A method and system for predicting the transformation risk of arsenic species in the Yellow River estuary based on a water-sediment dynamic constraint causal diagram attention network

Abstract

The present application relates to the technical field of causal diagram attention prediction, in particular to a method and system for predicting arsenic speciation transformation risk in the Yellow River estuary based on a water-sediment dynamic constraint causal diagram attention network. The method comprises multi-medium arsenic speciation risk prediction state coding based on preprocessed data, including monitoring sample input and medium adaptation representation, arsenic speciation composition and relative abundance coding, toxicity-biological effectiveness combined risk state coding, environmental response gate fusion and state result generation; arsenic speciation state input and water-sediment disturbance feature extraction, water and sediment disturbance intensity calculation in the water and sediment regulation stage, medium arsenic migration driving force constraint construction, and water-sediment dynamic disturbance prediction constraint set generation; the complete process from multi-medium arsenic monitoring data acquisition, arsenic speciation state coding, water-sediment dynamic disturbance prediction constraint construction, causal diagram attention arsenic speciation transformation prediction to arsenic pollution risk classification prediction and control output is realized.

Description

Technical Field

[0001] This invention relates to the field of causal graph attention prediction technology, and in particular to a method and system for predicting the risk of arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network. Background Technology

[0002] The Yellow River Estuary is a typical estuarine region characterized by strong land-sea interaction, active sediment transport, and important ecological functions. During water and sediment regulation, runoff, sediment concentration, salinity gradient, hydrodynamic conditions, and sediment resuspension processes undergo significant changes, which can easily alter the migration and distribution of arsenic among seawater, suspended particulate matter, surface sediments, and columnar sediments. As a pollutant with ecotoxicity and potential bioaccumulation risk, the pollution risk of arsenic depends not only on its total amount but also on its chemical forms, bioavailability, migration and transformation pathways in different environmental media, and future trends under hydrodynamic disturbances.

[0003] Existing studies on arsenic pollution in the Yellow River Estuary and nearshore waters mostly focus on monitoring total arsenic concentrations in water or sediments, or analyzing changes in arsenic content in seawater, suspended particulate matter, and sediments separately. While these methods can reflect pollution levels at a given time or in a localized area, they are insufficient to reveal the speciation of arsenic across multiple media under water and sediment regulation disturbances, and they also struggle to predict subsequent pollution risk changes based on current hydrodynamic conditions and arsenic speciation.

[0004] In arsenic speciation analysis, existing methods typically combine chemical extraction, speciation detection, and statistical correlation analysis to identify As(III), As(V), organic arsenic, and arsenic in different bound states in sediments. While these methods can obtain some speciation distribution information, they do not adequately consider the coupling effects of environmental factors such as hydrodynamics, suspended particulate matter concentration, salinity, pH, dissolved oxygen, redox potential, and sediment disturbance intensity. This makes it difficult to determine the influence of different factors on the direction and intensity of arsenic speciation transformation, and even more difficult to further translate this influence into risk prediction results tailored to the site, media, and water and sediment regulation stages.

[0005] In recent years, graph neural networks, attention mechanisms, and causal modeling methods have provided new technical pathways for the analysis of multi-media pollutant migration and transformation. By constructing different environmental media, arsenic speciation, and water and sediment environmental factors as graph nodes, and using graph attention mechanisms to learn the interaction strength between nodes, complex relationships between multiple media, speciations, and factors can be expressed. Furthermore, combining causal constraints can reduce misjudgments caused by relying solely on correlation for risk assessment. However, in existing technologies, relevant intelligent models are mostly used for monitoring and prediction of general environments, and have not fully incorporated the unique water and sediment dynamic disturbance processes during the Yellow River Estuary water and sediment regulation period, making it difficult to make targeted predictions of arsenic speciation transformation pathways, transformation directions, risk contributions, and subsequent pollution risk levels.

[0006] Therefore, there is an urgent need for an arsenic speciation risk prediction system for the Yellow River Estuary water and sediment regulation scenario. This system should be able to integrate arsenic speciation data from seawater, suspended particulate matter, surface sediments, and core sediments, combine water and sediment dynamics with environmental factors to construct constraints, and infer the transformation pathways, transformation probabilities, and risk contributions of arsenic among multiple media using a causal graph attention network. Furthermore, it should predict the arsenic pollution risk level, key risk media, and critical risk transformation pathways corresponding to different sampling stages, different stations, and different media, thereby providing technical support for the risk management of As pollution in the Yellow River Estuary and the ecological protection and restoration of estuarine wetlands. Summary of the Invention

[0007] To address the aforementioned problems, this invention provides a method and system for predicting the risk of arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network. Through multi-media arsenic monitoring data acquisition, arsenic speciation state encoding, hydrodynamic disturbance constraint construction, causal graph attention-based arsenic speciation transformation inference, and an arsenic pollution risk classification prediction output mechanism, this method achieves causal modeling and risk prediction of arsenic speciation migration and transformation relationships in the Yellow River Estuary under water and sediment regulation conditions, thereby improving the accuracy, interpretability, and applicability of As pollution risk prediction and control.

[0008] In a first aspect, the present invention provides a method for predicting the risk of arsenic speciation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network, comprising: Collect arsenic monitoring data in multiple media; Data preprocessing was performed based on the collected multi-media arsenic monitoring data; Based on the preprocessed data, multi-media arsenic speciation risk prediction state coding is performed, including monitoring sample input and media adaptation representation, arsenic speciation composition and relative abundance coding, toxicity-bioavailability joint risk state coding, environmental response gating fusion and state result generation. Based on the coding results, constraints for hydrodynamic disturbances are constructed, including arsenic speciation input and hydrodynamic disturbance feature extraction, disturbance intensity calculation during water and sediment regulation stages, construction of driving force constraints for arsenic migration between media, and generation of a set of hydrodynamic disturbance prediction constraints. A causal graph attention network under hydrodynamic constraints is constructed based on hydrodynamic perturbation constraints, including the construction of arsenic morphology causal nodes, generation of directed edges under hydrodynamic constraints, correction of causal direction and graph attention message passing, and generation of arsenic morphology transformation probability and graph representation. The arsenic pollution risk classification prediction results are output based on the arsenic speciation transformation diagram.

[0009] Secondly, a risk prediction system for arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network includes: The data acquisition module is configured as follows: Thirdly, the present invention provides a computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the aforementioned method for predicting the risk of arsenic speciation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network.

[0010] Fourthly, the present invention provides a terminal device, including a processor and a computer-readable storage medium, wherein the processor is used to implement various instructions; the computer-readable storage medium is used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor to provide a method for predicting the arsenic speciation risk in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network.

[0011] In summary, the present invention has the following beneficial technical effects: Compared to existing methods for predicting arsenic pollution risk in the Yellow River Estuary, which suffer from limitations such as reliance on single-media monitoring, insufficient synergistic characterization of arsenic speciation across multiple media, difficulty in quantifying the impact of hydrodynamic disturbances, and lack of causal explanation for arsenic speciation transformation pathways, this invention addresses the core need of "predicting arsenic speciation transformation and pollution risk in multiple media during water and sediment regulation periods." It proposes a Yellow River Estuary arsenic speciation transformation risk prediction system based on a hydrodynamically constrained causal graph attention network. This system achieves a complete workflow from multi-media arsenic monitoring data acquisition, arsenic speciation state encoding, construction of hydrodynamic disturbance prediction constraints, causal graph attention arsenic speciation transformation prediction, to arsenic pollution risk classification prediction and control output. This effectively improves the accuracy, interpretability, and applicability of As pollution risk prediction in the Yellow River Estuary.

[0012] First, this invention unifies and standardizes the collection of arsenic speciation data in seawater, suspended particulate matter, surface sediments, and columnar sediments at different stages before, during, and after water and sediment regulation. It organizes the sampling stage, monitoring station, sampling medium, arsenic speciation concentration, environmental factors, and sampling reliability into a unified monitoring sample, solving the problems of scattered multi-media data, inconsistent sampling stages, and difficulty in synchronously correlating environmental factors in traditional monitoring methods. This enables arsenic monitoring data from different media, different stations, and different stages of water and sediment regulation to be transformed into calculable, comparable, and traceable risk prediction inputs.

[0013] Secondly, this invention employs a multi-media arsenic speciation encoding mechanism to uniformly represent the states of As(III), As(V), organic arsenic, particulate-bound arsenic, weakly acid-extractable arsenic, reducible arsenic, oxidizable arsenic, and residual arsenic in seawater, suspended particulate matter, surface sediments, and columnar sediments. This is combined with relative abundance of speciation, toxicity weights, bioavailability weights, and environmental response gating to form an arsenic speciation state vector. This mechanism avoids the problem of risk prediction based solely on total arsenic concentration or the concentration of a single speciation, enabling the system to simultaneously characterize the arsenic speciation composition, ecological risk differences, and the moderating effects of environmental conditions, thereby improving the specificity of subsequent water and sediment disturbance prediction constraints.

[0014] Furthermore, this invention introduces a constraint construction mechanism for predicting hydrodynamic disturbances. Based on the arsenic speciation state vector, it extracts the response characteristics of flow disturbances, sediment disturbances, salinity disturbances, and redox disturbances. Combining this with the state differences between adjacent water and sediment regulation stages, it calculates the comprehensive disturbance intensity corresponding to different sampling stages, stations, and media. Simultaneously, this invention constructs constraints on the driving force of arsenic migration between media, the constraint weights for migration and transformation processes, and directional consistency constraints. This allows processes such as adsorption, desorption, sedimentation, resuspension, oxidation, and reduction to enter the graph attention prediction model in a directed constraint form, thereby reducing the risk of misjudgment of transformation direction caused by relying solely on statistical correlation analysis.

[0015] Finally, this invention establishes an intelligent prediction process for As pollution risk management in the Yellow River Estuary through a causal graph attention-based arsenic speciation prediction and risk grading prediction output mechanism. On the one hand, the system can construct the medium, arsenic speciation, and migration and transformation process as causal graph nodes and directed edges, and use hydrodynamic constraints and graph attention message passing to predict arsenic speciation transformation paths, transformation probabilities, and key risk channels. On the other hand, based on node risk response, path risk contribution, and comprehensive risk characterization at the station stage, the system can output predicted risk levels, key risk media, key risk arsenic speciations, main risk transformation paths, and management recommendations, making the arsenic pollution risk prediction results more interpretable and of practical management value.

[0016] In the application of arsenic pollution risk prediction during the Yellow River Estuary water and sediment regulation period, the system of this invention can improve the accuracy of multi-media arsenic speciation transformation pathway identification and enhance the ability to locate key risk media and critical risk transformation processes. Test results show that the accuracy of arsenic speciation transformation pathway identification by the system of this invention is more than 13% higher than that of traditional correlation analysis methods, the accuracy of risk level prediction is more than 11% higher than that of methods based on total arsenic concentration thresholds, and the consistency of key risk media identification is improved by about 9%. Therefore, this invention, while improving the accuracy of As pollution risk prediction and causal explanation ability in the Yellow River Estuary, can provide a reliable technical solution for arsenic pollution risk management and control, and for the ecological protection and restoration of estuarine wetlands during water and sediment regulation. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of a method for predicting the risk of arsenic speciation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network, according to Embodiment 1 of the present invention. Figure 2 This is a schematic diagram comparing the consistency of arsenic speciation transformation direction identification in various methods of Embodiment 1 of the present invention; Figure 3 This is a schematic diagram comparing the explanatory power of multi-media coupling contributions of various methods in Embodiment 1 of the present invention; Figure 4 This is a schematic diagram comparing the fitting errors of the water and sediment disturbance response of various methods in Embodiment 1 of the present invention; Figure 5 This is a schematic diagram comparing the recall rates of various methods for identifying bioavailable arsenic risks in Embodiment 1 of the present invention; Figure 6 This is a schematic diagram comparing the consistency of key risk location positioning in various methods of Embodiment 1 of the present invention; Figure 7 This is a schematic diagram comparing the accuracy of risk level determination of various methods in Embodiment 1 of the present invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings.

[0019] Example 1 Reference Figure 1 This embodiment of a method for predicting the risk of arsenic speciation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network includes: (1) Multi-media arsenic monitoring data acquisition module during water and sediment regulation period This module addresses the challenges of diverse sampling media, rapid changes in sampling stages, and the difficulty in synchronizing arsenic speciation data with water and sediment environmental factors during water and sediment regulation in the Yellow River Estuary. Its goal is to uniformly collect, standardize, and construct predictive inputs for arsenic speciation data and corresponding environmental parameters in seawater, suspended particulate matter, surface sediments, and core sediments. Through sampling stage division, station media combination, arsenic speciation detection matrix construction, synchronous recording of environmental factors, and sampling reliability calculation, this module generates a multi-media arsenic monitoring data set that can be directly used in subsequent arsenic speciation status coding and risk prediction models.

[0020] 1) Definition of sampling object To reflect the phased impact of water and sediment regulation on the migration and transformation of arsenic forms in the Yellow River Estuary, this module first establishes a set of sampling phases according to the time periods before, during, and after water and sediment regulation, and sets up multiple monitoring stations in each phase.

[0021] , in, This indicates the collection of samples during the water and sediment regulation sampling phase. Indicates the first Each sampling stage, Indicates the number of sampling stages.

[0022] , in, This indicates the collection of monitoring stations in the Yellow River Estuary area. Indicates the first One monitoring station, This indicates the number of monitoring stations.

[0023] To cover the distribution of arsenic in different environmental media, the sampling medium is represented as: , in, Represents the set of sampling media. Indicates the first The sampling media include seawater, suspended particulate matter, surface sediments, and columnar sediments. This indicates the quantity of sampling media. The above processing establishes the basic sampling targets for the water and sediment regulation stage, monitoring stations, and sampling media, providing a sampling index for subsequent arsenic speciation detection data recording and risk prediction input.

[0024] 2) Recording of arsenic speciation detection data in multiple media To differentiate the impact of total arsenic concentration and different arsenic speciation on subsequent risk prediction, this module performs graded detection of arsenic speciation in each sampling stage, each monitoring station, and each medium, and organizes the detection results into a unified speciation concentration vector.

[0025] Let the set of arsenic speciation categories be represented as: , in, Represents the set of arsenic speciation categories. Indicates the first Arsenic-like forms, including As(III), As(V), organic arsenic, particulate-bound arsenic, weakly acid-extractable arsenic, reducible arsenic, oxidizable arsenic, and residual arsenic. This indicates the number of arsenic species.

[0026] For the The sampling phase, the first The monitoring station and the first Using a sampling medium, construct an arsenic speciation concentration vector: , in, This represents the arsenic speciation concentration vector of the corresponding sampled object. Indicates the first Detection concentration of arsenic-like forms.

[0027] To reduce the impact of differences in testing batches and media units on subsequent risk prediction modeling, the arsenic speciation concentration vector was standardized: , in, Represents the standardized first Arsenic-like speciation concentration, Indicates the first In the first type of medium Mean concentration of arsenic-like speciation, Indicates the corresponding standard deviation. This represents the smallest positive number that prevents the denominator from being zero.

[0028] Further, the standardized arsenic speciation concentration vector was obtained: cc , in, This represents the standardized arsenic speciation concentration vector. Represents the standardized first Arsenic-like speciation concentrations. Following the above processing, standardized arsenic speciation detection results were obtained for each sampled object, providing a foundation for subsequent synchronous organization and risk prediction modeling with water and sediment environmental factors.

[0029] 3) Synchronous recording of water and sediment environmental factors In order to establish a correspondence between arsenic speciation data and hydrodynamic conditions during water and sediment regulation, this module synchronously records environmental factors such as flow rate, sediment concentration, suspended particulate matter concentration, salinity, pH, dissolved oxygen, redox potential, water depth, and sediment disturbance intensity, and constructs an environmental factor vector.

[0030] Let the set of environmental factor categories be represented as: , in, Represents a set of environmental factor categories. Indicates the first Environmental factors, This indicates the number of environmental factors.

[0031] For the The sampling phase, the first The monitoring station and the first Using a sampling medium, construct an environmental factor vector: , in, This represents the environmental factor vector of the corresponding sampled object. Indicates the first Observed values ​​of environmental factors.

[0032] To form a unified environmental factor input that can be used in risk prediction models, the environmental factor vector is normalized and mapped: , in, This represents the normalized environmental factor vector. Represents the linear transformation matrix of environmental factors. This represents the environmental factor bias vector.

[0033] To characterize the impact of sampling phase information on the water and sediment disturbance process, a combined periodic and phase coding approach was adopted for the water and sediment regulation phase: , in, Indicates the first The stage encoding vector for each sampling stage. Indicates an indicator function, , and These represent the stages before, during, and after water and sediment regulation, respectively. Through the above processing, environmental factor representations and stage codes synchronized with arsenic speciation detection results are obtained, providing a data source for subsequent construction of water and sediment dynamic disturbance constraints and risk prediction.

[0034] 4) Sampling confidence calculation and dataset generation To control the impact of missing samples, detection errors, and differences in station representativeness on the quality of risk prediction input, this module calculates sampling reliability based on sample integrity, detection repeatability, station representativeness, and the completeness of environmental records, and organizes arsenic speciation data, environmental factors, and sampling attributes into a unified monitoring sample.

[0035] First, construct the sampling quality vector: , in, Represents the sampling quality vector. Indicates the sample integrity score. This indicates the repeatability score of the test. This indicates the representativeness score of the position. The score indicates the completeness of environmental records.

[0036] Based on the sampling quality vector, arsenic speciation concentration vector, and environmental factor vector, the sampling confidence weights used for predicting input weighting are calculated: , in, Indicates the first The sampling phase, the first The monitoring station, the The sampling confidence weight of samples corresponding to different sampling media. , and Represents the weight vector. This represents the confidence bias term. This represents the Sigmoid function.

[0037] Further construct a unified monitoring sample: , in, Indicates a unified monitoring sample. Indicates the sampling phase. Indicates the monitoring station location. Indicates the sampling medium. Represents the standardized arsenic speciation concentration vector. Represents the normalized environmental factor vector. Represents the stage encoding vector. This represents the sampling confidence weight.

[0038] Finally, all uniformly monitored samples are organized into a single output for this module: , in, This represents a collection of multi-media arsenic monitoring data during the water and sediment regulation period. This represents a uniform monitoring sample under the corresponding sampling stage, monitoring station, and sampling medium. The output of this module will serve as the input for subsequent modules.

[0039] (2) Multi-media arsenic speciation risk prediction status coding module This module builds upon the multi-media arsenic monitoring data set from the previous module, which was generated during the water and sediment regulation period. Addressing the issue of inconsistent arsenic speciation, concentration dimensions, and risk implications across different media, its goal is to transform arsenic speciation data from seawater, suspended particulate matter, surface sediments, and core sediments into a unified risk prediction state code. This module generates a set of multi-media arsenic speciation state codes usable for subsequent water and sediment dynamic disturbance constraint construction and risk prediction modeling through media adaptation, speciation composition coding, toxicity and bioavailability weighting, environmental response gating, and state vector fusion.

[0040] 1) Monitoring Sample Input and Media Adaptation Indication To enable the standardized monitoring samples generated in the previous module to enter the risk prediction state coding process, this module first receives the sample data and constructs a medium-adaptive representation based on the differences in the sampling medium. This process does not repeatedly collect the original samples; instead, it performs state coding based on the standardized arsenic speciation concentration, environmental factors, stage coding, and sampling reliability obtained in the previous module.

[0041] For any unified monitoring sample output by the previous module, it is represented as: , in, Indicates the first The sampling phase, the first The monitoring station and the first The unified monitoring sample corresponding to the sampling medium. Represents the standardized arsenic speciation concentration vector. Represents the normalized environmental factor vector. Represents the stage encoding vector. This represents the sampling confidence weight.

[0042] To differentiate the applicability of arsenic speciation in different media, a media-species adaptation matrix is ​​constructed: , in, Represents the medium-morphology adaptation matrix. Indicates the first Type 1 sampling media is suitable for the first Arsenic-like speciation coding Indicates that it is not applicable. Indicates the quantity of sampling media. This indicates the number of arsenic species.

[0043] Based on the medium-morphology adaptation matrix, the concentration of arsenic speciation in the samples was screened for adaptation: , in, This represents the arsenic speciation concentration vector after medium adaptation. Represents the medium-morphology adaptation matrix The OK, This indicates element-wise multiplication. After the above processing, a medium-adapted arsenic speciation concentration vector is obtained, which is used for subsequent speciation encoding and risk prediction feature construction.

[0044] 2) Arsenic speciation and relative abundance coding To avoid the problem of difficulty in comparing different media due to relying solely on absolute concentration, this module further calculates the relative abundance of arsenic speciation, enabling the system to characterize the composition and phased changes of arsenic speciation in different media, providing speciation distribution characteristics for subsequent risk prediction.

[0045] Based on the arsenic speciation concentration after media adaptation, the first... Relative abundance of arsenic-like species: , in, Indicates the first The sampling phase, the first The monitoring station, the The first type of sampling medium The relative abundance of arsenic-like species Indicates the first after media adaptation Arsenic-like speciation concentration, This represents the smallest positive number that prevents the denominator from being zero.

[0046] Further, the relative abundance vector of arsenic speciation was obtained: , in, This represents the vector of relative abundance of arsenic species in the corresponding sample. Indicates the first Relative abundance of arsenic-like species.

[0047] To enable subsequent risk prediction to utilize both absolute concentration levels and relative composition, an arsenic speciation encoding vector is constructed: , in, This represents the encoding vector for the arsenic speciation. The morphological composition mapping matrix represents the shape. Represents the morphological composition mapping bias vector. Presentation layer normalization operation, This indicates a vector concatenation operation. After the above processing, an arsenic speciation encoding containing both speciation concentration and relative abundance information is obtained, which is used for subsequent weighting of toxicity and bioavailability.

[0048] 3) Toxicity-Bioavailability Joint Risk Status Coding To reflect the varying contributions of different arsenic speciations to pollution risk prediction results, the module introduces speciation toxicity weights and bioavailability weights to enhance the risk semantics of arsenic speciation encoding. This process can distinguish between highly toxic, highly mobile, and highly bioavailable arsenic speciations, avoiding risk prediction based solely on total concentration.

[0049] Let the arsenic speciation toxicity weight vector and bioavailability weight vector be represented as follows: , in, This represents the arsenic speciation toxicity weight vector. Indicates the first Toxicity weight of arsenic-like forms This represents the weight vector of arsenic speciation bioavailability. Indicates the first Bioavailability weights of arsenic-like forms.

[0050] Based on the relative abundance of arsenic speciation, toxicity weight, and bioavailability weight, a speciation risk response vector for risk prediction is calculated: , in, Represents the morphological risk response vector. This represents the vector of relative abundance of arsenic species. This represents the arsenic speciation toxicity weight vector. This represents the weight vector for the bioavailability of arsenic speciation.

[0051] To integrate arsenic speciation information with speciation risk response information, a joint toxicity-bioavailability risk status code is constructed: , in, This represents the joint state encoding of toxicity and bioavailability. and Represents the state mapping matrix, and This represents the state mapping bias vector. This represents a nonlinear activation function. After the above processing, a joint state code containing differences in morphological composition, toxicity, and bioavailability is obtained, which is used for subsequent environmental response gating fusion.

[0052] 4) Environmental response gating fusion and state result generation To ensure that arsenic speciation coding reflects the impact of environmental changes on risk prediction during water and sediment regulation, this module further incorporates environmental factors, stage coding, and sampling reliability into the gating fusion process. This process characterizes the moderating effects of factors such as salinity, pH, dissolved oxygen, redox potential, sediment concentration, and sedimentary disturbance on the expression of arsenic speciation risk status.

[0053] Based on environmental factors, stage coding, and sampling confidence, an environmental response coding system is constructed to modulate the expression of risk state. , in, Indicates environmental response coding. Represents the environmental response mapping matrix. This represents the environmental response mapping bias vector. Represents the normalized environmental factor vector. Represents the stage encoding vector. This represents the sampling confidence weight.

[0054] Environmental gating weights are calculated based on the joint state coding of toxicity-bioavailability and environmental response coding: , in, Represents the environmental gating weight vector. Represents the environment gating mapping matrix. This represents the environment gating mapping bias vector. This represents the Sigmoid function.

[0055] Based on environmental gating weights, a multi-media arsenic speciation state vector is generated: , in, Indicates the first The sampling phase, the first The monitoring station and the first Arsenic speciation vector corresponding to the sampling medium and Represents the state fusion mapping matrix. and This represents the state fusion bias vector.

[0056] Finally, all arsenic speciation state vectors are organized into a single output for this module: , in, This represents the set of arsenic speciation state encoding results in multiple media. This represents the arsenic speciation vector corresponding to the sampling stage, monitoring station, and sampling medium. The output of this module serves as the input for subsequent modules.

[0057] (3) Construction module for water and sediment dynamic disturbance constraint This module builds upon the multi-media arsenic speciation encoding results from the previous module. Addressing the challenge of quantifying the impact of factors such as sudden flow changes, increased sediment concentration, variations in suspended particulate matter concentration, salinity gradient reconstruction, and sediment resuspension during water and sediment regulation on arsenic speciation migration and transformation, the module aims to construct hydrodynamic perturbation constraints to guide subsequent causal graph attention predictions. Through hydrodynamic perturbation feature extraction, stage-specific response calculation, inter-media migration driving force modeling, and constraint weight generation, this module transforms the effects of water and sediment regulation processes on arsenic adsorption, desorption, sedimentation, resuspension, redox reactions, and vertical migration into a computable set of predictive constraints.

[0058] 1) Arsenic speciation input and water and sediment disturbance feature extraction To enable the arsenic speciation state coding results output from the previous module to be incorporated into the hydrodynamic disturbance constraint construction process, this module first receives the state vector and extracts response features related to hydrodynamic disturbances. This process does not repeat the interpretation of arsenic speciation and toxicity coding; instead, it obtains potential hydrodynamic and sediment response representations from the state vectors that characterize the impact of water and sediment regulation disturbances, providing a disturbance feature basis for subsequent risk prediction.

[0059] The state vector of any arsenic species output by the previous module is represented as: , in, Indicates the first The sampling phase, the first The monitoring station and the first Arsenic speciation vector corresponding to the sampling medium This represents the set of arsenic speciation state encoding results in multiple media.

[0060] To extract water and sediment disturbance response features from the arsenic speciation state vector for risk prediction constraints, the following mapping is constructed: , in, This indicates the response characteristics of water and sediment disturbance. and This represents the characteristic mapping matrix of water and sediment disturbance. and This represents the corresponding bias vector. Represents a non-linear activation function. Presentation layer normalization operation.

[0061] To distinguish between different sources of influence such as flow disturbance, sediment disturbance, salinity disturbance, and redox disturbance, the response characteristics of water and sediment disturbance are decoupled into components: , in, This represents the flow disturbance component. This indicates the component of sediment disturbance. This represents the salinity perturbation component. This represents the redox perturbation component. This represents the perturbation component splitting matrix. This represents the bias vector for splitting the disturbance components. After the above processing, a representation of the water and sediment disturbance components for the water and sediment regulation process is obtained, which is used for subsequent disturbance intensity calculation and prediction constraint generation.

[0062] 2) Calculation of disturbance intensity during water and sediment regulation stage To characterize the impact of changes in the water and sediment environment on arsenic speciation during different stages before, during, and after water and sediment regulation, this module calculates the comprehensive disturbance intensity based on disturbance components and differences in sampling stages. This process is used to characterize the response differences of different stations and media during water and sediment regulation and to provide staged disturbance constraints for subsequent risk prediction.

[0063] For the same monitoring station and the same sampling medium, construct the state differences between adjacent sampling stages: , in, Indicates the first The change in arsenic speciation in each sampling phase relative to the previous sampling phase. and These represent the arsenic speciation state vectors of two adjacent sampling stages, respectively. Indicates the number of sampling stages.

[0064] Based on the perturbation components and state differences, construct the stage perturbation prediction constraint input vector: , in, This represents the stage perturbation input vector. This indicates a vector concatenation operation.

[0065] To calculate the overall disturbance intensity, a gated mapping is performed on the staged disturbance input vector: , in, This indicates the stage perturbation gating weight. This represents the stage perturbation gating mapping matrix. This represents the stage perturbation gating bias vector.

[0066] Based on the stage-based perturbation gating weights, the comprehensive perturbation intensity of water and sediment regulation used for prediction constraints is calculated: , in, Indicates the first The sampling phase, the first The monitoring station and the first The comprehensive disturbance intensity of water and sediment regulation corresponding to the sampling medium. This represents the perturbation intensity mapping matrix. This represents the perturbation intensity mapping bias vector. This represents the disturbance intensity score vector. This indicates the bias term for the disturbance intensity score. This indicates element-wise multiplication. After the above processing, the comprehensive disturbance intensity corresponding to different sampling stages, stations, and media is obtained, which is used for subsequent modeling of inter-media migration driving forces and construction of risk prediction constraints.

[0067] 3) Construction of driving force constraints for arsenic migration between media To describe the migration and transformation trends of arsenic among seawater, suspended particulate matter, surface sediments, and columnar sediments during water and sediment regulation, this module constructs the inter-media migration driving forces based on the differences in arsenic speciation in different media and the intensity of water and sediment disturbance. This process is used to quantify the constraints of adsorption, desorption, sedimentation, resuspension, and vertical migration processes on the direction of arsenic speciation transformation and subsequent risk prediction.

[0068] For any two types of media \(m_j\) and \(m_l\) under the same sampling stage and the same monitoring station, construct the state difference vector between the media: , in, Indicates the first Type of media and the first State difference vector between media types and These represent the arsenic speciation state vectors corresponding to the two types of media, respectively.

[0069] To incorporate the impact of water and sediment regulation disturbances on the migration relationships between media, a media disturbance coupling vector is constructed: , in, This represents the inter-medium disturbance coupling vector. and These represent the combined disturbance intensity corresponding to the two types of media.

[0070] Based on the inter-medium perturbation coupling vector, the arsenic migration driving force used to constrain the prediction of arsenic speciation is calculated: , in, Indicates the first Type of medium points to the first The driving force for arsenic migration in media-like substances Represents the migration driving force mapping matrix. This represents the migration driving force mapping bias vector. This represents the migration driving force score vector. This indicates the migration drive scoring bias term. This represents the hyperbolic tangent function.

[0071] To differentiate the constraints imposed on risk prediction by different migration and transformation processes, the constraint weights for each process type are calculated: , in, Indicates the first Constraint weights for migration and transformation processes, including adsorption, desorption, sedimentation, resuspension, oxidation, reduction, and vertical migration. Indicates the number of migration and transformation process types. , , and Indicates the first The mapping parameters corresponding to the process type are obtained. After the above processing, the driving force of arsenic migration between media and the process type constraint weights are obtained, which are used for the generation of constraint edge weights in the subsequent causal graph attention network.

[0072] 4) Generation of constraint sets for predicting water and sediment dynamic disturbances To transform the intensity of hydro-sediment dynamic disturbances and the driving forces of inter-medium migration into predictive constraint information usable by the subsequent causal graph attention network, this module further generates directional constraint weights between media and organizes them into a unified set of hydro-sediment dynamic disturbance prediction constraints. This set is used to constrain the direction, intensity, and process type of edges in the subsequent graph structure, providing constraints for the prediction of arsenic speciation transformation paths and risk levels.

[0073] Calculate the directed constraint weights between media based on migration driving force and process type constraint weights: , in, Indicates from the first Type of medium to the first Type of medium in the first Directed constraint weights in the class migration and transformation process , , and This represents the constraint fusion coefficient.

[0074] To avoid conflicts in subsequent causal graphs caused by migration processes in opposite directions, directional consistency constraints are calculated: , in, Indicates from the first Type of medium to the first Orientation consistency constraints for media-like materials and These represent the directed constraint weights in two opposite directions. This represents the smallest positive number that prevents the denominator from being zero.

[0075] Based on directional constraint weights and directional consistency constraints, a single water and sediment dynamic disturbance prediction constraint record is generated: , in, Indicates the first The sampling phase, the first The monitoring station is under the first Type of medium points to the first Constraint records for predicting hydrodynamic disturbances in similar media. and These represent the starting medium and the target medium, respectively.

[0076] Finally, all water and sediment dynamic disturbance prediction constraints are recorded and organized into a single output for this module: , in, This represents the set of constraints for predicting hydrodynamic disturbances. This represents the disturbance prediction constraint record between the corresponding sampling stage, monitoring station, and medium pair. The unique output of this module will serve as the input for subsequent modules.

[0077] (4) Causal graph attention arsenic morphology transformation prediction module This module inherits the set of hydrodynamic disturbance prediction constraints output from the previous module and calls the arsenic speciation state vectors corresponding to its indices. Addressing the problems of unclear transformation directions, difficulty in quantifying transformation intensity, and the inability to explain multi-process coupling effects among seawater, suspended particulate matter, surface sediments, and core sediments, the module aims to construct a causal graph attention network under hydrodynamic constraints. Through arsenic speciation causal node construction, hydrodynamic constraint directed edge generation, causal direction consistency correction, and graph attention message passing, this module predicts the transformation paths, transformation probabilities, and key driving relationships of arsenic speciation across multiple media, generating an arsenic speciation transformation graph representation.

[0078] 1) Construction of arsenic speciation causal nodes To transform the hydrodynamic disturbance prediction constraints obtained in the previous module into a graph prediction structure, this module first extracts the corresponding arsenic speciation state vector based on the sampling stage, monitoring station, and medium index in the constraint record, and encodes the medium category and arsenic speciation category together as graph nodes. These nodes represent the state of a specific arsenic speciation in a specific stage, specific station, and specific medium.

[0079] The constraint record for any hydrodynamic disturbance prediction output from the previous module is represented as follows: , in, Indicates the first The sampling phase, the first The monitoring station is under the first Type of medium points to the first Constraint records for predicting hydrodynamic disturbances in similar media. This represents the set of constraints for predicting hydrodynamic disturbances.

[0080] Based on the index of the constraint record, extract the corresponding arsenic state vector: , in, Indicates the first The sampling phase, the first The monitoring station and the first Arsenic speciation vector corresponding to the medium Indicates the first sampling stage and monitoring station. Arsenic speciation vector corresponding to the medium This represents the set of arsenic speciation state encoding results in multiple media.

[0081] To map media category and arsenic speciation category to a unified node space, media coding and arsenic speciation coding are constructed separately: , in, Indicates the first Media encoding vector of the medium type, Indicates the first One-hot encoded vectors for media-like materials Represents the medium encoding matrix, This represents the media encoding bias vector.

[0082] , in, Indicates the first The morphological encoding vector of arsenic-like morphologies. Indicates the first The unique thermal coding vector of arsenic-like speciation, Represents the arsenic speciation coding matrix. This represents the arsenic speciation encoding bias vector.

[0083] Based on the arsenic speciation state vector, medium encoding, and arsenic speciation encoding, initial features for causal graph nodes used for transformation prediction are generated: , in, Indicates the first The sampling phase, the first The monitoring station, the Class media and the first Initial features of causal graph nodes corresponding to arsenic-like morphologies Represents the node feature mapping matrix. This represents the node feature mapping bias vector. Presentation layer normalization operation, This indicates a vector concatenation operation.

[0084] Further construct the causal node set of arsenic speciation: , in, Represents the set of causal nodes for arsenic speciation. This represents the causal graph nodes corresponding to the sampling stage, monitoring station, medium, and arsenic speciation. After the above processing, a set of arsenic speciation causal nodes is obtained, which is used for subsequent construction of water and sediment-constrained directed edges.

[0085] 2) Generation of water- and sediment-constrained directed edges To express the potential transformation relationships of arsenic in different media and in different forms, this module constructs directed edges from the arsenic form nodes in the source medium to the arsenic form nodes in the target medium, based on the hydrodynamic disturbance prediction constraint records obtained in the previous module. This process simultaneously considers migration driving forces, directional consistency, process type constraints, and form transformation compatibility.

[0086] For the Class migration and transformation process, construction process type encoding: , in, Indicates the first The process encoding vector of the class transfer transformation process, Indicates the first One-hot encoded vector of the process class, Represents the process encoding matrix, This represents the process encoding bias vector.

[0087] In order to determine the first Arsenic-like forms in the first Class process downwards The degree of adaptation of arsenic-like form transformation, and the calculation of form transformation compatibility weights: , in, Indicates the first Arsenic-like forms after the first Class process transformed into the first Compatibility weights of arsenic-like forms and These represent the encoding vectors for the source arsenic form and the target arsenic form, respectively. This represents a morphological transformation compatible mapping matrix. This represents the bias vector for morphological transformation compatible mapping. Represents the compatibility rating vector. This indicates the compatibility rating bias.

[0088] Based on the source node, target node, morphological transformation compatibility weights, and hydrodynamic disturbance constraints, the candidate strength of directed edges is calculated: , in, Indicates that by node Pointing to node And corresponding to the first Candidate strength of directed edges in the class transition process This represents the directed constraint weights obtained from the previous module. This indicates a directional consistency constraint. , , and Represents the fusion coefficient. Indicates the first The edge association matrix corresponding to the class process, This represents the edge strength bias term.

[0089] Directed edges constrained by water and sediment used for transformation prediction are selected based on candidate edge strength: , in, This represents a set of directed edges constrained by water and sand. Indicates from the source node Point to target node The directed edge, This represents the directed edge filtering threshold. After the above processing, a set of directed edges containing process type and water-sediment constraint strength is obtained, which is used for subsequent causal direction correction and graph attention message passing.

[0090] 3) Causal Direction Correction and Graph Attention Message Passing To avoid generating spurious transformation relationships with opposite directions based solely on correlation or local similarity, this module performs causal direction consistency correction on the candidate strengths of directed edges and executes graph attention message passing on the corrected graph structure. This process enhances edge relationships that conform to the direction of hydrodynamic disturbances and the arsenic speciation transformation mechanism, providing causal constraints for transformation probability prediction.

[0091] For a pair of candidate edges in opposite directions, calculate the causal orientation advantage: , in, Indicates from node Pointing to node Compared to the directional advantage of the opposite side, This indicates the strength of the reverse candidate edge.

[0092] Calculate causal direction correction weights based on directional advantage: , in, Indicates the causal direction correction weight. Indicates the direction-corrected temperature coefficient. This represents the smallest positive number that prevents the denominator from being zero.

[0093] Based on the directed edge candidate strength and causal direction correction weights, the causal constraint edge weights are obtained: , in, This represents the edge weight after the causal direction is corrected. Indicates the candidate strength of directed edges. This indicates the causal direction correction weight.

[0094] To aggregate neighboring node information for transformation probability prediction on the causal constraint graph, nodes are defined. The set of neighboring neighborhoods: , in, Represents a node The set of neighboring regions, This indicates a neighboring node that points to the target node. This indicates the type of the corresponding migration and transformation process.

[0095] Let the initial message passing state be: , in, Represents a node The initial message passing state, This indicates the initial characteristics of the node.

[0096] In the During round-robin message passing, calculate the graph attention weights of neighboring nodes to the target node: , in, Indicates the first Neighboring nodes in round-robin message passing Through the first Class process for target node Graph attention weights, Represents the query mapping matrix. Indicates the first The key mapping matrix corresponding to the class procedure, Represents the attention score vector. This represents a linear activation function with leakage correction.

[0097] Update node state based on graph attention weights: , in, Indicates the first The node status after round-robin message passing, Indicates the first The value mapping matrix corresponding to the class process. After the above processing, the node state after causal direction correction and graph attention message passing is obtained, which is used for subsequent transformation probability prediction.

[0098] 4) Arsenic speciation conversion probability and graphical representation generation To transform the causal graph attention prediction results into interpretable arsenic speciation relationships, this module calculates the transformation probability based on the final node state and causal constraint edge weights, and organizes it into an arsenic speciation transformation graph representation. This graph representation describes the directed transformation paths between different media, different arsenic speciations, and different migration and transformation processes, and serves as the graph structure input for subsequent risk prediction.

[0099] Let the number of message rounds be . The final node state is represented as: , in, Represents a node The final graph attention state, Indicates the number of message rounds.

[0100] Based on the final node state and causal constraint edge weights, predict the arsenic form transformation probability: , in, Indicates the first The sampling phase, the first At each monitoring station, arsenic was... The first in the class of media Class morphology after the first Class process transformation to the first The first in the class of media The probability of transformation of class form, Indicates the first The transformation probability mapping matrix corresponding to the class process.

[0101] Filter key conversion paths based on conversion probability: , in, This represents the set of key arsenic speciation transformation pathways. This represents an arsenic speciation transformation pathway that meets the screening criteria. This represents the conversion probability threshold.

[0102] To form a node-level transformation representation for risk prediction, the outgoing transformation probability and incoming transformation probability of the aggregation node are: , in, This represents a node-level transformation representation for risk prediction. This represents the transformation representation mapping matrix. This represents the transformation representation mapping bias vector.

[0103] Finally, the set of nodes, the set of directed edges, the set of key transformation paths, and the node-level transformation representation are organized into the unique output of this module: , in, This represents the arsenic speciation diagram obtained from the attention prediction of the causal graph. Represents the set of causal nodes for arsenic speciation. This represents a set of directed edges constrained by water and sand. This represents the set of key arsenic speciation transformation pathways. This represents the node-level transformation. The output of this module will serve as the input for subsequent modules.

[0104] (5) Arsenic Pollution Risk Classification Prediction and Control Output Module This module builds upon the arsenic speciation transformation diagram output from the previous module. Addressing the challenges of predicting arsenic pollution risk from a single concentration indicator, the significant differences in risk contribution among different arsenic speciations, and the difficulty in identifying key transformation pathways and major risk media, this module aims to classify and predict arsenic pollution risk at different stations and stages during the Yellow River Estuary water and sediment regulation period based on causal graph attention prediction results. It then generates key risk stations, major risk media, key transformation pathways, and control recommendations. The module's final output is derived from node risk response calculations, quantification of risk contribution from transformation pathways, comprehensive risk classification prediction at station stages, and the generation of control results.

[0105] 1) Transformation diagram input and node risk response calculation To enable the arsenic speciation transformation diagram obtained in the previous module to be incorporated into the risk prediction process, this module first receives the arsenic speciation transformation diagram representation and extracts the node-level transformation characteristics and key transformation paths. This process no longer reconstructs the causal graph; instead, it directly calculates the risk responses corresponding to different media and different arsenic speciations based on the graph prediction results.

[0106] The arsenic speciation transformation diagram output by the previous module is shown as follows: , in, The diagram shows the arsenic speciation transformation. Represents the set of causal nodes for arsenic speciation. This represents a set of directed edges constrained by water and sand. This represents the set of key arsenic speciation transformation pathways. Indicates the first The sampling phase, the first The monitoring station, the Class media and the first Node-level transformation characterization of arsenic-like morphologies.

[0107] To reflect the differences in toxicity, bioavailability, and migration-release capacity of different arsenic speciations, a speciation risk attribute vector is constructed: , in, Indicates the first Risk attribute vector of arsenic-like forms Indicates toxicity weight, Indicates biological effectiveness weight, This represents the migration release capability weight.

[0108] Based on the morphological risk attribute vector, generate the arsenic morphological risk attribute code: , in, Indicates the first Risk attribute coding of arsenic-like forms, and Represents the risk attribute mapping matrix. and This represents the risk attribute mapping bias vector. Represents a non-linear activation function. Presentation layer normalization operation.

[0109] Based on node-level transformation characterization and arsenic speciation risk attribute coding, the node risk response value for risk prediction is calculated: , in, This represents the arsenic pollution risk response value used for risk prediction at the corresponding node. Represents the node risk mapping matrix. This represents the node risk mapping bias vector. Represents the risk score vector. This indicates the risk scoring bias term. This represents the Sigmoid function. After the above processing, the node risk response value is obtained, which is used for subsequent quantification of the risk contribution of the conversion path.

[0110] 2) Quantification of Conversion Path Risk Contribution To identify arsenic speciation transformation pathways that significantly contribute to pollution risk prediction during water and sediment regulation, this module further combines the key transformation pathway set, transformation probability, and node risk response values ​​obtained in the previous module to calculate the risk contribution of each pathway. This process can distinguish between high-probability but low-risk pathways and low-probability but high-toxicity pathways.

[0111] For any path in the set of key conversion paths, it is represented as: , in, This represents a pathway for the transformation of arsenic forms. Indicates the source node, Indicates the target node. Indicates the type of migration and transformation process. This represents the corresponding conversion probability.

[0112] To quantify the combined effect of source node risk, target node risk, and conversion probability on risk prediction results, a path risk input vector is constructed: , in, Represents the path risk input vector. This represents the risk response value of the source node. Indicates the risk response value of the target node. This represents the conversion probability.

[0113] Based on the path risk input vector, the contribution of the transformation path to the risk prediction result is calculated: , in, This indicates the contribution value of the corresponding arsenic form transformation pathway to the risk prediction result. Represents the path risk mapping matrix. This represents the path risk mapping bias vector. Represents the path risk score vector. This represents the bias term in the path risk scoring.

[0114] To identify key risk transformation paths that significantly contribute to risk prediction results, a path set is constructed: , in, This represents the set of key risk transformation paths. This represents the threshold for path risk contribution. After the above processing, a set of key risk transformation paths is obtained, which is used for comprehensive risk classification prediction in subsequent positioning stages.

[0115] 3) Comprehensive risk classification and prediction at the stationing stage To transform node-level and path-level risk results into site-specific risk levels applicable to regional management, this module jointly aggregates multi-media and multi-form node and path risks at the same sampling stage and monitoring station. This process is used to predict the comprehensive arsenic pollution risk at each monitoring station during different water and sediment regulation stages.

[0116] For the The sampling phase and the first For each monitoring station, calculate the aggregated risk value for each node: , in, This represents the aggregated risk value of the nodes. This represents the aggregated weight of node risk. Indicates the node risk response value. This represents the smallest positive number that prevents the denominator from being zero.

[0117] The node risk aggregation weight is represented as follows: , in, This represents the aggregated weight of node risk. This represents the temperature coefficient for node risk aggregation.

[0118] Based on the set of key risk conversion paths, calculate the aggregate value of path risks: , in, This represents the aggregated value of path risk. Indicates the first The sampling phase, the first A subset of key risk transformation paths corresponding to each monitoring station. This indicates the number of paths in the subset.

[0119] To obtain a comprehensive risk prediction characterization for the station phase, node risk and path risk are jointly coded: , in, This indicates a comprehensive risk profile at the positioning stage. Represents the comprehensive risk mapping matrix. This represents the bias vector of the comprehensive risk mapping.

[0120] Let the set of risk levels be represented as: , in, This represents a set of arsenic pollution risk levels. Indicates the first Each risk level Indicates the number of risk levels.

[0121] Based on the comprehensive risk prediction characterization, the prediction probability of each risk level is calculated: , in, Indicates the first The sampling phase, the first The monitoring station belongs to the first The probability of predicting a risk level. This represents the risk level mapping matrix. This represents the risk level mapping bias vector. Indicates the first A scoring vector for each risk level. This indicates the corresponding rating bias item.

[0122] Risk level prediction during the site location determination phase: , in, Indicates the first The sampling phase, the first The predicted risk level of each monitoring station. This represents the probability of the corresponding risk level. After the above processing, the comprehensive predicted risk level for the station phase is obtained, which is used for subsequent identification and output generation of controlled objects.

[0123] 4) Identification of controlled objects and generation of results To transform the risk classification prediction results into outputs applicable to As pollution risk management and wetland ecological restoration in the Yellow River Estuary, this module further identifies key risk media, key risk arsenic forms, and major risk transformation pathways based on node risk response values, and generates structured management results. These results can indicate the predicted risk sources, risk propagation links, and priority control targets.

[0124] For the The sampling phase and the first One monitoring station is used to identify key risk media: , in, Indicates key risk media, Represents the set of sampling media. This represents the node's risk response value.

[0125] Identify key risk arsenic forms based on node risk response values: , in, This indicates key risk arsenic forms. This represents the set of arsenic speciation categories.

[0126] Identify the main risk conversion paths based on the path risk contribution value: , in, This represents the set of main risk transformation paths. Indicates taking the first The screening threshold corresponding to the risk contribution value of a larger path.

[0127] To generate control recommendations, rule mapping is performed on the predicted risk level, key risk media, key risk arsenic forms, and main risk transformation paths: , in, Indicates the first The sampling phase, the first Control recommendations for each monitoring station This represents the mapping function for control recommendations.

[0128] Further generate the predicted risk management results for individual station stages: , in, Indicates the first The sampling phase, the first The predicted risk management results of each monitoring station Indicates the sampling phase. Indicates the monitoring station location. This indicates the confidence level of the risk level.

[0129] Finally, all risk prediction results from each station phase are organized into a single output for this module: , in, This represents the set of output results for arsenic pollution risk classification prediction and control. This indicates the risk control results at the corresponding sampling stage and monitoring station. This module's output results... This is used to represent the final output results of the Yellow River Estuary water and sediment regulation period arsenic pollution risk classification prediction and control generated by this system.

[0130] Experimental verification To verify the effectiveness and reliability of the Yellow River Estuary arsenic speciation risk prediction system based on hydro-sediment dynamics-constrained causal graph attention network proposed in this invention in terms of multi-media arsenic speciation collaborative characterization, hydro-sediment dynamic disturbance constraint construction, arsenic speciation direction identification, key risk media location, and pollution risk classification prediction, a multi-media arsenic speciation monitoring and risk prediction dataset for the Yellow River Estuary water and sediment regulation period was constructed. This dataset consists of field monitoring samples from different stages before, during, and after water and sediment regulation in the Yellow River Estuary, including arsenic speciation detection data in seawater, suspended particulate matter, surface sediments, and columnar sediments, as well as environmental factors such as flow rate, sediment concentration, suspended particulate matter concentration, salinity, pH, dissolved oxygen, redox potential, water depth, and sediment disturbance intensity. It covers typical samples at different stations, in different media, with different hydro-sediment disturbance intensities, and at different arsenic pollution risk levels.

[0131] ① Seawater samples mainly include data on dissolved total arsenic, As(III), As(V), organic arsenic, and particulate arsenic, used to verify changes in arsenic speciation and the characteristics of dissolved-particulate distribution in the water. ② Suspended particulate matter samples mainly include data on particulate-bound arsenic, iron-manganese oxide-bound arsenic, and organically bound arsenic, used to verify the impact of particulate adsorption, desorption, and transport processes on arsenic migration during water and sediment regulation. ③ Surface and columnar sediment samples mainly include data on weakly acid-extractable, reducible, oxidizable, and residual arsenic, used to verify sediment resuspension, vertical migration, and potential release risks. All samples underwent laboratory speciation testing and expert review, resulting in labeled data including sampling stage, station coordinates, medium type, arsenic speciation, environmental factors, risk level, and key transformation pathways.

[0132] The multi-media monitoring data in the dataset were first standardized and uniformly coded, and then divided into three task levels according to the arsenic speciation risk prediction task: ① Multi-media arsenic speciation characterization task, including standardization of arsenic speciation concentration in different media, calculation of relative abundance of speciation, toxicity-bioavailability weighting, and environmental response coding; ② Water and sediment disturbance constraint modeling task, including calculation of differences in water and sediment regulation stages, flow disturbance, sediment disturbance, salinity disturbance, redox disturbance, and inter-media migration driving force; ③ Arsenic pollution risk prediction task, including identification of arsenic speciation transformation direction, screening of key risk pathways, location of key risk media, identification of key risk stations, and prediction of pollution risk level. A total of 1260 sets of multi-media arsenic speciation monitoring samples were constructed, of which 840 sets were used for model training and parameter adjustment, and 420 sets were used for validation and testing. The test samples covered different water and sediment regulation stages, different media combinations, different water and sediment disturbance intensities, and different risk levels.

[0133] To comprehensively evaluate the system performance, four comparative methods were selected, with the system of this invention serving as the method to be verified: ① Total-As Threshold is a risk discrimination method based on a total arsenic concentration threshold, which determines risk based on whether the total arsenic concentration in water or sediment exceeds the threshold, but does not distinguish between arsenic speciation and inter-media transformation relationships; ② Correlation Analysis is a traditional correlation analysis method, which judges the influence relationship through the correlation coefficient between arsenic speciation concentration and environmental factors, but it is difficult to identify the transformation direction and causal effect; ③ Random Forest is a random forest risk classification method, which can classify risk levels using multiple environmental factors and arsenic speciation characteristics, but lacks multi-media graph structure and transformation path expression; ④ GAT-NoHydro is a graph attention network method without introducing water and sediment dynamic constraints, which can perform attention fusion on multi-media nodes, but does not explicitly constrain the migration direction and process type under water and sediment regulation disturbances; ⑤ The system of this invention realizes arsenic pollution risk prediction during the Yellow River Estuary water and sediment regulation period through multi-media arsenic speciation state encoding, water and sediment dynamic disturbance constraint construction, causal graph attention arsenic speciation transformation prediction, and risk classification prediction and control output mechanism.

[0134] All methods were evaluated under the same monitoring samples, the same risk labeling, and a unified training and testing framework. To verify the system's stability under complex disturbance conditions related to water and sediment regulation, three levels of complexity were set up: ① Low-disturbance samples, where water and sediment dynamics change weakly and arsenic speciation changes are small, mainly used to verify conventional risk identification capabilities; ② Medium-disturbance samples, where flow rate, sediment concentration, salinity, and suspended particulate matter concentration change significantly, mainly used to verify multi-media coupling modeling and disturbance response identification capabilities; ③ High-disturbance samples, where there are strong water and sediment transport, sediment resuspension, fluctuations in redox conditions, and rapid arsenic speciation transformations, mainly used to verify the stability of water and sediment constrained causal graph transformation prediction and risk classification prediction under complex disturbances.

[0135] The following evaluation indicators were used: ① Arsenic speciation direction identification consistency (TDC) measures the consistency between the arsenic speciation direction identified by the system and the results of expert annotation or mechanism judgment; ② Multi-media coupling contribution explanation degree (MCE) measures the system's ability to explain the risk contribution relationship among seawater, suspended particulate matter, surface sediments, and columnar sediments; ③ Water and sediment disturbance response fitting error (HRE) measures the system's fitting deviation to the effects of flow rate, sediment concentration, salinity, redox conditions, and resuspension disturbance; ④ Bioavailable arsenic risk identification recall rate (BAR) measures the system's completeness in identifying high-mobility, high-bioavailable arsenic speciation risk samples; ⑤ Key risk station location consistency (RSL) measures the consistency between the key risk stations identified by the system and the results of manual verification; ⑥ Risk level prediction accuracy (R-ACC) measures the system's accuracy in predicting the arsenic pollution risk level for different stations and different water and sediment regulation stages.

[0136] Experimental results show that the system of the present invention exhibits good comprehensive performance in the task of predicting the risk of arsenic speciation transformation during the Yellow River Estuary water and sediment regulation period. The Total-As Threshold method can quickly identify samples with excessive total arsenic levels, achieving a risk level prediction accuracy of 72.4%. However, due to the lack of differentiation between As(III), As(V), weakly acid-extractable, and reducible forms, the recall rate for identifying bioavailable arsenic risk is only 0.681. The Correlation Analysis method can reveal the statistical relationship between some environmental factors and arsenic form concentrations, with a multi-media coupling contribution explanatory power of 0.706. However, it is difficult to determine the direction of form transformation, and the consistency in identifying the arsenic form transformation direction is 0.642. The Random Forest method can integrate multiple environmental factors and arsenic form characteristics for risk classification, improving the risk level prediction accuracy to 82.7%. However, due to the lack of inter-media migration path modeling, the consistency in locating key risk stations is only 0.748. The GAT-NoHydro method can improve the expression of inter-node correlations using multi-media graph structures, achieving a multi-media coupling contribution explanatory power of 0.792 and a consistency in locating key risk stations of 0.806. However, due to the lack of hydro-sediment dynamic constraints, its fitting error for hydro-sediment disturbance response is still 0.174. In comparison, the consistency of arsenic speciation transformation direction identification in this invention reaches 0.835, the explanatory power of multi-media coupling contribution reaches 0.864, the fitting error of water and sediment disturbance response is reduced to 0.136, the recall rate of bioavailable arsenic risk identification reaches 0.847, the consistency of key risk station location positioning reaches 0.879, and the accuracy of risk level prediction reaches 89.6%, which is significantly better than comparative methods such as Total-As Threshold, Correlation Analysis, Random Forest, and GAT-NoHydro.

[0137] Under high-disturbance sample testing conditions, the system of this invention maintains good stability. Because this invention incorporates the differences in water and sediment regulation stages, flow disturbances, sediment disturbances, salinity disturbances, redox disturbances, and inter-medium migration driving forces as constraint information into the causal graph attention network, it can reduce misjudgments caused by relying solely on concentration thresholds or statistical correlations under conditions of strong water and sediment transport, sediment resuspension, and rapid arsenic speciation. Test results show that the risk level prediction accuracy of the system of this invention decreases by less than 5.8% in high-disturbance samples, while the decreases in Total-As Threshold, RandomForest, and GAT-NoHydro are approximately 14.6%, 10.3%, and 8.1%, respectively, indicating that the system of this invention has stronger robustness and risk interpretation capabilities under strong water and sediment regulation disturbance scenarios.

[0138] In summary, the experimental results validate the effectiveness of the system of this invention in encoding arsenic speciation in multiple media, constructing constraints for hydrodynamic disturbances during the Yellow River Estuary's water and sediment regulation period, identifying arsenic speciation transformation directions, locating key risk media, and predicting pollution risk classification. This system can not only identify key transformation directions of arsenic among seawater, suspended particulate matter, and surface sediments, but also improve the ability to identify bioavailable arsenic risk samples, key risk sites, and major risk pathways. It provides a stable, interpretable, and engineering-approved technical solution for arsenic pollution risk management in the Yellow River Estuary and for the ecological protection and restoration of estuarine wetlands.

[0139] Table 1. Comparison of data from different methods under six major indicators. In the performance evaluation, the following indicators were positive: arsenic speciation transformation direction identification consistency (TDC), multi-media coupling contribution explanatory power (MCE), bioavailable arsenic risk identification recall rate (BAR), key risk station location consistency (RSL), and risk level prediction accuracy (R-ACC). Specifically, TDC measures the consistency between the system's identified arsenic speciation transformation direction and expert annotations or mechanistic judgments; MCE measures the system's ability to explain the risk contribution relationships among seawater, suspended particulate matter, surface sediments, and columnar sediments; BAR measures the system's completeness in identifying high-mobility, highly bioavailable arsenic speciation risk samples; RSL measures the consistency between the system's identified key risk stations and human verification results; and R-ACC measures the system's accuracy in predicting arsenic pollution risk levels at different stations and during different water and sediment regulation stages. The water and sediment disturbance response fitting error (HRE) was a negative indicator, measuring the system's fitting deviation to the effects of flow rate, sediment concentration, salinity, redox conditions, and resuspension disturbance.

[0140] The experimental results are shown in Table 1. Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 , Figure 7As shown, in the task of predicting the risk of arsenic speciation transformation during the Yellow River Estuary water and sediment regulation period, the risk discrimination method based on the total arsenic concentration threshold can quickly identify samples with excessive total arsenic concentration, but it does not distinguish between different arsenic speciations such as As(III), As(V), weakly acid-extractable arsenic, and reducible arsenic, making it difficult to reflect the actual risk of bioavailable arsenic. Traditional correlation analysis methods can reveal the statistical relationship between some environmental factors and arsenic speciation concentrations, but it is difficult to determine the transformation direction of arsenic between water bodies, suspended particulate matter, and sediments. The random forest method can classify risk levels by integrating multiple environmental factors and arsenic speciation characteristics, but it lacks the expression of migration paths and process constraints between media. The graph attention network method, which does not introduce water and sediment dynamic constraints, can improve the expression of node associations by using multi-media graph structures, but it is insufficient in constraining the migration direction under the influence of flow, sediment, salinity, and redox disturbances during water and sediment regulation. In contrast, the system of this invention, by introducing multi-media arsenic speciation encoding, water and sediment dynamic disturbance prediction constraint construction, causal graph attention arsenic speciation transformation prediction, and risk classification prediction and control output mechanism, realizes a complete processing flow from arsenic speciation characterization, water and sediment disturbance quantification, transformation path prediction to pollution risk classification prediction. It has better transformation interpretation ability and risk prediction stability in the complex disturbance scenario of water and sediment regulation in the Yellow River Estuary.

[0141] The results show that the consistency of arsenic speciation transformation direction identification (TDC) of the system of this invention reaches 0.835, which is about 6.8% higher than that of GAT-NoHydro; the multi-media coupling contribution explanation (MCE) reaches 0.864, which is about 9.1% higher than that of GAT-NoHydro; the water and sediment disturbance response fitting error (HRE) is reduced to 0.136, which is about 28.8% lower than that of Correlation Analysis and about 21.8% lower than that of GAT-NoHydro; the bioavailable arsenic risk identification recall rate (BAR) reaches 0.847, which is about 7.8% higher than that of Random Forest; the consistency of key risk station location (RSL) reaches 0.879, which is about 9.1% higher than that of GAT-NoHydro; and the risk level prediction accuracy (R-ACC) reaches 89.6%, which is about 8.3% higher than that of Random Forest and about 4.9% higher than that of GAT-NoHydro. The results show that the system of the present invention has significant advantages in identifying the direction of arsenic speciation during the Yellow River Estuary water and sediment regulation period, analyzing the contribution of multi-media coupling, fitting the water and sediment disturbance response, identifying the risk of bioavailable arsenic, and predicting pollution risk classification. This verifies its engineering feasibility and practical application potential in the task of As pollution risk management and ecological protection and restoration of estuary wetlands in the Yellow River Estuary.

[0142] Example 2 This embodiment provides a risk prediction system for arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network.

[0143] A computer-readable storage medium storing a plurality of instructions adapted for loading and execution by a processor of a terminal device of the method for predicting the risk of arsenic speciation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network.

[0144] A terminal device includes a processor and a computer-readable storage medium, the processor being used to implement various instructions; the computer-readable storage medium being used to store multiple instructions, the instructions being adapted to be loaded and executed by the processor as described in the "A Method for Predicting Arsenic Speciation Risk in the Yellow River Estuary Based on a Hydro-Sediment Dynamic Constrained Causal Graph Attention Network".

[0145] The above are all preferred embodiments of the present invention and are 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 the risk of arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network, characterized in that, include: Collect arsenic monitoring data in multiple media; Data preprocessing was performed based on the collected multi-media arsenic monitoring data; Based on the preprocessed data, multi-media arsenic speciation risk prediction state coding is performed, including monitoring sample input and media adaptation representation, arsenic speciation composition and relative abundance coding, toxicity-bioavailability joint risk state coding, environmental response gating fusion and state result generation. Based on the coding results, constraints for hydrodynamic disturbances are constructed, including arsenic speciation input and hydrodynamic disturbance feature extraction, disturbance intensity calculation during water and sediment regulation stages, construction of driving force constraints for arsenic migration between media, and generation of a set of hydrodynamic disturbance prediction constraints. A causal graph attention network under hydrodynamic constraints is constructed based on hydrodynamic perturbation constraints, including the construction of arsenic morphology causal nodes, generation of directed edges under hydrodynamic constraints, correction of causal direction and graph attention message passing, and generation of arsenic morphology transformation probability and graph representation. The arsenic pollution risk classification prediction results are output based on the arsenic speciation transformation diagram.

2. The method for predicting the risk of arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network as described in claim 1, characterized in that, The data preprocessing based on the collected multi-media arsenic monitoring data includes hierarchical detection of arsenic speciation in each sampling stage, each monitoring station, and each medium, and organizing the detection results into a unified speciation concentration vector; for the first... The sampling phase, the first The monitoring station and the first A sample medium was used to construct an arsenic speciation concentration vector. To reduce the impact of differences in different test batches and media units on subsequent risk prediction modeling, the arsenic speciation concentration vector was standardized to obtain a standardized arsenic speciation concentration vector. To establish a correspondence between arsenic speciation data and hydrodynamic conditions during water and sediment regulation, environmental factors such as flow rate, sediment concentration, suspended particulate matter concentration, salinity, pH, dissolved oxygen, redox potential, water depth, and sediment disturbance intensity were recorded simultaneously, and an environmental factor vector was constructed. To control the impact of sampling omissions, detection errors, and differences in station representativeness on the quality of risk prediction input, sampling reliability was calculated based on sample integrity, detection repeatability, station representativeness, and the completeness of environmental records. Arsenic speciation data, environmental factors, and sampling attributes were organized into a unified monitoring sample.

3. The method for predicting the risk of arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network as described in claim 2, characterized in that, The monitoring sample input and media adaptation representation includes, firstly, representing any uniform monitoring sample as follows: in, Indicates the first The sampling phase, the first The monitoring station and the first The unified monitoring sample corresponding to the sampling medium. Represents the standardized arsenic speciation concentration vector. Represents the normalized environmental factor vector. Represents the stage encoding vector. This represents the sampling confidence weight; to distinguish the applicability of arsenic speciation in different media, a media-species adaptation matrix is ​​constructed: in, Represents the medium-morphology adaptation matrix. Indicates the first Type 1 sampling media is suitable for the first Arsenic-like speciation coding Indicates that it is not applicable. Indicates the quantity of sampling media. This indicates the number of arsenic speciation categories; and based on the medium-species adaptation matrix, the concentration of arsenic speciation in the sample is adapted and screened. in, This represents the arsenic speciation concentration vector after medium adaptation. Represents the medium-morphology adaptation matrix The OK, This represents element-wise multiplication, leading to the media-adapted arsenic speciation concentration vector; arsenic speciation composition and relative abundance encoding, including calculating the arsenic speciation concentration based on the media-adapted arsenic speciation concentration. Relative abundance of arsenic-like species: in, Indicates the first The sampling phase, the first The monitoring station, the The first type of sampling medium The relative abundance of arsenic-like species Indicates the first after media adaptation Arsenic-like speciation concentration This represents a very small positive number to prevent the denominator from being zero, thus yielding the vector of relative abundance of arsenic species: in, This represents the vector of relative abundance of arsenic species in the corresponding sample. Indicates the first The relative abundance of arsenic-like species; finally, to enable subsequent risk prediction to utilize both absolute concentration levels and relative compositional structure, an arsenic species composition encoding vector is constructed: in, This represents the encoding vector for the arsenic speciation. The morphological composition mapping matrix represents the shape. Represents the morphological composition mapping bias vector. Presentation layer normalization operation, This indicates a vector concatenation operation.

4. The method for predicting the risk of arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network as described in claim 3, characterized in that, The toxicity-bioavailability joint risk status coding includes introducing speciation toxicity weights and bioavailability weights to enhance the risk semantics of the arsenic speciation composition coding in order to reflect the differences in the contribution of different arsenic speciations to the pollution risk prediction results. Based on the relative abundance of arsenic speciation, toxicity weight, and bioavailability weight, a speciation risk response vector for risk prediction is calculated. To integrate arsenic speciation information with speciation risk response information, a joint toxicity-bioavailability risk status code is constructed: in, This represents the joint state encoding of toxicity and bioavailability. and Represents the state mapping matrix, and This represents the state mapping bias vector. The nonlinear activation function is used to obtain a joint state code that includes differences in morphological composition, toxicity, and bioavailability. Environmental response gating fusion and state outcome generation include constructing an environmental response code to modulate risk state expression based on environmental factors, stage coding, and sampling confidence. in, Indicates environmental response coding. Represents the environmental response mapping matrix. This represents the environmental response mapping bias vector. Represents the normalized environmental factor vector. Represents the stage encoding vector. The sampling confidence weight is represented; and the environmental gating weight is calculated based on the joint state coding of toxicity-bioavailability and the environmental response coding. in, Represents the environmental gating weight vector. Represents the environment gating mapping matrix. This represents the environment gating mapping bias vector. This represents the Sigmoid function; simultaneously, based on environmental gating weights, it generates multi-media arsenic speciation state vectors, and finally organizes all arsenic speciation state vectors into a unique output of this module.

5. The method for predicting the risk of arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network as described in claim 4, characterized in that, The arsenic speciation state input and water and sediment disturbance feature extraction include first receiving a state vector and then extracting response features related to water and sediment disturbance from it. Any arsenic speciation state vector is represented as follows: in, Indicates the first The sampling phase, the first The monitoring station and the first Arsenic speciation vector corresponding to the sampling medium This represents the set of multi-media arsenic speciation state encoding results; a mapping is constructed to extract water and sediment disturbance response features from the arsenic speciation state vectors for risk prediction constraints. in, This indicates the response characteristics of water and sediment disturbance. and This represents the characteristic mapping matrix of water and sediment disturbance. and This represents the corresponding bias vector. Represents a non-linear activation function. The representation layer is normalized; to distinguish the different sources of influence such as flow disturbance, sediment disturbance, salinity disturbance, and redox disturbance, the water and sediment disturbance response characteristics are decoupled into components: in, This represents the flow disturbance component. This indicates the component of sediment disturbance. This represents the salinity perturbation component. This represents the redox perturbation component. This represents the perturbation component splitting matrix. The perturbation component is decomposed into a bias vector, thus obtaining the water and sediment perturbation component representation for the water and sediment regulation process; the perturbation intensity calculation for the water and sediment regulation stage includes constructing the state difference between adjacent sampling stages for the same monitoring station and the same sampling medium: in, Indicates the first The change in arsenic speciation in each sampling phase relative to the previous sampling phase. and These represent the arsenic speciation state vectors of two adjacent sampling stages, respectively. This indicates the number of sampling stages; based on the perturbation components and state differences, a stage perturbation prediction constraint input vector is constructed: in, This represents the stage perturbation input vector. This represents a vector concatenation operation; to calculate the overall disturbance strength, a gated mapping is applied to the stage disturbance input vector: in, This indicates the stage perturbation gating weight. This represents the stage perturbation gating mapping matrix. This represents the stage perturbation gating bias vector; finally, based on the stage perturbation gating weights, the comprehensive perturbation intensity for water and sediment regulation used to predict constraints is calculated: in, Indicates the first The sampling phase, the first The monitoring station and the first The comprehensive disturbance intensity of water and sediment regulation corresponding to the sampling medium. This represents the perturbation intensity mapping matrix. This represents the perturbation intensity mapping bias vector. This represents the disturbance intensity score vector. This indicates the bias term for the disturbance intensity score. This indicates element-wise multiplication.

6. The method for predicting the risk of arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network as described in claim 5, characterized in that, The construction of the inter-medium arsenic migration driving force constraint includes, to describe the migration and transformation trend of arsenic among seawater, suspended particulate matter, surface sediments, and columnar sediments during water and sediment regulation, constructing the inter-medium migration driving force based on the differences in arsenic speciation and water and sediment disturbance intensity in different media. First, for any two types of media at the same sampling stage and the same monitoring station, an inter-medium state difference vector is constructed. To incorporate the influence of water and sediment regulation disturbance on the inter-medium migration relationship, an inter-medium disturbance coupling vector is constructed. in, This represents the inter-medium disturbance coupling vector. and The combined perturbation intensities for the two types of media are represented respectively; and based on the perturbation coupling vector between the media, the arsenic migration driving force used to constrain the prediction of arsenic speciation is calculated: in, Indicates the first Class medium points to the first The driving force for arsenic migration in media-like substances Represents the migration driving force mapping matrix. This represents the migration driving force mapping bias vector. This represents the migration driving force score vector. This indicates the migration drive scoring bias term. The hyperbolic tangent function is represented; finally, to distinguish the constraints of different migration and transformation processes on risk prediction, the process type constraint weights are calculated; the set of constraints for hydro-sediment dynamic disturbance prediction is generated, including the calculation of the directed constraint weights between media based on the migration driving force and process type constraint weights to transform the hydro-sediment dynamic disturbance intensity and inter-media migration driving force into predictive constraint information usable by the subsequent causal graph attention network. in, Indicates from the first Type of medium to the first Type of medium in the first Directed constraint weights in the class migration and transformation process , , and The constraint fusion coefficient is represented; to avoid conflicts in the subsequent causal graph caused by migration processes in opposite directions, the directional consistency constraint is calculated; based on the directed constraint weight and the directional consistency constraint, a single water and sediment dynamic disturbance prediction constraint record is generated; finally, all water and sediment dynamic disturbance prediction constraint records are organized into a unique output.

7. The method for predicting the risk of arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network as described in claim 6, characterized in that, The construction of the arsenic speciation causal node includes transforming the hydrodynamic disturbance prediction constraints into a graph prediction structure. First, based on the sampling stage, monitoring station, and medium index in the constraint record, the corresponding arsenic speciation state vector is extracted. Then, the medium category and arsenic speciation category are jointly encoded as graph nodes. Any hydrodynamic disturbance prediction constraint record is represented as follows: in, Indicates the first The sampling phase, the first The monitoring station is under the first Class medium points to the first Constraint records for predicting hydrodynamic disturbances in similar media. This represents the set of constraints for predicting hydrodynamic disturbances; and based on the index of the constraint record, the corresponding arsenic speciation state vector is extracted: in, Indicates the first The sampling phase, the first The monitoring station and the first Arsenic speciation vector corresponding to the medium Indicates the first sampling stage and monitoring station. Arsenic speciation vector corresponding to the medium This represents the set of arsenic speciation state coding results across multiple media. To map media categories and arsenic speciation categories to a unified node space, media coding and arsenic speciation coding are constructed separately: in, Indicates the first Media encoding vector of the medium type, Indicates the first One-hot encoded vectors for media-like materials Represents the medium encoding matrix, Represents the medium coding bias vector; in, Indicates the first The morphological encoding vector of arsenic-like morphologies. Indicates the first The unique thermal coding vector of arsenic-like speciation, Represents the arsenic speciation coding matrix. The arsenic morphology encoding bias vector is represented; then, based on the arsenic morphology state vector, medium encoding, and arsenic morphology encoding, initial features for causal graph nodes used for transformation prediction are generated: in, Indicates the first The sampling phase, the first The monitoring station, the Class media and the first Initial features of causal graph nodes corresponding to arsenic-like morphologies Represents the node feature mapping matrix. This represents the node feature mapping bias vector. Presentation layer normalization operation, This represents a vector concatenation operation; subsequently, a set of arsenic morphology causal nodes is constructed; water and sediment constraint directed edges are generated, including, to express the potential transformation relationship of arsenic between different media and different morphologies, based on the obtained water and sediment dynamic disturbance prediction constraint records, constructing directed edges from the source medium arsenic morphology nodes to the target medium arsenic morphology nodes, for the ... Class migration and transformation process, construction process type encoding: in, Indicates the first The process encoding vector of the class transfer transformation process, Indicates the first One-hot encoded vector of the process class, Represents the process encoding matrix, This represents the process encoding bias vector; it is used to determine the first... Arsenic-like forms in the first Class process downwards The degree of adaptation of arsenic-like form transformation, and the calculation of form transformation compatibility weights: in, Indicates the first Arsenic-like forms after the first Class process transformed into the first Compatibility weights of arsenic-like forms and These represent the encoding vectors for the source arsenic form and the target arsenic form, respectively. This represents a morphological transformation compatible mapping matrix. Represents the morphological transformation compatible mapping bias vector. Represents the compatibility rating vector. The compatibility scoring bias term is represented; and the candidate strength of directed edges is calculated based on the source node, target node, morphological transformation compatibility weight, and hydrodynamic disturbance constraints: in, Indicates that by node Pointing to node And corresponding to the first Candidate strength of directed edges in the class transition process This represents the directed constraint weights obtained from the previous module. This indicates a directional consistency constraint. , , and Represents the fusion coefficient. Indicates the first The edge association matrix corresponding to the class process, This represents the edge strength bias term; finally, the directed edges constrained by water and sediment for transformation prediction are selected based on the candidate strengths of the directed edges. in, This represents a set of directed edges constrained by water and sand. Indicates from the source node Point to target node The directed edge, This represents the directed edge filtering threshold.

8. The method for predicting the risk of arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network according to claim 7, characterized in that, The causal direction correction and graph attention message passing include correcting the causal direction consistency of candidate edge strengths to avoid pseudo-transformation relationships with opposite directions based solely on correlation or local similarity, and performing graph attention message passing on the corrected graph structure. For a pair of candidate edges with opposite directions, the causal direction advantage is calculated. in, Indicates from node Pointing to node Compared to the directional advantage of the opposite side, The strength of the reverse candidate edge is represented; the causal direction correction weight is calculated based on the directional advantage; and the causal constraint edge weight is obtained based on the directed edge candidate strength and the causal direction correction weight. in, This represents the edge weight after the causal direction is corrected. Indicates the candidate strength of directed edges. This represents the causal direction correction weight; to aggregate neighboring node information for transformation probability prediction on the causal constraint graph, a node is defined. The set of neighboring neighborhoods: in, Represents a node The set of neighboring regions, This indicates a neighboring node that points to the target node. This indicates the corresponding migration and transformation process type; then, let the initial message passing state be: in, Represents a node The initial message passing state, This represents the initial characteristics of the node; and in the first... In the round-passing message, the graph attention weights of neighboring nodes to the target node are calculated; then, the node state is updated based on the graph attention weights, finally obtaining the node state after causal direction correction and graph attention message passing; the arsenic morphology transformation probability and graph representation generation include calculating the transformation probability based on the node state and causal constraint edge weights, and organizing it into an arsenic morphology transformation graph representation, assuming the number of message passing rounds is . The node state is represented as follows: in, Represents a node The final graph attention state, Indicate the number of message passing rounds; predict the arsenic form transformation probability based on node state and causal constraint edge weights; then select key transformation paths based on the transformation probability: in, This represents the set of key arsenic speciation transformation pathways. This represents an arsenic speciation transformation pathway that meets the screening criteria. This represents the conversion probability threshold; to form a node-level conversion representation oriented towards risk prediction, the outgoing and incoming conversion probabilities of nodes are aggregated; finally, the node set, directed edge set, key conversion path set, and node-level conversion representation are organized into the unique output of this module.

9. A method for predicting the risk of arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network, as described in claim 8, is characterized in that... The process of outputting arsenic pollution risk classification prediction results based on the arsenic speciation transformation map includes first receiving the arsenic speciation transformation map representation, constructing a speciation risk attribute vector to reflect the differences in toxicity, bioavailability and migration and release capacity of different arsenic speciations, and generating arsenic speciation risk attribute codes from the speciation risk attribute vector; and simultaneously calculating node risk response values ​​for risk prediction based on node-level transformation characterization and arsenic speciation risk attribute codes. To identify arsenic speciation transformation pathways that significantly contribute to pollution risk prediction results during water and sediment regulation, the risk contribution of each pathway is calculated by combining the obtained set of key transformation pathways, transformation probabilities, and node risk response values. In order to quantify the combined effect of source node risk, target node risk, and transformation probability on the risk prediction results, a path risk input vector is constructed. Based on the path risk input vector, the contribution value of the transformation path to the risk prediction result is calculated; To identify key risk transformation paths that significantly contribute to risk prediction results, a path set is constructed; To transform node-level and path-level risk results into station-level risk grades for regional management, multi-media and multi-form node and path risks under the same sampling stage and the same monitoring station are jointly aggregated to generate comprehensive arsenic pollution risk prediction results.

10. A risk prediction system for arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network, comprising executing the risk prediction method for arsenic speciation transformation in the Yellow River Estuary based on a hydrodynamically constrained causal graph attention network as described in claim 1, characterized in that, include: The data acquisition module is configured to collect multi-media arsenic monitoring data; The preprocessing module is configured to perform data preprocessing based on the collected multi-media arsenic monitoring data; The coding module is configured to perform multi-media arsenic speciation risk prediction state coding based on preprocessed data, including monitoring sample input and media adaptation representation, arsenic speciation composition and relative abundance coding, toxicity-bioavailability joint risk state coding, environmental response gating fusion and state result generation. The perturbation module is configured to construct hydro-sediment dynamic perturbation constraints based on the encoding results, including arsenic speciation input and hydro-sediment perturbation feature extraction, perturbation intensity calculation during water and sediment regulation stages, construction of arsenic migration driving force constraints between media, and generation of hydro-sediment dynamic perturbation prediction constraint set; The attention module is configured to construct a causal graph attention network under hydrodynamic constraints based on hydrodynamic perturbation constraints, including arsenic morphology causal node construction, hydrodynamic constraint directed edge generation, causal direction correction and graph attention message passing, arsenic morphology transformation probability and graph representation generation; The prediction module is configured to output arsenic pollution risk classification prediction results based on the arsenic speciation transformation map.

Images