Water pollutant tracing method and system coupled with pollution trend

Abstract

The invention discloses a pollution trend-coupled water pollutant tracing method and system, and belongs to the technical field of water pollution tracing. The existing water pollutant tracing method is difficult to realize quick response in pollution emergencies and cannot meet actual requirements. According to the water pollutant tracing method coupled with the pollution trend, a dynamic sampling strategy driven by the concentration change rate is introduced, the pollutant concentration change trend is utilized, the time-varying prior distribution of pollution source parameters is generated, and the sampling step length is determined, so that the sampling precision can be improved in a severe pollution fluctuation stage, and the sampling accuracy is improved. In the stationary stage, the convergence speed is increased, the parameter inversion efficiency and the result stability are effectively improved, and the problems that a traditional water pollutant tracing method is slow in convergence, inaccurate in positioning and the like in an emergent pollution scene are solved; and high-precision dynamic traceability analysis of sudden water body pollution events can be realized, so that the method can be suitable for emergency scenes requiring minute-level decision support.

Description

Technical Field

[0001] This invention relates to a method and system for tracing the source of water pollutants coupled with pollution trends, belonging to the field of water pollution source tracing technology. Background Technology

[0002] Chinese literature (Zhu Song, Liu Guohua, Wang Lizhong, et al. Bayesian method for pollution source identification in hydrodynamic-water quality coupling model [J]. Journal of Sichuan University (Engineering Science Edition), 2009, 41, (5): 30-35.) discloses a method for tracing the source of water pollutants. For the hydrodynamic-water quality coupling model, it establishes a mathematical model for pollutant point source identification based on Bayesian inference. Through Markov chain Monte Carlo (MCMC) posterior sampling, it obtains the posterior probability distribution and estimators of the pollution source location and intensity, effectively handling the uncertainty and nonlinearity of the model. The above examples show that the Bayesian inference method combined with MCMC sampling can well describe and solve the inverse problem of pollution source identification under hydrodynamic-water quality coupling conditions.

[0003] The aforementioned schemes and existing pollutant source tracing methods mainly rely on hydrodynamic and water quality models combined with the Markov Chain Monte Carlo (MCMC) algorithm for inversion. This method obtains possible pollution source parameters through continuous sampling and iteration. However, existing MCMC algorithms have significant limitations in practical applications, mainly in two aspects: First, it suffers from low computational efficiency: the standard MCMC algorithm requires a large number of samples to converge, especially under complex hydrodynamic conditions and high-dimensional parameter spaces, resulting in extremely slow convergence and making it difficult to achieve rapid response in pollution emergencies. This fails to meet the practical needs of emergency scenarios requiring minute-level or hour-level decision support.

[0004] Secondly, the prior distribution relies on manual setting: existing methods often require experts to set prior distributions for parameters such as the location and intensity of pollution sources based on experience. However, sudden pollution events often lack sufficient information, leading to inaccurate prior distributions and consequently affecting the reliability and accuracy of the source tracing results. This reliance on manual setting can easily cause significant deviations in actual environmental emergencies.

[0005] The information disclosed in this background section is only for understanding the background of the inventive concept, and therefore may include information that does not constitute prior art. Summary of the Invention

[0006] To address the aforementioned technical problems, one objective of this invention is to provide a method and system for tracing the source of water pollutants coupled with pollution trends. This innovatively introduces a dynamic sampling strategy driven by the rate of concentration change, utilizing the trend of pollutant concentration changes to screen the pollution source search area and determine the sampling step size. This improves sampling accuracy during periods of drastic pollution fluctuations and accelerates convergence speed during stable periods, effectively enhancing parameter inversion efficiency and result stability. It overcomes the problems of slow convergence and inaccurate positioning in traditional MCMC algorithms during sudden pollution events. Furthermore, it enables high-precision dynamic source tracing analysis of sudden water pollution events, making it suitable for emergency scenarios requiring minute-level or hour-level decision support.

[0007] To address the aforementioned technical problems, another objective of this invention is to provide a method and system for tracing the source of water pollutants coupled with pollution trends. The prior distribution no longer adopts a fixed form or a manual input form, but instead generates a time-varying prior distribution of pollution source parameters through the pollutant concentration change rate and historical pollution data. At the same time, an adaptive sampling step size is adopted, which can improve computational efficiency, reduce prior dependence, and achieve deep integration with the digital twin system while ensuring inversion accuracy. This enables rapid, accurate, and automatic source tracing of water pollution.

[0008] To address the aforementioned technical problems, another objective of this invention is to provide a method and system for tracing the source of water pollutants coupled with pollution trends. By setting up a water area simulation model, a pollution diffusion model, a pollution trend prediction model, and a pollution source tracing model, pollution is traced, simulated, and inverted to form a feedback closed loop. This enables high-precision dynamic source tracing analysis of sudden water pollution events. The solution is scientific, practical, and feasible.

[0009] To achieve one of the above objectives, the first technical solution of the present invention is as follows: A method for tracing the source of water pollutants coupled with pollution trends includes the following: Based on the topographic and hydrological data of a certain body of water, construct a digital twin object of the body of water; By using pollution monitoring data and historical pollution data, the pollution change trend in digital twin objects of water bodies is predicted, and the pollutant concentration change rate is obtained. Based on the pollutant concentration change rate and historical pollution data, a time-varying prior distribution of pollution source parameters is generated to screen the pollution source search area; at the same time, based on the adjustment requirements of sampling sensitivity, an inhibition coefficient is set. Based on the correspondence between the inhibition coefficient, the intensity of concentration change, and the sampling sensitivity, a mapping function from the pollutant concentration change rate to the weight is established; Using mapping functions and smoothing coefficients, a dynamic weighting factor is calculated to quantify the impact of pollutant concentration change rate on the sampling strategy; The sampling step size of the pollution source parameters is determined based on the dynamic weighting factor and the basic step size. Based on the sampling step size, a search is performed in the time-varying prior distribution to obtain pollutant source information, thereby enabling the tracing of pollutants in water bodies.

[0010] This invention innovatively introduces a dynamic sampling strategy driven by the rate of concentration change. By utilizing the trend of pollutant concentration changes, it filters the search area for pollution sources and determines the sampling step size. This improves sampling accuracy during periods of drastic pollution fluctuations and accelerates convergence during periods of stability, effectively enhancing parameter inversion efficiency and result stability. It overcomes the problems of slow convergence and inaccurate positioning in traditional MCMC algorithms under sudden pollution scenarios. Consequently, it enables high-precision dynamic source tracing analysis of sudden water pollution events, making it suitable for emergency scenarios requiring minute-level or hour-level decision support.

[0011] Furthermore, the prior distribution of this invention no longer adopts a fixed form or a manual input form, but generates a time-varying prior distribution of pollution source parameters through the pollutant concentration change rate and pollution history data. At the same time, it adopts an adaptive sampling step size, which can improve the computational efficiency, reduce prior dependence, and achieve deep integration with the digital twin system while ensuring inversion accuracy. This enables rapid, accurate, and automatic source tracing of water pollution, which can meet practical needs.

[0012] As a preferred technical measure: The method for constructing a digital twin object of a water body based on its topographic and hydrological data is as follows: Acquire topographic and hydrological data for a specific body of water. Topographic data includes the cross-sectional structure of the river channel, shoreline morphology, location of tributary confluences, and distribution of obstacles. Hydrological data includes longitudinal flow velocity, water depth, discharge, water level changes, and hydrodynamic parameters. Based on topographic data, the water area is discretized into a one-dimensional or two-dimensional grid, and depth and slope are associated on each grid cell; Hydrological data is incorporated into a one-dimensional or two-dimensional grid to obtain a digital twin object of the water area.

[0013] As a preferred technical measure: The following method is used to predict pollution change trends in digital twin objects of water bodies and obtain pollutant concentration change rates by utilizing pollution monitoring data and historical pollution data: Obtain environmental impact factors, including water temperature, rainfall, river flow, wind speed, and wind direction; A distributed water quality sensor network will be used to cover key monitoring sections in the upper, middle and lower reaches of the river, and historical pollution data, including chemical oxygen demand, ammonia content, nitrogen content and total phosphorus content, will be collected through the distributed water quality sensor network. Environmental impact factors and historical pollution data are normalized to obtain a multivariate observation mapping sequence; at the same time, several prediction time steps are set according to the source tracing requirements. The multivariate observation mapping sequence corresponding to a certain prediction time step is input into a pre-built pollution trend prediction model to predict the changing trend of pollutants and obtain the corresponding pollutant concentration change rate.

[0014] As a preferred technical measure: The method for constructing a pollution trend prediction model is as follows: Receive the standardized multivariate observation mapping sequence, which includes n pollution indicators and m environmental factors; Set up a gating mechanism, which includes a forget gate, an input gate, and an output gate; The forget gate is used to delete normal concentration data, the input gate is used to update the current environmental factors, and the output gate is used to control the contribution of the hidden state to the output, so that the pollution trend prediction model focuses on the core factors affecting the change of pollutant concentration. Based on the gating mechanism, a pollution trend prediction model is established using the hidden layers in the long short-term memory network; The pollution trend prediction model can learn the long-period dependence in the multivariate observation mapping sequence and map the output of the hidden layer to a single predicted value, namely the rate of change of pollutant concentration at a certain time step. The long-term dependence includes the lag variation of pollution concentration with water temperature, rainfall, river flow, wind speed, and wind direction. Then, the pollution trend prediction model is trained and optimized using historical pollution data, thereby completing the construction of the pollution trend prediction model.

[0015] As a preferred technical measure: The method for training and optimizing pollution trend prediction models using historical pollution data is as follows: Several sets of input and output samples were extracted from historical pollution data, covering different pollution scenarios and different hydrological conditions. The mean squared error formula is used as the loss function to measure the deviation between the predicted pollutant concentration change rate and the actual pollutant concentration change rate. The adaptive learning rate function is used as the optimizer, and the number of iterations is adjusted according to the convergence. The input and output samples are divided into training and validation sets in a 4:1 ratio. The training set is input into the pollution trend prediction model to train the pollution trend prediction model and obtain the prediction results. The prediction results are compared with the validation set to obtain the prediction error; By adjusting the parameters of the pollution trend prediction model based on the prediction error, the pollution trend prediction model can be optimized. Furthermore, if the prediction error increases, the number of hidden units in the hidden layer is reduced or regularization is increased; if the prediction error is too high, historical data is added to contaminate the data.

[0016] As a preferred technical measure: The method for generating the time-varying prior distribution of pollution source parameters based on pollutant concentration change rate and historical pollution data is as follows: Step 1: Based on historical pollution data, construct a pollution source parameter space, which includes the geographical coordinates of the pollution source, the pollution emission start time, and the time variation function of emission intensity. Step 2: Based on the pollution source parameter space, set static and dynamic features; Static characteristics include the spatial distribution of sewage outlets, river topography, and the frequency of historical pollution incidents; dynamic characteristics include real-time rainfall and the rate of change in pollutant concentrations. Step 3: Perform hierarchical processing and feature fusion on static and dynamic features, and extract the correlation patterns between static and dynamic features to obtain the fused multimodal features; Step 4: Map the fused multimodal features to the probability space of the pollution source parameters to generate a probability mapping result, which is a probability distribution of the pollution source location, emission start time, and emission intensity as a function of time. Step 5: Update the probability mapping results in real time according to changes in time and environment to form a time-varying prior distribution, so as to screen the search area for pollution sources.

[0017] As a preferred technical measure: The method for determining the sampling step size of pollution source parameters based on the pollutant concentration change rate is as follows: Based on the adjustment requirements of sampling sensitivity, set the suppression coefficient; Based on the correspondence between the inhibition coefficient, the intensity of concentration change, and the sampling sensitivity, a mapping function from the pollutant concentration change rate to the weight is established; Using mapping functions and smoothing coefficients, a dynamic weighting factor is calculated to quantify the impact of pollutant concentration change rate on the sampling strategy; The sampling step size of pollution source parameters is determined based on the dynamic weighting factor and the basic step size.

[0018] As a preferred technical measure: The method for obtaining pollutant source information by searching within a time-varying prior distribution based on the sampling step size is as follows: In time-varying prior distributions, candidate parameters are generated through Gaussian perturbation; Based on the sampling step size, sample from the candidate parameters to obtain candidate pollution source parameters; By inputting the parameters of the candidate pollution sources into a pre-built pollution diffusion model, new simulated values ​​of pollutant concentrations are obtained. Based on the new simulated pollutant concentrations, a log-likelihood function is constructed. The acceptance probability of candidate pollution source parameters is calculated using the log-likelihood function. Based on the acceptance probability, the sampling and screening process is continuously executed until the set convergence criteria are reached, thereby obtaining a high-confidence posterior sample set of pollution source parameters and outputting pollutant source information, which includes at least the spatial location of the pollution source, emission intensity curve, and start time.

[0019] As a preferred technical measure: The method for constructing a pollution diffusion model is as follows: Based on the digital twin object of the water area, boundary conditions are constructed, including the upstream boundary, the downstream boundary, and the shoreline; The upstream boundary is set as the inflow boundary, the downstream boundary is the outflow boundary, and the shoreline is set as the no-flux boundary. The boundary conditions and candidate pollution source parameters are used as input source terms for pollution diffusion. Based on the input source terms, the grid cells in the digital twin object of the water area are set as pollution release sources; Calculate the pollutant release amount of each grid cell at different time steps based on emission intensity; The finite difference method or finite volume method is used to solve the one-dimensional convection-diffusion equation, so as to transform the continuous differential equation into a discrete system of algebraic equations. Based on the algebraic equations and pollutant release, the pollution diffusion data of pollutants between grid cells is calculated in each time step. Based on pollution diffusion data, the pollutant concentration of each grid cell at different time steps is calculated iteratively, and the simulated pollutant concentration values ​​of each monitoring section are finally output, thereby realizing the construction of the pollution diffusion model.

[0020] To achieve one of the above objectives, the second technical solution of the present invention is as follows: A water pollutant source tracing system coupled with pollution trends includes a water area simulation model, a pollution diffusion model, a pollution trend prediction model, and a pollution source tracing model; A water area simulation model is used to construct a digital twin object of a water area based on its topographic and hydrological data. A pollution diffusion model is used to process digital twin objects of water bodies to obtain simulated pollutant concentration values ​​at multiple monitoring sections in order to simulate the pollutant diffusion process. Pollution trend prediction models are used to predict the changing trends of pollutants and obtain the rate of change of pollutant concentrations. The pollution source tracing model generates time-varying prior distributions and candidate pollution source parameters based on the pollutant concentration change rate. It sets suppression coefficients based on the adjustment requirements of sampling sensitivity. A mapping function from the pollutant concentration change rate to weights is established based on the correspondence between the suppression coefficient, the intensity of concentration change, and the sampling sensitivity. Using the mapping function and a smoothing coefficient, a dynamic weighting factor is calculated to quantify the impact of the pollutant concentration change rate on the sampling strategy. The sampling step size for pollution source parameters is determined based on the dynamic weighting factor and the basic step size. Finally, an inversion is performed using a pollution diffusion model to form a feedback loop, thereby obtaining pollutant source information and achieving source tracing of water pollutants.

[0021] This invention establishes a water area simulation model, a pollution diffusion model, a pollution trend prediction model, and a pollution source tracing model to trace, simulate, and invert pollution, forming a feedback loop. This enables high-precision dynamic source tracing analysis of sudden water pollution events, and is therefore applicable to the construction of a smart water environment monitoring and emergency source tracing system platform.

[0022] Furthermore, this invention fully integrates topographic and hydrological data to construct a digital twin object of water area that is highly consistent with the real water environment, serving as the physical support base for the pollution source inversion module, thereby improving the overall inference accuracy and scenario adaptability.

[0023] Meanwhile, based on the pollution source parameter inversion results and diffusion simulation feedback, this invention iteratively corrects the prior probability distribution of pollution sources, forming a closed-loop feedback chain for data simulation and source tracing. This enables the improvement of computational efficiency, reduction of prior dependence, and deep integration with digital twin systems while ensuring inversion accuracy. Thus, it can achieve rapid, accurate, and intelligent source tracing of water pollution.

[0024] Compared with existing technical solutions, the present invention has the following beneficial effects: This invention innovatively introduces a dynamic sampling strategy driven by the rate of concentration change. By utilizing the trend of pollutant concentration changes, it filters the search area for pollution sources and determines the sampling step size. This improves sampling accuracy during periods of drastic pollution fluctuations and accelerates convergence during periods of stability, effectively enhancing parameter inversion efficiency and result stability. It overcomes the problems of slow convergence and inaccurate positioning in traditional MCMC algorithms under sudden pollution scenarios. Consequently, it enables high-precision dynamic source tracing analysis of sudden water pollution events, making it suitable for emergency scenarios requiring minute-level or hour-level decision support.

[0025] Furthermore, the prior distribution of this invention no longer adopts a fixed form or a manual input form, but generates a time-varying prior distribution of pollution source parameters through the pollutant concentration change rate and pollution history data. At the same time, it adopts an adaptive sampling step size, which can improve computational efficiency, reduce prior dependence, and achieve deep integration with the digital twin system while ensuring inversion accuracy. This enables rapid, accurate, and automatic source tracing of water pollution.

[0026] Furthermore, by setting up a water area simulation model, a pollution diffusion model, a pollution trend prediction model, and a pollution source tracing model, this invention can trace, simulate, and invert pollution to form a feedback loop, enabling high-precision dynamic source tracing analysis of sudden water pollution events. Therefore, it can be applied to the construction of a smart water environment monitoring and emergency source tracing system platform. Attached Figure Description

[0027] Figure 1 This is a schematic flowchart of the water pollutant source tracing method of the present invention; Figure 2 This is a structural block diagram of the water pollutant tracing system of the present invention; Figure 3 This is another schematic diagram of the water pollutant source tracing method of the present invention. Detailed Implementation

[0028] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application. This invention covers any substitutions, modifications, equivalent methods, and solutions made within the spirit and scope of the invention as defined by the claims.

[0029] like Figure 1 As shown, this is the first specific embodiment of the water pollutant source tracing method coupled with pollution trends of the present invention: A method for tracing the source of water pollutants coupled with pollution trends includes the following: Based on the topographic and hydrological data of a certain body of water, construct a digital twin object of the body of water; By using pollution monitoring data and historical pollution data, the pollution change trend in digital twin objects of water bodies is predicted, and the pollutant concentration change rate is obtained. Based on pollutant concentration change rates and historical pollution data, a time-varying prior distribution of pollution source parameters is generated to screen pollution source search areas. Simultaneously, the sampling step size for pollution source parameters is determined according to the pollutant concentration change rate. This determination includes: setting a suppression coefficient based on the adjustment requirements of sampling sensitivity; establishing a mapping function from pollutant concentration change rate to weights based on the suppression coefficient, the correlation between concentration change intensity and sampling sensitivity; calculating a dynamic weighting factor using the mapping function and a smoothing coefficient to quantify the impact of pollutant concentration change rate on the sampling strategy; and finally, determining the sampling step size for pollution source parameters based on the dynamic weighting factor and the base step size. Based on the sampling step size, a search is performed in the time-varying prior distribution to obtain pollutant source information, thereby enabling the tracing of pollutants in water bodies.

[0030] A second specific embodiment of the water pollutant source tracing method coupled with pollution trends of the present invention: A method for tracing the source of water pollutants coupled with pollution trends includes the following: Construct water area modeling units, process the topographic and hydrological data of a certain water area, and generate a digital twin object of the water area; A pollution diffusion model was constructed, and digital twin objects of water bodies and pollution monitoring data were processed to obtain simulated values ​​of pollutant concentrations at multiple monitoring sections in order to simulate the pollutant diffusion process. A pollution trend prediction model was constructed, and the change trend of pollutants was predicted by using simulated pollutant concentration values ​​and historical pollution data, thus obtaining the pollutant concentration change rate. A pollution source tracing model is constructed, and the search area for pollution sources is screened based on the pollutant concentration change rate. An inhibition coefficient is set based on the adjustment requirements of sampling sensitivity. According to the correspondence between the inhibition coefficient, the intensity of concentration change, and the sampling sensitivity, a mapping function from the pollutant concentration change rate to the weights is established. Using the mapping function and the smoothing coefficient, a dynamic weighting factor is calculated to quantify the impact of the pollutant concentration change rate on the sampling strategy. Based on the dynamic weighting factor and the basic step size, the sampling step size of the pollution source parameters is determined. Combined with the simulated pollutant concentration values, the pollutant source information is obtained, realizing the source tracing of water pollutants.

[0031] like Figure 2 As shown, a specific embodiment of the water pollutant source tracing system coupled with pollution trends of the present invention is as follows: A water pollutant source tracing system coupled with pollution trends includes a water area simulation model, a pollution diffusion model, a pollution trend prediction model, and a pollution source tracing model; A water area simulation model is used to construct a digital twin object of a water area based on its topographic and hydrological data. A pollution diffusion model is used to process digital twin objects of water bodies to obtain simulated pollutant concentration values ​​at multiple monitoring sections in order to simulate the pollutant diffusion process. Pollution trend prediction models are used to predict the changing trends of pollutants and obtain the rate of change of pollutant concentrations. The pollution source tracing model is used to generate time-varying prior distributions and candidate pollution source parameters based on the pollutant concentration change rate, and to combine it with the pollution diffusion model for inversion to form a feedback closed loop, thereby obtaining pollutant source information and realizing the source tracing of water pollutants.

[0032] A third specific embodiment of the water pollutant source tracing method coupled with pollution trends of the present invention: A method for tracing the source of water pollutants coupled with pollution trends includes the following steps: S1: Deploy multi-parameter water quality sensors and meteorological data acquisition equipment in the target river section to collect real-time pollutant indicators such as chemical oxygen demand (COD), ammonia nitrogen content, and heavy metal content, as well as meteorological data such as rainfall, wind speed, and temperature. Perform anomaly detection and missing value repair on the raw data, combining rule-based thresholds and multi-parameter linkage to determine anomaly types and avoid mishandling pollution mutations. Simultaneously, perform normalization standardization processing on all types of data to construct a unified, standardized time-series dataset as input for subsequent modeling.

[0033] S2: Based on river cross-section information, current water depth conditions, upstream and downstream conditions (pollution source location, emission concentration, emission time), rainfall forecast, river friction coefficient, infiltration coefficient, and other data, a pollution diffusion model is constructed. This model can output the pollutant concentration distribution at each monitoring section based on any input set of pollution source parameters (including location, emission amount, duration, etc.), supporting sampling acceptance criteria and model inversion.

[0034] S3: Based on historical monitoring data, a Long Short-Term Memory (LSTM) network is trained to establish a pollution trend prediction model for predicting the rate of change in pollutant concentrations. This pollution trend prediction model supports online operation and can predict pollutant change trends for the next period based on the latest time-series data, providing real-time adjustment signals for the dynamic sampling strategy of Markov Chain Monte Carlo (MCMC).

[0035] S4: Based on the pollution source parameter vector, a MCMC pollution source tracing model is constructed. The pollution source parameter vector includes spatial location, emission start time, and emission intensity variation function over time. The prior distribution is no longer presented in a fixed or manually input form, but rather through a neural network model. Time-varying prior distribution of pollution source parameters generated by dynamic learning of multimodal features Based on this, parameter space sampling is performed using the MCMC algorithm to fit the pollution diffusion model simulation results with actual observation data, gradually converging to the posterior distribution.

[0036] S5: Construct a dynamic coupling mechanism between a Long Short-Term Memory (LSTM) network and MCMC. Based on the LSTM network's predictive ability of pollutant concentration change rates, the concentration change rate output by the LSTM module is introduced into the MCMC sampling process, constructing a dynamic weight and step size adjustment mechanism. The predicted value output by the LSTM network... Dynamic weighting factor for dynamically adjusting MCMC This allows for the control of the sampling step size, achieving an adaptive balance between sampling speed and accuracy.

[0037] S6: Based on the high spatiotemporal resolution capability of the pollution diffusion model, after each round of MCMC sampling generates candidate pollution source parameters, the diffusion model is called to simulate the spatiotemporal concentration distribution of pollutants and generate simulated pollutant concentration values; and the simulated pollutant concentration values ​​are compared with the observed values ​​as input for the acceptance criteria.

[0038] S7: Based on a large number of valid sampling results, a posterior distribution of pollution source location parameters is constructed, and a heat map of the spatial probability distribution of pollution sources is generated to assist decision-makers in quickly locating high-risk emission sources. Finally, the most probable geographical location of the pollution source, the start time of emission, the emission intensity, and their confidence intervals are output, providing a scientific and intuitive basis for tracing the source in response to sudden water pollution events.

[0039] In this embodiment S1: By deploying a water quality monitoring sensor network in rivers and pipe networks, a monitoring grid unit is constructed based on the location of sensor points in the study area. A data matrix is ​​constructed on each monitoring unit node to collect key water quality parameters of water pollutants in real time. The expression is as follows:

[0040] in, Number the nodes. To Within the time node The values ​​of these attributes are specifically chemical oxygen demand (COD), ammonia nitrogen (NH3-N), total phosphorus (TP), and total nitrogen (TN), etc. .

[0041] Simultaneously, meteorological data such as rainfall, wind speed, wind direction, and temperature are integrated. The sensor network employs a distributed deployment strategy, supporting the integration of heterogeneous data (such as automatic water quality monitoring stations, mobile drone monitoring, and remote sensing data) to form a multi-source data acquisition system. After preliminary processing by edge computing nodes, the collected data is transmitted in real-time to the cloud data center via wireless communication technologies (such as the NB-IoT wireless communication protocol and 5G networks) as input for subsequent model prediction and source tracing analysis. Specifically, it includes the following: S1.1 Deploy a multi-point distributed sensor node network in the river water body of the study area. The network includes multiple water quality sensing units, meteorological monitoring units, and communication modules.

[0042] Multiple water quality sensing units are used to monitor chemical oxygen demand (COD), ammonia nitrogen (NH3-N), total phosphorus (TP), conductivity (EC), and dissolved oxygen (DO). The EC sensor has a range of 0-20 mS / cm and an accuracy of ±2%. Meteorological monitoring units are used to acquire local meteorological information, such as rainfall, wind speed / direction, temperature / humidity, etc. The communication module supports the wireless communication protocols NB-IoT or LoRa, enabling long-distance, low-power, real-time data transmission.

[0043] Sensor deployment should be optimized by considering factors such as water flow velocity, hydraulic node arrangement, and hydrological topography. Common methods include equidistant deployment, upstream, midstream, and downstream zone deployment, or deployment at key cross-sections, with typical spacing ranging from 500m to 1km. A sampling frequency of at least once every 5 minutes is recommended to meet the timeliness requirements for responding to pollution emergencies.

[0044] Before being uploaded to the cloud platform, the water quality and meteorological data collected in S1.2 must first undergo integrity and reliability verification. The system employs an anomaly detection method combining rule-based thresholds and time-series analysis to identify and process obvious outliers and missing values ​​in the data. For extreme outliers caused by factors such as equipment failure or communication interference, the system handles them through interpolation repair or removal. For data with drastic fluctuations that may be caused by sudden pollution events, the system introduces a pollution-sensing anomaly identification mechanism to prevent real pollution signals from being misjudged as invalid data during the preprocessing stage.

[0045] This anomaly detection mechanism includes the following: The system performs multi-parameter linkage analysis and calculates the rate of change of key water quality indicators (such as COD, NH3-N, TP) in real time. If a single indicator is detected to rise sharply in a short period of time (such as an increase in chemical oxygen demand exceeding 50 mg / L) and other related indicators rise simultaneously, it is preliminarily determined that there may be a sudden pollution outbreak. Spatial consistency is verified. If similar or related indicators in adjacent monitoring sections (within a range of about 500m upstream and downstream) also show abrupt changes within a similar time period, the confidence level of the pollution event judgment is further enhanced.

[0046] S1.3 transmits the processed data to the central gateway node via its built-in wireless communication module, and then uploads it to the cloud data management platform via the 5G network. The uploaded data adopts a standardized structure (such as JSON or CSV format), with fields including timestamp, device ID, latitude and longitude, water quality parameters (such as COD, NH3-N, TP, DO, etc.) and meteorological parameters (such as temperature, rainfall, etc.) to meet the needs of multi-dimensional data fusion.

[0047] Uploaded data is stored chronologically according to sampling time and indexed in the database based on site number, enabling efficient data retrieval and horizontal / vertical comparative analysis. To eliminate the influence of different dimensions between indicators, the system performs unified normalization preprocessing on various parameters. Specifically, this includes using the Min-Max algorithm to normalize the data to the [0,1] interval or using the Z-score to standardize the data, providing a unified input format for subsequent concentration prediction models based on Long Short-Term Memory (LSTM) networks and MCMC sampling algorithms. Simultaneously, all collected data is automatically labeled with status fields (such as "original," "interpolated," "repaired," etc.) for dynamic weighting in the source tracing algorithm, improving the robustness of pollution event identification. To ensure data security and continuity, the platform also incorporates an automatic backup mechanism and a time-series caching strategy, enabling breakpoint resumption after network interruption recovery, effectively avoiding data loss and information lag.

[0048] In this embodiment S2: Based on data such as river cross-section information, current water depth conditions, upstream and downstream conditions of the river (pollution source location, emission concentration, emission time), rainfall forecast, river friction coefficient, and infiltration coefficient, a pollution diffusion model is constructed to calculate the concentration distribution of pollutants at each monitoring point when given pollution source parameters, and output the theoretical value of pollutant concentration.

[0049] The pollution diffusion model is a numerical simulation tool based on a one-dimensional pollutant convection and diffusion equation. It discretizes river channel cross-sections into a one-dimensional grid through interpolation to solve for the spatiotemporal distribution of pollutant concentrations within the river channel, simulating the pollutant diffusion process. This model comprehensively considers the effects of water flow velocity field, water depth distribution, boundary conditions, and source-sink terms to ensure a high degree of simulation of the pollutant migration process. Its expression is as follows:

[0050] In the formula, C is the pollutant concentration, u is the longitudinal flow velocity, D is the longitudinal diffusion coefficient, k is the overall degradation coefficient, and t is the time.

[0051] In this embodiment S3: the pollution trend prediction model can predict the pollutant concentration and its changing trend for the next period based on the latest time-series data, and output predicted values ​​and uncertainty indicators, providing real-time adjustment signals as a dynamic sampling strategy for MCMC. Its main steps include: S3.1 Extracts multivariate historical time-series data from the sensor network, including pollutant concentrations such as COD, NH3-N, and TP, and related environmental factors such as water temperature, rainfall, and flow velocity. Normalizes each variable. To train the model and capture the dynamic evolution of pollutant concentrations, constructs an input vector from the multivariate observation data at each time step τ. Its expression is as follows:

[0052] in: These represent the concentrations of COD, NH3-N, and TP at time τ (unit: mg / L). : Water temperature at time τ; Rainfall at time τ; River flow at time τ; Input feature vector; Time step index.

[0053] The input vectors from T consecutive time steps are combined to form the input sequence. Its expression is as follows:

[0054] The corresponding output target is the concentration change rate at the next time step. Its expression is as follows:

[0055] A training sample sequence set is constructed using a sliding window mechanism, with the input being the environmental state and pollutant concentration sequences from the past T steps. The output is the rate of change of concentration at future times. .

[0056] S3.2 Design the Long Short-Term Memory (LSTM) network structure and construct a deep sequence model containing 1 to 2 layers of LSTM. The typical structure includes an input layer, a hidden layer, and a fully connected output layer.

[0057] The input layer has dimensions [T, n], where n is the number of features, including pollutant concentrations such as COD, NH4-N, and TP, as well as environmental factors such as water temperature, rainfall, and flow rate. The hidden layers of the Long Short-Term Memory (LSTM) network contain several hidden units (e.g., 64–128), which capture long-term dependencies in the time series through forget gates, input gates, and output gates, learning the dynamic evolution of pollutant concentrations with changes in environmental factors.

[0058] The fully connected output layer is used to receive the hidden state output of the Long Short-Term Memory (LSTM) network and predict the rate of change of pollutant concentration at the next time step. .

[0059] The completed Long Short-Term Memory (LSTM) network model receives real multivariate input sequences, updates historical data through a sliding window mechanism, and outputs the future pollutant concentration change trend in real time, providing adjustment signals for MCMC dynamic sampling.

[0060] In this embodiment S4: A pollution source tracing model is constructed based on pollution source parameter vectors, including spatial location, emission start time, and emission intensity variation function over time. Unlike traditional fixed priors, this model utilizes neural networks to dynamically learn multimodal features, combining historical observation data and Long Short-Term Memory (LSTM) networks to predict future pollutant concentration change rates. This generates a time-varying prior distribution of pollution source parameters, thereby enabling dynamic adjustment of the prior distribution. During the MCMC sampling process, pollution source parameters are randomly sampled in each iteration, driving the pollution diffusion model to simulate the pollutant concentration sequence.

[0061] The simulated concentration change rate is then compared with the future concentration change rate predicted by the Long Short-Term Memory (LSTM) network to construct a likelihood function and guide the sampling strategy, ensuring that the exploration of the parameter space prioritizes matching the dynamic evolution trend of pollutant concentration. As the iteration proceeds, the sampling gradually converges to the posterior distribution, ultimately outputting the spatial location, emission time, and intensity curves of the pollution source.

[0062] By incorporating a Long Short-Term Memory (LSTM) network to predict future concentration change rates, the model can adjust the MCMC sampling strategy in real time, achieving dynamic optimization and trend consistency fitting of pollution source location, thus improving source tracing accuracy and response speed. Specifically, it includes the following steps: S4.1: Parameter space definition and dynamic prior construction, defining the pollution source parameter vector to be inverted as:

[0063] In the formula, , Geographic coordinates of the pollution source; The time when the pollution began; This is a time-varying function of the emission intensity of the pollution source.

[0064]

[0065] In the formula, 'a' represents the emission intensity, indicating the pollution source's... Maximum emissions at any given time; t is the current time; The time when the pollution began; The standard deviation of emission duration controls the "width" of the emission peak; the larger the value, the longer the emission duration.

[0066] S4.2: Constructing the time-varying prior distribution of pollution source parameters based on multimodal data and future concentration change rates. The input features include not only factory distribution F and rainfall intensity Historical accident frequency (H), topographic factor (G), etc., were also incorporated into the prediction of future pollutant concentration changes using a Long Short-Term Memory (LSTM) network. The specific steps are as follows: Step 1, construct the input features, which include the following: Multimodal features are extracted from historical environmental data, including factory distribution (F), rainfall intensity (R), historical accident frequency (H), and topographic factors (G); the future pollutant concentration change rate generated by the prediction module is obtained. All features are normalized or standardized to encode spatial information into a processable format; environmental features at continuous time steps are combined with future concentration change rates to form a multidimensional input vector.

[0067] Step 2: Perform pattern mapping on the input features to form a time-varying prior distribution, which includes the following: The input vector is processed hierarchically and its features are fused using the pre-dated fir tree mechanism to extract the correlation patterns between multimodal features. The fused internal representation is mapped to the probability space of pollution source parameters to generate the probability distribution of the pollution source location, emission start time, and emission intensity as a function of time. The mapping results can be updated in real time with time and environmental changes to form a time-varying prior distribution.

[0068] Step 3 involves training and optimizing the historical pollution event sample set, which includes the following: The system is trained using a historical pollution event sample set, with each sample containing environmental features, future concentration change rates, and actual pollution source parameters. The pre-projection fir tree mechanism adjusts its internal mapping parameters iteratively to make the generated probability distribution as close as possible to the actual inverted distribution of pollution source parameters. After training, the mechanism can receive the input features at the current moment in real time and output the time-varying prior distribution of pollution source parameters, providing intelligent guidance for MCMC sampling.

[0069] In this embodiment S5: a dynamic coupling mechanism between the Long Short-Term Memory Network (LSTM) and the MCMC is constructed. Based on the prediction ability of the LSTM for the rate of change of pollutant concentration, the future concentration change rate output by the LSTM module is introduced into the MCMC sampling process to form a dynamic weight and step size adjustment mechanism, thereby achieving an adaptive balance between sampling speed and accuracy.

[0070] In the pollution source tracing model, each pollution source parameter (including location, emission start time, emission intensity, etc.) is introduced with an independent dynamic weighting factor. This is used to adjust the search step size during the MCMC sampling process. The dynamic weights are calculated from the predicted future concentration change rate output by the Long Short-Term Memory (LSTM) network module. The calculation formula is as follows:

[0071] In the formula, This is the smoothing coefficient, usually set to 0.8. 0.95; The weights from the previous iteration; This is a mapping function from the rate of change to the weight, used to reflect the correspondence between the intensity of concentration change and the sampling sensitivity. It is often taken in the form shown below:

[0072] In the formula, k is the suppression coefficient, used to control the range of weight variation. When the concentration changes rapidly, the weight increases; when the change is gradual, the weight decreases, thus achieving dynamic adjustment of sampling sensitivity.

[0073] Based on dynamic weights, parameter sampling step size The calculation formula is as follows:

[0074] in, This is the base step size for this parameter. Weight By directly controlling the sampling amplitude of each parameter, adaptive regulation of the high-dimensional parameter space can be achieved.

[0075] When the rate of change of future pollutant concentration predicted by the Long Short-Term Memory (LSTM) network is large, the system determines that the pollution evolution is in a stage of drastic change, increases the dynamic weight, and accelerates the corresponding sampling step size to improve the exploration speed of the parameter space. When the rate of change of concentration is small, the weight decreases, the step size is shortened, the sampling is more refined, and the stability and inversion accuracy of the results are improved.

[0076] In this embodiment S6: Based on the high spatiotemporal resolution capability of the pollution diffusion model, after generating candidate pollution source parameters in each round of MCMC sampling, the diffusion model is invoked to simulate the spatiotemporal concentration distribution of pollutants and generate simulated pollutant concentration values; these are compared with observed values ​​and used as input for the acceptance criterion. This includes the following steps: S6.1: Convert each set of pollution source parameter vectors Inputting the pollution diffusion model yields a sequence of simulated pollutant concentrations at each monitoring section. The calculation formula is as follows:

[0077] In the formula, i represents the i-th sampling or simulation iteration; n represents the predicted concentration value at the n-th monitoring point; and E represents the environmental input variables (flow velocity field, water temperature, rainfall, etc.). This is a calculation function for a numerical model of pollution diffusion.

[0078] S6.2: Construct the likelihood function corresponding to the pollution diffusion model, and convert each set of pollution source parameter vectors... Input the pollution diffusion model to obtain a set of predicted concentration values. A likelihood function is constructed by comparing the pollutant concentration observations (Data) collected from actual monitoring points. The calculation formula is as follows:

[0079] In the formula: Data is the set of actual pollutant concentration observations collected at all monitoring points; Let J be the observed concentration at the j-th monitoring point; These are the simulated values ​​from the model; The standard deviation of the observation error; The confidence weight for the j-th monitoring point can be assigned based on the historical stability or importance of the monitoring point.

[0080] S6.3: Using the Metropolis acceptance criterion, calculate the acceptance probability of candidate sample X′. The calculation formula is as follows:

[0081] Where P(X) is a time-varying prior distribution (generated by the pre-fir tree mechanism), if random numbers When the time comes, the candidate sample X′ is accepted as the new state. Otherwise, maintain the current state. This represents the distribution of a continuous random variable that takes values ​​with equal probability in the range of 0 to 1; Candidate samples The probability value under a time-varying prior distribution; For the current state sample The probability value under a time-varying prior distribution.

[0082] S6.4: Dynamic weight feedback and step size adjustment. After each sampling, the pollutant concentration change rate prediction is based on the latest Long Short-Term Memory (LSTM) network model output. Update the dynamic weights for the current round. With parameter step size The calculation formula is as follows:

[0083] Through real-time feedback adjustment, the sampling mechanism can quickly respond to the trend of pollutant concentration changes. That is, when the predicted rate of change is large, the sampling step size is automatically widened to speed up the search; when the rate of change stabilizes, the step size is automatically narrowed to improve the local search accuracy.

[0084] S6.5: Iteration and Convergence Determination. Repeat the above sampling steps and evaluation process until the preset convergence condition of posterior variance stability is met, ultimately obtaining the posterior sample set of pollution source parameters. This sample set contains key parameters such as the geographical coordinates of the pollution source, start time, emission intensity, and emission duration, as well as corresponding simulated concentration values, likelihood values, and acceptance probabilities. By performing statistical analysis on the posterior sample set, the probability distribution characteristics of the pollution source parameters can be obtained, thereby achieving high-confidence inversion and uncertainty quantification of the spatial location, emission history, and intensity of pollution sources.

[0085] This function is used to determine whether the candidate parameters generated in each sampling are "reasonable". The parameter group is only used if the error of the current sampling result is less than a preset threshold. Retain the selected parameters; otherwise, discard them or continue iterating. Through the accumulation of a large number of valid samples, the retained parameter set will gradually approximate the spatial location and start-time characteristics of the actual pollution source, ultimately outputting a high-confidence pollution source inversion result. By introducing the above sampling and screening mechanism, the robustness and accuracy of the model in identifying pollution sources under complex pollution diffusion conditions can be effectively improved.

[0086] In this embodiment S7: After completing multiple rounds of MCMC sampling and meeting the convergence evaluation conditions, the system uses the posterior sample set of the sampled pollution source parameters. Each sample is used to construct the spatial probability distribution map of the pollution source. Specifically, this includes the following steps: S7.1: Extract all pollution source location parameters from the MCMC sampling results ,in, The coordinates of the pollution source on the main axis of the river (unit: meters, cumulative distance from the starting point along the river) are represented by N, which is the number of samplings.

[0087] S7.2: One-dimensional kernel density estimation is used to perform smoothing statistics on the pollution source location samples. Each MCMC sampling location xi is regarded as a sample point, and its corresponding confidence weight is used. As weighting coefficients, the samples are smoothed using a Gaussian kernel function.

[0088] By summing the weighted averages of all sampling points, the posterior probability density function of the pollution source at any location x along the river's length can be obtained. Its expression is as follows:

[0089] Where: N: number of valid samples; : The location of the i-th sampling point (coordinates along the river channel); : The confidence weight of this sample; : Bandwidth parameter, used to control the smoothness; The Gaussian kernel function is defined as follows:

[0090] Where z is the distance between the standardized sample point and the location to be estimated.

[0091] The bandwidth parameter h can be adaptively determined based on the sample standard deviation σ and the sample size N using an empirical formula, the expression of which is as follows:

[0092] To ensure that the density estimation results accurately reflect the reliability of the pollution source location samples, a weighting mechanism based on multi-factor fusion is adopted to determine the weighting coefficient for each sample point. .

[0093] For the i-th sample, its initial weights Based on the performance indicators of the MCMC sampling phase, the calculation formula is as follows:

[0094] In the formula: The probability of a sample being accepted during the MCMC sampling process; The likelihood function value corresponding to this sample is used to characterize the degree of fit between the model simulation value and the measured concentration. The confidence coefficient of the monitoring point associated with the sample can be determined based on the historical stability of the monitoring point or the integrity of the data. : Weighting adjustment coefficient, used to balance the influence of various factors, usually with a value range of 0.3 to 0.5.

[0095] The initial weights are normalized to obtain the final weight values. The calculation formula is as follows:

[0096] The normalized weights satisfy .

[0097] S7.3 determines the location of the most likely pollution source and the confidence interval, through... Function to calculate maximum a posteriori position Its expression is as follows:

[0098] in, A function is used to find the value of the independent variable that makes a given expression or function reach its maximum value.

[0099] Calculate the 95% confidence interval based on the kernel density curve P(x), i.e., find the interval [xL, xH], which must satisfy the following integration condition:

[0100] S7.4: Generate a posterior probability curve for the pollution source, with the horizontal axis representing the river location (m) and the vertical axis representing the posterior probability density. Mark the maximum posterior position. And the error range; it can be superimposed on the digital twin river section model to show the feedback consistency between the pollution propagation simulation results and the pollution source identification results.

[0101] This invention forms a feedback loop by tracing, simulating, and inverting pollution sources, enabling high-precision dynamic source tracing analysis of sudden water pollution events. Therefore, it is applicable to the construction of intelligent water environment monitoring and emergency source tracing system platforms.

[0102] This invention fully integrates the topography, hydrological and hydraulic boundaries, river cross-sectional structure, and historical pollution event data of the target river section to construct a digital twin model of pollutant diffusion that is highly consistent with the real water environment. The model supports simulation of multi-factor transport processes of pollutants, can access online monitoring water quality data and meteorological boundary conditions in real time, and performs pollution diffusion simulation. It serves as the physical support base for the pollution source inversion module, improving the overall inference accuracy and scenario adaptability.

[0103] This invention constructs a dynamic weighted adaptive MCMC sampling mechanism: it innovatively introduces a dynamic sampling strategy driven by the concentration change rate, using a Long Short-Term Memory (LSTM) network to predict pollutant concentration changes in real time, driving the dynamic adjustment of the sampling step size and acceptance probability of the MCMC model. This improves sampling accuracy during periods of drastic pollution fluctuations and accelerates convergence speed during stable periods, effectively enhancing parameter inversion efficiency and result stability, overcoming the problems of slow convergence and inaccurate localization in traditional MCMC algorithms under sudden pollution scenarios.

[0104] This invention, based on pollution source parameter inversion results and diffusion simulation feedback, iteratively corrects the prior probability distribution of pollution sources, forming a closed-loop feedback chain of "data-simulation-source tracing". A dynamic prior model is constructed using the MCMC algorithm, enhancing the ability to integrate and utilize multi-source heterogeneous information, including factory discharge records, historical accidents, and high-frequency water quality changes. Finally, it outputs a spatial distribution probability map of pollution sources, start time, and confidence intervals for emission intensity, achieving intelligent decision support that enables visualization, quantification, and traceability of pollution events.

[0105] The fourth specific embodiment of the water pollutant source tracing method coupled with pollution trends of the present invention: A method for tracing the source of water pollutants coupled with pollution trends includes an inflow forecasting module, a water demand prediction module, a water resource allocation module, an intelligent optimization module, a water transfer scheduling module, and a results display module. (Refer to...) Figure 3 The specific steps are as follows: Step 1: Deploy multi-parameter water quality sensors and meteorological data acquisition equipment in the target river section to collect real-time pollutant indicators such as COD, ammonia nitrogen, and heavy metals, as well as meteorological data such as rainfall, wind speed, and temperature. Perform anomaly detection and missing value repair on the raw data, combining rule-based thresholds and multi-parameter linkage to determine the type of anomaly and avoid mishandling pollution mutations. Simultaneously, perform normalization standardization processing on all types of data to construct a unified, standardized time-series dataset as input for subsequent modeling.

[0106] The second step involves constructing a pollution diffusion model based on data such as river cross-section information, current water depth conditions, upstream and downstream conditions (pollution source location, emission concentration, emission time), rainfall forecasts, river friction coefficient, and infiltration coefficient. The model supports inputting any set of pollution source parameters (including location, emission amount, duration, etc.) and outputs the pollutant concentration distribution at each monitoring cross-section, which supports sampling acceptance criteria and model inversion.

[0107] Step 3: Train a Long Short-Term Memory (LSTM) network based on historical monitoring data to establish a pollution trend prediction model based on the pollutant concentration change rate. This model supports online operation and can predict pollutant change trends for the next period based on the latest time-series data, providing real-time adjustment signals for the dynamic sampling strategy of MCMC.

[0108] Step 4: Construct an MCMC pollution source tracing model based on pollution source parameter vectors, including spatial location, emission start time, and emission intensity variation function over time. The prior distribution is no longer presented in a fixed or manually input form, but rather through a neural network model. Time-varying prior distribution of pollution source parameters generated by dynamic learning of multimodal features Based on this, parameter space sampling is performed using the MCMC algorithm to fit the pollution diffusion model simulation results with actual observation data, gradually converging to the posterior distribution.

[0109] Step 5: Construct a dynamic coupling mechanism between the Long Short-Term Memory (LSTM) network and MCMC. Based on the LSTM network's ability to predict pollutant concentration change rates, the concentration change rate output by the LSTM module is introduced into the MCMC sampling process, constructing a dynamic weight and step size adjustment mechanism. The predicted value output by the LSTM network... Dynamic weighting factor for dynamically adjusting MCMC This allows for the control of the sampling step size, achieving an adaptive balance between sampling speed and accuracy.

[0110] Step 6: Based on the high spatiotemporal resolution capability of the pollution diffusion model, after generating candidate pollution source parameters in each round of MCMC sampling, the diffusion model is called to simulate the spatiotemporal concentration distribution of pollutants and generate simulated pollutant concentration values; these are compared with observed values ​​and used as input for the acceptance criteria.

[0111] Step 7: Based on a large number of valid sampling results, construct the posterior distribution of pollution source location parameters and generate a heat map of the spatial probability distribution of pollution sources to assist decision-makers in quickly locating high-risk emission sources. Finally, output the most probable geographical location of the pollution source, the start time of emission, the emission intensity, and its confidence interval, providing a scientific and intuitive basis for tracing the source in response to sudden water pollution events.

[0112] In the first step of this embodiment, the methods for constructing the dataset and preprocessing the dataset are as follows: A comprehensive and fast-response water quality monitoring sensor network was constructed for a certain lake and its main inflow rivers, forming a multi-point distributed sensing system. The system's sensor deployment was optimized according to the actual hydrological structure and pollution propagation characteristics. Sensor nodes were mainly deployed at key cross-sections in the upper, middle, and lower reaches, at lake inlets, tributary confluences, and points of potential pollution risk, with an average spacing of 500m to 1km. The deployed sensor units include water quality monitoring modules, meteorological sensing modules, and communication modules.

[0113] The water quality monitoring module is used to collect key indicators in real time, such as chemical oxygen demand (COD), ammonia nitrogen (NH3-N), total phosphorus (TP), conductivity (EC, range 0-20mS / cm, accuracy ±2%), and dissolved oxygen (DO).

[0114] The meteorological sensing module is used to collect meteorological factors such as rainfall, wind speed / direction, temperature, and humidity within the area.

[0115] The communication module can be an NB-IoT or LoRa communication unit to enable communication between nodes, support long-distance low-power transmission, and achieve stable data reporting.

[0116] Data collected at the sensor end The expression is as follows: , in, Number the nodes. To Within the time node The values ​​of these attributes are specifically chemical oxygen demand (COD), ammonia nitrogen (NH3-N), total phosphorus (TP), and total nitrogen (TN), etc. .

[0117] In this embodiment, the collected data is processed as follows: Based on a strategy combining rule-based thresholds and time-series change analysis, extreme values, jumps, and missing data in the collected data are identified. For sudden pollution events, multi-parameter linkage judgment rules are set. For example, if the chemical oxygen demand (COD) change rate exceeds 50 mg / L, and indicators such as ammonia (NH3), nitrogen (N), and total phosphorus (TP) rise significantly and synchronously, exhibiting spatially linked abrupt changes across adjacent sections, it is determined to be a real pollution event to prevent false deletion. For abnormal data caused by equipment fluctuations or signal interference, interpolation completion, time smoothing, or direct removal are used for correction, and data status labels (such as "original," "interpolated," "repaired," etc.) are added for subsequent model weighting.

[0118] After local preprocessing, the data is uploaded to the cloud platform in real time via a 5G network. The transmitted data uses a standardized format (such as JSON structure) and includes fields such as timestamps, device numbers, geographic coordinates, and water quality and meteorological parameter values. The system performs unified archiving and indexing of the data in the cloud, building an index tree by site number and sampling time to enable rapid retrieval and regional comparative analysis. To ensure consistency of input for various water quality and meteorological indicators during the modeling phase, the platform uniformly adopts a normalization processing strategy, supporting minimum and maximum normalization and Z-score processing methods, and retains copies of the original values ​​in the database to support source tracing and error analysis.

[0119] In addition, the platform has a built-in automatic backup mechanism and breakpoint resume strategy, which can automatically re-transmit data after network interruption and recovery, ensuring continuity and integrity, and providing stable, standardized, and high-quality data input for subsequent Long Short-Term Memory Network (LSTM) concentration prediction and MCMC pollution source tracing analysis.

[0120] In the second step of this embodiment, a one-dimensional pollution diffusion model based on the actual topography and hydrodynamic characteristics of the river channel is constructed to extrapolate the spatiotemporal distribution of pollutants in the water body. Based on cross-sectional mapping data of the river channel surrounding a lake, hydrological monitoring data, and candidate pollution source parameters, the model numerically solves for the migration and diffusion process of pollutants within the river channel, outputting the theoretical concentration values ​​of pollutants at each monitoring point at the corresponding time. Specifically, it includes the following: First, based on the river cross-section information collected in the field, the river morphology is interpolated to discretize the continuous river channel into one-dimensional equidistant grid cells, establishing the river channel computational domain. On this basis, current water depth data, river friction coefficient, infiltration coefficient, boundary flow velocity, rainfall intensity, and other environmental conditions are collected and incorporated as input parameters for pollution diffusion simulation.

[0121] The pollution diffusion model uses a one-dimensional convection diffusion equation as its basic mathematical expression, comprehensively considering the transport and diffusion mechanisms of pollutants with water flow. The model equations are as follows:

[0122] In the formula, C is the pollutant concentration, u is the longitudinal flow velocity, D is the longitudinal diffusion coefficient, and k is the overall degradation coefficient.

[0123] The model boundary conditions are dynamically set based on the upstream inflow velocity and downstream water level of a certain lake, and the initial conditions are set based on the background concentration to ensure that the simulation has realistic constraints.

[0124] In practical applications, the model can receive a set of given pollution source parameters, including the location of the pollution source. Emission concentration Emission start time By combining river hydrological boundaries and meteorological forecast information (such as rainfall intensity), the dynamic simulation of pollutant concentrations is completed within the current period, and the theoretical curves of concentration changes over time at each monitoring section are output.

[0125] This model, deeply integrated with the pollution source sampling algorithm (MCMC), rapidly simulates the pollution concentration distribution after each round of parameter sampling and compares it with measured data, providing a computational basis for the sampling acceptance mechanism. Through dynamic grid adjustment and high-precision numerical difference methods, the model achieves efficient and accurate reproduction of the pollutant migration process, serving as a key physical simulation support module for the entire source tracing inference chain.

[0126] In the third step of this embodiment, a pollution change trend prediction model is constructed based on a Long Short-Term Memory (LSTM) network to predict the rate of change of pollutant concentration. The model first constructs normalized time-series training samples based on historical pollutant concentration data (including COD, NH3-N, TP, etc.) collected by a sensor network and related environmental factors (such as water temperature, rainfall, flow velocity, etc.), and generates model input-output pairs using a sliding window method.

[0127] Subsequently, a deep neural network structure containing one to two layers of Long Short-Term Memory (LSTM) network units was designed and trained using historical monitoring data to predict future pollutant concentrations or rates of change in concentration. The trained model was deployed on a cloud platform, capable of receiving real-time data input online and quickly outputting pollution trends over several future time steps, serving as a dynamic adjustment basis for MCMC sampling during pollution source parameter inversion. This model can adaptively adjust the sampling strategy according to the pollution change rate, effectively improving the timeliness and accuracy of pollution event response and enhancing the intelligence and robustness of the pollution source tracing system.

[0128] The Long Short-Term Memory (LSTM) network structure constructed in this embodiment is as follows: The input feature dimension is [12,6], where the number 12 indicates that the model receives the past 12 time steps each time (e.g., one data point every 5 minutes, which corresponds to the data in the past 1 hour); the number 6 indicates that the input variables include: chemical oxygen demand (COD), ammonia nitrogen (NH3-N), total phosphorus (TP), water temperature, rainfall, and water flow velocity. Long Short-Term Memory (LSTM) network hidden layers: contain several hidden units (e.g., 64-128); Fully connected output layer: Predicts the rate of change in pollutant concentration within the next time step (5 minutes). It is used to guide the MCMC dynamic sampling strategy.

[0129] The completed pollution trend prediction model receives real-time input data and outputs the future concentration change rate. .

[0130] In the fourth step of this embodiment, an MCMC pollution source tracing model based on pollution source parameter vectors is constructed, and a data-driven dynamic prior generation mechanism is introduced to enhance the model's adaptability to actual pollution scenarios and its inversion accuracy.

[0131] First, define the pollution source parameter vector to be inverted. It includes the following elements:

[0132] In the formula, , Geographic coordinates of the pollution source; The time when the pollution began; Let be a time-varying function of the emission intensity of the pollution source, and its expression is as follows:

[0133] In the formula, 'a' represents the emission intensity, indicating the pollution source's... Maximum emissions at any given time; t is the current time; The time when the pollution began; The standard deviation of emission duration controls the "width" of the emission peak; the larger the value, the longer the emission duration.

[0134] Secondly, a prior distribution step is dynamically constructed based on the multimodal external features of the pollution scenario. The specific process is as follows: Constructing a dynamic prior generative model Its expression is as follows:

[0135] Where: F represents the spatial distribution of factories / sewage discharge outlets within the region; : Meteorological data at the current time t, such as rainfall intensity; H: frequency of historical pollution incidents or sensitive areas; G: topographic and hydrological characteristics of the study area; The multilayer perceptron (MLP) built using the PyTorch framework contains 3 hidden layers, each with 64 neurons, and uses the ReLU activation function. It outputs the expectation and covariance of the pollution source parameter vector for subsequent sampling.

[0136] Training the prior model includes the following: Using samples from a historical pollution event database Compared with actual pollution source parameters Construct a supervised learning dataset and train a prior generative model. The optimization objective is to make the probability distribution of pollution source parameters predicted by the model as close as possible to the actual distribution corresponding to the actual pollution source parameters.

[0137] After the model is trained, it is used in actual operation based on the input features at the current time. Time-varying prior distribution of pollution source parameters , which serves as the initial distribution for MCMC sampling.

[0138] In this embodiment, the Metropolis-Hastings algorithm is used to implement Markov chain sampling in the parameter space: Candidate solutions Generated by symmetric Gaussian perturbation; the predicted concentration distribution corresponding to the current sampling solution is calculated using a pollution diffusion model; and the likelihood function is calculated based on the predicted concentration distribution and actual monitoring data; acceptance probability. The calculation formula is as follows:

[0139] Based on the calculated acceptance probability, a decision is made on whether to accept the current sample, thus forming a sample chain that converges to the posterior distribution.

[0140] The posterior sample set of pollution source parameters output by the model can be used to subsequently draw spatial probability heat maps and support emergency decision analysis.

[0141] In the fifth step of this embodiment, the dynamic coupling sampling process includes the following: Using the MCMC model and the Long Short-Term Memory (LSTM) network model initialized with prior distribution, the system receives real-time temporal characteristic inputs such as pollutant concentration, meteorological, and hydrological data from each monitoring point during the online operation phase, and predicts the pollutant concentration change rate in future periods, including ΔCOD and ΔNH3-N.

[0142] During MCMC sampling, the pollutant concentration change rate output by the Long Short-Term Memory (LSTM) network model is used as a dynamic control signal and embedded into the MCMC step size adjustment and acceptance probability control mechanism. The system introduces a dynamic weighting factor. The weighting factor is used to measure the temporal deviation between the predictions of the Long Short-Term Memory (LSTM) network and the actual observations. This factor is incorporated into the MCMC acceptance criterion to dynamically adjust the acceptance probability of candidate solutions in each round of sampling. When the LSTM network detects a significant rate of change in pollutant concentration, indicating that the current pollution process is in a phase of rapid change, the sampling step size of the MCMC is dynamically reduced, and the local sampling density is increased to improve the inversion accuracy in high-change areas. Conversely, when the LSTM network output indicates that pollutant changes are stable, the sampling pace is appropriately accelerated, and the step size is increased to improve computational efficiency. Through this dynamic coupling mechanism, the LSTM network's real-time control over the MCMC sampling behavior is achieved, enhancing the source tracing model's response capability to sudden pollution events and its inference efficiency in steady-state scenarios. The calculation formula is as follows:

[0143] In the formula, This is the smoothing coefficient, usually set to 0.8. 0.95; The predicted rate of change of future pollutant concentrations output by the Long Short-Term Memory (LSTM) network model; In the sixth step of this embodiment, a Bayesian inversion framework for pollution source parameters is constructed based on a high spatiotemporal resolution pollution diffusion model, and a Metropolis-Hastington algorithm sampling mechanism is embedded to achieve high-precision identification of parameters such as the spatial location, emission intensity, and start time of pollution sources. Specifically, the following steps are included: Step 1 involves simulating pollution diffusion and predicting concentrations, which includes the following: Each round of MCMC sampling generates a candidate pollution source parameter vector. Then, input the pollution diffusion model, and the corresponding output will be the simulated concentration sequence of pollutants at each monitoring section under the current parameters. The calculation formula is as follows:

[0144] In the formula This represents the pollutant concentration value at the nth monitoring point simulated by the parameters generated from the i-th sampling.

[0145] Step 2: Compare the simulated pollutant concentration values ​​with the actual observed concentrations. Comparison, constructing the log-likelihood function The calculation formula is as follows:

[0146] In the formula: The standard deviation of the observation error; For the first The confidence weight of each monitoring point (can be set based on sensor stability or historical accuracy). This represents the total number of observation sections.

[0147] Step 3: Using the likelihood function described above, calculate the acceptance probability of the candidate parameter set. The calculation formula is as follows:

[0148] In the formula: Candidate samples The probability value under the time-varying prior distribution at the current time t; Candidate samples The probability value under the time-varying prior distribution at the current time t; Candidate samples The likelihood value; Candidate samples The likelihood value.

[0149] Sample a random number u Uniform(0,1), if Then accept the candidate sample. Then proceed to the next iteration and update the state; otherwise, retain the current state.

[0150] Step 4: During the sampling process, the pollutant concentration change rate output by the Long Short-Term Memory (LSTM) network model is introduced. As a dynamic control signal, the sampling step size and acceptance criteria are dynamically adjusted according to the drastic changes in pollutants, including the following: like Reduce the step size, increase the judgment threshold, and enhance the local sampling density; If the pollution concentration changes steadily, the acceptance criteria should be relaxed appropriately, the step size increased, and the convergence speed accelerated.

[0151] Step 5: Continue executing the above sampling and screening process until the set convergence criteria are met (e.g., the change in posterior variance is less than a certain threshold). The convergence time is 30 minutes. Finally, a high-confidence posterior sample set of pollution source parameters is obtained, and the spatial location and emission intensity curve of the pollution source are output. With start time Wait for the inversion results.

[0152] In the seventh step of this embodiment, the results are output and visualized, which includes the following: After completing the pollution source inversion calculation (based on MCMC multi-round sampling) and satisfying the posterior distribution convergence condition, a posterior probability distribution model of the pollution source location parameters is constructed, generating a spatial distribution probability heatmap and visualization charts to assist decision-makers in quickly identifying the locations of high-risk emission sources. Key pollution source parameters and their confidence intervals are also output, providing visual support for emergency response to pollution incidents. The main steps are as follows: Step 1: Extract all pollution source location parameters from the sample set of MCMC sampling output by the pollution source tracing module to obtain the pollution source location parameter sample, the expression of which is as follows:

[0153] In the formula, This indicates the estimated location of the pollution source on the main axis of the river in the i-th sampling (with the river's starting point as 0, in meters). This represents the total number of valid samples.

[0154] Step 2, based on kernel density estimation, construct the spatial distribution of pollution sources, which includes the following: A one-dimensional kernel density estimation method is used to smooth the pollution source location samples and estimate the continuous distribution function P(x) of the "presence probability" of the pollution source in the river direction. This reflects the posterior probability distribution of pollution sources in the spatial dimension of the river channel, that is, the probability that a certain location is a real pollution source. The calculation formula is as follows:

[0155] Where: N is the number of samples; K is the Gaussian kernel function; h is the bandwidth parameter; i is the index of the i-th sample point; x is a point at a certain location along the length of the river channel; It is the location of the sample point.

[0156] Step 3, based on the posterior probability density function Extract the most likely pollution source location and confidence interval.

[0157] Location of most likely pollution source The formula for calculating the (maximum a posteriori estimate) is as follows:

[0158] Calculate the confidence interval, i.e. solve the interval It satisfies the following expression:

[0159] This indicates a 95% confidence level that the pollution source is located within this range.

[0160] Step 4: Visualize the results and overlay a map for display. This includes the following: The system generates and outputs the following visualization charts: Generate a spatial probability density curve of pollution sources. The horizontal axis of the curve represents the river channel length (x, in meters), and the vertical axis represents the posterior probability density. The curve represents the distribution of uncertainty in the location of the pollution source; and the most likely pollution source location and confidence interval are marked.

[0161] The spatial heat map of pollution sources is overlaid, that is, the kernel density result is overlaid as a probability heat map of pollution sources onto the river GIS model of a water environment digital twin system of a certain lake area, and displayed in conjunction with the pollution diffusion simulation animation to achieve consistency verification between pollution propagation path and pollution source location tracing.

[0162] The final system outputs pollution source inversion results, which include the geographical coordinates of the most likely pollution source and the start time of emission. Emission intensity function Parameter 95% confidence interval, spatial posterior distribution map and heat map.

[0163] This invention constructs a Bayesian inversion framework for the spatiotemporal parameters of pollution sources by integrating multi-source sensing data with a pollution diffusion mechanism model. It introduces a Long Short-Term Memory (LSTM) neural network to predict pollutant concentration trends and dynamically adjusts the MCMC sampling strategy to achieve an adaptive balance between sampling efficiency and inversion accuracy. By constructing a multimodal prior model and a pollution simulation feedback loop, the model's adaptability and stability to complex pollution situations are enhanced. This method can output a posterior probability distribution map of pollution sources and confidence intervals for key parameters, providing visual and quantitative support for environmental emergency decision-making. It significantly improves the response speed, diagnostic accuracy, and control effectiveness in water pollution events, effectively addressing the problems of difficult pollution source location, low spatiotemporal inversion accuracy, and model response lag in sudden water pollution events.

[0164] A server embodiment applying the method of the present invention: A server comprising: One or more processing units; Storage device for storing one or more programs; When the one or more programs are executed by the one or more processing units, the one or more processing units implement the above-described method for tracing the source of water pollutants coupled with pollution trends.

[0165] The storage device can be internal memory, external memory, cache memory, or other special memory. The processing unit has signal processing capabilities and can be a general-purpose processor, a digital signal processor, an application-specific integrated circuit, an off-the-shelf programmable gate array, or other programmable logic device.

[0166] An embodiment of a device applying the method of the present invention: An electronic device is provided with a computer-readable storage medium on which a computer program is stored. When the program is executed by a processing unit, it implements the above-described method for tracing the source of water pollutants coupled with pollution trends.

[0167] Computer-readable storage media refers to physical carriers capable of storing computer-recognizable data, instructions, or programs. These media must meet the core characteristic of being "readable by a computer" (i.e., the data exists in the form of electrical, magnetic, or optical signals and can be converted into binary information that a computer can process through appropriate devices). The physical carrier can be a magnetic storage medium, optical storage medium, semiconductor storage medium, or other storage media.

[0168] Those skilled in the art will understand that the embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects.

[0169] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features within the scope of the technology disclosed in the present invention; and these modifications or substitutions will not cause the substance of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. Any modifications or equivalent substitutions that do not deviate from the spirit and scope of the present invention should be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for tracing the source of water pollutants coupled with pollution trends, characterized in that: Includes the following: Based on the topographic and hydrological data of a certain body of water, construct a digital twin object of the body of water; By using pollution monitoring data and historical pollution data, the pollution change trend in digital twin objects of water bodies is predicted, and the pollutant concentration change rate is obtained. Based on the pollutant concentration change rate and historical pollution data, a time-varying prior distribution of pollution source parameters is generated to screen the pollution source search area; at the same time, based on the adjustment requirements of sampling sensitivity, an inhibition coefficient is set. Based on the correspondence between the inhibition coefficient, the intensity of concentration change, and the sampling sensitivity, a mapping function from the pollutant concentration change rate to the weight is established; Using mapping functions and smoothing coefficients, a dynamic weighting factor is calculated to quantify the impact of pollutant concentration change rate on the sampling strategy; The sampling step size of the pollution source parameters is determined based on the dynamic weighting factor and the basic step size. Based on the sampling step size, a search is performed in the time-varying prior distribution to obtain pollutant source information, thereby enabling the tracing of pollutants in water bodies.

2. The method for tracing the source of water pollutants coupled with pollution trends as described in claim 1, characterized in that: The method for constructing a digital twin object of a water body based on its topographic and hydrological data is as follows: Acquire topographic and hydrological data for a specific body of water. Topographic data includes the cross-sectional structure of the river channel, shoreline morphology, location of tributary confluences, and distribution of obstacles. Hydrological data includes longitudinal flow velocity, water depth, discharge, water level changes, and hydrodynamic parameters. Based on topographic data, the water area is discretized into a one-dimensional or two-dimensional grid, and depth and slope are associated on each grid cell; Hydrological data is incorporated into a one-dimensional or two-dimensional grid to obtain a digital twin object of the water area.

3. The method for tracing the source of water pollutants coupled with pollution trends as described in claim 1, characterized in that: The following method is used to predict pollution change trends in digital twin objects of water bodies and obtain pollutant concentration change rates by utilizing pollution monitoring data and historical pollution data: Obtain environmental impact factors, including water temperature, rainfall, river flow, wind speed, and wind direction; A distributed water quality sensor network will be used to cover key monitoring sections in the upper, middle and lower reaches of the river, and historical pollution data, including chemical oxygen demand, ammonia content, nitrogen content and total phosphorus content, will be collected through the distributed water quality sensor network. Environmental impact factors and historical pollution data are normalized to obtain a multivariate observation mapping sequence; at the same time, several prediction time steps are set according to the source tracing requirements. The multivariate observation mapping sequence corresponding to a certain prediction time step is input into a pre-built pollution trend prediction model to predict the changing trend of pollutants and obtain the corresponding pollutant concentration change rate.

4. The method for tracing the source of water pollutants coupled with pollution trends as described in claim 3, characterized in that: The method for constructing a pollution trend prediction model is as follows: Receive the standardized multivariate observation mapping sequence, which includes n pollution indicators and m environmental factors; Set up a gating mechanism, which includes a forget gate, an input gate, and an output gate; The forget gate is used to delete normal concentration data, the input gate is used to update the current environmental factors, and the output gate is used to control the contribution of the hidden state to the output, so that the pollution trend prediction model focuses on the core factors affecting the change of pollutant concentration. Based on the gating mechanism, a pollution trend prediction model is established using the hidden layers in the long short-term memory network; The pollution trend prediction model can learn the long-period dependence in the multivariate observation mapping sequence and map the output of the hidden layer to a single predicted value, namely the rate of change of pollutant concentration at a certain time step. The long-term dependence includes the lag variation of pollution concentration with water temperature, rainfall, river flow, wind speed, and wind direction. Then, the pollution trend prediction model is trained and optimized using historical pollution data, thereby completing the construction of the pollution trend prediction model.

5. The method for tracing the source of water pollutants coupled with pollution trends as described in claim 4, characterized in that: The method for training and optimizing pollution trend prediction models using historical pollution data is as follows: Several sets of input and output samples were extracted from historical pollution data, covering different pollution scenarios and different hydrological conditions. The mean squared error formula is used as the loss function to measure the deviation between the predicted pollutant concentration change rate and the actual pollutant concentration change rate. The adaptive learning rate function is used as the optimizer, and the number of iterations is adjusted according to the convergence. The input and output samples are divided into training and validation sets in a 4:1 ratio. The training set is input into the pollution trend prediction model to train the pollution trend prediction model and obtain the prediction results. The prediction results are compared with the validation set to obtain the prediction error; By adjusting the parameters of the pollution trend prediction model based on the prediction error, the pollution trend prediction model can be optimized.

6. The method for tracing the source of water pollutants coupled with pollution trends as described in claim 1, characterized in that: The method for generating the time-varying prior distribution of pollution source parameters based on pollutant concentration change rate and historical pollution data is as follows: Step 1: Based on historical pollution data, construct a pollution source parameter space, which includes the geographical coordinates of the pollution source, the pollution emission start time, and the time variation function of emission intensity. Step 2: Based on the pollution source parameter space, set static and dynamic features; Static characteristics include the spatial distribution of sewage outlets, river topography, and the frequency of historical pollution incidents; dynamic characteristics include real-time rainfall and the rate of change in pollutant concentrations. Step 3: Perform hierarchical processing and feature fusion on static and dynamic features, and extract the correlation patterns between static and dynamic features to obtain the fused multimodal features; Step 4: Map the fused multimodal features to the probability space of the pollution source parameters to generate a probability mapping result, which is a probability distribution of the pollution source location, emission start time, and emission intensity as a function of time. Step 5: Update the probability mapping results in real time according to changes in time and environment to form a time-varying prior distribution, so as to screen the search area for pollution sources.

7. The method for tracing the source of water pollutants coupled with pollution trends as described in claim 1, characterized in that: The method for obtaining pollutant source information by searching within a time-varying prior distribution based on the sampling step size is as follows: In time-varying prior distributions, candidate parameters are generated through Gaussian perturbation; Based on the sampling step size, sample from the candidate parameters to obtain candidate pollution source parameters; By inputting the parameters of the candidate pollution sources into a pre-built pollution diffusion model, new simulated values ​​of pollutant concentrations are obtained. Based on the new simulated pollutant concentrations, a log-likelihood function is constructed. The acceptance probability of candidate pollution source parameters is calculated using the log-likelihood function. Based on the acceptance probability, the sampling and screening process is continuously executed until the set convergence criteria are reached, thereby obtaining a high-confidence posterior sample set of pollution source parameters and outputting pollutant source information, which includes at least the spatial location of the pollution source, emission intensity curve, and start time.

8. The method for tracing the source of water pollutants coupled with pollution trends as described in claim 7, characterized in that: The method for constructing a pollution diffusion model is as follows: Based on the digital twin object of the water area, boundary conditions are constructed, including the upstream boundary, the downstream boundary, and the shoreline; The upstream boundary is set as the inflow boundary, the downstream boundary is the outflow boundary, and the shoreline is set as the no-flux boundary. The boundary conditions and candidate pollution source parameters are used as input source terms for pollution diffusion. Based on the input source terms, the grid cells in the digital twin object of the water area are set as pollution release sources; Calculate the pollutant release amount of each grid cell at different time steps based on emission intensity; The finite difference method or finite volume method is used to solve the one-dimensional convection-diffusion equation, so as to transform the continuous differential equation into a discrete system of algebraic equations. Based on the algebraic equations and pollutant release, the pollution diffusion data of pollutants between grid cells is calculated in each time step. Based on pollution diffusion data, the pollutant concentration of each grid cell at different time steps is calculated iteratively, and the simulated pollutant concentration values ​​of each monitoring section are finally output, thereby realizing the construction of the pollution diffusion model.

9. A water pollutant source tracing system coupled with pollution trends, characterized in that: This includes water area simulation models, pollution diffusion models, pollution trend prediction models, and pollution source tracing models; A water area simulation model is used to construct a digital twin object of a water area based on its topographic and hydrological data. A pollution diffusion model is used to process digital twin objects of water bodies to obtain simulated pollutant concentration values ​​at multiple monitoring sections in order to simulate the pollutant diffusion process. Pollution trend prediction models are used to predict the changing trends of pollutants and obtain the rate of change of pollutant concentrations. A pollution source tracing model is used to generate time-varying prior distributions and candidate pollution source parameters based on the pollutant concentration change rate, and to set suppression coefficients based on the adjustment requirements of sampling sensitivity. Based on the correspondence between the inhibition coefficient, the intensity of concentration change, and the sampling sensitivity, a mapping function from the pollutant concentration change rate to the weight is established. Using the mapping function and the smoothing coefficient, a dynamic weighting factor is calculated to quantify the impact of the pollutant concentration change rate on the sampling strategy. Based on the dynamic weighting factor and the basic step size, the sampling step size of the pollution source parameters is determined. In the process, the pollution diffusion model is combined with inversion to form a feedback loop, thereby obtaining pollutant source information and realizing the source tracing of pollutants in water bodies.

Images