Urban atmospheric pollution real-time monitoring and tracing method based on reinforcement learning

Abstract

The invention discloses an urban atmospheric pollution real-time monitoring and tracing method based on reinforcement learning, and relates to the technical field of atmospheric environment monitoring, and the method comprises the following steps: S1, collecting a pollution source data set; s2, constructing a Gaussian plume fidelity simulation model; s3, constructing a CFD fidelity simulation model; s4, constructing a multi-fidelity simulation model based on a Gaussian process proxy model; s5, constructing a DPPO reinforcement learning model; s6, using an improved multi-fidelity Bayesian optimization algorithm to continuously train and iterate the DPPO reinforcement learning model; and S7, generating a standardized pollution source real-time traceability analysis report. According to the method, the limitations of much manual intervention, low efficiency and poor real-time performance in a traditional urban pollution monitoring and tracing method are overcome, and an efficient and accurate solution is provided for intelligent real-time monitoring of urban environmental pollution and accurate and rapid tracing of pollution sources.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of atmospheric environment monitoring, and particularly relates to a city atmospheric pollution real-time monitoring and tracing method based on reinforcement learning. BACKGROUND

[0002] With the increasing demand for urban air quality monitoring and pollution tracing management, traditional pollution diffusion monitoring and pollution source positioning methods are facing unprecedented challenges. In the field of modern environmental management, real-time monitoring and rapid tracing of pollutants are not only important means to ensure environmental safety, but also key factors to improve environmental governance efficiency. However, current urban atmospheric pollution monitoring and pollution source tracing methods mainly rely on manual monitoring, fixed monitoring sites or traditional diffusion model methods, which are complex and inefficient. Traditional pollution source positioning and tracing methods usually use sparse pollution data collected by fixed monitoring stations, and use manual experience or empirical formula to speculate the diffusion path and estimate the source. Although these methods can meet the needs in some simple situations, they lack the ability to adapt to real-time data and complex environments, and are difficult to meet the monitoring and tracing needs of dynamic changes in pollution diffusion in urban environments.

[0003] The limitations of traditional pollution diffusion monitoring and tracing methods are slow response and insufficient accuracy. Existing methods often rely on fixed rules and manual analysis to predict and alarm pollution diffusion trends, which not only takes a long time, but also cannot timely and accurately capture the diffusion path and dynamic changes of pollutants in complex urban environments. In the processing of urban environments with dense buildings, complex terrain and variable weather conditions, the traditional methods have very limited performance and adaptability to pollution diffusion dynamics. Especially for the complexity, multi-source and spatio-temporal dynamic changes of urban atmospheric pollution, manual monitoring and traditional diffusion model methods lack efficient and accurate processing means, resulting in serious deficiencies in the accuracy, real-time performance and dynamic adaptability of pollution source positioning.

[0004] In addition, traditional pollution monitoring and tracing methods often ignore the dynamic correlation between multi-source pollution data and complex environmental factors, and are difficult to adaptively adjust the layout and frequency of monitoring sites. For example, in the process of identifying spatial distribution of pollutant concentration field and analyzing diffusion path, traditional methods cannot effectively capture the detailed features of pollution diffusion or identify hidden pollution diffusion patterns, resulting in insufficient timeliness of tracing analysis, and even missing important pollution source information. Even if some methods use machine learning techniques to analyze pollution diffusion, they usually do not fully integrate the deep information of multi-modal data (such as meteorological, remote sensing and air quality monitoring data), making it difficult for the model to accurately identify the pollution diffusion pattern in complex environments, and thus affecting the accuracy and rapid response capability of the pollution source positioning and tracing system.

[0005] Therefore, how to utilize modern reinforcement learning technology to realize real-time accurate monitoring of urban air pollution and intelligent and rapid source tracing of pollution has become a key technical problem to be solved in the current environmental governance field. SUMMARY

[0006] The application provides a real-time monitoring and source tracing method for urban air pollution based on reinforcement learning. The method can not only collect multi-modal pollutant concentration, meteorological conditions and remote sensing data in real time, automatically extract pollutant diffusion characteristics, but also dynamically predict the position of a pollution source and the trend of pollution diffusion through a DPPO reinforcement learning strategy network, thereby significantly improving the accuracy and timeliness of urban pollution monitoring and source tracing. The method overcomes the limitations of traditional urban pollution monitoring and source tracing methods, such as human intervention, low efficiency and poor real-time performance, and provides an efficient and accurate solution for intelligent real-time monitoring of urban environmental pollution and accurate and rapid source tracing of pollution.

[0007] According to the real-time monitoring and source tracing method for urban air pollution based on reinforcement learning, the following steps are included:

[0008] S1, real-time collection of pollution source data sets, pre-processing to obtain pollution diffusion multi-modal feature vectors, uploading to a remote management center and storing as historical pollution source data sets;

[0009] S2, based on the historical pollution source data sets, setting Gaussian plume diffusion model parameters, calculating pollutant concentration spatial distribution prediction data and constructing a Gaussian plume fidelity simulation model;

[0010] S3, based on the historical pollution source data sets and city three-dimensional space structure data, using CFD simulation software to calculate pollutant concentration spatial distribution prediction data and constructing a CFD fidelity simulation model;

[0011] S4, based on the Gaussian plume fidelity simulation model and the CFD fidelity simulation model, using a random forest regression algorithm to establish a multi-fidelity mapping relationship and constructing a multi-fidelity simulation model based on a Gaussian process proxy model;

[0012] S5, based on the pollution diffusion multi-modal feature vectors, constructing a DPPO reinforcement learning model;

[0013] S6, based on the DPPO reinforcement learning model, using an improved multi-fidelity Bayesian optimization algorithm for continuous training iterations, and finally outputting an optimized DPPO reinforcement learning model;

[0014] S7, real-time input of the pollution diffusion multi-modal feature vector into the optimized DPPO reinforcement learning model, output of the optimal pollution source positioning path decision, pollution source positioning result and credibility evaluation index through the policy network, and output of a standardized pollution source real-time tracing analysis report.

[0015] Optionally, the pollution source dataset includes geographical position coordinates of the pollution source, pollutant types, emission intensity values, emission rate values, emission height values, wind speed, wind direction, temperature, humidity, atmospheric pressure, and atmospheric stability parameters; and the preprocessing includes missing value processing, data noise reduction processing, abnormal data detection and elimination, data synchronization registration processing, feature extraction and fusion.

[0016] Optionally, the Gaussian plume fidelity simulation model specifically includes:

[0017] Based on the historical pollution source dataset, geographical position coordinates of the pollution source, pollutant types, emission intensity values, emission rate values, and emission height values are extracted, and emission information of the Gaussian plume diffusion model is set;

[0018] Taking the geographical position coordinates of the pollution source as the center, a pollution diffusion simulation calculation region is defined and divided into a plurality of standard grid units with a side length of 1 km;

[0019] Based on the historical pollution source dataset, wind speed, wind direction, temperature, humidity, atmospheric pressure, and atmospheric stability parameters are extracted, and wind field conditions of the Gaussian plume diffusion model are set;

[0020] The diffusion coefficient of the Gaussian plume diffusion model is set as an empirical diffusion coefficient, and the effects of terrain undulations and building flow and shielding during pollutant diffusion are ignored;

[0021] The emission information, wind field conditions, and empirical diffusion coefficient are input into the Gaussian plume diffusion formula, the pollutant concentration values of each standard grid unit are calculated, the pollutant concentration spatial distribution prediction data is output in the GeoTIFF standard raster data format, and the Gaussian plume fidelity simulation model is constructed.

[0022] Optionally, the CFD fidelity simulation model specifically includes:

[0023] Based on the urban three-dimensional space structure data, a Cityengine three-dimensional modeling software is used to construct an urban three-dimensional calculation domain model including building three-dimensional geometric contours, road space layout, and terrain elevation, and a structured grid method is used for grid division operation to obtain a three-dimensional calculation domain grid;

[0024] The three-dimensional calculation domain grid is imported into the OpenFOAM simulation software, the inlet boundary, the outlet boundary, the wall boundary and the initial pollutant concentration condition are set, the pollutant source emission intensity, the emission rate, the emission height and the pollutant type are set as the emission boundary condition based on the historical pollutant source data set, the wind speed, the wind direction, the temperature, the humidity, the atmospheric pressure and the atmospheric stability parameter are input as the simulation wind field condition, and the CFD pollutant diffusion numerical calculation model is constructed;

[0025] The steady-state numerical solver simpleFoam based on the finite volume method in OpenFOAM is selected, the continuity equation, the momentum conservation equation and the pollutant diffusion transport equation are solved, and the pollutant concentration spatial distribution prediction data is obtained;

[0026] The pollutant concentration spatial distribution prediction data is post-processed by using the interpolation and smoothing processing method, and the pollutant concentration spatial distribution prediction data with the same spatial resolution as the Gaussian plume fidelity simulation model is obtained.

[0027] The pollutant concentration spatial distribution prediction data is output in the GeoTIFF standard raster data format, and the CFD fidelity simulation model is constructed.

[0028] Optionally, the random forest regression algorithm specifically includes:

[0029] The pollutant concentration spatial distribution prediction data output by the Gaussian plume fidelity simulation model is subjected to data normalization processing, and an input feature vector is obtained;

[0030] The pollutant concentration spatial distribution prediction data output by the CFD fidelity simulation model is subjected to data normalization processing, and a target feature vector is obtained;

[0031] Based on the input feature vector and the target feature vector, a plurality of sub-training data sets are generated by using a random sampling method with replacement;

[0032] A corresponding decision tree model is established based on each sub-training data set, a part of features is randomly extracted when each decision tree model is split, the optimal splitting of the tree node is performed according to the variance reduction principle of the features, and a plurality of decision trees are generated;

[0033] The regression prediction results of all the decision tree models are summarized, the mean value of the prediction results of each decision tree is calculated as the final prediction result of the random forest regression model, and the random forest regression model is output;

[0034] The new Gaussian plume fidelity simulation model pollutant concentration spatial distribution data is input, the corresponding CFD fidelity simulation model pollutant concentration spatial distribution data is predicted by the random forest regression model, and the multi-fidelity model mapping prediction result is output.

[0035] Optionally, the multi-fidelity simulation model specifically includes:

[0036] For both Gaussian plume fidelity simulation models and CFD fidelity simulation models, a unified spatiotemporal reference coordinate system and a unified timestamp standard are defined.

[0037] Establish data input and output interface specifications for Gaussian plume fidelity simulation model and CFD fidelity simulation model, and define specific data format specifications for standardized pollution source location coordinates, pollutant emission intensity values, emission rate values, emission height values, wind speed, wind direction, temperature, humidity, atmospheric pressure, atmospheric stability parameters, and spatial distribution data of pollutant concentration field;

[0038] Spatial distribution data of pollutant concentration fields generated by Gaussian plume fidelity simulation model and CFD fidelity simulation model are spatially interpolated, aligned and normalized under a unified spatiotemporal reference coordinate system and a unified timestamp to construct a multi-fidelity simulation model dataset with consistent input and output data formats.

[0039] The random forest regression algorithm is used, with the spatial distribution prediction data of pollutant concentration output by the Gaussian plume simulation model as input and the spatial distribution prediction data of pollutant concentration output by the CFD simulation model as target. A mapping function between the Gaussian plume simulation model and the CFD simulation model is established as a multi-fidelity mapping function.

[0040] The random forest model is used to predict and output a multifidelity model mapping dataset, which is used as the initial dataset for the Gaussian process surrogate model. The initial kernel function type of the Gaussian process surrogate model is set to RBF kernel function, and the initial values ​​of hyperparameters are set.

[0041] Calculate the covariance matrix of the Gaussian process surrogate model, solve for the posterior mean and prediction variance of the Gaussian process, and obtain the Gaussian process surrogate model.

[0042] Define a prediction variance threshold. When the prediction variance is lower than the prediction variance threshold, call the Gaussian plume fidelity simulation model for evaluation; when the prediction variance is higher than the prediction variance threshold, call the CFD fidelity simulation model for evaluation and obtain the evaluation result.

[0043] The evaluation results are added to the training dataset of the Gaussian process surrogate model. The kernel function parameters and hyperparameters of the Gaussian process surrogate model are re-optimized using the new data. The updated posterior distribution is calculated, and the updated Gaussian process surrogate model is output.

[0044] The Gaussian process surrogate model is continuously updated and iterated until the prediction variance converges.

[0045] Based on the multi-fidelity mapping function, the calling conditions, model switching rules and decision thresholds of the pollution diffusion model are defined, a unified calling interface is constructed, and a multi-fidelity simulation model based on the Gaussian process proxy model is built.

[0046] Optionally, the DPPO reinforcement learning model specifically comprises:

[0047] The pollution diffusion multi-modal feature vector is defined as a state space, and each state corresponds to a feature set of pollution diffusion at a specific time;

[0048] The specific position adjustment of the monitoring device in the spatial grid unit and the discrete adjustment of the monitoring frequency level are defined as an action space;

[0049] The reward function is defined by using a linear combination of the reciprocal of the pollution source positioning error and the reciprocal of the traceability time consumption, and a positive reward is given when the positioning error and the traceability time consumption are both lower than the respective preset thresholds, otherwise a negative reward proportional to the error size or the time consumption length is given;

[0050] Based on the defined state space, action space and reward function, a DPPO reinforcement learning strategy network including two layers of convolutional neural networks to extract spatial features, one layer of long short-term memory network to extract time sequence features and two layers of fully connected network to output action probability, and a DPPO value network including a combination of convolutional neural network and fully connected network are constructed, to obtain the DPPO reinforcement learning model.

[0051] Optionally, the improved multi-fidelity Bayesian optimization algorithm specifically comprises:

[0052] The hyperparameter search space of the DPPO reinforcement learning model is defined;

[0053] Based on the multi-fidelity simulation model, historical data and real-time feedback data are obtained, and the Gaussian plume fidelity simulation model and the CFD fidelity simulation model are run respectively to output corresponding pollutant concentration spatial distribution prediction data;

[0054] After aligning the pollutant concentration spatial distribution prediction data output by the Gaussian plume fidelity simulation model and the CFD fidelity simulation model, the corresponding mean, covariance and standard deviation are calculated, and the Pearson correlation coefficient is obtained by dividing the covariance by the product of the standard deviations of the two models;

[0055] A correlation threshold is set, and the combination of the Gaussian plume fidelity simulation model and the CFD fidelity simulation model with a Pearson correlation coefficient higher than the correlation threshold is selected;

[0056] An active learning strategy is introduced, and based on the multi-fidelity simulation model, the prediction variance of the multi-fidelity Bayesian optimization proxy model is calculated in real time, and when the prediction variance is lower than the threshold, the Gaussian plume fidelity simulation model is used for evaluation, otherwise the CFD fidelity simulation model is actively called for evaluation;

[0057] According to the evaluation result, iteratively update the agent model, and dynamically adjust the calling frequency and switching time of the Gaussian plume fidelity simulation model and the CFD fidelity simulation model until the prediction variance converges;

[0058] A multi-stage hyperparameter optimization mechanism is introduced, based on the Gaussian plume fidelity simulation model, and the Bayesian optimization method is used to quickly search in the effective hyperparameter space to filter out candidate hyperparameter combinations;

[0059] The initial candidate hyperparameter combination is input into the CFD fidelity simulation model, and the pollution source positioning error and source tracing time consumption index corresponding to each candidate combination are calculated to obtain the evaluation result of the CFD fidelity simulation model;

[0060] The evaluation result of the CFD fidelity simulation model is added to the Gaussian process agent model training data set as a new observation data point, and the covariance matrix and posterior distribution of the Gaussian process agent model are recalculated to update the prediction mean and variance of the Gaussian process agent model until the prediction variance converges;

[0061] During the Bayesian optimization iteration process, the model prediction difference of the Gaussian plume fidelity simulation model and the CFD fidelity simulation model is monitored in real time;

[0062] A difference threshold is set, and when the model prediction difference exceeds the difference threshold, the latest correlation analysis result between the Gaussian plume fidelity simulation model and the CFD fidelity simulation model is automatically triggered for re-evaluation;

[0063] According to the latest correlation analysis result, dynamically correct the multi-fidelity model combination strategy, adjust the switching rules and calling frequency of the Gaussian plume fidelity simulation model and the CFD fidelity simulation model, and retrain the Gaussian process agent model;

[0064] Continue to iteratively train until the Gaussian process agent model converges, obtain the optimal hyperparameter combination of the DPPO reinforcement learning model, and retrain the DPPO reinforcement learning strategy network using the optimal hyperparameter combination until the strategy network converges, and obtain the optimized DPPO reinforcement learning model.

[0065] Optionally, the S7 specifically includes:

[0066] S71, input the pollution diffusion multi-modal feature vector into the strategy network of the optimized DPPO reinforcement learning model, and output the optimal pollution source positioning path decision including the sensor device spatial grid coordinates and the monitoring frequency adjustment scheme;

[0067] S72, based on the optimal pollution source positioning path decision, real-time adjust the sensor device position and monitoring frequency, real-time determine the geographical position coordinates of the pollution source, the pollution source type and the emission intensity, and obtain the pollution source positioning result;

[0068] S73, calculate the uncertainty distribution of the pollution source positioning result by using the Monte Carlo method, and generate a credibility evaluation index;

[0069] S74, based on the pollution source positioning path decision, the pollution source positioning result and the credibility evaluation index, generate a standardized pollution source real-time tracing analysis report and upload it to the remote management center.

[0070] The beneficial effects of the present application are:

[0071] (1) The present application successfully overcomes the shortcomings of traditional urban atmospheric pollution real-time monitoring and tracing methods in dealing with dynamic and complex pollution diffusion scenarios by combining the DPPO reinforcement learning model and the improved multi-fidelity Bayesian optimization algorithm. Traditional methods usually rely on fixed monitoring points and manual analysis experience, and these methods perform poorly in real-time capturing of pollution diffusion dynamic changes and rapid and accurate source tracing. By introducing the DPPO reinforcement learning model, the present application can dynamically predict the pollution diffusion trend and the pollution source location in real time, and automatically adjust the sensor layout and monitoring frequency through the strategy network of reinforcement learning, significantly improving the efficiency and accuracy of pollution tracing. The improved multi-fidelity Bayesian optimization algorithm introduces an active learning strategy and a multi-stage hyperparameter optimization mechanism to adaptively and dynamically adjust the calling frequency and switching time of different fidelity models, effectively solving the problem of high computational cost, precision and speed balance in traditional methods, and significantly improving the performance of the pollution monitoring and tracing system. This method provides an efficient and accurate intelligent solution for urban atmospheric pollution real-time monitoring and accurate source tracing, which can respond to pollution events in real time, locate pollution sources in a timely manner, and promote the progress of environmental monitoring and management technology.

[0072] (2) The present application realizes accurate monitoring and rapid tracing of urban atmospheric pollution diffusion and pollution source positioning by integrating the DPPO reinforcement learning model, the improved multi-fidelity Bayesian optimization algorithm and the multi-modal data fusion technology. The pollution diffusion process in urban environment is usually accompanied by complex meteorological conditions, terrain characteristics and diversified pollution source characteristics, and traditional methods are difficult to accurately and real-time capture the influence of these complex environmental information. Through the multi-modal data fusion technology, the present application can collect air quality sensor data, meteorological data and remote sensing image data in real time, and extract the spatio-temporal feature information of pollution diffusion, and efficiently predict the pollution diffusion path and pollution source location through the DPPO reinforcement learning model. In addition, the improved multi-fidelity Bayesian optimization algorithm effectively handles the simulation accuracy problem under different pollution diffusion scenarios, automatically selects the appropriate hyperparameter combination and simulation model calling strategy, and further improves the system prediction accuracy and response speed. Through this method, the urban atmospheric pollution diffusion trend and pollution source location can be identified in a timely and accurate manner, effectively improving the response capability and management efficiency of pollution events, and providing important technical support for environmental protection. BRIEF DESCRIPTION OF DRAWINGS

[0073] The accompanying drawings are included to provide a further understanding of the application and are incorporated in and constitute a part of this specification, illustrate embodiments of the application and are meant to explain the present application and are not intended to limit the application. In the drawings:

[0074] Fig. 1 A flowchart of a city atmospheric pollution real-time monitoring and tracing method based on reinforcement learning proposed by the present application;

[0075] Fig. 2 A flowchart of an improved multi-fidelity Bayesian optimization algorithm for city atmospheric pollution real-time monitoring based on reinforcement learning proposed by the present application. DETAILED DESCRIPTION

[0076] The present application will now be described in further detail with reference to the drawings. These drawings show only the essential features of the application and are therefore to be regarded only as a schematic representation of the basic structure of the application.

[0077] Reference Figs. 1-2 A city atmospheric pollution real-time monitoring and tracing method based on reinforcement learning, comprising the following steps:

[0078] S1, real-time collection of pollution source data set, pre-processing to obtain pollution diffusion multi-modal feature vector, uploading to remote management center, storing as historical pollution source data set;

[0079] S2, based on the historical pollution source data set, setting the Gaussian plume diffusion model parameters, calculating the pollutant concentration spatial distribution prediction data, constructing the Gaussian plume fidelity simulation model;

[0080] S3, based on the historical pollution source data set and the city three-dimensional space structure data, using CFD simulation software, calculating the pollutant concentration spatial distribution prediction data, constructing the CFD fidelity simulation model;

[0081] S4, based on the Gaussian plume fidelity simulation model and the CFD fidelity simulation model, using the random forest regression algorithm to establish the multi-fidelity mapping relationship, constructing the multi-fidelity simulation model based on the Gaussian process proxy model;

[0082] S5, based on the pollution diffusion multi-modal feature vector, constructing the DPPO reinforcement learning model;

[0083] S6, based on the DPPO reinforcement learning model, using the improved multi-fidelity Bayesian optimization algorithm to continuously train and iterate, finally outputting the optimized DPPO reinforcement learning model;

[0084] S7, input the pollution diffusion multi-modal feature vector into the optimized DPPO reinforcement learning model in real time, output the optimal pollution source positioning path decision, pollution source positioning result and credibility evaluation index through the strategy network, and output the standardized pollution source real-time tracing analysis report.

[0085] In the embodiment, the pollution source dataset includes geographical position coordinates of the pollution source, pollutant types, emission intensity values, emission rate values, emission height values, wind speed, wind direction, temperature, humidity, atmospheric pressure, and atmospheric stability parameters; and the preprocessing includes missing value processing, data noise reduction processing, abnormal data detection and elimination, data synchronization registration processing, feature extraction and fusion.

[0086] In the embodiment, the preprocessing includes missing value processing, which effectively fills in data gaps and guarantees data integrity. The data noise reduction processing effectively reduces noise interference in the data and improves data quality and stability. The abnormal data detection and elimination accurately removes abnormal values in the data and enhances the accuracy and credibility of the data. The data synchronization registration processing precisely aligns different sources of data in space and time, providing a unified data basis for subsequent data fusion analysis. The feature extraction and fusion extracts and fuses key information of multi-modal data into feature vectors, further improving the recognition accuracy and generalization ability of the pollution diffusion monitoring and pollution source positioning model.

[0087] In the embodiment, the Gaussian plume fidelity simulation model specifically includes:

[0088] Based on the historical pollution source dataset, the geographical position coordinates of the pollution source, the pollutant types, the emission intensity values, the emission rate values, and the emission height values are extracted, and the emission information of the Gaussian plume diffusion model is set;

[0089] Taking the geographical position coordinates of the pollution source as the center, the pollution diffusion simulation calculation region is defined and divided into a plurality of standard grid cells with a side length of 1 km;

[0090] Based on the historical pollution source dataset, the wind speed, the wind direction, the temperature, the humidity, the atmospheric pressure, and the atmospheric stability parameters are extracted, and the wind field conditions of the Gaussian plume diffusion model are set;

[0091] The diffusion coefficient of the Gaussian plume diffusion model is set as an empirical diffusion coefficient, and the effects of terrain undulations and building flow and shielding in the pollution diffusion process are ignored;

[0092] The emission information, the wind field conditions, and the empirical diffusion coefficient are input into the Gaussian plume diffusion formula, the pollutant concentration values of each standard grid cell are calculated, the pollutant concentration spatial distribution prediction data is output in the GeoTIFF standard raster data format, and the Gaussian plume fidelity simulation model is constructed.

[0093] The embodiment realizes rapid simulation and prediction of the spatial distribution of pollutant concentration by constructing and parameterizing the Gaussian plume diffusion model, and provides efficient and accurate basic data support for analysis of the diffusion trend of urban air pollution. The emission information, wind field conditions and empirical diffusion coefficient are input into the Gaussian plume diffusion formula to calculate the pollutant concentration value of each standard grid unit, and the pollutant concentration spatial distribution prediction data is output in the GeoTIFF standard grid data format, a Gaussian plume faithful simulation model is constructed, and the standardization, ease of use and convenience of data sharing of pollution diffusion prediction are significantly improved.

[0094] In the embodiment, the CFD faithful simulation model specifically comprises:

[0095] Based on the urban three-dimensional spatial structure data, a three-dimensional calculation domain model of the city including the three-dimensional geometric contour of the building, the spatial layout of the road and the terrain elevation is constructed by using the CityEngine three-dimensional modeling software, and a structured grid method is used to perform a grid division operation to obtain a three-dimensional calculation domain grid;

[0096] The three-dimensional calculation domain grid is imported into the OpenFOAM simulation software, the inlet boundary, the outlet boundary, the wall boundary and the initial pollutant concentration condition are set, the emission intensity, the emission rate, the emission height and the pollutant type of the pollution source are set as the emission boundary condition based on the historical pollution source data set, the wind speed, the wind direction, the temperature, the humidity, the atmospheric pressure and the atmospheric stability parameter are input as the simulation wind field condition, and a CFD pollutant diffusion numerical calculation model is constructed;

[0097] The steady-state numerical solver simpleFoam based on the finite volume method in OpenFOAM is selected, the continuity equation, the momentum conservation equation and the pollutant diffusion transport equation are solved, and the pollutant concentration spatial distribution prediction data is obtained;

[0098] The pollutant concentration spatial distribution prediction data is post-processed by using the interpolation and smoothing processing method to obtain the pollutant concentration spatial distribution prediction data with the same spatial resolution as the Gaussian plume faithful simulation model;

[0099] The pollutant concentration spatial distribution prediction data is output in the GeoTIFF standard grid data format, and the CFD faithful simulation model is constructed.

[0100] The embodiment accurately captures the detailed characteristics of pollutant diffusion in a complex urban environment by constructing a refined CFD fidelity simulation model, providing high-precision data support for accurate pollution source positioning and tracing. The steady-state numerical solver simpleFoam based on the finite volume method in OpenFOAM is selected, and by solving the continuity equation, momentum conservation equation and pollutant diffusion transport equation, the pollutant concentration spatial distribution prediction data is obtained, which accurately simulates the concentration spatial distribution of pollutants in the urban environment during the diffusion process, ensuring the high precision and reliability of the simulation prediction results. The interpolation and smoothing processing method is used for post-processing of the pollutant concentration spatial distribution prediction data, and the same spatial resolution as the Gaussian plume fidelity simulation model is obtained, which effectively improves the spatial compatibility and data fusion convenience of different fidelity model results.

[0101] In the embodiment, the random forest regression algorithm specifically includes:

[0102] The pollutant concentration spatial distribution prediction data output by the Gaussian plume fidelity simulation model is subjected to data normalization processing to obtain an input feature vector;

[0103] The pollutant concentration spatial distribution prediction data output by the CFD fidelity simulation model is subjected to data normalization processing to obtain a target feature vector;

[0104] Based on the input feature vector and the target feature vector, a plurality of sub-training data sets are generated by using a random sampling method with replacement;

[0105] Based on each sub-training data set, a corresponding decision tree model is established, and each decision tree model randomly extracts part of the features when splitting the nodes, and performs optimal splitting of the tree nodes according to the variance reduction principle of the features to generate a plurality of decision trees;

[0106] The regression prediction results of all decision tree models are summarized, the mean value of each decision tree prediction result is calculated as the final prediction result of the random forest regression model, and the random forest regression model is output;

[0107] The new Gaussian plume fidelity simulation model pollutant concentration spatial distribution data is input, the corresponding CFD fidelity simulation model pollutant concentration spatial distribution data is predicted by the random forest regression model, and the multi-fidelity model mapping prediction result is output.

[0108] The embodiment constructs a random forest regression model, accurately establishes a data mapping relationship between a Gaussian plume fidelity model and a CFD fidelity model, and significantly improves the prediction accuracy and generalization performance of the multi-fidelity simulation environment. Based on the input feature vector and the target feature vector, a random sampling method with replacement is used to generate multiple sub-training data sets, which significantly improves the data diversity and robustness of the model training process, and effectively reduces the variance and overfitting risk of the model.

[0109] In the embodiment, the multi-fidelity simulation model specifically includes:

[0110] For the Gaussian plume fidelity simulation model and the CFD fidelity simulation model, a unified space-time reference coordinate system and a unified timestamp standard are defined;

[0111] The data input and output interface specifications of the Gaussian plume fidelity simulation model and the CFD fidelity simulation model are established, and the specific data format specifications of the pollution source position coordinates, the pollution emission intensity values, the emission rate values, the emission height values, the wind speed, the wind direction, the temperature, the humidity, the atmospheric pressure, the atmospheric stability parameters and the pollutant concentration field spatial distribution data are defined;

[0112] The pollutant concentration field spatial distribution data generated by the Gaussian plume fidelity simulation model and the CFD fidelity simulation model are spatially interpolated, aligned and normalized under the unified space-time reference coordinate system and the unified timestamp, and a multi-fidelity simulation model data set with consistent input and output data formats is constructed;

[0113] A random forest regression algorithm is used, the pollutant concentration spatial distribution prediction data output by the Gaussian plume fidelity simulation model is taken as the input, and the pollutant concentration spatial distribution prediction data output by the CFD fidelity simulation model is taken as the target, a mapping function between the Gaussian plume fidelity simulation model and the CFD fidelity simulation model is established, and the mapping function is used as a multi-fidelity mapping function;

[0114] The random forest model is used to predict and output the multi-fidelity model mapping data set as the initial data set of the Gaussian process proxy model, the initial kernel function type of the Gaussian process proxy model is set to RBF kernel function, and the initial hyperparameter value is set;

[0115] The covariance matrix of the Gaussian process proxy model is calculated, the posterior mean and the prediction variance of the Gaussian process are solved, and the Gaussian process proxy model is obtained;

[0116] A prediction variance threshold is defined, when the prediction variance is lower than the prediction variance threshold, the Gaussian plume fidelity simulation model is called for evaluation; when the prediction variance is higher than the prediction variance threshold, the CFD fidelity simulation model is called for evaluation, and an evaluation result is obtained;

[0117] The evaluation result is added to the training data set of the Gaussian process proxy model, the kernel function parameters and hyperparameters of the Gaussian process proxy model are re-optimized using the newly added data, the updated posterior distribution is calculated, and the updated Gaussian process proxy model is output;

[0118] The iterative Gaussian process proxy model is continuously updated until the prediction variance converges.

[0119] Based on the multi-fidelity mapping function, the calling conditions, model switching rules and decision thresholds of the pollution diffusion model are defined, a unified calling interface is constructed, and a multi-fidelity simulation model based on the Gaussian process proxy model is constructed.

[0120] The embodiment realizes intelligent adaptive calling of the Gaussian plume fidelity simulation model and the CFD fidelity simulation model by constructing a multi-fidelity simulation model based on the Gaussian process proxy model, effectively balancing the calculation efficiency and accuracy of simulation prediction. The evaluation result is added to the training data set of the Gaussian process proxy model, the kernel function parameters and hyperparameters of the Gaussian process proxy model are re-optimized using the newly added data, the updated posterior distribution is calculated, and the updated Gaussian process proxy model is output, realizing dynamic adaptive updating of the Gaussian process proxy model and ensuring continuous optimization of the proxy model.

[0121] In the embodiment, the DPPO reinforcement learning model specifically includes:

[0122] The pollution diffusion multi-modal feature vector is defined as a state space, and each state corresponds to a feature set of pollution diffusion at a specific time;

[0123] The specific position adjustment of the monitoring device in the spatial grid unit and the discrete adjustment of the monitoring frequency level are defined as an action space;

[0124] The reward function is defined by the linear combination of the reciprocal of the pollution source positioning error and the reciprocal of the time consumption of source tracing, and a positive reward is given when the positioning error and the time consumption of source tracing are both lower than the respective preset thresholds, otherwise a negative reward proportional to the error size or the time consumption length is given;

[0125] Based on the defined state space, action space and reward function, a DPPO reinforcement learning strategy network including two layers of convolutional neural networks to extract spatial features, one layer of long short-term memory network to extract time sequence features, and two layers of fully connected network to output action probability, and a DPPO value network including a combination of convolutional neural network and fully connected network are constructed to obtain the DPPO reinforcement learning model.

[0126] The embodiment realizes real-time dynamic analysis and decision of pollution diffusion characteristic information by constructing a DPPO-based reinforcement learning model, effectively improving the real-time performance and intelligent level of the pollution source positioning and tracing process. The pollution diffusion multi-modal feature vector is defined as the state space, each state corresponds to a specific time pollution diffusion feature set, which clearly defines the model state information and ensures that the reinforcement learning model can accurately perceive the dynamic change characteristics of the pollution diffusion process. The specific position adjustment of the monitoring device in the spatial grid unit and the discrete adjustment of the monitoring frequency level are defined as the action space, which clearly defines the range and granularity of the decision action, making the model decision result easy to be actually executed and applied. The reward function is defined by the linear combination of the reciprocal of the pollution source positioning error and the reciprocal of the tracing time consumption. When the positioning error and the tracing time consumption are both lower than the respective preset threshold, a positive reward is given, otherwise a negative reward proportional to the error size or the time consumption length is given, effectively encouraging the reinforcement learning model to choose more accurate and faster pollution source positioning strategies.

[0127] In the embodiment, the improved multi-fidelity Bayesian optimization algorithm specifically includes:

[0128] Defining the hyperparameter search space of the DPPO reinforcement learning model;

[0129] Based on the multi-fidelity simulation model, historical data and real-time feedback data are obtained, and Gaussian plume fidelity simulation model and CFD fidelity simulation model are run respectively to output corresponding pollutant concentration spatial distribution prediction data;

[0130] After aligning the pollutant concentration spatial distribution prediction data output by the Gaussian plume fidelity simulation model and the CFD fidelity simulation model, the corresponding mean, covariance and standard deviation are calculated, and the Pearson correlation coefficient is obtained by dividing the covariance by the product of the standard deviations of the two models;

[0131] Set a correlation threshold, and select the combination of Gaussian plume fidelity simulation model and CFD fidelity simulation model whose Pearson correlation coefficient is higher than the correlation threshold;

[0132] Introducing an active learning strategy, based on the multi-fidelity simulation model, the prediction variance of the multi-fidelity Bayesian optimization proxy model is calculated in real time, when the prediction variance is lower than the threshold, the Gaussian plume fidelity simulation model is evaluated, otherwise the CFD fidelity simulation model is called actively for evaluation;

[0133] According to the evaluation result, the proxy model is iteratively updated, and the calling frequency and switching time of the Gaussian plume fidelity simulation model and the CFD fidelity simulation model are dynamically adjusted until the prediction variance converges;

[0134] A multi-stage hyperparameter optimization mechanism is introduced, a Gaussian plume fidelity simulation model is used, and a Bayesian optimization method is used to quickly search in the effective hyperparameter space to screen out candidate hyperparameter combinations;

[0135] The initial candidate hyperparameter combination is input into the CFD fidelity simulation model, the positioning error and the source tracing time of each candidate combination are calculated, and the evaluation results of the CFD fidelity simulation model are obtained;

[0136] The evaluation results of the CFD fidelity simulation model are added to the training data set of the Gaussian process proxy model as new observation data points, the covariance matrix and the posterior distribution of the Gaussian process proxy model are recalculated, and the prediction mean and variance of the Gaussian process proxy model are updated until the prediction variance converges;

[0137] In the Bayesian optimization iteration process, the model prediction difference between the Gaussian plume fidelity simulation model and the CFD fidelity simulation model is monitored in real time;

[0138] A difference threshold is set, and when the model prediction difference exceeds the difference threshold, the latest correlation analysis results between the Gaussian plume fidelity simulation model and the CFD fidelity simulation model are automatically triggered for reevaluation;

[0139] According to the latest correlation analysis results, the multi-fidelity model combination strategy is dynamically modified, the switching rules and calling frequency of the Gaussian plume fidelity simulation model and the CFD fidelity simulation model are adjusted, and the Gaussian process proxy model is retrained;

[0140] The training is continuously iterated until the Gaussian process proxy model converges, the optimal hyperparameter combination of the DPPO reinforcement learning model is obtained, the optimal hyperparameter combination is used to retrain the DPPO reinforcement learning strategy network, and the training is iterated until the strategy network converges, and the optimized DPPO reinforcement learning model is obtained.

[0141] The embodiment realizes precise optimization and efficient search of reinforcement learning hyperparameters by combining the multi-fidelity Bayesian optimization algorithm and the DPPO reinforcement learning model, significantly improves the decision accuracy and generalization ability of the DPPO model in the task of urban air pollution monitoring and source tracing. The active learning strategy is introduced, the prediction variance of the multi-fidelity Bayesian optimization proxy model is calculated in real time based on the multi-fidelity simulation model, when the prediction variance is lower than the threshold, the Gaussian plume fidelity simulation model is evaluated, otherwise the CFD fidelity simulation model is called actively for evaluation, the proxy model is iteratively updated according to the evaluation results, and the calling frequency and switching time of the Gaussian plume fidelity simulation model and the CFD fidelity simulation model are dynamically adjusted until the prediction variance converges, realizing the dynamic balance of model prediction accuracy and calculation cost, and effectively improving the model running efficiency and self-adaptation ability.

[0142] In the embodiment, the S7 specifically comprises:

[0143] S71, input the pollution diffusion multi-modal feature vector into the strategy network of the optimized DPPO reinforcement learning model, and output an optimal pollution source positioning path decision containing a sensor device spatial grid coordinate and a monitoring frequency adjustment scheme;

[0144] S72, based on the optimal pollution source positioning path decision, real-time adjustment of the sensor device position and monitoring frequency, real-time determination of the geographical position coordinates of the pollution source, the pollution source type and the emission intensity, and obtaining the pollution source positioning result;

[0145] S73, using the Monte Carlo method to calculate the uncertainty distribution of the pollution source positioning result, and generating a credibility evaluation index;

[0146] S74, based on the pollution source positioning path decision, the pollution source positioning result and the credibility evaluation index, generating a standardized pollution source real-time tracing analysis report and uploading it to the remote management center.

[0147] The embodiment realizes real-time intelligent decision and dynamic precise adjustment of the pollution source positioning path through the strategy network of the optimized DPPO reinforcement learning model, effectively improving the accuracy, timeliness and management convenience of pollution source tracing. Based on the optimal pollution source positioning path decision, the sensor device position and monitoring frequency are adjusted in real time, the geographical position coordinates of the pollution source, the pollution source type and the emission intensity are determined in real time, and the pollution source positioning result is obtained, realizing the rapid and accurate real-time positioning of the pollution source, providing direct and effective data support for pollution prevention and control. The Monte Carlo method is used to calculate the uncertainty distribution of the pollution source positioning result, and a credibility evaluation index is generated, which clearly reflects the reliability of the pollution source positioning result, effectively improving the transparency and risk assessment ability of the decision-making process.

[0148] Example 1:

[0149] In order to verify the feasibility of the present application in real-time monitoring of urban air pollution and pollution source tracing, the present application is applied to the pollution diffusion monitoring and pollution source positioning system of a city environmental monitoring and pollution source tracing management platform (hereinafter referred to as "platform W"). In the traditional pollution monitoring and tracing system, fixed monitoring sites and diffusion analysis methods based on empirical rules are usually used. These methods not only have a complex process and slow response, but also are difficult to cope with the dynamic changes and multi-source complex characteristics of pollution diffusion in urban environment, resulting in insufficient positioning accuracy of pollution sources and low monitoring efficiency. In order to solve the above problems, platform W decides to use the urban air pollution real-time monitoring and tracing method based on the DPPO reinforcement learning algorithm and the multi-fidelity Bayesian optimization algorithm proposed by the present application.

[0150] In the implementation process, the platform W first collects particulate matter concentration data, nitrogen dioxide data, carbon monoxide data, ozone data, and volatile organic compound data, as well as wind speed, wind direction, temperature, humidity, atmospheric pressure, and atmospheric stability parameters in real time through air quality sensors, meteorological sensors, and remote sensing equipment in urban areas, performs data cleaning, interpolation, noise reduction, data synchronization registration, and multi-modal feature extraction and fusion, and forms a unified pollution diffusion multi-modal feature vector. Through the optimized DPPO reinforcement learning model, the platform W realizes real-time and accurate prediction of pollution diffusion trends and pollution source positioning paths. The DPPO model extracts spatial and temporal features through convolutional neural networks and long short-term memory networks, dynamically optimizes the spatial layout and monitoring frequency of sensors through the strategy network, and effectively improves the real-time and accuracy of pollution source positioning. Through the multi-fidelity Bayesian optimization algorithm, the platform W realizes accurate optimization of the DPPO model hyperparameter combination. The multi-fidelity Bayesian optimization model accurately balances the calling frequency and switching timing of the Gaussian plume model and the CFD model through the Gaussian process proxy model and the active learning strategy, and constructs an accurate mapping relationship between the Gaussian plume model and the CFD model through the random forest regression, achieving efficient consideration of model prediction accuracy and computational efficiency.

[0151] By combining the DPPO reinforcement learning algorithm and the multi-fidelity Bayesian optimization algorithm, the platform W successfully realizes real-time and accurate monitoring of urban pollution diffusion and rapid tracing of pollution sources, significantly improving pollution event response efficiency and environmental management capability. In the implementation process, the technical team of the platform W found that compared with traditional methods, the method of the present application significantly improves the accuracy of pollution source tracing and the accuracy of pollution diffusion trend prediction.

[0152] To further verify the effectiveness of the method of the present application, the platform W compared the performance of the method of the present application with that of traditional methods, and the specific comparison data are shown in Table 1:

[0153] Table 1 Performance comparison table of pollution monitoring and tracing management system of platform W

[0154]

[0155]

[0156] As can be seen from Table 1, under the application of the method of the present application, the pollution source positioning accuracy, pollution diffusion trend prediction accuracy and overall operation efficiency of the urban atmospheric pollution real-time monitoring and source tracing system are all significantly improved. The pollution source positioning accuracy is improved from 73.5% of the traditional method to 95.4%, the pollution diffusion trend prediction accuracy is improved from 71.2% to 93.7%, significantly enhancing the real-time identification and prediction ability of the system for pollution events and diffusion trends. The average pollution source tracing time is shortened from 120 minutes to 20 minutes, with an improvement of 83.3%. The system data processing capacity is increased from 250 per second to 800 per second, and the number of monitored areas is increased from 2 to 8, significantly improving the system concurrent processing capacity and data analysis efficiency. The monthly attendance frequency of the operation personnel is reduced from 30 times to 10 times, significantly reducing the need for manual intervention. The overall monthly environmental operation cost is also reduced from 1.2 million yuan to 0.65 million yuan, with a cost saving rate of 45.8%.

[0157] Through the method of the present application, the platform W can more accurately and efficiently complete the real-time monitoring and rapid tracing of urban atmospheric pollution, track the pollution diffusion path in real time and locate the pollution source in advance, greatly improving the safety and reliability of environmental monitoring and management. The present application not only improves the degree of automation and system operation efficiency, significantly reduces the cost of manual operation, but also enhances the stability and robustness of the monitoring system, providing strong technical support for intelligent monitoring and management of urban environment.

[0158] The above is only the preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art within the technical scope disclosed by the present application, according to the technical solution and inventive concept of the present application, can make equivalent replacement or change, which should be covered within the protection scope of the present application.

Claims

1. A method for real-time monitoring and source tracing of urban air pollution based on reinforcement learning, characterized in that, Includes the following steps: S1. Collect pollution source datasets in real time, preprocess them to obtain multimodal feature vectors of pollution diffusion, and upload them to the remote management center for storage as historical pollution source datasets. S2. Based on historical pollution source datasets, set the parameters of the Gaussian plume diffusion model, calculate the predicted data of spatial distribution of pollutant concentrations, and construct a Gaussian plume fidelity simulation model. S3. Based on historical pollution source datasets and urban three-dimensional spatial structure data, CFD simulation software is used to calculate the spatial distribution prediction data of pollutant concentrations and construct a CFD simulation model. S4. Based on the Gaussian plume fidelity simulation model and the CFD fidelity simulation model, a multi-fidelity mapping relationship is established using the random forest regression algorithm, and a multi-fidelity simulation model based on the Gaussian process surrogate model is constructed. S5. Construct a DPPO reinforcement learning model based on the multimodal feature vectors of pollution diffusion; S6. Based on the DPPO reinforcement learning model, the improved multi-fidelity Bayesian optimization algorithm is used to continuously train and iterate, and finally the optimized DPPO reinforcement learning model is output. S7. Input the multimodal feature vector of pollution diffusion into the optimized DPPO reinforcement learning model in real time, and output the optimal pollution source location path decision, pollution source location result and credibility assessment index through the policy network, and output a standardized real-time pollution source tracing analysis report.

2. The method for real-time monitoring and source tracing of urban air pollution based on reinforcement learning according to claim 1, characterized in that, The pollution source dataset includes the geographical coordinates of the pollution source, pollutant type, emission intensity value, emission rate value, emission height value, wind speed, wind direction, temperature, humidity, atmospheric pressure, and atmospheric stability parameters; the preprocessing includes missing value processing, data noise reduction processing, abnormal data detection and removal, data synchronization and registration processing, and feature extraction and fusion.

3. The method for real-time monitoring and source tracing of urban air pollution based on reinforcement learning according to claim 1, characterized in that, The Gaussian plume fidelity simulation model specifically includes: Based on historical pollution source datasets, we extract the geographical coordinates of pollution sources, pollutant types, emission intensity values, emission rate values, and emission height values, and set the emission information for the Gaussian plume diffusion model. The pollution diffusion simulation calculation area is defined with the geographical coordinates of the pollution source as the center and divided into multiple standard grid cells with a side length of 1km. Based on historical pollution source datasets, wind speed, wind direction, temperature, humidity, atmospheric pressure, and atmospheric stability parameters are extracted, and wind field conditions for the Gaussian plume diffusion model are set. The diffusion coefficient of the Gaussian plume diffusion model is set as the empirical diffusion coefficient, and the effects of topographic undulations and building flow around and shielding during the pollutant diffusion process are ignored. Emission information, wind field conditions, and empirical diffusion coefficients are input into the Gaussian plume diffusion formula to calculate the pollutant concentration value for each standard grid cell. The spatial distribution prediction data of pollutant concentration is output in GeoTIFF standard raster data format to construct a Gaussian plume fidelity simulation model.

4. The method for real-time monitoring and source tracing of urban air pollution based on reinforcement learning according to claim 1, characterized in that, The CFD high-fidelity simulation model specifically includes: Based on urban three-dimensional spatial structure data, a three-dimensional computational domain model of the city, including the three-dimensional geometric outline of buildings, road spatial layout and terrain elevation, was constructed using CityEngine three-dimensional modeling software. The structured mesh method was used to perform mesh division operation to obtain the three-dimensional computational domain mesh. The three-dimensional computational domain mesh was imported into the OpenFOAM simulation software. Inlet boundary, outlet boundary, wall boundary and initial pollutant concentration conditions were set. Based on the historical pollution source dataset, the emission intensity, emission rate, emission height and pollutant type of the pollution source were set as emission boundary conditions. Wind speed, wind direction, temperature, humidity, atmospheric pressure and atmospheric stability parameters were input as simulation wind field conditions to construct a CFD pollutant diffusion numerical calculation model. We selected simpleFoam, a steady-state numerical solver based on the finite volume method in OpenFOAM, and obtained the spatial distribution prediction data of pollutant concentration by solving the continuity equation, momentum conservation equation and pollutant diffusion and transport equation. The spatial distribution prediction data of pollutant concentration is post-processed using interpolation and smoothing methods to obtain spatial distribution prediction data of pollutant concentration with the same spatial resolution as the Gaussian plume fidelity simulation model. The spatial distribution prediction data of pollutant concentration is output using the GeoTIFF standard raster data format, and a CFD simulation model is constructed.

5. The method for real-time monitoring and source tracing of urban air pollution based on reinforcement learning according to claim 1, characterized in that, The random forest regression algorithm specifically includes: The spatial distribution prediction data of pollutant concentrations output by the Gaussian plume fidelity simulation model are normalized to obtain the input feature vector. The spatial distribution prediction data of pollutant concentrations output by the CFD fidelity simulation model are normalized to obtain the target feature vector. Based on the input feature vector and the target feature vector, multiple sub-training datasets are generated using random sampling with replacement. A corresponding decision tree model is built based on each sub-training dataset. Each decision tree model randomly extracts some features when splitting nodes, and performs optimal splitting of tree nodes according to the principle of reducing feature variance, generating multiple decision trees. Summarize the regression prediction results of all decision tree models, calculate the mean of the prediction results of each decision tree as the final prediction result of the random forest regression model, and output the random forest regression model. Input the spatial distribution data of pollutant concentrations from the new Gaussian plume fidelity simulation model, predict the spatial distribution data of pollutant concentrations from the corresponding CFD fidelity simulation model using the random forest regression model, and output the multi-fidelity model mapping prediction results.

6. The method for real-time monitoring and source tracing of urban air pollution based on reinforcement learning according to claim 1, characterized in that, The multi-fidelity simulation model specifically includes: For both Gaussian plume fidelity simulation models and CFD fidelity simulation models, a unified spatiotemporal reference coordinate system and a unified timestamp standard are defined. Establish data input and output interface specifications for Gaussian plume fidelity simulation model and CFD fidelity simulation model, and define specific data format specifications for standardized pollution source location coordinates, pollutant emission intensity values, emission rate values, emission height values, wind speed, wind direction, temperature, humidity, atmospheric pressure, atmospheric stability parameters, and spatial distribution data of pollutant concentration field; Spatial distribution data of pollutant concentration fields generated by Gaussian plume fidelity simulation model and CFD fidelity simulation model are spatially interpolated, aligned and normalized under a unified spatiotemporal reference coordinate system and a unified timestamp to construct a multi-fidelity simulation model dataset with consistent input and output data formats. The random forest regression algorithm is used, with the spatial distribution prediction data of pollutant concentration output by the Gaussian plume simulation model as input and the spatial distribution prediction data of pollutant concentration output by the CFD simulation model as target. A mapping function between the Gaussian plume simulation model and the CFD simulation model is established as a multi-fidelity mapping function. The random forest model is used to predict and output a multifidelity model mapping dataset, which is used as the initial dataset for the Gaussian process surrogate model. The initial kernel function type of the Gaussian process surrogate model is set to RBF kernel function, and the initial values ​​of hyperparameters are set. Calculate the covariance matrix of the Gaussian process surrogate model, solve for the posterior mean and prediction variance of the Gaussian process, and obtain the Gaussian process surrogate model. Define a prediction variance threshold. When the prediction variance is lower than the prediction variance threshold, call the Gaussian plume fidelity simulation model for evaluation; when the prediction variance is higher than the prediction variance threshold, call the CFD fidelity simulation model for evaluation and obtain the evaluation result. The evaluation results are added to the training dataset of the Gaussian process surrogate model. The kernel function parameters and hyperparameters of the Gaussian process surrogate model are re-optimized using the new data. The updated posterior distribution is calculated, and the updated Gaussian process surrogate model is output. The Gaussian process surrogate model is continuously updated and iterated until the prediction variance converges. Based on the multi-fidelity mapping function, the calling conditions, model switching rules and decision thresholds of the pollution diffusion model are defined, a unified calling interface is constructed, and a multi-fidelity simulation model based on the Gaussian process proxy model is built.

7. The method for real-time monitoring and source tracing of urban air pollution based on reinforcement learning according to claim 1, characterized in that, The DPPO reinforcement learning model specifically includes: The multimodal feature vector of pollution diffusion is defined as a state space, where each state corresponds to a set of features of pollution diffusion at a specific time. The specific location adjustment of the monitoring equipment within the spatial grid unit and the discretization adjustment of the monitoring frequency level are defined as the action space; A reward function is defined by a linear combination of the reciprocal of the pollution source location error and the reciprocal of the source tracing time. When both the location error and the source tracing time are lower than their respective preset thresholds, a positive reward is given; otherwise, a negative reward proportional to the error magnitude or the time duration is given. Based on the defined state space, action space, and reward function, a DPPO reinforcement learning policy network is constructed, which includes two layers of convolutional neural network to extract spatial features, one layer of long short-term memory network to extract temporal features, and two layers of fully connected network to output action probabilities. A DPPO value network, which includes a combination of convolutional neural network and fully connected network, is also constructed, resulting in the DPPO reinforcement learning model.

8. The method for real-time monitoring and source tracing of urban air pollution based on reinforcement learning according to claim 1, characterized in that, The improved multi-fidelity Bayesian optimization algorithm includes the following steps: Define the hyperparameter search space for the DPPO reinforcement learning model; Based on the multi-fidelity simulation model, historical data and real-time feedback data are acquired, and the Gaussian plume fidelity simulation model and CFD fidelity simulation model are run respectively to output the corresponding spatial distribution prediction data of pollutant concentration. After aligning the spatial distribution prediction data of pollutant concentrations output by the Gaussian plume fidelity simulation model and the CFD fidelity simulation model, the corresponding mean, covariance and standard deviation are calculated respectively. The corresponding Pearson correlation coefficient is obtained by dividing the covariance by the product of the standard deviations of the two models. Set a correlation threshold and select a combination of Gaussian plume fidelity simulation model and CFD fidelity simulation model with a Pearson correlation coefficient higher than the correlation threshold. An active learning strategy is introduced. Based on a multi-fidelity simulation model, the prediction variance of the multi-fidelity Bayesian optimized surrogate model is calculated in real time. When the prediction variance is lower than the threshold, the Gaussian plume fidelity simulation model is used for evaluation; otherwise, the CFD fidelity simulation model is actively called for evaluation. The proxy model is iteratively updated based on the evaluation results, and the frequency of calling and switching of the Gaussian plume fidelity simulation model and the CFD fidelity simulation model are dynamically adjusted until the prediction variance converges. A multi-stage hyperparameter optimization mechanism is introduced. Based on the Gaussian plume fidelity simulation model, the Bayesian optimization method is used to quickly search within the effective hyperparameter space and select candidate hyperparameter combinations. The initial candidate hyperparameter combinations are input into the CFD simulation model, and the pollution source location error and source tracing time corresponding to each candidate combination are calculated to obtain the evaluation results of the CFD simulation model. The evaluation results of the CFD fidelity simulation model are added as new observation data points to the training dataset of the Gaussian process surrogate model. The covariance matrix and posterior distribution of the Gaussian process surrogate model are recalculated, and the predicted mean and variance of the Gaussian process surrogate model are updated until the predicted variance converges. During the Bayesian optimization iteration process, the difference between the model prediction outputs of the Gaussian plume fidelity simulation model and the CFD fidelity simulation model is monitored in real time. Set a difference threshold. When the difference in model predictions exceeds the difference threshold, automatically trigger a re-evaluation of the latest correlation analysis results between the Gaussian plume fidelity simulation model and the CFD fidelity simulation model. Based on the latest correlation analysis results, the multi-fidelity model combination strategy was dynamically revised, the switching rules and calling frequency of the Gaussian plume fidelity simulation model and the CFD fidelity simulation model were adjusted, and the Gaussian process proxy model was retrained. The training continues iteratively until the Gaussian process surrogate model converges, obtaining the optimal hyperparameter combination of the DPPO reinforcement learning model. The DPPO reinforcement learning policy network is then retrained and iterated using the optimal hyperparameter combination until the policy network converges, resulting in the optimized DPPO reinforcement learning model.

9. The method for real-time monitoring and source tracing of urban air pollution based on reinforcement learning according to claim 1, characterized in that, Specifically, S7 includes: S71. Input the multimodal feature vector of pollution diffusion into the policy network of the optimized DPPO reinforcement learning model, and output the optimal pollution source location path decision including the spatial grid coordinates of sensor equipment and the monitoring frequency adjustment scheme. S72. Based on the optimal pollution source location path decision, the location and monitoring frequency of sensor devices are adjusted in real time to determine the geographical coordinates, pollution source type and emission intensity of the pollution source in real time, and the pollution source location result is obtained. S73. Calculate the uncertainty distribution of pollution source location results using the Monte Carlo method and generate a credibility assessment index; S74. Based on the pollution source location path decision, pollution source location results and credibility assessment indicators, generate a standardized real-time pollution source tracing analysis report and upload it to the remote management center.

Images