Floodplain chemical industry area multi-medium pollution monitoring and early warning method and system, storage medium and computer equipment

Abstract

The present application belongs to the technical field of environmental monitoring and early warning and environmental pollution source analysis, and particularly relates to a kind of flood plain chemical area multi-medium pollution monitoring and early warning method, system, storage medium and computer equipment. Including steps: S1, data acquisition and pretreatment;S2, dynamic structure causal diagram construction;S3, SCM-IML hybrid modeling;S4, causal effect quantification;S5, differential early warning instruction generation and output. Can improve the precision of flood plain chemical area multi-medium pollution monitoring and early warning.

Description

Technical Field

[0001] This invention belongs to the field of environmental monitoring and early warning and environmental pollution source analysis technology, specifically involving a multi-media pollution monitoring and early warning method, system, storage medium and computer equipment for floodplain chemical industrial zones. Background Technology

[0002] As a typical "water-land-life" interaction sensitive zone, the multi-media pollution (water-soil-air-sediment) processes in floodplain chemical industrial zones are driven by both hydrodynamics and human activities, exhibiting significant spatiotemporal heterogeneity. Floodplains are mostly located on first-level river terraces, characterized by high media permeability and intense surface water-groundwater interaction, providing a complex environmental carrier for the multi-media migration of characteristic pollutants such as phosphorus, sulfur, and titanium. Rivers in my country's monsoon climate zone generally exhibit dramatic flow variations between wet and dry seasons. This hydrological dynamic can alter the energy conditions and pathways for pollutant migration, leading to nonlinear abrupt changes in pollution processes and exacerbating the difficulty of dynamic control of regional environmental risks. Therefore, accurately monitoring the spatiotemporal patterns of multi-media pollution in floodplain chemical industrial zones, timely identifying pollution source contributions, and issuing source-specific and phased early warnings are of significant engineering value for precise pollution control and ecological environment risk prevention at the watershed scale.

[0003] Currently, the mainstream technologies for multi-media pollution monitoring and early warning in the Hemantan chemical industrial zone mainly rely on traditional environmental monitoring networks and static statistical models. At the monitoring level, periodic sampling is typically conducted using fixed groundwater monitoring wells, surface water monitoring sections, and soil sampling points, with pollutant concentration data obtained through manual testing or online monitoring. At the analytical level, statistical methods such as multiple linear regression, principal component analysis, or positive definite matrix factor decomposition are mainly used for source apportionment and pollution prediction.

[0004] However, traditional methods have the following shortcomings when dealing with multi-media pollution problems in the dynamic hydrological environment of floodplains: First, traditional monitoring networks lack the ability to couple with hydrological dynamics and cannot automatically switch monitoring strategies and early warning thresholds according to changes in river flow during the wet and dry seasons, making it difficult to differentiate between industrial seepage risks during the dry season and agricultural non-point source runoff risks during the wet season; Second, traditional source apportionment techniques rely heavily on statistical correlation analysis, simplifying the complex causal relationship between pollution sources and pollution receptors into a linear response, and failing to capture the nonlinear threshold effect and migration path switching mechanism caused by drastic flow changes.

[0005] The shortcomings of traditional methods directly lead to a bottleneck in pollution monitoring and early warning in floodplain chemical industrial zones: First, there is a lack of quantitative analysis of the dynamic coupling relationship between "hydrological dynamics-migration path-pollution pattern," making it impossible to clarify the transmission mechanism by which industrial emissions, agricultural activities, and river flow indirectly regulate the distribution of multi-media pollution by influencing pollutant migration paths, resulting in early warning releases lagging behind dynamic changes in pollution; Second, it is difficult to identify key pollution driving factors at different sources and stages from a causal mechanism perspective, and existing estimates are mostly based on "static assumptions" or "empirical models," failing to develop differentiated source contribution analysis and early warning strategies for dry and wet seasons, resulting in insufficient targeting and timeliness; Third, the "black box" status of cross-media migration paths has not yet been broken, and the migration process of pollutants from emission sources to receptors still relies on empirical assumptions, especially lacking verification of the real-time regulation mechanism of hydrological dynamics on path flux allocation, leading to insufficient scientific rigor in pollution risk assessment. The unique hydrogeological conditions of the floodplain chemical industrial zone further amplify these problems. Dramatic changes in flow during the wet and dry seasons lead to dynamic transformations in the direction of the hydraulic gradient (vertical / lateral). Under low flow conditions, the vadose zone retention effect forms a migration bottleneck, while under high runoff conditions, surface runoff activates lateral diffusion. Traditional methods have failed to fully incorporate these nonlinear influencing factors of hydrological-pollution coupling, and lack the technical architecture to embed dynamic structural causal inference and interpretable machine learning into the monitoring and early warning system. This results in poor accuracy and delayed response of multi-media pollution monitoring and early warning, which seriously restricts the effectiveness of precise pollution control and ecological environment risk prevention and control at the watershed scale. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method, system, storage medium and computer equipment for multi-media pollution monitoring and early warning in floodplain chemical industrial zones, which can improve the accuracy of multi-media pollution monitoring and early warning in floodplain chemical industrial zones.

[0007] The technical solution adopted by this invention to solve its technical problem is: a multi-media pollution monitoring and early warning method for floodplain chemical industrial zones, comprising the following steps: S1. Data Acquisition and Preprocessing: Acquire environmental monitoring data, enterprise production data, agricultural activity data, and hydrological and meteorological data from the floodplain chemical industrial zone to form a multi-source spatiotemporal dataset. Environmental monitoring data includes groundwater pollutant concentration data, surface water pollutant concentration data, and soil pollutant concentration data. Enterprise production data includes sulfuric acid production data, phosphate fertilizer production data, and characteristic pollutant emission log data. Agricultural activity data includes fertilizer application data and spatial distribution data of planting structure. Hydrological and meteorological data includes daily flow records from watershed hydrological stations and cumulative precipitation data. Preprocess the multi-source spatiotemporal dataset to obtain a standardized spatiotemporal grid dataset. S2, Construction of dynamic structural causal graph: Construct an initial directed acyclic graph containing industrial emission nodes, agricultural input nodes, flow nodes, surface water concentration nodes, and groundwater concentration nodes; identify significant causal edges; identify strong causal chains during the dry season and strong causal chains during the wet season, and generate a dynamic structural causal graph; S3, SCM-IML Hybrid Modeling: Based on dynamic structural causal graphs, a nonlinear structural equation for groundwater pollutant concentration is constructed; a hydrological period-based modeling strategy is adopted to construct a hybrid model of structural causal model and interpretable machine learning, training sub-models for wet and dry periods respectively, fitting the nonlinear relationship between pollutant concentration and various data in the multi-source spatiotemporal dataset; the interpretability of the hybrid model is achieved through SHAP value decomposition and partial dependency graphs to diagnose hydrological threshold effects; pollution prediction results and source contribution decomposition values ​​are output. S4, Causal Effect Quantification: Based on a mixed model, the stratified average treatment effect analysis method and the mediation effect analysis method are used to quantify the change in pollutant concentration, the proportion of pollutant migration path flux, and the contribution ratio of each pollution source when the flow rate increases by one unit. S5, Differentiated Early Warning Command Generation and Output: Based on the results of automatic hydrological period division, pollution prediction results, and the contribution ratio of each pollution source, differentiated early warning commands are generated by source and period, and the pollution prediction results, the contribution ratio of each pollution source, the proportion of pollutant migration path flux, and the hydrological threshold effect diagnosis results are output.

[0008] Furthermore, step S1 involves preprocessing the multi-source spatiotemporal dataset, including spatiotemporal alignment, spatial grid standardization, automatic hydrological period division based on flow thresholds, and construction of features derived from lag variables and interaction terms.

[0009] Furthermore, step S2 specifically includes the following steps: S201, Initial Directed Acyclic Graph Construction: Based on prior knowledge of the pollution process in the floodplain chemical zone, an initial directed acyclic graph is constructed, including industrial emission nodes, agricultural input nodes, flow nodes, surface water concentration nodes, and groundwater concentration nodes. S202, PC Algorithm Data-Driven Optimization: The PC algorithm is used to perform data-driven optimization on the initial directed acyclic graph, and significant causal edges are identified through Gaussian conditional independence test; S203, Path selection for differences between wet and dry seasons: The PC algorithm is used to automatically select paths with significant differences between wet and dry seasons, identify strong causal chains in the dry season and strong causal chains in the wet season, and generate a dynamic causal graph.

[0010] Furthermore, step S3 specifically includes the following steps: S301, Construction of Nonlinear Structural Equations: Using dynamic structural causal graphs as the theoretical framework, nonlinear structural equations for groundwater pollutant concentrations are constructed. S302, Hydrological Period Sub-model Training: A hydrological period modeling strategy is adopted to construct a hybrid model of structural causal model and interpretable machine learning. Based on the sample data of the wet season and the sample data of the dry season, the wet season sub-model and the dry season sub-model are trained respectively. S303, SHAP value decomposition: By decomposing the causal effects of each pollution source's contribution through SHAP value, the differentiated contributions of industrial and agricultural sources in different hydrological periods are quantified. S304, Partial Dependency Plot Diagnosis: Visualizing the marginal effect of flow on pollutant concentration using partial dependency plots to diagnose hydrological threshold effects; S305, Interaction Term Calculation: Embedded interaction term calculation module, outputs pollution prediction results and source contribution decomposition values.

[0011] Furthermore, step S4 specifically includes the following steps: S401, Stratified average treatment effect analysis: Based on the mixed model, the stratified average treatment effect analysis method is used to compare the causal effect of flow changes during the wet and dry seasons on pollutant concentrations. By calculating the difference between the predicted values ​​of the high flow scenario and the baseline scenario, the change in pollutant concentration for each unit increase in flow is quantified. S402, Mediation effect analysis: Based on the mixed model, the mediation effect analysis method is used to quantify the flux proportion of pollutant migration paths, construct a linear model and regression equation for mediating variables, and calculate the contribution rate of the mediating effect of pollutant migration paths. S403, Spatial Distribution of Counterfactual Source Contributions: The spatial distribution of pollution source contributions is quantified through counterfactual analysis, and the contribution ratio of each pollution source is quantified through a counterfactual scenario of zero industrial emissions.

[0012] Furthermore, step S403 is followed by step S404, residual spatial clustering: combining residual spatial clustering to identify pollution hotspots not explained by the hybrid model, performing spatial clustering analysis on the residuals of the hybrid model, and identifying the overlap between the hotspots not explained by the hybrid model and the groundwater level anomaly areas.

[0013] Furthermore, the differentiated early warning instruction in step S5 specifically includes: When it is determined to be a dry season and the flow rate is below the critical threshold, an industrial point source leakage early warning instruction is generated. When it is determined to be a high-water season and the flow rate is higher than the high-water threshold, an agricultural non-point source runoff early warning instruction is generated.

[0014] The multi-media pollution monitoring and early warning system for the Hemantan Chemical Zone adopts a multi-media pollution monitoring and early warning method for the Hemantan Chemical Zone, including a data acquisition and preprocessing module, a dynamic structural cause-effect graph construction module, an SCM-IML hybrid modeling module, a cause-effect quantification module, and a differentiated early warning command generation and output module. The data acquisition and preprocessing module is used to acquire environmental monitoring data, enterprise production data, agricultural activity data, and hydrological and meteorological data from the floodplain chemical industrial zone to form a multi-source spatiotemporal dataset. The environmental monitoring data includes groundwater pollutant concentration data, surface water pollutant concentration data, and soil pollutant concentration data. The enterprise production data includes sulfuric acid production data, phosphate fertilizer production data, and characteristic pollutant emission logs. The agricultural activity data includes fertilizer application data and spatial distribution data of planting structures. The hydrological and meteorological data includes daily flow records from watershed hydrological stations and cumulative precipitation data. The multi-source spatiotemporal dataset is preprocessed to obtain a standardized spatiotemporal grid dataset. The dynamic structural causal graph construction module is used to construct an initial directed acyclic graph containing industrial emission nodes, agricultural input nodes, flow nodes, surface water concentration nodes, and groundwater concentration nodes; identify significant causal edges; identify strong causal chains during the dry season and strong causal chains during the wet season; and generate a dynamic structural causal graph. The SCM-IML hybrid modeling module is used to construct a nonlinear structural equation for groundwater pollutant concentration based on a dynamic structural causal graph. It employs a hydrological period-based modeling strategy to construct a hybrid model combining structural causal modeling and interpretable machine learning, training sub-models for both high-water and low-water periods to fit the nonlinear relationship between pollutant concentration and various data points in a multi-source spatiotemporal dataset. The interpretability of the hybrid model is achieved through SHAP value decomposition and partial dependency graphs, diagnosing hydrological threshold effects. Finally, it outputs pollution prediction results and source contribution decomposition values. The causal effect quantification module is used to quantify the change in pollutant concentration, the proportion of pollutant migration path flux, and the contribution ratio of each pollution source when the flow rate increases by one unit based on a hybrid model, using the stratified average treatment effect analysis method and the mediation effect analysis method. The differentiated early warning command generation and output module is used to generate differentiated early warning commands by source and period based on the results of automatic hydrological period division, pollution prediction results and the contribution ratio of each pollution source, and output pollution prediction results, contribution ratio of each pollution source, pollutant migration path flux ratio and hydrological threshold effect diagnosis results.

[0015] A storage medium storing a computer program, which, when executed by a processor, implements the steps of a multi-media pollution monitoring and early warning method for a river floodplain chemical industrial zone.

[0016] A computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of a multi-media pollution monitoring and early warning method for a river floodplain chemical industrial zone.

[0017] Compared with the prior art, the beneficial effects of the present invention are: the present invention provides a method, system, storage medium and computer equipment for multi-media pollution monitoring and early warning in floodplain chemical industrial zones, which can improve the accuracy of multi-media pollution monitoring and early warning in floodplain chemical industrial zones.

[0018] Specifically, through data collection and preprocessing, the system achieves spatiotemporal alignment and standardization of heterogeneous data from multiple sources, including groundwater, surface water, soil, enterprise production, agricultural activities, and hydrological and meteorological data, thus constructing a unified data foundation. Through dynamic structural causal graph construction, it integrates physical cognition of "source-hydrology-path-sink" with PC algorithm data-driven optimization to achieve automatic identification and dynamic updating of causal paths differing between wet and dry seasons. Through SCM-IML hybrid modeling, it preserves physical interpretability with a structural causal model, uses gradient boosting trees to handle nonlinear responses and threshold effects, and combines SHAP value decomposition to achieve transparent model analysis. Through causal effect quantification, it achieves full-chain causal quantification, revealing the flux allocation mechanism of migration paths. Through differentiated early warning command generation and output, based on automatic hydrological period division and pollution prediction results, it dynamically switches between source- and period-specific differentiated early warning strategies, focusing on industrial point source leakage during the dry season and agricultural non-point source runoff risk during the wet season, achieving a leap from passive, lagging monitoring to proactive, real-time early warning. Ultimately, this invention achieves a dual unity of physical interpretability and nonlinear fitting accuracy, causal mechanism characterization and real-time early warning application, providing scientific and engineering-practical technical support for precise pollution control and ecological environment risk prevention at the watershed scale. Attached Figure Description

[0019] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a diagram of the overall system architecture of the present invention; Figure 3 This is a flowchart of the dynamic structure causal graph construction and SCM-IML hybrid modeling in this invention; Figure 4 This is a flowchart illustrating the quantification of causal effects and the generation and output of differentiated early warning instructions in this invention. Figure 5 This is a schematic diagram of the differential warning instruction generation logic in this invention. Detailed Implementation

[0020] The following is in conjunction with the appendix Figure 1 , Figure 2 , Figure 3 , Figure 4 and Figure 5 The invention is further illustrated by the embodiments.

[0021] The flow rate in this article specifically refers to the flow rate at the upstream and downstream sections of the river and the flow rate at sewage outlets.

[0022] like Figure 1 As shown, the multi-media pollution monitoring and early warning method for the river floodplain chemical industrial zone includes the following steps: S1. Data Acquisition and Preprocessing: Acquire environmental monitoring data, enterprise production data, agricultural activity data, and hydrological and meteorological data from the floodplain chemical industrial zone to form a multi-source spatiotemporal dataset. Environmental monitoring data includes groundwater pollutant concentration data, surface water pollutant concentration data, and soil pollutant concentration data. Enterprise production data includes sulfuric acid production data, phosphate fertilizer production data, and characteristic pollutant emission log data. Agricultural activity data includes fertilizer application data and spatial distribution data of planting structure. Hydrological and meteorological data includes daily flow records from watershed hydrological stations and cumulative precipitation data. Preprocess the multi-source spatiotemporal dataset to obtain a standardized spatiotemporal grid dataset.

[0023] Specifically, step S1 involves preprocessing the multi-source spatiotemporal dataset, including spatiotemporal alignment, spatial grid standardization, automatic hydrological period division based on flow thresholds, and feature construction of lagged variables and interaction terms.

[0024] like Figure 3 As shown in Figure S2, dynamic structural causal graph construction: construct an initial directed acyclic graph containing industrial emission nodes, agricultural input nodes, flow nodes, surface water concentration nodes, and groundwater concentration nodes; identify significant causal edges; identify strong causal chains during the dry season and strong causal chains during the wet season, and generate a dynamic structural causal graph.

[0025] Step S2 specifically includes the following steps: S201, Initial Directed Acyclic Graph Construction: Based on prior knowledge of the pollution process in the floodplain chemical zone, an initial directed acyclic graph is constructed, including industrial emission nodes, agricultural input nodes, flow nodes, surface water concentration nodes, and groundwater concentration nodes; among them, the flow nodes explicitly express their regulatory role on pollutant migration paths by adjusting edge attributes. S202, PC Algorithm Data-Driven Optimization: The PC algorithm is used to perform data-driven optimization on the initial directed acyclic graph, and significant causal edges are identified through Gaussian conditional independence test; S203, Path selection for differences between wet and dry seasons: The PC algorithm is used to automatically select paths with significant differences between wet and dry seasons, identify strong causal chains in the dry season and strong causal chains in the wet season, and generate a dynamic causal graph.

[0026] like Figure 3As shown, S3, SCM-IML hybrid modeling (i.e., hybrid modeling of structural causal model and interpretable machine learning): Based on the dynamic structural causal graph, a nonlinear structural equation for groundwater pollutant concentration is constructed; a hydrological period-based modeling strategy is adopted to construct a hybrid model of structural causal model and interpretable machine learning, and sub-models for the wet season and dry season are trained separately to fit the nonlinear relationship between pollutant concentration and each data in the multi-source spatiotemporal dataset; the interpretability of the hybrid model is achieved through SHAP value decomposition and partial dependency graph to diagnose the hydrological threshold effect; and the pollution prediction results and source contribution decomposition values ​​are output.

[0027] Groundwater pollutant concentration refers to the concentration of pollutants in groundwater from different sources.

[0028] Step S3 specifically includes the following steps: S301, Construction of Nonlinear Structural Equations: Using a dynamic structural causal graph as the theoretical framework, a nonlinear structural equation for groundwater pollutant concentration is constructed; where the nonlinear function f is fitted by the gradient boosting tree algorithm. S302, Hydrological Period Sub-model Training: A hydrological period modeling strategy is adopted to construct a hybrid model of structural causal model and interpretable machine learning. Based on the sample data of the wet season and the sample data of the dry season, the wet season sub-model and the dry season sub-model are trained respectively to capture the mechanism switching caused by the drastic change in flow and to improve the model accuracy. S303, SHAP value decomposition: By decomposing the causal effects of each pollution source's contribution through SHAP value, the differentiated contributions of industrial and agricultural sources in different hydrological periods are quantified. S304, Partial Dependency Plot Diagnosis: Visualizing the marginal effect of flow on pollutant concentration using partial dependency plots to diagnose hydrological threshold effects; S305, Interaction Term Calculation: Embedded interaction term calculation module, outputs pollution prediction results and source contribution decomposition values, capturing the synergistic effect between hydrodynamics and pollution sources.

[0029] like Figure 4 As shown in Figure S4, quantification of causal effects: Based on a hybrid model, the stratified average treatment effect analysis method and the mediation effect analysis method are used to quantify the change in pollutant concentration, the proportion of pollutant migration path flux, and the contribution ratio of each pollution source when the flow rate increases by one unit.

[0030] Step S4 specifically includes the following steps: S401, Stratified average treatment effect analysis: Based on the mixed model, the stratified average treatment effect analysis method is used to compare the causal effect of flow changes during the wet and dry seasons on pollutant concentrations. By calculating the difference between the predicted values ​​of the high flow scenario and the baseline scenario, the change in pollutant concentration for each unit increase in flow is quantified. S402, Mediation effect analysis: Based on the mixed model, the mediation effect analysis method is used to quantify the flux proportion of pollutant migration paths, construct a linear model and regression equation for mediating variables, and calculate the contribution rate of the mediating effect of pollutant migration paths. S403, Spatial Distribution of Counterfactual Source Contributions: The spatial distribution of pollution source contributions is quantified through counterfactual analysis, and the contribution ratio of each pollution source is quantified through a counterfactual scenario of zero industrial emissions.

[0031] Preferably, step S403 is followed by step S404, residual spatial clustering: combining residual spatial clustering to identify pollution hotspots not explained by the hybrid model, performing spatial clustering analysis on the residuals of the hybrid model, and identifying the overlap between the hotspots not explained by the hybrid model and the groundwater level anomaly areas.

[0032] like Figure 4 , Figure 5 As shown, S5, Differentiated Early Warning Command Generation and Output: Based on the results of automatic hydrological period division, pollution prediction results and the contribution ratio of each pollution source, differentiated early warning commands by source and period are generated, and the pollution prediction results, the contribution ratio of each pollution source, the proportion of pollutant migration path flux and the hydrological threshold effect diagnosis results are output.

[0033] The differentiated early warning instructions in step S5 specifically include: When it is determined to be a dry season and the flow rate is below the critical threshold, an industrial point source leakage early warning instruction is generated. When it is determined to be a high-water season and the flow rate is higher than the high-water threshold, an agricultural non-point source runoff early warning instruction is generated.

[0034] like Figure 2 As shown, the multi-media pollution monitoring and early warning system for the Hemantan Chemical Zone adopts a multi-media pollution monitoring and early warning method for the Hemantan Chemical Zone, including a data acquisition and preprocessing module, a dynamic structural cause-effect graph construction module, an SCM-IML hybrid modeling module, a cause-effect quantification module, and a differentiated early warning command generation and output module. The data acquisition and preprocessing module is used to acquire environmental monitoring data, enterprise production data, agricultural activity data, and hydrological and meteorological data from the floodplain chemical industrial zone to form a multi-source spatiotemporal dataset. The environmental monitoring data includes groundwater pollutant concentration data, surface water pollutant concentration data, and soil pollutant concentration data. The enterprise production data includes sulfuric acid production data, phosphate fertilizer production data, and characteristic pollutant emission logs. The agricultural activity data includes fertilizer application data and spatial distribution data of planting structures. The hydrological and meteorological data includes daily flow records from watershed hydrological stations and cumulative precipitation data. The multi-source spatiotemporal dataset is preprocessed to obtain a standardized spatiotemporal grid dataset. The dynamic structural causal graph construction module is used to construct an initial directed acyclic graph containing industrial emission nodes, agricultural input nodes, flow nodes, surface water concentration nodes, and groundwater concentration nodes; identify significant causal edges; identify strong causal chains during the dry season and strong causal chains during the wet season; and generate a dynamic structural causal graph. The SCM-IML hybrid modeling module is used to construct a nonlinear structural equation for groundwater pollutant concentration based on a dynamic structural causal graph. It employs a hydrological period-based modeling strategy to construct a hybrid model combining structural causal modeling and interpretable machine learning, training sub-models for both high-water and low-water periods to fit the nonlinear relationship between pollutant concentration and various data points in a multi-source spatiotemporal dataset. The interpretability of the hybrid model is achieved through SHAP value decomposition and partial dependency graphs, diagnosing hydrological threshold effects. Finally, it outputs pollution prediction results and source contribution decomposition values. The causal effect quantification module is used to quantify the change in pollutant concentration, the proportion of pollutant migration path flux, and the contribution ratio of each pollution source when the flow rate increases by one unit based on a hybrid model, using the stratified average treatment effect analysis method and the mediation effect analysis method. The differentiated early warning command generation and output module is used to generate differentiated early warning commands by source and period based on the results of automatic hydrological period division, pollution prediction results and the contribution ratio of each pollution source, and output pollution prediction results, contribution ratio of each pollution source, pollutant migration path flux ratio and hydrological threshold effect diagnosis results.

[0035] A storage medium storing a computer program, which, when executed by a processor, implements the steps of a multi-media pollution monitoring and early warning method for a river floodplain chemical industrial zone.

[0036] A computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of a multi-media pollution monitoring and early warning method for a river floodplain chemical industrial zone.

[0037] Example 1: This study focuses on a typical phosphorus, sulfur, and titanium chemical industrial park in the middle and lower reaches of the Yangtze River. Located on the first terrace of the Yangtze River at an altitude of 50–70 meters, the park's soil is primarily sandy loam with a permeability coefficient of 5–10 m / d, exhibiting intense interaction between surface water and groundwater. The park covers approximately 12 km², comprising 12 titanium dioxide and phosphate fertilizer production enterprises and supporting storage facilities. Surrounding it are 5.3 km² of farmland, primarily used for rice and rapeseed cultivation. The region has a subtropical monsoon climate with an average annual flow of 120 m³ / s, exceeding 200 m³ / s during the wet season and often falling below 20 m³ / s during the dry season. The deployment period is from 2015 to 2023, and the specific implementation steps are as follows: S1, Data Acquisition and Preprocessing: Construct a multi-dimensional, multi-media pollution monitoring network and clarify the observation elements and deployment standards: Groundwater monitoring sensors were deployed in the park and surrounding areas using 52 monitoring wells, each 15-25m deep, covering a sandy loam aquifer. SO4²⁻ samples were collected monthly. - Concentration data of pollutants such as TP and Ti were determined by ion chromatography and inductively coupled plasma mass spectrometry. Surface water monitoring sensors were deployed at eight monitoring sections along the downstream river of the industrial park to monitor SO4² weekly. - The concentrations of TP and Ti, as well as conventional physicochemical indicators, were monitored. Three sections were located 200 m downstream of the enterprise's discharge outlet and were used for key monitoring of industrial emission diffusion. Soil sampling point sensors are set up with 100 sampling points in a 500 m × 500 m grid to collect 0-20 cm of topsoil and determine the total amount and available content of pollutants; The enterprise production data acquisition terminal connects with the online monitoring equipment through the enterprise pollution discharge permit system to obtain sulfuric acid production, phosphate fertilizer production, and emission logs of characteristic pollutants (SO2, TP) from 12 enterprises, with a time resolution of daily. The agricultural activity data collection terminal obtains data on fertilizer application (nitrogen fertilizer 180-220 kg / ha·a, phosphorus fertilizer 90-120 kg / ha·a) and spatial distribution of planting structure (Shapefile format) of a surrounding 5.3 km² of farmland through farmland survey and UAV remote sensing interpretation. The hydrological and meteorological data acquisition terminal accesses daily flow records (accuracy 0.1 m³ / s) and 3-month cumulative precipitation data (accuracy 1 mm) from the watershed hydrological stations.

[0038] Spatiotemporal alignment processing is performed on multi-source spatiotemporal datasets. Specifically, monthly groundwater data, weekly surface water data, daily enterprise data, and daily hydrological data are uniformly mapped to a monthly scale, and the time window averaging method is used to process data with different time resolutions. Spatial grid standardization processing is performed on the multi-source spatiotemporal dataset. Specifically, a 500m×500m spatial grid is used to process the multi-source data to a uniform resolution, and the point monitoring data is interpolated to the grid using the inverse distance weighting method. The system automatically divides the hydrological period into multiple spatiotemporal datasets based on flow thresholds. Specifically, it uses the flow P75 threshold of 110 m³ / s as the threshold for determining the wet season and the critical flow for the dry season as 10 m³ / s. The system automatically divides the data from 2015 to 2023 into wet season samples (flow > 110 m³ / s) and dry season samples (flow ≤ 10 m³ / s). We construct features derived from lagged variables and interaction terms for multi-source spatiotemporal datasets. Specifically, we construct lagged industrial emissions (lag1_emission) to reflect the cumulative effect of pollutant migration, and construct the interaction term between flow and SO2 emission intensity (flow×emission) to capture the synergistic effect between hydrodynamics and industrial sources.

[0039] S2, Construction of a dynamic cause-effect graph: S201, Initial construction of the directed acyclic graph: Based on the physical understanding of the "source-pathway-sink" structure of the floodplain chemical industrial zone, an initial DAG (Directed Acyclic Graph) was constructed, comprising nodes for industrial emissions, agricultural inputs, river flow, vadose zone retention, surface water concentration, and groundwater concentration. The river flow nodes explicitly express their regulatory role on pollutant migration pathways by adjusting edges (→) that point to the vadose zone retention and surface water nodes.

[0040] S202, PC algorithm data-driven optimization: The initial DAG is optimized using the PC algorithm, with a Gaussian conditional independence test significance level of α=0.01. The system automatically traverses all node pairs, removes redundant edges, and identifies strong causal connections.

[0041] S203, Pathway Screening Based on Differences Between High and Low Water Seasons: The PC algorithm was run on samples from both the dry season and the wet season to automatically filter differential paths with Δp << 0.05. The system identified a strong causal chain of "industry → vadose zone → groundwater" (p << 0.001) in the dry season data and a strong causal chain of "agriculture → surface water → groundwater" (p << 0.001) in the wet season data, generating a dynamic causal graph in GraphML format and embedding the effect of hydrological season on path weights.

[0042] S3, SCM-IML hybrid modeling: S301, Construction of nonlinear structural equations: Using a dynamic structural causal diagram as the theoretical framework, a groundwater SO4²2 model was constructed. - Nonlinear structural equation for concentration: Groundwater SO4² - = f(industrial, agriculture, flow, lag1_emission, flow×emission); The nonlinear function f is fitted using the Gradient Boosting Tree (GBRT) algorithm, with a tree depth of 8, a learning rate of 0.05, and 500 iterations. In the formula, Industrial represents industrial emissions, agriculture represents agricultural input, flow represents flow rate, lag1_emission represents the lagged industrial emissions, and flow×emission represents the interaction term between flow rate and SO2 emission intensity.

[0043] S302, Hydrological Period Sub-model Training: Sub-models were trained separately for samples from the wet season and the dry season. The dry season sub-model (dry_model) focuses on fitting the nonlinear response of the vadose zone retention effect under low flow conditions. The wet season sub-model (wet_model) focuses on fitting the linear growth trend of surface runoff activation under high runoff conditions.

[0044] S303, SHAP value decomposition: The causal effects of each source's contribution are decomposed using SHAP values. The visualization module outputs the impact of industrial emissions during the dry season on groundwater SO4²⁻. - The SHAP value of agricultural inputs accounted for 68.5%, agricultural inputs contributed 12.3%, and river flow contributed 15.8%. During the high-water season, agricultural activities accounted for 52.1% of the TP SHAP value, while industrial emissions contributed 28.4% and river flow contributed 16.3%.

[0045] S304, Partial Dependency Graph Diagnosis: The marginal effect of flow rate on pollutant concentration was visualized using PDP (Partial Dependence Plot) curves. The system identified an S-shaped threshold response in the PDP curves during the dry season: when the flow rate was below 10 m³ / s, SO₄²⁻... - The concentration increases sharply as the flow rate decreases, resulting in a sudden change in pollutant retention; when the flow rate exceeds 10 m³ / s, the rate of concentration decrease slows down. During the high-water season, the PDP curve shows a linear growth pattern, with TP concentration increasing by approximately 21% for every 50 m³ / s increase in flow rate.

[0046] S305, Interaction Item Calculation: The interaction term between flow and SO2 emission (flow×SO2_emission) is calculated and used as an important input feature of the gradient boosting tree model to capture the synergistic amplification effect when high runoff and industrial emissions overlap during the wet season.

[0047] S4, Quantification of Causal Effects: S401, Stratified average treatment effect analysis: A stratified average treatment effect analysis method was used to compare the effects of flow changes during wet and dry seasons on groundwater SO4².- Causal effect of concentration. Counterfactual analysis shows that for every 1 m³ / s increase in flow rate during the dry season, groundwater SO₄²⁻... - The concentration decreased by 0.38 mg / L; during the high-water season, the concentration increased by 0.12 mg / L for every 1 m³ / s increase in flow rate. This opposite trend reveals the nonlinear regulatory mechanism of drastic flow rate changes on the pollution process.

[0048] S402, Mediation Effect Analysis: An intermediate pathway model of "industrial emissions → surface water → groundwater" was constructed. During the dry season, the contribution rate of the intermediate effect of this pathway was calculated to be 70%, meaning that industrial emissions mainly enter the groundwater through the vertical infiltration pathway of the vadose zone. During the wet season, due to the dominance of surface runoff, the contribution rate of this pathway decreased to 32%, while the contribution rate of the lateral diffusion pathway increased to 58%.

[0049] S403, Spatial distribution of counterfactual source contributions: By constructing a counterfactual scenario of zero industrial emissions, the impact of industrial sources on groundwater SO4² during the dry season is quantified. - The spatial contribution rate of the pollution plume reached 89%, and the pollution hotspots were distributed in strips along the river, which highly coincided with the location of the enterprise's sewage outlet.

[0050] S404, Residual Space Clustering: Spatial cluster analysis of the model residuals identified pollution hotspots not explained by the model. These areas overlapped with areas of abnormal groundwater level rise by 91%, suggesting that local hydrogeological conditions formed hidden migration channels.

[0051] S5, Differentiated Early Warning Command Generation and Output: Based on the automatic hydrological period division results and model output, differentiated early warning instructions are generated according to source and period: Dry season warning (January 2023, average daily flow rate 8.5 m³ / s): The current dry season is determined, and the flow rate is below the critical threshold of 10 m³ / s, triggering an industrial point source leakage warning. The warning information dissemination terminal sends instructions to the park's environmental protection department: conduct key inspections of the titanium dioxide workshop's wastewater pool and the phosphogypsum storage area's seepage prevention facilities; initiate vadose zone high-frequency monitoring (increasing the monitoring frequency from monthly to weekly); and add temporary monitoring at a cross-section 200m downstream of the enterprise's discharge outlet.

[0052] High-water season warning (July 2023, average daily flow 185 m³ / s): The system is determined to be in the high-water season with a flow exceeding 110 m³ / s, triggering an agricultural non-point source runoff warning. The warning information dissemination terminal sends instructions to the watershed management agency: vegetation interception is activated in the 20m buffer zone between farmland and the river channel; permeable reactive barriers on both banks of the river are put on standby; and the monitoring frequency of surface water sections is increased from weekly to daily.

[0053] Simultaneously outputs pollution prediction results (SO4²⁺ for the next 30 days). - / TP concentration prediction), contribution ratio of each pollution source (industrial source / agricultural source ratio), pollutant migration path flux ratio (vertical infiltration / lateral diffusion ratio), and hydrological threshold effect diagnosis report (10m³ / s critical threshold trigger record).

[0054] The verification results show that the SCM-IML hybrid model (i.e., the hybrid model of structural causal model and interpretable machine learning) of the present invention has an overall explanatory power of R²=0.82 for the multi-media pollution pattern, which is significantly improved compared with the traditional multiple linear regression model (R²=0.41). It can accurately capture the switching law of pollution mechanism during the wet and dry seasons, and meet the engineering needs of real-time monitoring and dynamic early warning of multi-media pollution in the floodplain chemical zone.

[0055] Example 2: This embodiment uses a chemical industrial park in a floodplain of the upper Yangtze River, contaminated with a complex mixture of organic and heavy metal pollution, as the application object to verify the universality and adaptability of the system under different pollution types and geological conditions. The park is located on a floodplain of a tributary of the Jialing River, at an altitude of 120-150 m, with the strata mainly composed of clay loam soil and a permeability coefficient of 1-3 m / d. The interaction between surface water and groundwater is weaker than in Implementation Case 1. The park covers an area of ​​approximately 8 km², including 8 fine chemical enterprises and 3.2 km² of surrounding farmland. The dominant pollutants are COD, ammonia nitrogen, and heavy metals such as Cd and Pb. The average annual flow rate in the area is 80 m³ / s, exceeding 150 m³ / s during the wet season and often below 15 m³ / s during the dry season. The system was deployed from 2022 to 2024, forming a two-dimensional comparison with Implementation Case 1 in terms of pollution type and geological conditions.

[0056] S1, Data Acquisition and Preprocessing: Construct a multi-dimensional, multi-media pollution monitoring network and adjust monitoring elements according to the pollution characteristics of this industrial park: Thirty-eight groundwater monitoring wells were deployed, focusing on monitoring Cd, Pb, COD, and ammonia nitrogen. Six surface water monitoring sections were set up, two of which are located downstream of enterprises that emit heavy metals; 80 soil sampling points were set up in a 500m×500m grid. Enterprise production data, including emission logs from 8 enterprises, with a daily time resolution; Hydrological and meteorological data are integrated into the daily flow records of hydrological stations in the tributary basin.

[0057] Based on the hydrological characteristics of this park, the threshold for hydrological period division has been adjusted: the threshold for the wet season has been adjusted to a flow rate of 95 m³ / s at P75, and the critical flow rate for the dry season has been adjusted to 8 m³ / s (because the permeability coefficient of clay loam is low, the retention effect is significantly enhanced at lower flow rates).

[0058] Spatial grid standardization still uses a 500m×500m grid, and the derived feature construction adds a heavy metal-organic interaction term (heavy_metal×organic) to capture the complexation effect of organic matter on heavy metal migration in clay loam soil.

[0059] S2, Construction of a dynamic cause-effect graph: In the dry season samples, a strong causal chain of "heavy metal emission → clay loam adsorption → heavy metals in groundwater" was identified (p<<0.001). Because the adsorption capacity of clay loam is higher than that of sandy loam, the path weight is reduced by about 15% compared with the sandy loam scenario in Case 1. In the wet season samples, a strong causal chain of "agricultural input → surface runoff → river sediment resuspension" was identified (p<<0.001). Because the upstream river flow velocity is low, sediment resuspension becomes an important indirect pathway for agricultural non-point source pollutants to enter groundwater.

[0060] S3, SCM-IML hybrid modeling: Due to the low permeability coefficient of clay loam soil and the slow vertical migration rate of pollutants, the GBRT parameters were adjusted: the tree depth was increased to 10 to capture more complex nonlinear hysteresis effects, and the number of iterations was increased to 800. SHAP value decomposition showed that industrial emissions during the dry season contributed 55.3% to groundwater Cd (compared to SO4² in Case Study 1). - The contribution of agricultural non-point sources to ammonia nitrogen decreased to 68.5%, reflecting the adsorption and buffering effect of clay loam soil. During the wet season, the contribution of agricultural non-point sources to ammonia nitrogen was 48.6%.

[0061] PDP curve diagnosis shows that the critical flow threshold shifts to 8 m³ / s during the dry season. Below this value, the clay loam soil becomes saturated with adsorption, and the Cd concentration shows a nonlinear jump. During the wet season, for every 30 m³ / s increase in flow, the ammonia nitrogen concentration increases by about 18%, and the response slope is gentler than that of the sandy loam soil scenario in Implementation Case 1.

[0062] S4, Quantification of Causal Effects: The stratified average treatment effect analysis results showed that: during the dry season, for every 1 m³ / s increase in flow rate, the Cd concentration in groundwater decreased by 0.25 mg / L; during the wet season, for every 1 m³ / s increase in flow rate, the ammonia nitrogen concentration increased by 0.09 mg / L.

[0063] Mediation effect analysis showed that the mediating effect contribution rate of the "industrial emissions → vadose zone → groundwater" pathway was 55% during the dry season (down from 70% in Case 1), while the contribution rate of the "agricultural input → surface water → groundwater" pathway decreased to 28% during the wet season, reflecting that clay loam soil weakened the rapid interaction between surface water and groundwater.

[0064] S5, Differentiated Early Warning Command Generation and Output: Adjusting early warning thresholds and strategies based on the geological characteristics of this park: Dry season warning (February 2024, average daily flow rate 6.8 m³ / s): Triggered an industrial heavy metal leakage warning, instructing key monitored enterprises with Cd-containing wastewater treatment facilities to set up temporary monitoring wells in areas with clay loam soil thickness less than 3 m (high-risk areas for vadose zone retention), and to increase the monitoring frequency to weekly.

[0065] High-water season warning (June 2024, average daily flow 132 m³ / s): Triggers a joint warning for agricultural non-point source pollution and river sediment resuspension, instructing the addition of organic matter adsorbent materials in farmland buffer zones to intercept heavy metals, while monitoring the resuspension flux of riverbed sediments.

[0066] Dual-case comparative verification demonstrates that the multi-media pollution monitoring and early warning system for floodplain chemical industrial zones of this invention can adaptively adjust hydrological thresholds, model parameters, and early warning strategies based on different geological conditions (sandy loam vs. clay loam) and pollution types (inorganic salts vs. heavy metal-organic composites). The dynamic structural causal graph construction module automatically identifies differentiated dominant causal paths, accurately captures nonlinear migration patterns in media with different permeability coefficients through SCM-IML hybrid modeling, and the early warning generation module outputs differentiated control instructions matched to geological conditions. In Implementation Case 2, the overall model explanatory power reached R²=0.79, which, although slightly lower than the sandy loam scenario (R²=0.82), is still significantly higher than the traditional static regression model (R²<<0.45), verifying the system's universality and engineering adaptability under complex geological and composite pollution conditions.

[0067] The specific embodiments described are preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. All equivalent changes made to the structure, shape, and principle of the present invention should be covered within the scope of protection of the present invention.

Claims

1. A multi-media pollution monitoring and early warning method for floodplain chemical industrial zones, characterized in that, Including the following steps: S1. Data Acquisition and Preprocessing: Acquire environmental monitoring data, enterprise production data, agricultural activity data, and hydrological and meteorological data from the floodplain chemical industrial zone to form a multi-source spatiotemporal dataset. Environmental monitoring data includes groundwater pollutant concentration data, surface water pollutant concentration data, and soil pollutant concentration data. Enterprise production data includes sulfuric acid production data, phosphate fertilizer production data, and characteristic pollutant emission log data. Agricultural activity data includes fertilizer application data and spatial distribution data of planting structure. Hydrological and meteorological data includes daily flow records from watershed hydrological stations and cumulative precipitation data. Preprocess the multi-source spatiotemporal dataset to obtain a standardized spatiotemporal grid dataset. S2, Construction of dynamic structural causal graph: Construct an initial directed acyclic graph containing industrial emission nodes, agricultural input nodes, flow nodes, surface water concentration nodes, and groundwater concentration nodes; identify significant causal edges; identify strong causal chains during the dry season and strong causal chains during the wet season, and generate a dynamic structural causal graph; S3, SCM-IML Hybrid Modeling: Based on dynamic structural causal graphs, a nonlinear structural equation for groundwater pollutant concentration is constructed; a hydrological period-based modeling strategy is adopted to construct a hybrid model of structural causal model and interpretable machine learning, training sub-models for wet and dry periods respectively, fitting the nonlinear relationship between pollutant concentration and various data in the multi-source spatiotemporal dataset; the interpretability of the hybrid model is achieved through SHAP value decomposition and partial dependency graphs to diagnose hydrological threshold effects; pollution prediction results and source contribution decomposition values ​​are output. S4, Causal Effect Quantification: Based on a mixed model, the stratified average treatment effect analysis method and the mediation effect analysis method are used to quantify the change in pollutant concentration, the proportion of pollutant migration path flux, and the contribution ratio of each pollution source when the flow rate increases by one unit. S5, Differentiated Early Warning Command Generation and Output: Based on the results of automatic hydrological period division, pollution prediction results, and the contribution ratio of each pollution source, differentiated early warning commands are generated by source and period, and the pollution prediction results, the contribution ratio of each pollution source, the proportion of pollutant migration path flux, and the hydrological threshold effect diagnosis results are output.

2. The multi-media pollution monitoring and early warning method for floodplain chemical industrial zones as described in claim 1, characterized in that, Step S1 involves preprocessing the multi-source spatiotemporal dataset, including spatiotemporal alignment, spatial grid standardization, automatic hydrological period division based on flow thresholds, and construction of features derived from lag variables and interaction terms.

3. The multi-media pollution monitoring and early warning method for floodplain chemical industrial zones as described in claim 1, characterized in that, Step S2 specifically includes the following steps: S201, Initial Directed Acyclic Graph Construction: Based on prior knowledge of the pollution process in the floodplain chemical zone, an initial directed acyclic graph is constructed, including industrial emission nodes, agricultural input nodes, flow nodes, surface water concentration nodes, and groundwater concentration nodes. S202, PC Algorithm Data-Driven Optimization: The PC algorithm is used to perform data-driven optimization on the initial directed acyclic graph, and significant causal edges are identified through Gaussian conditional independence test; S203, Path selection for differences between wet and dry seasons: The PC algorithm is used to automatically select paths with significant differences between wet and dry seasons, identify strong causal chains in the dry season and strong causal chains in the wet season, and generate a dynamic causal graph.

4. The multi-media pollution monitoring and early warning method for floodplain chemical industrial zones as described in claim 1, characterized in that, Step S3 specifically includes the following steps: S301, Construction of Nonlinear Structural Equations: Using dynamic structural causal graphs as the theoretical framework, nonlinear structural equations for groundwater pollutant concentrations are constructed. S302, Hydrological Period Sub-model Training: A hydrological period modeling strategy is adopted to construct a hybrid model of structural causal model and interpretable machine learning. Based on the sample data of the wet season and the sample data of the dry season, the wet season sub-model and the dry season sub-model are trained respectively. S303, SHAP value decomposition: By decomposing the causal effects of each pollution source's contribution through SHAP value, the differentiated contributions of industrial and agricultural sources in different hydrological periods are quantified. S304, Partial Dependency Plot Diagnosis: Visualizing the marginal effect of flow on pollutant concentration using partial dependency plots to diagnose hydrological threshold effects; S305, Interaction Term Calculation: Embedded interaction term calculation module, outputs pollution prediction results and source contribution decomposition values.

5. The multi-media pollution monitoring and early warning method for floodplain chemical industrial zones as described in claim 1, characterized in that, Step S4 specifically includes the following steps: S401, Stratified average treatment effect analysis: Based on the mixed model, the stratified average treatment effect analysis method is used to compare the causal effect of flow changes during the wet and dry seasons on pollutant concentrations. By calculating the difference between the predicted values ​​of the high flow scenario and the baseline scenario, the change in pollutant concentration for each unit increase in flow is quantified. S402, Mediation effect analysis: Based on the mixed model, the mediation effect analysis method is used to quantify the flux proportion of pollutant migration paths, construct a linear model and regression equation for mediating variables, and calculate the contribution rate of the mediating effect of pollutant migration paths. S403, Spatial Distribution of Counterfactual Source Contributions: The spatial distribution of pollution source contributions is quantified through counterfactual analysis, and the contribution ratio of each pollution source is quantified through a counterfactual scenario of zero industrial emissions.

6. The multi-media pollution monitoring and early warning method for floodplain chemical industrial zones as described in claim 5, characterized in that, Step S403 is followed by step S404, residual spatial clustering: combined with residual spatial clustering, pollution hotspots not explained by the hybrid model are identified, and spatial clustering analysis is performed on the residuals of the hybrid model to identify the overlap between the hotspots not explained by the hybrid model and the groundwater level anomaly areas.

7. The multi-media pollution monitoring and early warning method for floodplain chemical industrial zones as described in claim 1, characterized in that, The differentiated early warning instructions in step S5 specifically include: When it is determined to be a dry season and the flow rate is below the critical threshold, an industrial point source leakage early warning instruction is generated. When it is determined to be a high-water season and the flow rate is higher than the high-water threshold, an agricultural non-point source runoff early warning instruction is generated.

8. A multi-media pollution monitoring and early warning system for floodplain chemical industrial zones, characterized in that: The method for monitoring and early warning of multi-media pollution in floodplain chemical industrial zones as described in any one of claims 1-7 includes a data acquisition and preprocessing module, a dynamic structural causal graph construction module, an SCM-IML hybrid modeling module, a causal effect quantification module, and a differentiated early warning instruction generation and release module. The data acquisition and preprocessing module is used to acquire environmental monitoring data, enterprise production data, agricultural activity data, and hydrological and meteorological data from the floodplain chemical industrial zone to form a multi-source spatiotemporal dataset. The environmental monitoring data includes groundwater pollutant concentration data, surface water pollutant concentration data, and soil pollutant concentration data. The enterprise production data includes sulfuric acid production data, phosphate fertilizer production data, and characteristic pollutant emission logs. The agricultural activity data includes fertilizer application data and spatial distribution data of planting structures. The hydrological and meteorological data includes daily flow records from watershed hydrological stations and cumulative precipitation data. The multi-source spatiotemporal dataset is preprocessed to obtain a standardized spatiotemporal grid dataset. The dynamic structural causal graph construction module is used to construct an initial directed acyclic graph containing industrial emission nodes, agricultural input nodes, flow nodes, surface water concentration nodes, and groundwater concentration nodes; identify significant causal edges; identify strong causal chains during the dry season and strong causal chains during the wet season; and generate a dynamic structural causal graph. The SCM-IML hybrid modeling module is used to construct a nonlinear structural equation for groundwater pollutant concentration based on a dynamic structural causal graph. It employs a hydrological period-based modeling strategy to construct a hybrid model combining structural causal modeling and interpretable machine learning, training sub-models for both high-water and low-water periods to fit the nonlinear relationship between pollutant concentration and various data points in a multi-source spatiotemporal dataset. The interpretability of the hybrid model is achieved through SHAP value decomposition and partial dependency graphs, diagnosing hydrological threshold effects. Finally, it outputs pollution prediction results and source contribution decomposition values. The causal effect quantification module is used to quantify the change in pollutant concentration, the proportion of pollutant migration path flux, and the contribution ratio of each pollution source when the flow rate increases by one unit based on a hybrid model, using the stratified average treatment effect analysis method and the mediation effect analysis method. The differentiated early warning instruction generation and release module is used to generate differentiated early warning instructions by source and period based on the results of automatic hydrological period division, pollution prediction results and the contribution ratio of each pollution source, and output pollution prediction results, contribution ratio of each pollution source, pollutant migration path flux ratio and hydrological threshold effect diagnosis results.

9. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the multi-media pollution monitoring and early warning method for the river floodplain chemical industrial zone as described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the multi-media pollution monitoring and early warning method for the river floodplain chemical industrial zone as described in any one of claims 1-7.

Images