Water pollution source tracking and atmospheric particulate distribution analysis method based on machine learning

Abstract

The invention discloses a water pollution source tracking and atmospheric particulate distribution analysis method based on machine learning, comprising the following steps: S1, collecting and preprocessing multi-source data to generate an environment data set; s2, pollution diffusion characteristics are extracted, pollution sources are reasoned, and a water pollution tracking result is generated; s3, constructing a deep Bayesian neural network model, and outputting particulate matter concentration prediction and a confidence interval; s4, introducing a zebra optimization algorithm, and jointly optimizing a model structure and hyper-parameters; s5, fusing the water pollution tracking result and the environmental characteristics, and outputting a prediction result; s6, collecting real-time monitoring data, calculating a prediction residual error and feeding back the prediction residual error to the optimization model; and S7, generating a pollution thermodynamic diagram and a trend prediction diagram, and outputting a pollution level and an early warning suggestion. According to the invention, by fusing water pollution source tracking and atmospheric particulate distribution modeling, intelligent linkage analysis of pollution propagation path identification and air quality accurate prediction is realized.

Description

Technical Field

[0001] The present invention relates to the field of machine learning technology, and in particular to a method for tracing water pollution sources and analyzing atmospheric particulate matter distribution based on machine learning. Background Art

[0002] In recent years, with the rapid development of industrialization and urbanization, environmental pollution has become increasingly serious. In particular, the co-evolution of water pollution and atmospheric particulate matter pollution has had a compounding impact on ecosystems and public health. Studies have shown that after water pollution incidents, certain regions experience elevated concentrations of atmospheric particulate matter. This phenomenon is particularly pronounced in industrial clusters and urban river networks near large water bodies. Traditionally, water pollution and air pollution have been viewed as separate environmental issues. However, a growing body of environmental monitoring data and empirical research indicates that coupled migration mechanisms, driven by physical diffusion, evaporation and volatilization, and meteorological factors, exist between the two. Therefore, tracing water pollution sources and predicting their impact on the distribution of atmospheric particulate matter are crucial for improving comprehensive environmental governance and achieving precise pollution warning and control.

[0003] In existing technologies, water pollution source tracing methods primarily include physical simulation methods based on traceability models and reverse reasoning methods based on sensor data. The former, such as the Advance–Dispersion model and hydrodynamic simulation models, emphasize the physical diffusion of pollutants in water bodies and have good theoretical support. However, these models are highly dependent on parameters and require strict hydrological boundary conditions, making them difficult to adapt to the complexities of actual hydrology. The latter, based primarily on concentration data from a limited number of monitoring points, estimates the location and intensity of pollution sources through inversion algorithms. Representative methods include least squares estimation, Bayesian inversion, and Markov chain Monte Carlo methods. However, these methods suffer from high computational complexity, sensitivity to abnormal monitoring values, and a general lack of dynamic modeling capabilities for pollution diffusion processes.

[0004] In terms of atmospheric particulate matter distribution prediction, existing technologies mostly use statistical models based on time series (such as ARIMA, SVR, etc.) or data-driven deep learning models (such as LSTM, CNN, Transformer, etc.) to predict regional pollution concentrations. These models can achieve a certain degree of short-term prediction, but their results usually rely on the historical concentration of the atmosphere itself, meteorological factors, etc., and lack the ability to model external pollution sources, especially inputs from water pollution events. In addition, atmospheric pollution prediction models usually only focus on a single target (such as PM2.5 concentration), ignoring the propagation path and spatiotemporal evolution characteristics of pollution sources. Therefore, it is difficult to effectively explain the causes of pollution events, and it is not conducive to policy response and the deployment of governance measures.

[0005] In recent years, with the rapid development of machine learning technology, a growing number of studies have attempted to incorporate neural networks, graphical models, and optimization algorithms into environmental modeling tasks. For example, some papers use graph neural networks to construct watershed network relationships and model pollution diffusion paths. Other studies have applied Bayesian neural networks to air quality prediction to obtain uncertainty estimates for prediction results. While these efforts have improved predictive performance and interpretability to some extent, they still face common challenges, such as difficulty integrating heterogeneous data from multiple sources, difficulty adaptively tuning model structure parameters, and a lack of feedback mechanisms.

[0006] First, water pollution source tracking and atmospheric pollution analysis involve multiple data types, including hydrological data, atmospheric and meteorological data, and geospatial data. These data are high-dimensional and heterogeneous in format. Existing methods lack an effective fusion mechanism for these multimodal data, resulting in insufficient model extraction of key features, affecting overall prediction accuracy. Second, current model structures are typically set based on manual experience and lack the ability to dynamically optimize their structures. This results in unstable model performance and weak generalization in different scenarios. This is especially true in complex areas, such as urban river networks, port areas, and industrial clusters, where fixed model structures struggle to adapt to the ever-changing dynamic characteristics of pollution sources. Third, existing technologies generally lack adaptive update mechanisms based on feedback residuals, making it impossible to achieve responsive optimization of real-time monitoring data. When errors occur in model predictions, the system cannot proactively correct model parameters or structure based on the residual information, resulting in a decline in long-term prediction capabilities and even failure during sudden pollution events.

[0007] Furthermore, while swarm intelligence optimization algorithms, such as particle swarm optimization, genetic algorithms, and differential evolution, have been widely used to optimize model parameters, most still employ static update strategies, making it difficult to integrate real-time data for feedback adjustment. Their limitations also stem from their single behavioral mechanism and lack of memory and anomaly detection capabilities. For example, the lack of multiple behavioral co-evolution strategies and a mechanism for remembering historical optimal solutions prevents them from balancing local search with global exploration, nor can they use anomalous samples to guide them out of local optima, limiting optimization performance and convergence speed.

[0008] Therefore, how to provide a method for water pollution source tracking and atmospheric particulate matter distribution analysis based on machine learning is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0009] One purpose of the present invention is to propose a method for tracking water pollution sources and analyzing the distribution of atmospheric particulate matter based on machine learning. The present invention comprehensively utilizes a deep Bayesian neural network, an improved zebra optimization algorithm, a multi-scale residual feedback mechanism, and a feature embedding and fusion strategy to construct a composite prediction system with spatiotemporal modeling capabilities, uncertainty quantification capabilities, and online adaptive optimization capabilities. By constructing a pollution diffusion feature extraction and tracking module, a water-air pollution linkage modeling mechanism, and a prediction error feedback optimization process, the full-process coupled modeling from pollution source tracking to particulate matter spatial distribution prediction is achieved. Compared with the existing technology, the present invention has the advantages of more accurate pollution perception, more comprehensive data fusion, more reliable prediction output, and more efficient model updating, and is suitable for pollution monitoring and decision support systems in complex environments.

[0010] The method for tracing water pollution sources and analyzing atmospheric particulate matter distribution based on machine learning according to an embodiment of the present invention includes the following steps:

[0011] S1. Collect multi-source data including water pollutant concentration data and atmospheric particulate matter concentration data, pre-process the multi-source data, and generate an environmental data set;

[0012] S2. Extract water pollutant concentrations, water spatial structure information, and water flow velocity data from environmental datasets, analyze pollutant diffusion paths, infer the location and intensity of upstream pollution sources, and generate water pollution tracking results;

[0013] S3. Based on the environmental dataset and water pollution tracking results, a deep Bayesian neural network model is constructed to output the predicted atmospheric particulate matter concentration and the corresponding confidence interval for each spatial grid in the target area.

[0014] S4. Introduce the zebra optimization algorithm to jointly optimize the structure and hyperparameters of the deep Bayesian neural network model to obtain the optimized deep Bayesian neural network model structure and hyperparameters;

[0015] S5. Input the generated water pollution tracking results into the optimized deep Bayesian neural network model, establish the impact relationship between water pollution and the spatial distribution of atmospheric particulate matter through feature embedding and fusion mechanism, and output the prediction results;

[0016] S6. Continuously collect real-time monitoring data, compare and analyze the actual observations with the predicted results, calculate the prediction residuals, update the individual zebra population status, and further optimize the deep Bayesian neural network model structure and hyperparameters;

[0017] S7. Generate a pollutant distribution heat map and trend forecast map based on the prediction results, and trigger a pollution warning when the predicted value exceeds the set pollution threshold, output pollution level information and regional pollution control recommendations.

[0018] Optionally, the multi-source data is preprocessed, specifically including data cleaning, missing value filling, normalization and spatiotemporal alignment, to improve the quality of the multi-source data and construct an environmental data set in a unified format.

[0019] Optionally, the S2 specifically includes:

[0020] S21, extracting water pollutant concentration monitoring data, water body geometric boundary information, and water flow velocity data of each area from the environmental data set;

[0021] S22. Divide the water body grid area according to the water body geometric boundary information, and establish the upstream and downstream flow direction relationship between each water body grid unit to form the flow path structure within the water body;

[0022] S23. Combine water pollutant concentration monitoring data with the flow path structure within the water body to analyze the spatiotemporal diffusion changes of water pollutants between grid cells and obtain the trajectory of water pollutants propagating from downstream to upstream;

[0023] S24. By analyzing the concentration trends of water pollutants at multiple downstream monitoring points over time, we can identify the path of concentration increase and trace back the source of water pollution.

[0024] S25. Locate the water pollution source in the backtracking path, and estimate the release intensity of the water pollution source based on the concentration peak and occurrence time near the water pollution source;

[0025] S26. Integrate the identified water pollution source location coordinates with the corresponding water pollution source release intensity to form a water pollution tracking result.

[0026] Optionally, the S3 specifically includes:

[0027] S31. Extract meteorological characteristic data including temperature, wind speed, humidity and rainfall, as well as spatial characteristic information of geographic coordinates, topography and land use type from the environmental data set to characterize the external environmental conditions that affect the distribution of atmospheric particulate matter;

[0028] S32. Extract water pollution characteristic data including the location of water pollution sources and the concentration intensity of water pollutants from the water pollution tracking results as atmospheric input variables characterizing the impact of pollution emissions;

[0029] S33, aligning all extracted features according to time and spatial positions, and performing normalization processing, constructing an input feature vector for each spatial grid point at a specific time point, and forming a training data set for the deep Bayesian neural network model;

[0030] S34. Construct a deep Bayesian neural network model, set the weight parameters in the deep Bayesian neural network to a probability distribution form, use the feature vector as input and the pollution concentration as the supervision output, and use the variational inference method to train the deep Bayesian neural network model;

[0031] S35. Input the environmental characteristics and pollution source characteristics into the trained deep Bayesian neural network model, perform inference on each spatial grid, and output the corresponding predicted value of atmospheric particulate matter concentration and the corresponding confidence interval;

[0032] S36. Organize and store the prediction results and confidence intervals of the deep Bayesian neural network model according to the spatial grid number and time sequence.

[0033] Optionally, all the features specifically include meteorological characteristics of temperature, wind speed, humidity, and rainfall, spatial characteristics of geographic coordinates, topography and land use type, and water pollution characteristics of water pollution source location and water pollutant concentration intensity, which are used to construct the input vector of the deep Bayesian neural network model.

[0034] Optionally, the S4 specifically includes:

[0035] S41. Construct individual representation vector Z of zebra optimization algorithm i =[L i ,N i ,D i ,A i ,μ i ,σ i ,η i ], used to jointly encode the structural parameters and hyperparameters of the deep Bayesian neural network model, where L i Indicates the number of network layers, N i Indicates the number of neurons in each layer, D i represents the Dropout rate, A i Indicates the activation function type encoding, μ i and σ i represents the prior distribution mean and standard deviation, η i represents the learning rate;

[0036] S42, initializing the zebra optimization algorithm, setting the population size N and the maximum number of iterations T, and dividing the zebra individuals into a guiding population, an adapting population, and a learning population, which are used to perform wide-area search, local convergence, and memory-based target revisiting, respectively;

[0037] S43. In each iteration, assign a behavior type b(i)∈{L,F,S,M} to each zebra individual, where L represents dominant behavior, F represents following behavior, S represents vigilance behavior, and M represents memory recall behavior;

[0038] S44. Build a behavior-driven adaptive step-size mechanism based on behavior types, and record the average fitness improvement value of each behavior type in the last k rounds of iterations. Calculate the step weight factor and perform the main behavior update:

[0039]

[0040] in, represents the position of the i-th zebra individual in the t-th iteration, represents the position of the i-th zebra individual in the t+1th iteration, λ0 is the initial step length coefficient, Represents the updated direction vector corresponding to the behavior performed by the i-th zebra individual in the t-th iteration, It represents the historical average fitness improvement value of the behavior type b(i) of zebra individual i in the tth iteration, represents a set of behavior types, including leading behavior, following behavior, alert behavior, and memory behavior. b′ is a traversal variable. represents the set of zebra individuals that perform behavior type b′ in round t;

[0041] S45, when the zebra individual x i When it belongs to the learning population and meets the memory revisit triggering conditions, the result is updated based on the main behavior obtained Execution memory recall correction update:

[0042]

[0043] Among them, x memory is the historical best individual in the currently selected memory pool, is the revisit step size factor, represents the position of zebra individual i after the memory recall correction in round t, δ is the sinusoidal perturbation amplitude coefficient, ω is the sinusoidal perturbation frequency parameter, and sin is the sine function;

[0044] S46. After all zebra individuals have completed the position update, for the zebra individuals that have not performed the memory recall, update the main behavior As the evaluation input, the zebra individuals that perform memory recall behavior are updated with the corrected results. Replace the original position and unify it as the final position of the zebra individual in this round to build the corresponding deep Bayesian neural network model. The final position of the zebra individual determines the structure and hyperparameter configuration of the constructed deep Bayesian neural network model, affecting the prediction performance index and fitness function value f(Z i ):

[0045] f(Z i )=α·RMSE(Zi )+β·NLL(Z i )+γ·UCIW(Z i );

[0046] Among them, RMSE(Z i ) represents the root mean square error of the deep Bayesian neural network model prediction, NLL(Z i ) represents the negative log-likelihood, UCIW(Z i ) represents the width of the prediction confidence interval, α, β, γ are non-negative weighting coefficients;

[0047] S47, respectively evaluate the fitness of the guiding population, the adapting population, and the learning population, and select the individual with the best current fitness from each population to add to the global experience pool;

[0048] S48. Based on the individual fitness differences and spatial distribution information entropy indicators in the global experience pool, the population affiliation of individuals is dynamically adjusted to guide the co-evolution and structural adaptive adjustment among the guiding population, the adapting population, and the learning population.

[0049] S49, after each iteration, record the current best fitness zebra individual Z * , the optimization process ends when any of the following termination conditions is met:

[0050] t≥T or

[0051] Among them, t represents the current iteration number, T is the maximum number of iterations, represents the fitness value of the global optimal individual in round t, represents the fitness value of the global optimal individual in the t-1th round, ∈ is the convergence threshold;

[0052] S410: Output the deep Bayesian neural network model structure and hyperparameter configuration represented by the current optimal individual as the final optimization result.

[0053] Optionally, the execution of the main behavior update refers to the standard position change operation performed by the zebra individual according to the behavior type and the behavior-driven rules in each round of iteration, specifically including the update path executed by the dominant behavior, following behavior and vigilance behavior, which is used to guide the zebra individual to move towards a better solution in the search space.

[0054] Optionally, the S5 specifically includes:

[0055] S51. Obtain water pollution tracking results and extract characteristics of pollution source location, water pollutant concentration intensity, and pollution diffusion direction;

[0056] S52, normalizing the extracted water pollution features and using them as pollution input vectors to construct a feature embedding interface;

[0057] S53, extracting the spatial characteristics of temperature, humidity, wind speed, rainfall, topography and land use type from the environmental data set to form an environmental input vector;

[0058] S54, based on the feature embedding fusion mechanism, the pollution input vector and the environment input vector are mapped to a unified embedding space, and the fusion encoding is completed;

[0059] S55, inputting the fused feature vector into the optimized deep Bayesian neural network model, executing the deep Bayesian neural network model inference process, and outputting the prediction result;

[0060] S56. Structuralize and format the generated prediction results to generate an atmospheric particulate matter distribution map for the target area, which serves as the basis for pollution assessment and decision-making analysis.

[0061] Optionally, the prediction results specifically include the predicted values ​​of atmospheric particulate matter concentrations for each spatial grid within the target area and the corresponding confidence interval information, which are used to reflect the water pollutant concentration level and prediction uncertainty.

[0062] Optionally, the S6 specifically includes:

[0063] S61, continuously collect the real-time monitoring data of atmospheric particulate matter concentration in each spatial grid in the target area as the actual observation value y j ;

[0064] S62. Extract the prediction results output in the current round and the real-time observation value y j Compare and construct the prediction error sequence, where Z * is the zebra individual with the best current fitness;

[0065] S63. Calculate the prediction residual of each sample point Based on different time scales T k and spatial clustering area S c Constructing the residual feedback matrix in is the residual mean, is the residual variance, k is the time scale number, and c is the spatial region number;

[0066] S64. Embed the residual feedback matrix and the feedback factor vector ψ into the fitness function of the zebra optimization algorithm to modify the fitness value:

[0067] f′(Z i )=f(Z i )+λ e ·||E (k,c) ·ψ||2;

[0068] Among them, λ e is the adjustment coefficient, ‖‖2 represents the second norm, f(Z i ) is the original fitness function, f′(Z i ) is the modified fitness function;

[0069] S65. If the prediction residual corresponding to the zebra individual satisfies the abnormal condition |e j |>θ, where θ is the abnormal trigger threshold, the abnormal driving mechanism is triggered and the individual zebra position is updated:

[0070]

[0071] Where η is the guiding factor, δ r is the perturbation coefficient, randn() represents the Gaussian perturbation function, is the current individual zebra position, The updated position of the abnormal driver. is the global optimal position;

[0072] S66. Determine the final position of the zebra individual in this round based on the priority logic If the zebra individual performs abnormal drive update, then Otherwise, if memory recall update is performed, Otherwise, take the main behavior update result in, Correct location for memory recall, Update the result for the main behavior;

[0073] S67, with the final position Build a deep Bayesian neural network model for the input and use the modified fitness function f′(Z i ) Evaluate and sort the individuals in the population, update the population status of the zebra optimization algorithm, and adjust the deep Bayesian neural network model structure and hyperparameters represented by the current optimal individual online as the configuration basis for the next round of pollution prediction.

[0074] The beneficial effects of the present invention are:

[0075] This paper addresses existing issues in water pollution tracking and atmospheric pollution modeling, such as fragmented modeling chains, insufficient data fusion capabilities, uncontrollable prediction uncertainty, and a lack of adaptive feedback optimization. By proposing a fusion approach with full-process modeling capabilities, this approach enables cross-media pollution modeling, from water pollution source identification and diffusion inference to atmospheric particulate matter concentration prediction. By introducing a deep Bayesian neural network model, this approach not only enables high-precision predictions of the spatial distribution of atmospheric particulate matter but also outputs uncertainty information, such as confidence intervals, enhancing prediction interpretability and supporting safety decision-making. A feature embedding fusion mechanism effectively integrates water pollution characteristics with heterogeneous data from multiple sources, including atmospheric meteorology and spatial geography, improving the accuracy of modeling the impact of pollution diffusion on the atmosphere. The Zebra optimization algorithm is introduced to jointly optimize the deep Bayesian neural network structure and hyperparameters. Incorporating innovative mechanisms such as behavior-driven, adaptive step size, memory recall, and abnormal perturbation, it enhances the model's structural adaptability and global convergence efficiency. By constructing a multi-scale spatial-temporal residual feedback mechanism, the optimization objective is automatically adjusted based on prediction error, driving the optimization algorithm to dynamically adjust its direction and population state, thereby enabling online model updates and adaptive enhancement of prediction performance. Overall, the present invention has achieved method innovation and performance improvement in many key links such as intelligent identification of pollution sources, pollution migration modeling and pollution impact prediction. It can effectively solve core problems such as low prediction accuracy, lagging update mechanism, and rough optimization strategy in traditional methods, and has good engineering feasibility and environmental application promotion value. BRIEF DESCRIPTION OF THE DRAWINGS

[0076] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0077] Figure 1 This is a flow chart of the method for water pollution source tracking and atmospheric particulate matter distribution analysis based on machine learning proposed in the present invention;

[0078] Figure 2 This is a flow chart of the joint optimization of the structure and hyperparameters of a deep Bayesian neural network using the zebra optimization algorithm for the water pollution source tracking and atmospheric particulate matter distribution analysis method based on machine learning proposed in the present invention. DETAILED DESCRIPTION

[0079] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0080] refer to Figure 1 and Figure 2 The method for tracing water pollution sources and analyzing atmospheric particulate matter distribution based on machine learning includes the following steps:

[0081] S1. Collect multi-source data including water pollutant concentration data and atmospheric particulate matter concentration data, pre-process the multi-source data, and generate an environmental data set;

[0082] S2. Extract water pollutant concentrations, water spatial structure information, and water flow velocity data from environmental datasets, analyze pollutant diffusion paths, infer the location and intensity of upstream pollution sources, and generate water pollution tracking results;

[0083] S3. Based on the environmental dataset and water pollution tracking results, a deep Bayesian neural network model is constructed to output the predicted atmospheric particulate matter concentration and the corresponding confidence interval for each spatial grid in the target area.

[0084] S4. Introduce the zebra optimization algorithm to jointly optimize the structure and hyperparameters of the deep Bayesian neural network model to obtain the optimized deep Bayesian neural network model structure and hyperparameters;

[0085] S5. Input the generated water pollution tracking results into the optimized deep Bayesian neural network model, establish the impact relationship between water pollution and the spatial distribution of atmospheric particulate matter through feature embedding and fusion mechanism, and output the prediction results;

[0086] S6. Continuously collect real-time monitoring data, compare and analyze the actual observations with the predicted results, calculate the prediction residuals, update the individual zebra population status, and further optimize the deep Bayesian neural network model structure and hyperparameters;

[0087] S7. Generate a pollutant distribution heat map and trend forecast map based on the prediction results, and trigger a pollution warning when the predicted value exceeds the set pollution threshold, output pollution level information and regional pollution control recommendations.

[0088] In this embodiment, the multi-source data preprocessing specifically includes data cleaning, missing value filling, normalization and spatiotemporal alignment, which are used to improve the quality of multi-source data and construct an environmental data set in a unified format.

[0089] In this embodiment, S2 specifically includes:

[0090] S21, extracting water pollutant concentration monitoring data, water body geometric boundary information, and water flow velocity data of each area from the environmental data set;

[0091] S22. Divide the water body grid area according to the water body geometric boundary information, and establish the upstream and downstream flow direction relationship between each water body grid unit to form the flow path structure within the water body;

[0092] S23. Combine water pollutant concentration monitoring data with the flow path structure within the water body to analyze the spatiotemporal diffusion changes of water pollutants between grid cells and obtain the trajectory of water pollutants propagating from downstream to upstream;

[0093] S24. By analyzing the concentration trends of water pollutants at multiple downstream monitoring points over time, we can identify the path of concentration increase and trace back the source of water pollution.

[0094] S25. Locate the water pollution source in the backtracking path, and estimate the release intensity of the water pollution source based on the concentration peak and occurrence time near the water pollution source;

[0095] S26. Integrate the identified water pollution source location coordinates with the corresponding water pollution source release intensity to form a water pollution tracking result.

[0096] In this embodiment, S3 specifically includes:

[0097] S31. Extract meteorological characteristic data including temperature, wind speed, humidity and rainfall, as well as spatial characteristic information of geographic coordinates, topography and land use type from the environmental data set to characterize the external environmental conditions that affect the distribution of atmospheric particulate matter;

[0098] S32. Extract water pollution characteristic data including the location of water pollution sources and the concentration intensity of water pollutants from the water pollution tracking results as atmospheric input variables characterizing the impact of pollution emissions;

[0099] S33, aligning all extracted features according to time and spatial positions, and performing normalization processing, constructing an input feature vector for each spatial grid point at a specific time point, and forming a training data set for the deep Bayesian neural network model;

[0100] S34. Construct a deep Bayesian neural network model, set the weight parameters in the deep Bayesian neural network to a probability distribution form, use the feature vector as input and the pollution concentration as the supervision output, and use the variational inference method to train the deep Bayesian neural network model;

[0101] S35. Input the environmental characteristics and pollution source characteristics into the trained deep Bayesian neural network model, perform inference on each spatial grid, and output the corresponding predicted value of atmospheric particulate matter concentration and the corresponding confidence interval;

[0102] S36. Organize and store the prediction results and confidence intervals of the deep Bayesian neural network model according to the spatial grid number and time sequence.

[0103] In this embodiment, all the characteristics specifically include meteorological characteristics of temperature, wind speed, humidity, and rainfall, spatial characteristics of geographic coordinates, topography and land use type, and water pollution characteristics of water pollution source location and water pollutant concentration intensity, which are used to construct the input vector of the deep Bayesian neural network model.

[0104] In this embodiment, the S4 specifically includes:

[0105] S41. Construct individual representation vector Z of zebra optimization algorithm i =[L i ,N i ,D i ,A i ,μ i ,σ i ,η i ], used to jointly encode the structural parameters and hyperparameters of the deep Bayesian neural network model, where L i Indicates the number of network layers, N i Indicates the number of neurons in each layer, D i represents the Dropout rate, A i Indicates the activation function type encoding, μ i and σ i represents the prior distribution mean and standard deviation, η i represents the learning rate;

[0106] S42, initializing the zebra optimization algorithm, setting the population size N and the maximum number of iterations T, and dividing the zebra individuals into a guiding population, an adapting population, and a learning population, which are used to perform wide-area search, local convergence, and memory-based target revisiting, respectively;

[0107] S43. In each iteration, assign a behavior type b(i)∈{L,F,S,M} to each zebra individual, where L represents dominant behavior, F represents following behavior, S represents vigilance behavior, and M represents memory recall behavior;

[0108] S44. Build a behavior-driven adaptive step-size mechanism based on behavior types, and record the average fitness improvement value of each behavior type in the last k rounds of iterations. Calculate the step weight factor and perform the main behavior update:

[0109]

[0110] in, represents the position of the i-th zebra individual in the t-th iteration, represents the position of the i-th zebra individual in the t+1th iteration, λ0 is the initial step length coefficient, Represents the updated direction vector corresponding to the behavior performed by the i-th zebra individual in the t-th iteration, It represents the historical average fitness improvement value of the behavior type b(i) of zebra individual i in the tth iteration, represents a set of behavior types, including leading behavior, following behavior, alert behavior, and memory behavior. b′ is a traversal variable. represents the set of zebra individuals that perform behavior type b′ in round t;

[0111] S45, when the zebra individual x i When it belongs to the learning population and meets the memory revisit triggering conditions, the result is updated based on the main behavior obtained Execution memory recall correction update:

[0112]

[0113] Among them, x memory is the historical best individual in the currently selected memory pool, is the revisit step size factor, represents the position of zebra individual i after the memory recall correction in round t, δ is the sinusoidal perturbation amplitude coefficient, ω is the sinusoidal perturbation frequency parameter, and sin is the sine function;

[0114] S46. After all zebra individuals have completed the position update, for the zebra individuals that have not performed the memory recall, update the main behavior As the evaluation input, the zebra individuals that perform memory recall behavior are updated with the corrected results. Replace the original position and unify it as the final position of the zebra individual in this round to build the corresponding deep Bayesian neural network model. The final position of the zebra individual determines the structure and hyperparameter configuration of the constructed deep Bayesian neural network model, affecting the prediction performance index and fitness function value f(Z i ):

[0115] f(Z i )=α·RMSE(Z i )+β·NLL(Z i )+γ·UCIW(Z i );

[0116] Among them, RMSE(Z i ) represents the root mean square error of the deep Bayesian neural network model prediction, NLL(Z i ) represents the negative log-likelihood, UCIW(Z i ) represents the width of the prediction confidence interval, α, β, γ are non-negative weighting coefficients;

[0117] S47, respectively evaluate the fitness of the guiding population, the adapting population, and the learning population, and select the individual with the best current fitness from each population to add to the global experience pool;

[0118] S48. Based on the individual fitness differences and spatial distribution information entropy indicators in the global experience pool, the population affiliation of individuals is dynamically adjusted to guide the co-evolution and structural adaptive adjustment among the guiding population, the adapting population, and the learning population.

[0119] S49, after each iteration, record the current best fitness zebra individual Z * , the optimization process ends when any of the following termination conditions is met:

[0120] t≥T or

[0121] Among them, t represents the current iteration number, T is the maximum number of iterations, represents the fitness value of the global optimal individual in round t, represents the fitness value of the global optimal individual in the t-1th round, ∈ is the convergence threshold;

[0122] S410: Output the deep Bayesian neural network model structure and hyperparameter configuration represented by the current optimal individual as the final optimization result.

[0123] In this embodiment, the execution of the main behavior update refers to the standard position change operation performed by the zebra individual according to the behavior driving rules in each round of iteration based on the behavior type to which it belongs, specifically including the update path executed by the dominant behavior, following behavior and vigilance behavior, which is used to guide the zebra individual to move towards a better solution in the search space.

[0124] In this embodiment, the S5 specifically includes:

[0125] S51. Obtain water pollution tracking results and extract characteristics of pollution source location, water pollutant concentration intensity, and pollution diffusion direction;

[0126] S52, normalizing the extracted water pollution features and using them as pollution input vectors to construct a feature embedding interface;

[0127] S53, extracting the spatial characteristics of temperature, humidity, wind speed, rainfall, topography and land use type from the environmental data set to form an environmental input vector;

[0128] S54, based on the feature embedding fusion mechanism, the pollution input vector and the environment input vector are mapped to a unified embedding space, and the fusion encoding is completed;

[0129] S55, inputting the fused feature vector into the optimized deep Bayesian neural network model, executing the deep Bayesian neural network model inference process, and outputting the prediction result;

[0130] S56. Structuralize and format the generated prediction results to generate an atmospheric particulate matter distribution map for the target area, which serves as the basis for pollution assessment and decision-making analysis.

[0131] In this embodiment, the prediction results specifically include the predicted values ​​of atmospheric particulate matter concentrations for each spatial grid within the target area and the corresponding confidence interval information, which are used to reflect the water pollutant concentration level and prediction uncertainty.

[0132] In this embodiment, S6 specifically includes:

[0133] S61, continuously collect the real-time monitoring data of atmospheric particulate matter concentration in each spatial grid in the target area as the actual observation value y j ;

[0134] S62. Extract the prediction results output in the current round and the real-time observation value y j Compare and construct the prediction error sequence, where Z * is the zebra individual with the best current fitness;

[0135] S63. Calculate the prediction residual of each sample point Based on different time scales T k and spatial clustering area S c Constructing the residual feedback matrix in is the residual mean, is the residual variance, k is the time scale number, and c is the spatial region number;

[0136] S64. Embed the residual feedback matrix and the feedback factor vector ψ into the fitness function of the zebra optimization algorithm to modify the fitness value:

[0137] f′(Z i )=f(Z i )+λ e ·||E (k,c) ·ψ||2;

[0138] Among them, λ e is the adjustment coefficient, ‖‖2 represents the second norm, f(Z i ) is the original fitness function, f′(Z i ) is the modified fitness function;

[0139] S65. If the prediction residual corresponding to the zebra individual satisfies the abnormal condition |e j |>θ, where θ is the abnormal trigger threshold, the abnormal driving mechanism is triggered and the individual zebra position is updated:

[0140]

[0141] Where η is the guiding factor, δ r is the perturbation coefficient, randn() represents the Gaussian perturbation function, is the current individual zebra position, The updated position of the abnormal driver. is the global optimal position;

[0142] S66. Determine the final position of the zebra individual in this round based on the priority logic If the zebra individual performs abnormal drive update, then Otherwise, if memory recall update is performed, Otherwise, take the main behavior update result in, Correct location for memory recall, Update the result for the main behavior;

[0143] S67, with the final position Build a deep Bayesian neural network model for the input and use the modified fitness function f′(Z i ) Evaluate and sort the individuals in the population, update the population status of the zebra optimization algorithm, and adjust the deep Bayesian neural network model structure and hyperparameters represented by the current optimal individual online as the configuration basis for the next round of pollution prediction.

[0144] Example 1:

[0145] In order to verify the feasibility of the present invention in practice, the present invention was applied to a certain industrial area. Due to long-term factory wastewater discharge and rainwater erosion, the water quality of the rivers in the area has deteriorated, accompanied by a significant increase in the concentration of particulate matter in the atmosphere. The local environmental protection department requires the establishment of an intelligent monitoring system that can track the location and intensity of water pollution sources in real time and predict the distribution of atmospheric particulate matter in the area so that pollution control measures can be taken in a timely manner. To this end, this embodiment selects the industrial park and the adjacent river system as the monitoring area, and uses the method of the present invention to model and predict the pollution phenomenon.

[0146] In practical applications, the system first collects multi-source data on river water, regional atmosphere, and meteorological data. The system collects real-time data from water quality sensors, meteorological monitoring stations, and remote sensing imagery, gathering environmental data such as water pollutant concentrations, atmospheric particulate matter concentrations, flow velocity, temperature, humidity, wind speed, rainfall, topography, and land use. Data collection spans June to November 2024, with a frequency of one set every 15 minutes, covering an area of ​​approximately 10 square kilometers. After data cleaning, normalization, and spatiotemporal alignment, a standardized environmental dataset is generated, providing the foundational input for subsequent model construction.

[0147] Based on this environmental dataset, the system first applies the Water Pollution Tracking module to dynamically model pollutant concentrations, flow rates, and river system topology in the river water, identifying pollution diffusion paths and the location and intensity of upstream pollution sources. For example, in early August 2024, the system identified a pollution source approximately 3 kilometers upstream from the industrial park, with peak pollutant concentrations reaching 150 mg per liter, while normal water concentrations are generally below 20 mg / L. This information is integrated into the Water Pollution Tracking results and serves as an important input for the subsequent Air Pollution Forecasting module.

[0148] Next, the present invention uses the optimized deep Bayesian neural network model, combined with the feature embedding and fusion mechanism, to jointly model the water pollution tracking results with the atmospheric environment data. The deep Bayesian model not only outputs the predicted value of the atmospheric particulate matter concentration in each spatial grid, but also provides confidence interval information to reflect the uncertainty of the prediction. After offline training and online updating, the system can realize pollution prediction for each grid in the area (such as a 1km×1km grid). Using the zebra optimization algorithm, the model structure and hyperparameters are dynamically adjusted to further improve the prediction accuracy. By introducing multi-scale residual feedback and anomaly driving mechanisms, the model can quickly respond to real-time monitoring data during the prediction process. When there is a large error between the actual observation value and the predicted value, the system will automatically update the model parameters to maintain long-term high-precision performance.

[0149] During actual monitoring operations, the system collected data and generated predictions as shown below. For example, on a specific day (September 15, 2024), 20 representative monitoring points were selected within the region. The corresponding real-time observations, model predictions, and calculated prediction residuals are shown in the table below.

[0150] Table 1 Comparison of atmospheric particulate matter concentration data at some monitoring points on September 15, 2024

[0151]

[0152] From the monitoring data collected on September 15, 2024, seven typical monitoring points in the industrial area were selected for analysis. It can be seen that the model proposed in this invention has good prediction ability and uncertainty expression ability in actual scenarios. At monitoring points 1 to 7, the difference between the actual observation value and the prediction value of the deep Bayesian neural network model is generally small, and the prediction residual is within ±3μg / m 3 Specifically, the actual value of monitoring point 1 is 85μg / m 3 The predicted value is 82 μg / m 3 , the residual is +3 μg / m 3 , within the confidence interval of ±5 μg / m 3The predicted value of monitoring point 2 is 88μg / m 3 , and the actual value is 90μg / m 3 The difference is only 2 μg / m 3 , the corresponding confidence interval is ±6 μg / m 3 , reflecting the stable prediction ability of the model in the medium and high concentration range.

[0153] Similarly, in monitoring sites 3 and 4, although there were slightly negative residuals (–2 μg / m 3 and +5 μg / m 3 ), but the model prediction value is still within the credible range, especially the actual value of monitoring point 4 is 110μg / m 3 , while the model output is 105μg / m 3 The predicted results are given as ±7μg / m 3 The confidence intervals of the deep Bayesian model fully demonstrate the ability of the deep Bayesian model to estimate uncertainty in highly polluted areas. The prediction residuals of monitoring points 5 to 7 are all +2 to +3 μg / m 3 interval, indicating that the model has consistent generalization ability in continuous spatial grids.

[0154] Data from these seven sample points demonstrate that the proposed pollution source tracking and particulate matter concentration prediction method can accurately reflect pollution diffusion trends in real-world environments and output prediction results with uncertainty estimates, improving the stability and reliability of the environmental prediction system. Furthermore, through an online feedback mechanism, the system triggers anomaly-driven optimization operations when monitoring points with large relative deviations occur, maintaining the model's adaptability. Overall, these results demonstrate the model's practicality and engineering feasibility in real-world deployment environments.

[0155] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for tracing water pollution sources and analyzing atmospheric particulate matter distribution based on machine learning, characterized by: The steps include: S1. Collect multi-source data including water pollutant concentration data and atmospheric particulate matter concentration data, pre-process the multi-source data, and generate an environmental data set; S2. Extract water pollutant concentrations, water spatial structure information, and water flow velocity data from environmental datasets, analyze pollutant diffusion paths, infer the location and intensity of upstream pollution sources, and generate water pollution tracking results; S3. Based on the environmental dataset and water pollution tracking results, a deep Bayesian neural network model is constructed to output the predicted atmospheric particulate matter concentration and the corresponding confidence interval for each spatial grid in the target area. S4. Introduce the zebra optimization algorithm to jointly optimize the structure and hyperparameters of the deep Bayesian neural network model to obtain the optimized deep Bayesian neural network model structure and hyperparameters; S5. Input the generated water pollution tracking results into the optimized deep Bayesian neural network model, establish the impact relationship between water pollution and the spatial distribution of atmospheric particulate matter through feature embedding and fusion mechanism, and output the prediction results; S6. Continuously collect real-time monitoring data, compare and analyze the actual observations with the predicted results, calculate the prediction residuals, update the individual zebra population status, and further optimize the deep Bayesian neural network model structure and hyperparameters; S7. Generate a pollutant distribution heat map and trend forecast map based on the prediction results, and trigger a pollution warning when the predicted value exceeds the set pollution threshold, output pollution level information and regional pollution control recommendations.

2. The method for tracing water pollution sources and analyzing atmospheric particulate matter distribution based on machine learning according to claim 1 is characterized in that: The multi-source data preprocessing specifically includes data cleaning, missing value filling, normalization and spatiotemporal alignment, which is used to improve the quality of multi-source data and construct an environmental data set in a unified format.

3. The method for tracing water pollution sources and analyzing atmospheric particulate matter distribution based on machine learning according to claim 1 is characterized in that: The S2 specifically includes: S21, extracting water pollutant concentration monitoring data, water body geometric boundary information, and water flow velocity data of each area from the environmental data set; S22. Divide the water body into grid areas based on the geometric boundary information of the water body, and establish upstream and downstream flow direction relationships between each water body grid unit to form a flow path structure within the water body; S23. Combine water pollutant concentration monitoring data with the flow path structure within the water body to analyze the spatiotemporal diffusion changes of water pollutants between grid cells and obtain the trajectory of water pollutants propagating from downstream to upstream; S24. By analyzing the concentration trends of water pollutants over time at multiple downstream monitoring points, we can identify the path of concentration increase and trace back the source of water pollution. S25. Locate the water pollution source in the backtracking path, and estimate the release intensity of the water pollution source based on the concentration peak and occurrence time near the water pollution source; S26. Integrate the identified water pollution source location coordinates with the corresponding water pollution source release intensity to form a water pollution tracking result.

4. The method for tracing water pollution sources and analyzing atmospheric particulate matter distribution based on machine learning according to claim 1 is characterized in that: The S3 specifically includes: S31. Extract meteorological characteristic data including temperature, wind speed, humidity and rainfall, as well as spatial characteristic information of geographic coordinates, topography and land use type from the environmental data set to characterize the external environmental conditions that affect the distribution of atmospheric particulate matter; S32. Extract water pollution characteristic data including the location of water pollution sources and the concentration intensity of water pollutants from the water pollution tracking results as atmospheric input variables characterizing the impact of pollution emissions; S33, aligning all extracted features according to time and spatial positions, and performing normalization processing, constructing an input feature vector for each spatial grid point at a specific time point, and forming a training data set for the deep Bayesian neural network model; S34. Construct a deep Bayesian neural network model, set the weight parameters in the deep Bayesian neural network to a probability distribution form, use the feature vector as input and the pollution concentration as the supervision output, and use the variational inference method to train the deep Bayesian neural network model; S35. Input the environmental characteristics and pollution source characteristics into the trained deep Bayesian neural network model, perform inference on each spatial grid, and output the corresponding predicted value of atmospheric particulate matter concentration and the corresponding confidence interval; S36. Organize and store the prediction results and confidence intervals of the deep Bayesian neural network model according to the spatial grid number and time sequence.

5. The method for tracing water pollution sources and analyzing atmospheric particulate matter distribution based on machine learning according to claim 4 is characterized in that: All the characteristics specifically include meteorological characteristics such as temperature, wind speed, humidity, and rainfall, spatial characteristics such as geographic coordinates, topography, and land use types, and water pollution characteristics such as the location of water pollution sources and the concentration intensity of water pollutants, which are used to construct the input vector of the deep Bayesian neural network model.

6. The method for tracing water pollution sources and analyzing atmospheric particulate matter distribution based on machine learning according to claim 1 is characterized in that: The S4 specifically includes: S41. Construct individual representation vector Z of zebra optimization algorithm i =[L i ,N i ,D i ,A i ,μ i ,σ i ,η i ], used to jointly encode the structural parameters and hyperparameters of the deep Bayesian neural network model, where L i Indicates the number of network layers, N i Indicates the number of neurons in each layer, D i represents the Dropout rate, A i Indicates the activation function type encoding, μ i and σ i represents the prior distribution mean and standard deviation, η i represents the learning rate; S42, initializing the zebra optimization algorithm, setting the population size N and the maximum number of iterations T, and dividing the zebra individuals into a guiding population, an adapting population, and a learning population, which are used to perform wide-area search, local convergence, and memory-based target revisiting, respectively; S43. In each iteration, assign a behavior type b(i)∈{L,F,S,M} to each zebra individual, where L represents dominant behavior, F represents following behavior, S represents vigilance behavior, and M represents memory recall behavior; S44. Build a behavior-driven adaptive step-size mechanism based on behavior types, and record the average fitness improvement value of each behavior type in the last k rounds of iterations. Calculate the step weight factor and perform the main behavior update: in, represents the position of the i-th zebra individual in the t-th iteration, represents the position of the i-th zebra individual in the t+1th iteration, λ0 is the initial step length coefficient, Represents the updated direction vector corresponding to the behavior performed by the i-th zebra individual in the t-th iteration, It represents the historical average fitness improvement value of the behavior type b(i) of zebra individual i in the tth iteration, represents a set of behavior types, including leading behavior, following behavior, alert behavior, and memory behavior. b′ is a traversal variable. represents the set of zebra individuals that perform behavior type b′ in round t; S45, when the zebra individual x i When it belongs to the learning population and meets the memory revisit triggering conditions, the result is updated based on the main behavior obtained Execution memory recall correction update: Among them, x memory is the historical best individual in the currently selected memory pool, is the revisit step size factor, represents the position of zebra individual i after the memory recall correction in round t, δ is the sinusoidal perturbation amplitude coefficient, ω is the sinusoidal perturbation frequency parameter, and sin is the sine function; S46. After all zebra individuals have completed the position update, for the zebra individuals that have not performed the memory recall, update the main behavior As the evaluation input, the zebra individuals that perform memory recall behavior are updated with the corrected results. Replace the original position and unify it as the final position of the zebra individual in this round to build the corresponding deep Bayesian neural network model. The final position of the zebra individual determines the structure and hyperparameter configuration of the constructed deep Bayesian neural network model, affecting the prediction performance index and fitness function value f(Z i ): f(Z i )=α·RMSE(Z i )+β·NLL(Z i )+γ·UCIW(Z i ); Among them, RMSE(Z i ) represents the root mean square error of the deep Bayesian neural network model prediction, NLL(Z i ) represents the negative log-likelihood, UCIW(Z i ) represents the width of the prediction confidence interval, α, β, γ are non-negative weighting coefficients; S47, respectively evaluate the fitness of the guiding population, the adapting population, and the learning population, and select the individual with the best current fitness from each population to add to the global experience pool; S48. Based on the individual fitness differences and spatial distribution information entropy indicators in the global experience pool, the population affiliation of individuals is dynamically adjusted to guide the co-evolution and structural adaptive adjustment among the guiding population, the adapting population, and the learning population. S49, after each iteration, record the current best fitness zebra individual Z * , the optimization process ends when any of the following termination conditions is met: t≥T or Among them, t represents the current iteration number, T is the maximum number of iterations, represents the fitness value of the global optimal individual in round t, represents the fitness value of the global optimal individual in the t-1th round, ∈ is the convergence threshold; S410: Output the deep Bayesian neural network model structure and hyperparameter configuration represented by the current optimal individual as the final optimization result.

7. The method for tracing water pollution sources and analyzing atmospheric particulate matter distribution based on machine learning according to claim 6 is characterized in that: The execution of the main behavior update refers to the standard position change operation performed by the zebra individual in each iteration according to the behavior driving rules based on the behavior type to which it belongs. Specifically, it includes the update path executed by the leading behavior, following behavior and vigilance behavior, which is used to guide the zebra individual to move towards a better solution in the search space.

8. The method for tracing water pollution sources and analyzing atmospheric particulate matter distribution based on machine learning according to claim 1 is characterized in that: The S5 specifically includes: S51. Obtain water pollution tracking results and extract characteristics of pollution source location, water pollutant concentration intensity, and pollution diffusion direction; S52, normalizing the extracted water pollution features and using them as pollution input vectors to construct a feature embedding interface; S53, extracting the spatial characteristics of temperature, humidity, wind speed, rainfall, topography and land use type from the environmental data set to form an environmental input vector; S54, based on the feature embedding fusion mechanism, the pollution input vector and the environment input vector are mapped to a unified embedding space, and the fusion encoding is completed; S55, inputting the fused feature vector into the optimized deep Bayesian neural network model, executing the deep Bayesian neural network model inference process, and outputting the prediction result; S56. Structuralize and format the generated prediction results to generate an atmospheric particulate matter distribution map for the target area, which serves as the basis for pollution assessment and decision-making analysis.

9. The method for tracing water pollution sources and analyzing atmospheric particulate matter distribution based on machine learning according to claim 8 is characterized in that: The prediction results specifically include the predicted values ​​of atmospheric particulate matter concentration for each spatial grid within the target area and the corresponding confidence interval information, which are used to reflect the water pollutant concentration level and prediction uncertainty.

10. The method for tracing water pollution sources and analyzing atmospheric particulate matter distribution based on machine learning according to claim 1, characterized in that: The S6 specifically includes: S61, continuously collect the real-time monitoring data of atmospheric particulate matter concentration in each spatial grid in the target area as the actual observation value y j ; S62. Extract the prediction results output in the current round and the real-time observation value y j Compare and construct the prediction error sequence, where Z * is the zebra individual with the best current fitness; S63. Calculate the prediction residual of each sample point Based on different time scales T k and spatial clustering area S c Constructing the residual feedback matrix in is the residual mean, is the residual variance, k is the time scale number, and c is the spatial region number; S64. Embed the residual feedback matrix and the feedback factor vector ψ into the fitness function of the zebra optimization algorithm to modify the fitness value: f′(Z i )=f(Z i )+λ e ·||E (k,c) ·ψ||2; Among them, λ e is the adjustment coefficient, ‖‖2 represents the second norm, f(Z i ) is the original fitness function, f′(Z i ) is the modified fitness function; S65. If the prediction residual corresponding to the zebra individual satisfies the abnormal condition |e j |>θ, where θ is the abnormal trigger threshold, the abnormal driving mechanism is triggered and the individual zebra position is updated: Where η is the guiding factor, δ r is the perturbation coefficient, randn() represents the Gaussian perturbation function, is the current individual zebra position, The updated position of the abnormal driver. is the global optimal position; S66. Determine the final position of the zebra individual in this round based on the priority logic If the zebra individual performs abnormal drive update, then Otherwise, if memory recall update is performed, Otherwise, take the main behavior update result in, Correct location for memory recall, Update the result for the main behavior; S67, with the final position Build a deep Bayesian neural network model for the input and use the modified fitness function f′(Z i ) Evaluate and sort the individuals in the population, update the population status of the zebra optimization algorithm, and adjust the deep Bayesian neural network model structure and hyperparameters represented by the current optimal individual online as the configuration basis for the next round of pollution prediction.

Images