemporal dynamic contribution of different pollution sources to the concentration of secretory interfering substances in the target area based on the internal mechanism and input characteristics of the prediction model. The dynamic pollution source analysis and contribution measurement module implements a hybrid source tracing method based on integral gradient and structural equation model; The visualization early warning and decision support module is used to visualize the prediction results and traceability information, and generate risk early warning and control suggestion reports; The visualization early warning and decision support module integrates geographic information system and dynamic dashboard technology.

5. The system for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs according to claim 4, characterized in that, In the multi-source heterogeneous data acquisition and fusion module, the hydro-meteorological monitoring unit is responsible for collecting data on precipitation, evaporation, runoff, water temperature, wind speed, and wind direction; the water quality parameter monitoring unit is responsible for collecting laboratory analysis concentrations of water body pH, dissolved oxygen, turbidity, chlorophyll a, and target endocrine disruptors; the human activity intensity monitoring unit obtains data on industrial wastewater discharge, urban domestic sewage discharge, agricultural fertilizer application, and livestock and poultry breeding by accessing regional statistical yearbooks, sewage discharge permit databases, and land use data; and the remote sensing inversion data acquisition unit inverts water body suspended solids concentration, colored dissolved organic matter concentration, and water temperature spatial distribution maps by processing multispectral and hyperspectral satellite images.

6. The system for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs according to claim 4, characterized in that, In the spatiotemporal feature extraction module guided by the physical mechanism, the water age distribution is obtained by solving the conservative material transport equation of the lake and reservoir water; the pollutant transport path is simulated based on the Lagrange particle tracking algorithm; the sedimentation rate is calculated based on Stokes' law and the adsorption characteristics of particulate matter; and the photolysis rate is determined based on the molar absorptivity of the target endocrine disruptor and the light intensity profile of the water body. The physical mechanism-guided spatiotemporal feature extraction module extracts statistical features and deep learning features from the fused dataset. The statistical features include the mean, variance, skewness, and kurtosis of each monitoring indicator within the time window, as well as the local Moran index in the spatial neighborhood. The deep learning features are extracted from the spatial distribution map of water quality parameters retrieved by remote sensing through a pre-trained convolutional autoencoder. All features are concatenated and organized into a four-dimensional tensor, with dimensions of time step, number of spatial grid rows, number of spatial grid columns, and number of feature channels.

7. The system for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs according to claim 4, characterized in that, In the causal inference-driven concentration prediction module, the graph structure learning layer adopts an attention-based mechanism to calculate the causal strength score between each pair of feature variables. Edges with scores exceeding a preset threshold of 0.05 are retained to form a dynamic causal graph. The hidden layer dimension of the long short-term memory network layer is set to 128. During model training, a physical consistency loss term is introduced, which penalizes the deviation between the predicted concentration and the concentration change derived from the mass conservation equation.

8. The system for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs according to claim 4, characterized in that, In the dynamic pollution source analysis and contribution measurement module, the integral gradient algorithm calculates the gradient integral of the prediction model output with respect to each input feature to obtain the initial contribution score of each feature to the final predicted concentration. The integral path adopts a straight line path. The structural equation model uses the initial contribution score as the observed variable and the pollution source emission intensity, hydro-meteorological conditions and geographical distance as latent variables to establish a linear structural relationship between the latent variables and the observed variables. The model parameters are solved by the maximum likelihood estimation method.

9. The system for predicting and tracing the concentration of endocrine disruptors in lakes and reservoirs according to claim 4, characterized in that, The visualization early warning and decision support module overlays the concentration prediction results onto a high-precision electronic map in the form of a heat map, with the color mapping range from blue to red; the pollution source contribution results are displayed in the form of a stacked area map and flow direction arrows; the early warning submodule automatically triggers graded early warning signals according to preset concentration thresholds; the decision support submodule has a built-in rule base, which contains suggestions for governance measures for different pollution source types and contribution levels.

Images