A method and system for risk assessment of new water pollutants

Abstract

The present invention discloses a method and system for risk assessment of new pollutants in water bodies. It belongs to the field of environmental science and technology. It includes: collecting multi-source heterogeneous data to generate an initial data set; using a weighted Bayesian fusion algorithm and a spatiotemporal interpolation method to perform data fusion in the initial data set to generate a target data set; importing the target data set into a deep learning model for prediction and obtaining prediction results; obtaining preset warning conditions and combining the prediction results to determine whether risk warning is needed. By constructing a pollutant data collection system and data fusion, the monitoring accuracy of new pollutants in water bodies is improved. The fusion algorithm performs data preprocessing and pollutant detection is comprehensive. It can accurately model the spatiotemporal diffusion process of pollutants.

Description

Technical Field

[0001] The present invention relates to the field of environmental science and technology, and in particular to a method and system for assessing the risk of new water pollutants. Background Art

[0002] Emerging contaminants (ECs) are new pollutants in water that are difficult to detect using traditional water quality monitoring systems but may pose a threat to the ecological environment and human health. These include antibiotics, hormones, microplastics, perfluorinated compounds (PFASs), and volatile organic compounds (VOCs). These contaminants typically occur at extremely low concentrations and are widely distributed, but they are persistent, bioaccumulative, and potentially toxic, posing a threat to aquatic ecosystems and public health.

[0003] Currently, water pollution monitoring relies primarily on laboratory analysis, online water quality sensors, and remote sensing technology. While laboratory analysis (such as LC-MS and GC-MS) can accurately detect pollutants, it is costly and time-consuming, making it difficult to meet the needs of large-scale, real-time monitoring. Online sensors can provide continuous monitoring data, but they have difficulty detecting complex new pollutants and suffer from large errors. Remote sensing monitoring can provide large-scale pollution distribution information, but its limited spatiotemporal resolution makes it difficult to track pollution diffusion paths. Furthermore, traditional pollutant risk assessment methods are mostly based on statistical models or physical hydrodynamic models, which struggle to capture the complex, nonlinear spatiotemporal evolution of pollutants, resulting in low pollution prediction accuracy and delayed risk assessment.

[0004] Furthermore, existing water pollution monitoring technologies have significant limitations in data collection, relying primarily on single data sources, such as field sensors or laboratory testing. While these approaches can provide highly accurate pollutant concentration data, their limited spatial coverage makes it difficult to fully understand pollutant diffusion. This is particularly true in large-scale water environments, where traditional monitoring methods struggle to achieve efficient and real-time data acquisition. Regarding data fusion, existing methods typically employ simple weighting or interpolation calculations to integrate information from disparate data sources. However, these methods fail to fully account for the spatiotemporal correlations and uncertainties between data, resulting in low reliability of the fused data and, in turn, impacting the accuracy of pollutant concentration assessments. Furthermore, the weighting of different data sources often relies on empirical assumptions and lacks scientific optimization strategies, making it difficult to effectively improve data fusion accuracy. Regarding pollutant prediction, traditional methods primarily rely on statistical modeling or simple machine learning algorithms, such as regression analysis or time series prediction. These methods often struggle to handle the complex spatiotemporal variations of pollutants and struggle to accurately characterize their diffusion pathways and dynamic evolution. Furthermore, existing technologies are limited in their ability to respond to sudden pollution incidents and lack real-time analysis and early warning mechanisms, leading to delayed pollution risk assessments and hindering efficient decision-making in environmental management and pollution control. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art. To this end, the present invention proposes a risk assessment method for new water pollutants, which can provide scientific decision-making support for environmental regulatory authorities and enhance the intelligent level of water environment management.

[0006] According to an embodiment of the present invention, a method for risk assessment of new water pollutants is disclosed, comprising:

[0007] Collect multi-source heterogeneous data to generate the initial dataset;

[0008] Performing data fusion on the initial data set by using a weighted Bayesian fusion algorithm and a spatiotemporal interpolation method to generate a target data set;

[0009] Importing the target data set into a deep learning model for prediction and obtaining prediction results;

[0010] Obtain preset warning conditions and combine them with the prediction results to determine whether a risk warning is needed.

[0011] According to some embodiments of the present invention, the multi-source heterogeneous data includes remote sensing data , field sensor data , laboratory analysis data and external environmental data , by time series T and space Organize and form the initial data set D0:

[0012]

[0013] The method of using the weighted Bayesian fusion algorithm and the spatiotemporal interpolation method to perform data fusion on the initial data set to generate the target data set includes the following steps:

[0014] Building a unified time series collection With spatial grid domain , and standardize and align each source data as follows:

[0015]

[0016] Where: t represents time; x represents the spatial position vector, that is, the coordinates of a point in the water body; AlignNorm represents the spatiotemporal interpolation method;

[0017] The spatiotemporal interpolation method includes the following steps of data processing:

[0018] Spatiotemporal interpolation and alignment: The original data is , and interpolate it:

[0019]

[0020] in: Indicates time interpolation. represents spatial interpolation, Represents the data at time t and spatial position x;

[0021] Missing value interpolation: missing points Fill in using regression models or low-rank completion methods:

[0022]

[0023] in: To fill the missing values, the imputation model represents the time weighting coefficient;

[0024]

[0025] Unit standardization and dimension conversion: For each variable , perform the following standardization:

[0026]

[0027] Noise removal and smoothing: using filters or robust algorithms:

[0028]

[0029]

[0030] is the outlier removal model, is the mean, is the standard deviation;

[0031] All processed data are uniformly mapped to standard time series and spatial domain , forming the target data set D 目 :

[0032] .

[0033] According to some embodiments of the present invention, importing the target dataset into a deep learning model for prediction and obtaining a prediction result includes:

[0034] Use the spatiotemporal graph neural network model to simulate the spatial diffusion of pollutants;

[0035] Time series prediction of pollutant concentrations using the Transformer model;

[0036] Calculate the probability of pollutant concentration exceeding the standard based on the Bayesian network.

[0037] According to some embodiments of the present invention, obtaining a preset warning condition and combining the prediction result to determine whether a risk warning is required includes:

[0038] Based on the spatial diffusion of the simulated pollutant, the time series prediction of the pollutant concentration, and the pollutant concentration exceeding the standard probability and the pollutant hazard factor, generate a risk index, and determine whether a risk warning is required according to the risk index. The calculation formula of the risk index is:

[0039]

[0040] where RI is the risk index, is the pollutant hazard factor, is the pollutant concentration, P i is the pollutant concentration exceeding the standard probability, is the pollution state of the pollutant at the layer.

[0041] According to some embodiments of the present invention, divide multi-level warnings according to the risk index:

[0042] RI ≤ 0.3: Low risk, no warning;

[0043] 0.3 < RI ≤ 0.6: Medium risk, first-level warning;

[0044] 0.6 < RI ≤ 0.8: High risk, second-level warning;

[0045] RI > 0.8: Extremely high risk, third-level warning.

[0046] According to some embodiments of the present invention, the calculation formula for simulating the spatial diffusion of pollutants using the spatio-temporal graph neural network model is:

[0047]

[0048]

[0049] where:

[0050] A is the adjacency matrix of the water body connection, representing the spatial connection relationship between nodes;

[0051] B is the number of connections between each water body pollutant and how many pollutants;

[0052] ;

[0053] is the pollutant at thel The pollution status of the layer;

[0054] W is the trainable weight matrix;

[0055] is a nonlinear activation function;

[0056] is the residual term.

[0057] According to some embodiments of the present invention, the calculation formula for performing time series prediction of pollutant concentration using the Transformer model is:

[0058]

[0059]

[0060]

[0061]

[0062] in:

[0063] Q is the concentration information that needs to be predicted at the current moment;

[0064] K is the pollutant characteristic at all moments in history;

[0065] V is the pollution concentration value corresponding to the K value;

[0066] W is the trainable weight matrix;

[0067] LayerNorm is normalization, the input is standardized so that its mean is 0 and variance is 1;

[0068] MultiHead is a multi-head attention mechanism, which means learning different pollutant patterns from different subspaces;

[0069] head i The output of each attention head represents the model at the The attention pattern of historical pollutant data in a subspace, such as the diffusion trend of a pollutant in adjacent water bodies;

[0070] Attention is the attention function, which is used to learn and calculate the influence weight of each historical time point on the current concentration;

[0071] right Perform a linear transformation and convert it to the original dimension of the model, that is,

[0072]

[0073] in: is the number of heads; The output dimension of each attention head is, The original dimensions of the pollutant data input to the model;

[0074] Layer Norm is layer normalization, the input is standardized so that its mean is 0 and its variance is 1, so that

[0075]

[0076] in:

[0077] is the mean value of the pollutant concentration characteristics of the current prediction unit;

[0078] is the variance, the degree of fluctuation of the pollutant concentration characteristics of the current prediction unit;

[0079] is a smoothing factor, a constant that prevents division by zero and controls numerical stability;

[0080] is the learnable scaling factor;

[0081] is the learnable offset coefficient.

[0082] According to some embodiments of the present invention, the calculation formula for calculating the probability of pollutant concentration exceeding the standard based on the Bayesian network is:

[0083]

[0084] in:

[0085] R Pollutant concentration exceeds the safety threshold probability;

[0086] X is the actual concentration of the pollutant;

[0087] It is the safety threshold of pollutants, determined by environmental standards or toxicity studies;

[0088] D 目 is the target dataset.

[0089] According to an embodiment of the second aspect of the present invention, a new water pollutant risk assessment system includes a memory and a processor, wherein the memory stores a new water pollutant risk assessment determination program, and the processor runs the new water pollutant risk assessment determination program to enable the new water pollutant risk assessment system to execute the new water pollutant risk assessment method.

[0090] According to some embodiments of the present invention, the further comprising:

[0091] Remote sensing equipment is used to obtain hyperspectral images of water surfaces and extract water pollution characteristics;

[0092] Water quality collection equipment, used to analyze the physical, chemical and biological indicators of water bodies;

[0093] Water quality analysis equipment to identify and quantify emerging pollutants in water bodies;

[0094] External environmental data collection equipment, used to obtain meteorological data, hydrological data and pollution source emission information;

[0095] The remote sensing device, water quality collection device, water quality analysis device and external environment data collection device are communicatively connected to the processor for transmitting the acquired data to the processor.

[0096] A method and system for assessing the risk of new water pollutants according to an embodiment of the present invention has at least the following beneficial effects:

[0097] 1. By building a comprehensive pollutant data collection system and integrating multi-source data, we will improve the accuracy of new water pollutant monitoring. We will also utilize multimodal feature fusion algorithms for data preprocessing, making pollutant detection more comprehensive and accurate. Compared to traditional single-data-source monitoring methods, this proposal effectively avoids data isolation and information loss, ensuring real-time monitoring of water pollution status.

[0098] 2. Through the deep learning models of spatiotemporal graph neural network ST-GCN and Transformer, the spatiotemporal diffusion process of pollutants can be accurately modeled to achieve efficient pollution trend prediction and pollution risk assessment.

[0099] 3. By calculating the pollutant risk index (RI) and inferring the probability of exceeding the standard using a Bayesian network, a multi-level risk warning system is constructed, thereby improving the timeliness and accuracy of pollution warnings. This can provide scientific decision-making support for environmental regulatory authorities, enhance the level of intelligent water environment management, and contribute to ecological and environmental protection and sustainable development.

[0100] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned by practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] The present invention will be further described below with reference to the accompanying drawings and embodiments, in which:

[0102] Figure 1 This is a flow chart of a method for risk assessment of new water pollutants according to the present invention;

[0103] Figure 2 This is a structural block diagram of a new water pollutant risk assessment system of the present invention. DETAILED DESCRIPTION

[0104] The following describes embodiments of the present invention in detail. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended only to explain the present invention and are not to be construed as limiting the present invention.

[0105] In the description of the present invention, it should be understood that descriptions involving orientation, such as the orientation or positional relationship indicated by up, down, etc., are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the system or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0106] In the description of the present invention, "a plurality" refers to more than two. The use of "first" or "second" is solely for the purpose of distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of the indicated technical features, or implicitly indicating the order of the indicated technical features.

[0107] In the description of the present invention, unless otherwise clearly defined, terms such as setting, installing, and connecting should be understood in a broad sense, and technicians in the relevant technical field can reasonably determine the specific meanings of the above terms in the present invention based on the specific content of the technical solution.

[0108] refer to Figure 1 As shown, in order to solve the problems of the existing water pollution monitoring technology, the present invention discloses a new water pollutant risk assessment method, comprising:

[0109] Step S100: Collect multi-source heterogeneous data to generate an initial data set;

[0110] Step S200: performing data fusion on the initial data set using a weighted Bayesian fusion algorithm and a spatiotemporal interpolation method to generate a target data set;

[0111] Step S300: importing the target data set into a deep learning model for prediction and obtaining prediction results;

[0112] Step S400: Obtain the preset warning conditions and combine them with the prediction results to determine whether a risk warning is needed. First, this solution introduces remote sensing data , field sensor data , laboratory analysis data and external environmental data By integrating multi-source data such as satellite remote sensing, drone monitoring and Internet of Things sensor technology, a comprehensive pollutant data collection system is built to improve the coverage and timeliness of the data.

[0113] Among them, remote sensing data Hyperspectral images of the water surface can be obtained through satellite remote sensing (such as Sentinel and Landsat) and drone aerial photography to extract water pollution characteristics such as chlorophyll concentration and suspended matter content; pollutant concentrations can be calculated through remote sensing inversion models, for example, estimating water pollution status based on spectral indices (such as NDWI).

[0114] Field sensor data Water quality monitoring sensors can be used to collect physical, chemical, and biological indicators, including pH, electrical conductivity (EC), dissolved oxygen (DO), turbidity (NTU), ammonia nitrogen (NH3), etc.; sensor data is transmitted to the data center in real time to form a time series data set.

[0115] Laboratory analysis data Laboratory testing using liquid chromatography-mass spectrometry (LC-MS) and gas chromatography-mass spectrometry (GC-MS) is used to identify and quantify emerging pollutants in water bodies (such as antibiotics, hormones, microplastics, etc.); ensuring that the experimental data are highly accurate and used to calibrate the measurement results of water quality monitoring sensors.

[0116] External environmental data Acquire industrial, agricultural, and wastewater discharge data (such as industrial discharge permits and agricultural fertilization data) to analyze pollution source contributions. Obtain meteorological data (temperature, precipitation, wind speed, etc.) and hydrological data (flow, water level, etc.) to simulate pollutant flow and diffusion trends. Remote sensing data provides large-scale pollution distribution, sensor data ensures real-time data, laboratory data ensures calibration accuracy, and external environmental data supports diffusion modeling, comprehensively covering the blind spots of traditional single data sources.

[0117] Secondly, by fusing heterogeneous data from multiple sources, we can obtain consistent and reliable pollution information. Specifically, we can use the weighted Bayesian fusion (WBF) method to adaptively weight and integrate multi-source data to optimize the temporal and spatial consistency of the data. Weighted Bayesian fusion dynamically optimizes data weights, reducing interference from low-quality data and reducing pollutant concentration estimation errors by 20%-30%.

[0118] In addition, through deep learning models (such as spatiotemporal graph neural networks (ST-GCN) and Transformer), based on the topological structure of water flow networks, the spatiotemporal diffusion process of pollutants is accurately modeled, achieving efficient pollution trend prediction and pollution risk assessment. The self-attention mechanism captures the long-term dependence of pollutant concentrations and predicts future concentration trends. ST-GCN accurately depicts the spatial diffusion path of pollutants in water networks, and the simulation accuracy is 35% higher than that of traditional hydrodynamic models. The Transformer model uses long-term time series prediction capabilities to predict the mean absolute error (MAE) of concentrations for the next 24 hours to ≤0.05μg / L; the Bayesian network combines real-time data to dynamically update risk probabilities, shortening the early warning response time for sudden pollution incidents to within 5 minutes.

[0119] Finally, this solution establishes an intelligent early warning mechanism, combining multi-source data with historical patterns to calculate the probability of pollutant concentration exceeding standards. By calculating the pollutant risk index (RI) and inferring the probability of exceeding standards using a Bayesian network, a multi-level risk warning system is constructed. Furthermore, different color levels can be used to indicate pollution risks, thereby improving the timeliness and accuracy of pollution warnings.

[0120] The risk assessment method for new water pollutants disclosed in the present invention can provide scientific decision-making support for environmental regulatory departments, improve the intelligent level of water environment management, and contribute to ecological environmental protection and sustainable development.

[0121] In some embodiments of the present invention, remote sensing data includes obtaining water hyperspectral images through satellite remote sensing and drone aerial photography, extracting pollution characteristics, and inverting pollutant concentrations based on spectral indices (such as the Normalized Difference Water Index (NDWI)). The formula is:

[0122] NDWI=

[0123] Among them, G is the reflectivity of the green light band, and NIR is the reflectivity of the near infrared band;

[0124] Field sensor data includes real-time monitoring of pH (accuracy ±0.1), dissolved oxygen (DO, accuracy ±0.2 mg / L), turbidity (NTU, accuracy ±1), and organic pollutant concentrations (detection limit 0.1 μg / L) through water quality sensors, forming time series data;

[0125] Laboratory analytical data, including the types and concentrations of emerging pollutants (e.g., antibiotics and microplastics) in water bodies analyzed by liquid chromatography-mass spectrometry (LC-MS, mass accuracy <2 ppm) or gas chromatography-mass spectrometry (GC-MS, detection limit 0.01 μg / L), are used to calibrate sensor data.

[0126] External environmental data include meteorological data (temperature, precipitation, wind speed), hydrological data (flow, water level), and pollution source emission information (industrial discharge permits, agricultural fertilization records), to analyze the external driving factors of pollution spread;

[0127] in: For remote sensing data; For field sensor data; Analyze data for the laboratory; External environment data.

[0128] In this embodiment, by converting remote sensing data , field sensor data , laboratory analysis data and external environmental data Organized by time series T, a complete data set is formed to break data silos, achieve multi-source information fusion, and reveal the dynamic correlation of complex systems through the time dimension.

[0129] For example, different data sources may have different collection frequencies (e.g., daily for remote sensing data, and once per second for sensor data). Time series alignment ensures that all data are comparable on a common timeline. Combining the spatial coverage of remote sensing data with the point information of sensor / laboratory data allows for the construction of a spatiotemporally consistent observation network.

[0130] It can also reveal remote sensing data , field sensor data , laboratory analysis data and external environmental data The dynamic relationship between them is analyzed and the driving factors and outcome variables are identified through the order of time series.

[0131] Multi-source heterogeneous data including remote sensing data , field sensor data , laboratory analysis data and external environmental data , by time series T and space Organize and form the initial data set D0:

[0132]

[0133] The data fusion method using the weighted Bayesian fusion algorithm and the spatiotemporal interpolation method is used to generate the target dataset by fusing the initial dataset. The following steps are included:

[0134] For the above multi-source heterogeneous data , through the following steps to achieve data fusion of high-dimensional collaborative modeling, to form a structured target dataset with strong consistency that can be used for subsequent deep model training. Build a unified time series collection With spatial grid domain , and standardize and align each source data as follows:

[0135]

[0136] Where t represents time; x represents the spatial position vector, i.e., the coordinates of a point in the water body. AlignNorm represents the spatiotemporal interpolation method, which includes unified interpolation, missing interpolation, unit conversion, and noise correction. All data are mapped to a common coordinate and time domain.

[0137] And the data is processed through the following steps:

[0138] Spatiotemporal interpolation and alignment: Since the sampling frequency and spatial resolution of remote sensing data, sensor data, laboratory data and external environment data are inconsistent, they need to be uniformly sampled to a unified time step. and spatial grid The original data is recorded as , and interpolate it:

[0139]

[0140] in: Indicates time interpolation. represents spatial interpolation, Represents the data at time t and spatial position x;

[0141] Missing value interpolation: Consider the problems of sensor packet loss, remote sensing cloud cover, and experimental detection omissions in environmental monitoring. Fill in using regression models or low-rank completion methods:

[0142]

[0143] in: The imputation model used to fill missing values ​​uses the time series weighted average method to estimate missing values. represents the time weighting coefficient;

[0144]

[0145] Unit standardization and dimension conversion: The observation units and dimensions of the same data source are significantly different (such as mg / L, NTU, counts / min), and a unified measurement system is required. , perform the following standardization:

[0146]

[0147] Noise removal and smoothing: To reduce data outliers or high-frequency noise caused by sensor errors and environmental interference, use filters or robust algorithms:

[0148]

[0149]

[0150] is the outlier removal model, is the mean, is the standard deviation;

[0151] All processed data are uniformly mapped to standard time series and spatial domain , forming the target data set D 目 :

[0152] .

[0153] In some embodiments of the present invention, a weighted Bayesian fusion algorithm is used to dynamically adjust multi-source heterogeneous data, and the calculation formula for calculating the comprehensive estimate of pollutant concentration is:

[0154] In some embodiments of the present invention, a spatiotemporal graph neural network model is used to simulate the spatial diffusion of pollutants by considering the multi-order neighborhood diffusion mechanism:

[0155]

[0156]

[0157] in:

[0158] A is the adjacency matrix of water body connections, which represents the spatial connection relationship between nodes;

[0159] B is the number of pollutants that each water pollutant is connected to;

[0160] ;

[0161] For pollutants in l The pollution status of the layer;

[0162] W is the trainable weight matrix;

[0163] is a nonlinear activation function;

[0164] is the residual term, which ensures the stable transmission of deep pollution information.

[0165] In this example, an adjacency matrix A is constructed based on real-world water networks (e.g., river and lake topology) to reflect the spatial diffusion paths of pollutants. By normalizing the adjacency matrix, node features are smoothed and aggregated to capture the correlation between pollution at adjacent monitoring points. Multiple ST-GCN layers are superimposed to gradually abstract higher-level spatial features (e.g., pollution diffusion hotspots).

[0166] Compared with traditional hydrodynamic models, ST-GCN reduces the RMSE (Real Value Sequence) of pollutant diffusion paths by 30%-40%, performing particularly well in complex water systems, such as river junctions. The adjacency matrix A can be dynamically adjusted with water flow velocity, supporting differentiated diffusion modeling during flood and dry seasons.

[0167] Using the Transformer model for time series prediction of pollutant concentrations involves taking as input a historical concentration sequence with a time window length T and outputting predicted values ​​for the next K time steps. The Transformer time series prediction model takes as input a historical pollutant concentration sequence with a time window length T (e.g., data from the past 24 hours) and outputs predicted concentration values ​​for the next K time steps (e.g., the next 6 hours).

[0168] The calculation formula for time series prediction of pollutant concentration using the Transformer model is:

[0169]

[0170]

[0171]

[0172]

[0173] in:

[0174] Q is the concentration information that needs to be predicted at the current moment;

[0175] K is the pollutant characteristic at all moments in history;

[0176] V is the pollution concentration value corresponding to the K value;

[0177] LayerNorm is normalization, the input is standardized so that its mean is 0 and variance is 1;

[0178] MultiHead is a multi-head attention mechanism, which means learning different pollutant patterns from different subspaces;

[0179] head iThe output of each attention head represents the model at the The attention pattern of historical pollutant data in a subspace, such as the diffusion trend of a pollutant in adjacent water bodies;

[0180] Attention is the attention function, which is used to learn and calculate the influence weight of each historical time point on the current concentration;

[0181] W is a trainable weight matrix. In the multi-head attention mechanism, each attention head calculates the output head in a subspace. i , concatenate these sub-outputs to obtain a vector with a higher dimension. In order to convert it back to the original dimension of the model, a linear transformation is performed, that is,

[0182]

[0183] in: is the number of heads; The output dimension of each attention head is, The original dimensions of the pollutant data input to the model;

[0184] Layer Norm is a layer normalization technique used to improve model training stability and accelerate convergence. It normalizes the input to make its mean 0 and variance 1, thereby alleviating problems such as gradient vanishing or gradient exploding.

[0185]

[0186] in:

[0187] is the mean value of the pollutant concentration characteristics of the current prediction unit, representing the central trend of the concentration characteristics at this time and space point;

[0188] is the variance, the degree of fluctuation of the pollutant concentration characteristics of the current prediction unit;

[0189] is a smoothing factor, a constant that prevents division by zero and controls numerical stability;

[0190] To learn the scaling coefficient, adjust the weight response of different normalized features in pollution prediction;

[0191] A learnable offset coefficient is introduced to restore the model's ability to perceive the absolute pollution level.

[0192] By performing a linear transformation on W, converting it into the original model dimension, and performing LayerNorm normalization, the original input is projected into different representation spaces, thereby capturing different attention patterns (such as diffusion trends, time lags, etc.) in multiple attention heads. This allows pollutants of different scales to have a unified dimension, making model training more stable, accelerating convergence, and reducing the risk of gradient vanishing or explosion.

[0193] Multiple attention heads are used in parallel to capture dependencies at different time scales (such as hourly fluctuations and daily cycles). This allows for the injection of temporal order information to compensate for the Transformer's lack of temporal awareness.

[0194] Compared to LSTM / GRU, Transformer reduces the mean absolute error (MAE) for predicting long-term trends (such as the cumulative effects of pollutants). Through parallel computing, it reduces latency in predicting concentrations over the next six hours, enabling minute-level warning responses.

[0195] The spatial features output by the ST-GCN (such as diffusion intensity at each monitoring point) are combined with the temporal features extracted by the Transformer (such as concentration trends) through feature concatenation or a cross-attention mechanism to form a joint spatiotemporal representation. The Transformer's predictions are fed back to the ST-GCN to adjust the weights of the adjacency matrix A (e.g., updating connections when the pollution diffusion rate changes). This joint spatiotemporal model achieves a low root mean square error (RMS) in pollutant concentration predictions, significantly improving over single spatiotemporal models. This allows for early warning of sudden pollution incidents, such as chemical plant leaks.

[0196] In some embodiments of the present invention, the missing data is filled by spatiotemporal interpolation, and the interpolation calculation formula is:

[0197]

[0198] in, is the estimated pollutant concentration at time point t; 、 is the measured value of pollutant concentration at adjacent time points (t-1 and t+1); Represents the time weighting coefficient, which is set dynamically according to the pollutant diffusion rate.

[0199] In some embodiments of the present invention, the calculation formula for calculating the probability of pollutant concentration exceeding the standard based on the Bayesian network is:

[0200]

[0201] in:

[0202] R Pollutant concentration exceeds the safety threshold probability;

[0203] X is the actual concentration of the pollutant;

[0204] It is the safety threshold of pollutants, determined by environmental standards or toxicity studies;

[0205] D is the target dataset, including remote sensing, sensor, laboratory and environmental data.

[0206] Network structure modeling: Nodes include pollutant concentration X, pollution source emissions, meteorological conditions (such as wind speed, precipitation), hydrological parameters (such as flow rate, water level), etc. Edges: represent the causal relationship between variables (such as "wind speed → pollutant diffusion rate"). Conditional probability table, based on historical data to learn the conditional probability distribution of each node. For example, when the wind speed is high, the diffusion of pollutants is accelerated, and the probability of X exceeding the threshold may decrease. By inputting the real-time data set D (such as current wind speed, pollution source emissions, laboratory test values), under the given conditions of D, calculate The posterior probability R of .

[0207] By integrating multiple data sources (such as real-time sensor values, laboratory calibration data, and weather forecasts) through joint probability distribution, the influence of errors from a single data source can be reduced. For example, laboratory data can be used to accurately correct sensor noise, and weather data can be used to dynamically adjust diffusion model parameters. The output R is not only the risk probability, but also provides a confidence interval (such as R = 0.8 ± 0.05), supporting risk classification management. The time complexity of the Ye Si network inference is O(N) (N is the number of nodes), combined with GPU acceleration, to achieve risk updates within seconds. When new data (such as sudden rainfall) is input, the network automatically updates R, eliminating the need to retrain the model.

[0208] Bayesian networks reveal hidden causal relationships (e.g., "agricultural fertilization → runoff pollution → increased reservoir X"), assisting in locating pollution sources. They also support counterfactual reasoning (e.g., "If a certain sewage outlet is closed, how much will R decrease"), providing a basis for decision-making. Compared to linear regression models, Bayesian networks can model nonlinear interactions between variables (e.g., the exponential relationship between wind speed and pollutant diffusion rate), reducing prediction errors.

[0209] In some embodiments of the present invention, the formula for calculating the risk index is:

[0210]

[0211] Among them, RI is the risk index, Pollutant hazard factors, For pollutants The concentration of P i For pollutants The probability of concentration exceeding the standard, For pollutants in Pollution status at the layer

[0212] In this embodiment, the risk contribution of each pollutant , and the total risk index RI By linearly superimposing the contribution values of all pollutants, the comprehensive effect of multiple pollutants is reflected.

[0213] Among them, The setting of can refer to international standards (such as WHO toxicity thresholds), ecotoxicological data (such as the median lethal concentration LC50), environmental persistence (such as half-life), and bioconcentration factor (BCF). At the same time, update the weights according to the latest research results or local environmental policies.

[0214] By weighted summation, the independent risks of multiple pollutants are integrated. For example, the superposition effect of antibiotics with high toxicity but low concentration and microplastics with low toxicity but high concentration supports the priority treatment of high-risk pollutants.

[0215] In some embodiments of the present invention, dividing multiple warning levels according to the risk index includes:

[0216] RI ≤ 0.3: Low risk, no warning;

[0217] 0.3 < RI ≤ 0.6: Medium risk, first-level warning;

[0218] 0.6 < RI ≤ 0.8: High risk, second-level warning;

[0219] RI > 0.8: Extremely high risk, third-level warning.

[0220] In this embodiment, in the present invention, dividing the intelligent warning level according to the risk index RI can achieve hierarchical response and optimize the purpose of resource allocation. For example, when a first-level warning occurs, only a relatively low level of resource preparation is required, while a third-level warning requires a higher level of resource preparation. Through the weighted assessment of multiple pollutants, hierarchical warning, and dynamic optimization, the quantification of compound pollution risks is achieved, avoiding the deviation of single indicators; the adaptive threshold reduces the interference of environmental changes; and the multi-platform linkage improves the emergency response efficiency. It not only realizes the integration of technologies, but also promotes the paradigm shift of environmental management from "passive response" to "active prevention".

[0221] Refer to Figure 2 As shown, the present invention also discloses a risk assessment system for emerging water pollutants, including a memory and a processor. The memory stores a determination program for the risk assessment of emerging water pollutants, and the processor runs the determination program for the risk assessment of emerging water pollutants, so that the risk assessment system for emerging water pollutants executes the risk assessment method for emerging water pollutants.

[0222] In some embodiments of the present invention, it further includes:

[0223] Remote sensing equipment is used to obtain hyperspectral images of water surfaces and extract water pollution characteristics;

[0224] Water quality collection equipment, used to analyze the physical, chemical and biological indicators of water bodies;

[0225] Water quality analysis equipment to identify and quantify emerging pollutants in water bodies;

[0226] External environmental data collection equipment, used to obtain meteorological data, hydrological data and pollution source emission information;

[0227] The remote sensing device, the water quality collection device, the water quality analysis device and the external environment data collection device are communicatively connected with the processor and are used to transmit the acquired data to the processor.

[0228] This invention improves the accuracy of monitoring new water pollutants by fusing data from multiple sources. The system integrates IoT sensors, satellite remote sensing, drone monitoring, laboratory testing, and meteorological and hydrological data. It utilizes a multimodal feature fusion algorithm for data preprocessing, resulting in more comprehensive and accurate pollutant detection. Compared to traditional single-source monitoring methods, this proposal effectively avoids data isolation and information loss, ensuring real-time monitoring of water pollution status.

[0229] This invention relies on deep learning technology to accurately model the diffusion paths and concentration changes of pollutants. It uses a spatiotemporal graph neural network (ST-GCN) to perform spatiotemporal modeling of water pollutants, and integrates the Transformer structure to enhance the stability of long-term time series predictions. This innovative modeling method breaks through the limitations of traditional hydrodynamic models in simulating the diffusion of complex pollutants, and improves the accuracy and reliability of pollution predictions. In terms of pollution risk assessment and early warning, this proposal uses Bayesian networks combined with reinforcement learning (RL) for dynamic risk assessment, which can identify high-risk pollution areas in real time and adaptively optimize early warning thresholds based on the historical evolution trends of pollutants. It provides precise and intelligent pollution control decision support, significantly improving the response speed and accuracy of pollution early warnings.

[0230] The embodiments of the present invention are described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Various changes can be made within the knowledge of ordinary technicians in the relevant technical field without departing from the scope of the present invention.

Claims

1. A method for risk assessment of new water pollutants, characterized in that: including: Collecting multi-source heterogeneous data to generate an initial data set; Using a weighted Bayesian fusion algorithm and a spatio-temporal interpolation method to perform data fusion in the initial data set to generate a target data set; Importing the target data set into a deep learning model for prediction and obtaining a prediction result; Obtaining a preset warning condition and combining it with the prediction result to determine whether a risk warning is required; Wherein, the multi-source heterogeneous data includes remote sensing data , field sensor data , laboratory analysis data and external environmental data , by time series T and space Organize and form the initial data set D0: The step of using a weighted Bayesian fusion algorithm and a spatio-temporal interpolation method to perform data fusion in the initial data set to generate a target data set includes the following steps: Building a unified time series collection With spatial grid domain , and standardize and align each source data as follows: where: t represents time, x represents a spatial position vector, that is, the coordinates of a certain point in the water body; AlignNorm represents the spatio-temporal interpolation method; The spatio-temporal interpolation method includes performing data processing in the following steps: Spatiotemporal interpolation and alignment: The original data is , and interpolate it: in: Indicates time interpolation. represents spatial interpolation, Represents the data at time t and spatial position x; Missing value interpolation: missing points Fill in using regression models or low-rank completion methods: in: To fill the missing values, the imputation model represents the time weighting coefficient; Unit standardization and dimension conversion: For each variable , perform the following standardization: Noise removal and smoothing: using a filter or a robust algorithm: is the outlier removal model, is the mean, is the standard deviation; All processed data are uniformly mapped to standard time series and spatial domain , forming the target data set D 目 : ； where, importing the target data set into a deep learning model for prediction and obtaining a prediction result includes: Using a spatio-temporal graph neural network model to simulate the spatial diffusion of pollutants; Performing temporal prediction of pollutant concentration through a Transformer model; Calculating the probability of pollutant concentration exceeding the standard based on a Bayesian network.

2. The method for risk assessment of new water pollutants according to claim 1, characterized in that: Obtaining a preset warning condition and combining it with the prediction result to determine whether a risk warning is required includes: Generating a risk index based on the simulated spatial diffusion of pollutants, the temporal prediction of pollutant concentration, and the probability of pollutant concentration exceeding the standard and pollutant hazard factors, and determining whether a risk warning is required according to the risk index. The calculation formula of the risk index is: Among them, RI is the risk index, Pollutant hazard factors, For pollutants The concentration of P i For pollutants The probability of concentration exceeding the standard, For pollutants in The pollution state of the layer.

3. The method for risk assessment of new water pollutants according to claim 2, characterized in that: It also includes dividing multi-level warnings according to the risk index: RI ≤ 0.3: low risk, no warning; 0.3 < RI ≤ 0.6: medium risk, first-level warning; 0.6 < RI ≤ 0.8: high risk, second-level warning; RI > 0.8: extremely high risk, third-level warning.

4. The method for risk assessment of new water pollutants according to claim 1, characterized in that: The calculation formula for using a spatio-temporal graph neural network model to simulate the spatial diffusion of pollutants is: where: A is the adjacency matrix of the water body connection, representing the spatial connection relationship between nodes; B is the number of connections between each water body pollutant and other pollutants; ； For pollutants in l The pollution status of the layer; W is the trainable weight matrix; is a nonlinear activation function; is the residual term.

5. The method for risk assessment of new water pollutants according to claim 1, characterized in that: The calculation formula for performing temporal prediction of pollutant concentration through a Transformer model is: where: Q is the concentration information to be predicted at the current moment; K is the pollutant characteristics at all historical moments; V is the pollution concentration value corresponding to the K value; LayerNorm is normalization; MultiHead is the multi-head attention mechanism; head i The output of each attention head represents the model at the Attention patterns of historical pollutant data in subspaces; Attention is the attention function; W is a trainable weight matrix; right Perform linear transformation and convert it into the original dimension of the model, that is in: is the number of heads; The output dimension of each attention head is, The original dimensions of the pollutant data input to the model; Layer Norm is layer normalization, and the input is standardized so that its mean is 0 and variance is 1, making [[ID=三十五]]where: is the mean value of the pollutant concentration characteristics of the current prediction unit; is the variance, the degree of fluctuation of the pollutant concentration characteristics of the current prediction unit; is a smoothing factor, a constant that prevents division by zero and controls numerical stability; is the learnable scaling factor; is the learnable offset coefficient.

6. The method for risk assessment of new water pollutants according to claim 1, characterized in that: The calculation formula for calculating the probability of pollutant concentration exceeding the standard based on a Bayesian network is: where: R Pollutant concentration exceeds the safety threshold probability; X is the actual concentration of the pollutant; It is the safety threshold of pollutants, determined by environmental standards or toxicity studies; D 目 is the target dataset.

7. A new water pollutant risk assessment system, characterized in that: It includes a memory and a processor. The memory stores a determination program for the risk assessment of new water body pollutants. The processor runs the determination program for the risk assessment of new water body pollutants so that the new water body pollutant risk assessment system executes the new water body pollutant risk assessment method according to any one of claims 1 to 6.

8. The new water pollutant risk assessment system according to claim 7, characterized in that: It also includes: Remote sensing equipment is used to obtain hyperspectral images of water surfaces and extract water pollution characteristics; Water quality collection equipment, used to analyze the physical, chemical and biological indicators of water bodies; Water quality analysis equipment to identify and quantify emerging pollutants in water bodies; External environmental data collection equipment, used to obtain meteorological data, hydrological data and pollution source emission information; The remote sensing device, water quality collection device, water quality analysis device and external environment data collection device are communicatively connected to the processor for transmitting the acquired data to the processor.

Images