Industry pollution emission contribution prediction method and device based on multi-modal model

Abstract

The application relates to the related technical field of pollution, and particularly relates to an industry pollution emission contribution prediction method and device based on a multi-modal model. The method comprises the following steps: collecting original data; wherein the original data comprises related data of PM2.5; a SwinLSTM model is used to fuse the multi-scale feature extraction capability of a Swin Transformer and the time series processing capability of an LSTM, so as to pre-process the original data and obtain training data; a multi-modal optimization deep learning model composed of a convolutional neural network and a graph attention network is constructed; the multi-modal optimization deep learning model is trained and optimized based on the training data; target industry information to be predicted is acquired; the trained multi-modal optimization deep learning model is used to perform prediction based on the target industry information, so as to obtain pollution emission prediction of a target industry located in a certain region.

Description

Technical Field

[0001] The present application relates to the field of pollution-related technologies, and specifically to a method and device for predicting industry pollution emission contributions based on a multimodal model. Background Art

[0002] In the field of air pollution control, accurately identifying the industry sources of PM2.5 components and quantifying their contributions is crucial for effective air pollution control and the formulation of scientific environmental policies. Previous research and practice have employed various traditional methods for PM2.5 source apportionment, such as the chemical mass balance (CMB) model and positive definite matrix factorization (PMF). These methods can, to a certain extent, identify the main pollution source categories and estimate their approximate contributions. However, these traditional methods have significant limitations. First, traditional models often analyze data at a single scale. Whether it is large-scale regional meteorological data or small-scale localized urban emissions data, they fail to fully reflect the complex atmospheric pollution processes. For example, at large scales, they cannot accurately capture the differences in pollutant diffusion within cities due to local topography and building layout; at small scales, they cannot account for the long-range transport of pollutants between regions. Furthermore, traditional methods have limited ability to process massive, complex, and dynamically changing data. With the increase in the number of environmental monitoring stations and the increase in monitoring frequency, the amount of data generated is growing exponentially. Traditional models struggle to quickly and accurately identify the complex relationships between various industries and PM2.5 components underlying these data. Summary of the Invention

[0003] In view of this, the embodiments of the present application are dedicated to providing a method and device for predicting industry pollution emission contributions based on a multimodal model.

[0004] This application provides a method for predicting industry pollution emission contributions based on a multimodal model, including:

[0005] Collecting raw data; wherein the raw data includes PM2.5 related data;

[0006] Using the SwinLSTM model, which integrates the multi-scale feature extraction capabilities of the Swin Transformer and the time series processing capabilities of the LSTM, we can deeply mine various data at different scales, extract key spatial features, and simultaneously capture the temporal evolution of data and its impact on PM2.5 concentrations. We then preprocess the raw data to generate training data.

[0007] A multimodal optimization deep learning model consisting of a convolutional neural network and a graph attention network was constructed. The convolutional neural network, through a combination of convolutional and pooling layers, efficiently extracted spatial features of the data, accurately captured the correlation between emission sources in various industries and the distribution of PM2.5 concentrations, and identified the impact patterns of industry emissions on PM2.5 pollution in different regions. The graph attention network used the attention mechanism to weight the characteristics of emission source nodes in various industries, deeply analyzing the relationships between emission sources in various industries and their impact weights on various PM2.5 pollution components. The multimodal optimization deep learning model was used to predict the contribution ratios of different industries to various PM2.5 pollution components in different regions, different time periods, and different meteorological conditions.

[0008] Training and tuning the multimodal optimization deep learning model based on the training data;

[0009] Obtain information on the target industry to be forecasted;

[0010] Using the trained multimodal optimization deep learning model, predictions are made based on the target industry information to obtain a prediction of pollution emissions when the target industry is located in a certain area.

[0011] In some embodiments, the raw data includes meteorological data, geographic information data, industrial emission data, traffic emission data, energy emission data, and PM2.5 component concentration monitoring data.

[0012] In some embodiments, tuning the multimodal optimization deep learning model based on the training data includes:

[0013] Inputting part of the training data into the multimodal optimization deep learning model to perform prediction to obtain preliminary prediction results;

[0014] The preliminary prediction results are compared with the true values ​​to obtain the prediction error. The back propagation algorithm is used to feed the prediction error back to each layer of the model layer by layer, and the model weights and parameters are dynamically adjusted. After calculating the loss function, the model parameters are continuously adjusted in the direction of minimizing the loss function according to the gradient descent method, so as to gradually narrow the difference between the prediction results and the true values.

[0015] In some embodiments, the target industry information includes: the industry category, the industry's emission data, the geographical information of the location, and meteorological conditions.

[0016] This application provides a device for predicting industry pollution emission contributions based on a multimodal model, comprising:

[0017] A collection module, configured to collect raw data, wherein the raw data includes PM2.5 related data;

[0018] The preprocessing module uses the SwinLSTM model, integrating the Swin Transformer's multi-scale feature extraction capabilities with the LSTM time series processing capabilities. This module deeply mines various data at different scales, extracts key spatial features, and captures the temporal evolution of data and its impact on PM2.5 concentration. It also preprocesses the raw data to generate training data.

[0019] A construction module is used to build a multimodal optimized deep learning model composed of a convolutional neural network and a graph attention network. The convolutional neural network uses a combination of convolutional layers and pooling layers to efficiently extract data spatial features, accurately capture the relationship between emission sources in various industries and the distribution of PM2.5 concentrations, and identify the impact patterns of industry emissions on PM2.5 pollution in different regions. The graph attention network uses the attention mechanism to weight the characteristics of emission source nodes in various industries, deeply analyze the mutual relationship between emission sources in various industries, and their impact weights on various PM2.5 pollution components. The multimodal optimized deep learning model is used to predict the contribution ratio of different industries to various PM2.5 pollution components in different regions, different time periods, and different meteorological conditions.

[0020] A tuning module, configured to train and tune the multimodal optimization deep learning model based on the training data;

[0021] An acquisition module is used to obtain target industry information to be predicted;

[0022] The prediction module is used to use the trained multimodal optimization deep learning model to make predictions based on the target industry information to obtain a prediction of pollution emissions when the target industry is located in a certain area.

[0023] In some embodiments, the raw data includes meteorological data, geographic information data, industrial emission data, traffic emission data, energy emission data, and PM2.5 component concentration monitoring data.

[0024] In some embodiments, tuning the multimodal optimization deep learning model based on the training data includes:

[0025] Inputting part of the training data into the multimodal optimization deep learning model to perform prediction to obtain preliminary prediction results;

[0026] The preliminary prediction results are compared with the true values ​​to obtain the prediction error. The back propagation algorithm is used to feed the prediction error back to each layer of the model layer by layer, and the model weights and parameters are dynamically adjusted. After calculating the loss function, the model parameters are continuously adjusted in the direction of minimizing the loss function according to the gradient descent method, so as to gradually narrow the difference between the prediction results and the true values.

[0027] In some embodiments, the target industry information includes: the industry category, the industry's emission data, the geographical information of the location, and meteorological conditions.

[0028] The present application provides an electronic device, including:

[0029] A processor, and a memory for storing a program executable by the processor;

[0030] The processor is used to implement the above-mentioned method for predicting industry pollution emission contributions based on a multimodal model by running the program in the memory.

[0031] The present application provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the processor executes the above-mentioned method for predicting industry pollution emission contributions based on a multimodal model.

[0032] The present application provides a method for predicting the contribution of industrial pollution emissions based on a multimodal model. First, the raw data is collected; wherein, the raw data includes PM2.5 related data; the SwinLSTM model is used to integrate the multi-scale feature extraction capability of SwinTransformer and the time series processing capability of LSTM to deeply mine various types of data from different scales, extract key spatial features, and capture the evolution of data over time and its impact on PM2.5 concentration. The raw data is preprocessed to obtain training data; a multimodal optimized deep learning model composed of a convolutional neural network and a graph attention network is constructed; wherein the convolutional neural network efficiently extracts data spatial features through a combination of convolutional layers and pooling layers, accurately captures the relationship between emission sources of various industries and PM2.5 The model uses a graph attention network to weight the characteristics of emission source nodes within each industry, analyzing the interrelationships between these sources and their impact on PM2.5 components. A multimodal optimization deep learning model is used to predict the contribution of different industries to each PM2.5 component across different regions, time periods, and meteorological conditions. The model is trained and fine-tuned based on the training data. Information about the target industry to be predicted is obtained. The trained multimodal optimization deep learning model is used to predict pollution emissions based on this information, yielding a predicted estimate of the target industry's location within a specific area. This setup, coupled with multiscale models and AI technology, enables accurate prediction of PM2.5 components, precisely tracing the contribution of each industry to PM2.5, and analyzing its spatial distribution patterns. The SwinLSTM model is used to deeply mine massive amounts of complex data, including meteorological, geographical, and industry-specific emissions data, to accurately extract key spatial features and capture temporal evolution patterns, significantly improving data processing efficiency and quality. The innovative multimodal optimization deep learning model (CNN+GAT) was constructed and subsequently optimized to not only accurately represent the changing trends in fine particulate matter composition and concentration at a single site, but also to construct accurate regional concentration models. This model accurately calculates the contribution of each industry to each PM2.5 component under different regions, time periods, and meteorological conditions, clearly defining each industry's responsibility for air pollution and providing clear guidance for the development of targeted governance strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The above and other purposes, features, and advantages of the present application will become more apparent through a more detailed description of the embodiments of the present application in conjunction with the accompanying drawings. The accompanying drawings are intended to provide a further understanding of the embodiments of the present application and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application and do not constitute a limitation of the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0034] Figure 1 This is a flow chart of a method for predicting industry pollution emission contributions based on a multimodal model provided in an embodiment of the present application.

[0035] Figure 2 This is a partial flow chart of a method provided in one embodiment of the present application.

[0036] Figure 3 This is a structural diagram of an industry pollution emission contribution prediction device based on a multimodal model provided in an embodiment of the present application.

[0037] Figure 4 This is a schematic diagram of the structure of an electronic device provided by an embodiment of the present application. DETAILED DESCRIPTION

[0038] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. All other embodiments obtained by ordinary technicians in this field based on the embodiments of the present invention without making any creative efforts shall fall within the scope of protection of the present invention.

[0039] Application Overview

[0040] In the field of air pollution control, accurately identifying the industry sources of PM2.5 components and quantifying their contributions is crucial for effective air pollution control and the formulation of scientific environmental policies. Previous research and practice have employed various traditional methods for PM2.5 source apportionment, such as the chemical mass balance (CMB) model and positive definite matrix factorization (PMF). These methods can, to a certain extent, identify the main pollution source categories and estimate their approximate contributions. However, these traditional methods have significant limitations. First, traditional models often analyze data at a single scale. Whether it is large-scale regional meteorological data or small-scale localized urban emissions data, they fail to fully reflect the complex atmospheric pollution processes. For example, at large scales, they cannot accurately capture the differences in pollutant diffusion within cities due to local topography and building layout; at small scales, they cannot account for the long-range transport of pollutants between regions. Furthermore, traditional methods have limited ability to process massive, complex, and dynamically changing data. With the increase in the number of environmental monitoring stations and the increase in monitoring frequency, the amount of data generated is growing exponentially. Traditional models struggle to quickly and accurately identify the complex relationships between various industries and PM2.5 components underlying these data. With the rapid development of big data and artificial intelligence technologies, as well as the increasing demand for more refined air pollution control, there is an urgent need for an innovative method that can integrate multi-scale data and, with the powerful data processing and analysis capabilities of artificial intelligence, accurately trace the industry contribution of each component of PM2.5, thereby providing solid technical support for more efficient and targeted air pollution control strategies.

[0041] Research prior to this study revealed that some researchers have used single-scale atmospheric dispersion models, such as the Gaussian plume model, to analyze the industry contributions of PM2.5 components. To apply this model, pollutant emission inventories for each industry within a specific region must be compiled, clearly defining the location, emission volume, and emission altitude of each industry's emission sources. Regional meteorological data, such as average wind speed and prevailing wind direction, must also be obtained. Based on this data, the Gaussian plume model's formula is used to calculate the contribution of pollutants emitted by each industry to the concentration of each PM2.5 component at the monitoring site, taking into account the relative location of the emission source and the impact of meteorological conditions on pollutant dispersion. For example, for a chemical industry emission source, the model calculates the contribution of sulfate formed by sulfur dioxide emissions in PM2.5 at surrounding monitoring sites. However, these single-scale models have numerous limitations. They only consider local, small-scale meteorological conditions and emission source information and fail to reflect the impact of large-scale regional meteorological variations on pollutant transport, such as the long-distance transport of pollutants caused by interregional atmospheric circulation. The model's ability to handle complex terrain and urban surfaces is limited, making it unable to accurately simulate pollutant diffusion paths and concentration distributions in mountainous areas and urban areas with densely populated high-rise buildings. Due to the model's relatively simplistic assumptions, it struggles to accurately estimate the complex chemical reactions between pollutants emitted by various industries and the secondary PM2.5 components produced. This results in an incomplete and inaccurate analysis of the industry contributions of PM2.5 components, making it unable to meet the current demand for refined air pollution control.

[0042] Some researchers have also used traditional factor analysis to explore the sources and industry contributions of PM2.5 components. They collected long-term data on PM2.5 chemical compositions from multiple monitoring sites, along with information on related potential pollution sources. After standardizing the data, they used factor analysis algorithms to reduce the numerous related variables to a small number of uncorrelated composite factors. During the analysis, the correlation between each factor and each variable was determined based on the factor loading matrix, thereby identifying the main pollution source categories, such as industrial sources, traffic sources, and dust sources. Then, based on the scores of each factor, the relative contributions of different pollution source categories to the various PM2.5 components were estimated. For example, if a factor is highly correlated with characteristic pollutants emitted by industry and has a high score for that factor, it can be inferred that industrial sources contribute significantly to certain PM2.5 components.

[0043] However, this traditional factor analysis method has significant shortcomings. It primarily relies on the statistical characteristics of the data and places extremely high demands on data quality. Noise or outliers in the data can severely impact the accuracy of the analysis results. Furthermore, this method cannot dynamically update and adapt to changing pollution sources, making it difficult to promptly and accurately incorporate newly emerging industries or emission sources into the analysis system. Furthermore, it can only provide a rough estimate of the contribution of each pollution source category to PM2.5 components, making it difficult to pinpoint specific industries and unable to meet the needs of precise source tracing and the development of targeted control measures.

[0044] This invention aims to develop a method for accurately tracing the contribution of each industry to PM2.5 components by using multi-scale model coupling and AI technology. At present, air pollution control faces severe challenges and many technical difficulties need to be overcome. On the one hand, the formation and distribution of PM2.5 are affected by multiple factors such as meteorology, geography, and emissions from various industries, making it difficult to accurately determine the specific contribution ratio of each industry to each PM2.5 pollution component in different regions, time periods, and meteorological conditions.

[0045] Traditional analytical methods are unable to cope with such complex interactions. On the other hand, faced with massive, multi-source and heterogeneous data, existing technologies lack the ability to efficiently integrate and deeply mine, making it difficult to fully extract key information, which seriously affects the accuracy of tracing the sources of PM2.5 components. Furthermore, even if some traceability information is obtained, there is no systematic solution to how to effectively transform it into a scientific and effective pollution prevention and control strategy. Based on this,

[0046] The present invention constructs a multimodal model system that integrates the SwinLSTM model, convolutional neural network (CNN) and graph attention network (GAT) to accurately analyze the contribution ratio of various industries to various PM2.5 pollution components. At the same time, based on the precise calculation results, the emission data of a certain industry, the geographical information of the location, meteorological conditions, etc. are input into the model to simulate its impact on the various PM2.5 components in the surrounding environment, and realize the prediction of pollution emissions after the industry is located in a certain area, providing core data support for air pollution control, and promoting air pollution control towards refinement and efficiency.

[0047] After introducing the basic principles of the present application, various non-limiting embodiments of the present application will be described in detail with reference to the accompanying drawings.

[0048] Exemplary Methods

[0049] like Figure 1 、 Figure 2 As shown, the method for predicting industry pollution emission contributions based on a multimodal model provided in this application includes the following contents.

[0050] Step S110, collecting raw data; wherein the raw data includes PM2.5 related data;

[0051] Specifically, the original data includes meteorological data, geographic information data, industrial emission data, traffic emission data, energy emission data and PM2.5 component concentration monitoring data.

[0052] Collecting raw data is the foundation of the entire invention, and this data will be used for subsequent model training and analysis. The collected raw data includes:

[0053] Meteorological data: such as temperature, humidity, wind speed, wind direction, etc. These data can help understand the movement and diffusion characteristics of the atmosphere and have an important impact on the transmission and diffusion of pollutants.

[0054] Geographic information data: including information on topography, landforms, land use types, etc. These data help analyze the impact of the geographical environment on the diffusion of pollutants, such as the blocking and guiding effects of mountains, rivers, and urban building layout on airflow.

[0055] Industrial emission data: covers information such as pollutant emissions, emission source locations, and emission heights of various industrial enterprises. Because industrial emissions are one of the important sources of PM2.5, the emission characteristics and impact ranges of different industries vary.

[0056] Traffic emission data: such as motor vehicle exhaust emissions, traffic flow, road distribution, etc. Traffic emissions have a significant impact on urban air quality, especially in traffic congested areas.

[0057] Energy emission data: involves pollutant emissions during energy production and use, such as exhaust emissions from thermal power plants. Energy-related emissions have a significant impact on regional air quality.

[0058] Monitoring data on the concentration of PM2.5 components: The concentration data of different chemical components in PM2.5 (such as sulfate, nitrate, organic carbon, etc.) obtained through environmental monitoring stations are the basic data for evaluating the contribution of various industries and directly reflect the current status of PM2.5 pollution in the atmosphere.

[0059] Step S120: Using the SwinLSTM model, integrating the multi-scale feature extraction capabilities of the Swin Transformer and the time series processing capabilities of the LSTM, the model deeply mines various types of data at different scales, extracts key spatial features, and simultaneously captures the temporal evolution of the data and its impact on PM2.5 concentration. The raw data is preprocessed to obtain training data.

[0060] Specifically, after collecting the raw data, it is necessary to use the SwinLSTM model to preprocess the data to obtain training data suitable for model training. The specific process is as follows:

[0061] Combining the Swin Transformer's multi-scale feature extraction capabilities with the LSTM's time series processing capabilities: The Swin Transformer extracts features from data at different scales, capturing spatial characteristics within the data. The LSTM excels at processing time series data, capturing patterns in data over time. Combining these two capabilities allows comprehensive mining of key information from various data types, including spatial and temporal characteristics, and their impact on PM2.5 concentrations.

[0062] Deeply mine various types of data and extract key spatial features: Through multi-scale analysis of raw data, key spatial features related to PM2.5 concentrations are extracted, such as the distribution characteristics of emission sources in different regions and the impact of geographical obstacles on pollutant diffusion. These features help understand the spatial distribution patterns of pollutants.

[0063] Capture the temporal evolution of data and its impact on PM2.5 concentrations: Analyze how changes in meteorological conditions, emission source activities, etc. over time affect PM2.5 concentrations, such as the impact of seasonal changes and day-night changes on pollutant generation and diffusion, so as to grasp the temporal evolution trend of PM2.5 concentrations.

[0064] Preprocess the raw data to obtain training data: After extracting key features and capturing evolution patterns, screen the data to remove invalid or abnormal data points, and then perform standardization and normalization to form an ordered data set. This serves as input data for subsequent model training, ensuring that the model can learn effective patterns from high-quality data.

[0065] Step S130: Construct a multimodal optimized deep learning model consisting of a convolutional neural network and a graph attention network. The convolutional neural network uses a combination of convolutional layers and pooling layers to efficiently extract spatial features of the data, accurately capture the correlation between emission sources of various industries and the distribution of PM2.5 concentrations, and identify the impact patterns of industry emissions on PM2.5 pollution in different regions. The graph attention network uses an attention mechanism to weight the features of emission source nodes of various industries, deeply analyze the relationships between emission sources of various industries, and their impact weights on various PM2.5 pollution components. The multimodal optimized deep learning model is used to predict the contribution ratios of different industries to various PM2.5 pollution components in different regions, different time periods, and different meteorological conditions.

[0066] Specifically, a multimodal optimization deep learning model consisting of a convolutional neural network (CNN) and a graph attention network (GAT) was constructed to predict the contribution ratio of each PM2.5 component in different industries under different conditions. The specific construction process is as follows:

[0067] Convolutional Neural Network (CNN): This approach efficiently extracts spatial features from the data through a combination of convolutional and pooling layers. Convolutional layers automatically learn local features in the data, while pooling layers reduce data dimensionality and computational complexity while retaining important features. This combination effectively extracts spatial features related to PM2.5 concentration distribution from preprocessed data, such as the intensity and distribution patterns of emission sources in different regions. It also accurately captures the correlation between emission sources across various industries and PM2.5 concentration distribution. By learning spatial features from the data, the CNN can identify the correlation between the location, intensity, and other characteristics of emission sources across different industries and the spatial distribution of PM2.5 concentrations, such as the correspondence between certain high-emission industrial areas and areas with high PM2.5 concentrations. It also identifies the impact of industrial emissions on PM2.5 pollution within different regions. This approach further analyzes the combined impact of emissions from various industries within different regions and identifies the spatial distribution patterns of PM2.5 pollution from different combinations of industrial emissions. For example, it can identify the pollution characteristics of different functional zones, such as urban industrial areas and traffic-intensive areas, on the surrounding environment.

[0068] Graph Attention Network (GAT) section: Utilizes the attention mechanism to weight the features of emission source nodes in each industry: Treat the emission sources of each industry as nodes in the graph, and use the attention mechanism to weight the features of emission sources from different industries based on the relationship and importance between the nodes, so that the model can pay more attention to emission sources from industries that contribute more to PM2.5. In-depth analysis of the relationships between emission sources in different industries: Consider the possible synergistic or antagonistic effects between emission sources in different industries. For example, the precursors emitted by certain industries undergo chemical reactions in the atmosphere to generate secondary pollutants, affecting the pollution contribution of other industries. By analyzing these relationships, the comprehensive impact of each industry can be more accurately assessed. Analyze the impact weights of emission sources from various industries on the various pollutant components of PM2.5: Determine the contribution weights of emission sources from different industries to different chemical components in PM2.5. For example, some industries may mainly contribute sulfate components, while other industries have a greater impact on organic carbon components. This is crucial for accurately quantifying the pollution contribution of each industry.

[0069] The overall function of the multimodal optimization deep learning model: It combines the spatial features extracted by CNN with the industry emission source relationships and weight information analyzed by GAT to form a multimodal model architecture that can comprehensively consider the characteristics of both spatial and industry relationships, thereby more accurately predicting the contribution ratio of different industries to various PM2.5 components in different regions, different time periods, and different meteorological conditions.

[0070] Step S140: training and tuning the multimodal optimization deep learning model based on the training data;

[0071] Step S150, obtaining target industry information to be predicted;

[0072] In step S160 , the trained multimodal optimization deep learning model is used to make predictions based on the target industry information to obtain a prediction of pollution emissions when the target industry is located in a certain area.

[0073] With this setup, through multi-scale model coupling and AI technology, we can achieve accurate predictions of PM2.5 components, accurately trace the contribution of each industry to PM2.5, and deeply analyze its spatial distribution patterns. With the help of the SwinLSTM model, we can deeply mine massive amounts of complex data such as meteorology, geography, and emissions from various industries to accurately extract key spatial features, capture temporal evolution patterns, and greatly improve the efficiency and quality of data processing. The innovative use of a multimodal optimization deep learning model (CNN+GAT), its construction and subsequent optimization can not only accurately present the changing trends of fine particulate matter composition and concentration at a single site, but also construct an accurate regional concentration model. On this basis, we can accurately calculate the contribution ratio of each industry to each PM2.5 component in different regions, time periods, and meteorological conditions, clearly define the responsibilities of each industry in air pollution, and provide a clear direction for the formulation of targeted governance strategies.

[0074] Specifically, tuning the multimodal optimization deep learning model based on the training data includes: inputting part of the training data into the multimodal optimization deep learning model to perform prediction to obtain a preliminary prediction result; comparing the preliminary prediction result with the true value to obtain a prediction error, using a backpropagation algorithm to feed back the prediction error to each level of the model layer by layer, dynamically adjusting the model weights and parameters, and after calculating the loss function, continuously adjusting the model parameters in the direction of minimizing the loss function according to the gradient descent method, so as to gradually narrow the difference between the prediction result and the true value.

[0075] The following are the specific steps to tune the multimodal optimization deep learning model based on training data:

[0076] Model training and preliminary predictions:

[0077] Data Preparation and Partitioning: The training data obtained in step S120 is divided into two parts: one part serves as the actual training set for initial model training, and the other part serves as the validation set for subsequent model evaluation and tuning. The training and validation sets are typically split in a certain ratio, such as 70% of the data as training and 30% as validation, or cross-validation is used to ensure the stability and generalization ability of the model on different data subsets.

[0078] Model Input and Preliminary Forecast: The training set data is fed into the multimodal optimization deep learning model. Based on the previously constructed architecture and initialized parameters, the model calculates and analyzes the input data and outputs preliminary forecast results. These preliminary forecasts include the contribution of each PM2.5 component by industry, across the different regions, time periods, and meteorological conditions corresponding to the training set.

[0079] Prediction error calculation and feedback:

[0080] Comparison and Error Calculation: The model's initial predictions are compared with the corresponding true values ​​in the training set (actual observations or known PM2.5 component contribution ratios). The difference between the two is calculated, representing the prediction error. Common error calculation methods include mean squared error (MSE) and mean absolute error (MAE). These error metrics can quantify the accuracy of model predictions and provide guidance for subsequent model tuning.

[0081] Backpropagation and Error Feedback: Using the backpropagation algorithm, the calculated prediction error is fed back to each layer of the model layer by layer. During this feedback process, the model weights and parameters are dynamically adjusted based on the error. Specifically, starting from the output layer, the gradient of each layer is calculated forward based on the error gradient, all the way to the input layer. This process allows the model to understand the impact of each parameter on the final prediction result, providing a basis for optimizing and adjusting the parameters.

[0082] Model parameter adjustment and optimization:

[0083] Loss Function Calculation and Gradient Descent: While feeding back the error, the model's loss function is calculated. This loss function is typically defined based on the prediction error, such as the mean squared error loss function. Then, using the gradient descent method, the model parameters are adjusted in the direction that minimizes the loss function, following the gradient of the loss function with respect to the model parameters. Gradient descent iteratively updates the parameters, gradually reducing the loss function and thereby improving the model's predictive performance.

[0084] Learning rate and adjustment step size: During parameter tuning, the learning rate is a critical hyperparameter, determining the step size for each parameter update. Excessively large learning rates can cause the model to oscillate or diverge during optimization, while excessively small learning rates can slow down the training process. Therefore, it's important to appropriately set the learning rate based on the model's training progress, or employ an adaptive learning rate adjustment strategy, to ensure the model quickly and stably converges to optimal or near-optimal parameter states.

[0085] Model evaluation and validation:

[0086] Validation set evaluation: After completing a round of parameter adjustments, the model is evaluated using the validation set. The prediction error and loss function value of the model on the validation set are calculated to verify the model's generalization ability and performance on data that was not used in training. If the model performs poorly on the validation set, such as a large error or overfitting (i.e., performing well on the training set but poorly on the validation set), further adjustments to the model structure or parameters are needed, such as adding regularization terms, adjusting the number of network layers or neurons, to improve the model's generalization performance.

[0087] Model optimization iteration: The above process (model training, error calculation, parameter adjustment, and validation set evaluation) is repeated multiple times, with each iteration retraining and evaluating based on the new parameter state until the model's performance on the validation set reaches a satisfactory level, or the loss function value stabilizes and no longer decreases significantly. At this point, the model is considered to have been fully tuned and training can be stopped. The purpose of the entire tuning process is to enable the model to learn effective patterns and regularities from the training data, so that when faced with new, unseen data, it can accurately predict the contribution of different industries to the various PM2.5 components, providing reliable decision-making support for air pollution control.

[0088] Specifically, information about the target industry to be predicted is obtained; this information includes the industry category, emissions data, geographic information about the location, and meteorological conditions. A trained multimodal optimization deep learning model is used to perform a prediction based on this information, yielding a predicted pollution emission estimate for the target industry located in a specific location.

[0089] The following are the specific steps for obtaining target industry information and using the trained multimodal optimization deep learning model to predict pollution emissions in the target industry:

[0090] Obtain target industry information:

[0091] Determine the industry category: Identify the category of the target industry to be predicted, such as chemical, steel, cement, transportation and other different industry types. Because different industries have significant differences in production processes, raw material usage and pollutant emission characteristics, their contributions to the various components of PM2.5 are also different.

[0092] Collect industry emission data: Obtain detailed emission data for the target industry, including the emission volume of major pollutants (such as sulfur dioxide, nitrogen oxides, and particulate matter), the location of emission sources (latitude and longitude coordinates, etc.), emission height (such as chimney height), and the emission time pattern (continuous or intermittent). This data can be obtained from the company's environmental impact assessment report, environmental monitoring data, government environmental statistics departments, or through on-site monitoring.

[0093] Obtain geographic information about the location: Collect geographic information about the target industry's location, such as topography (e.g., whether it is located in a valley, plain, or coastal area), surrounding land use types (e.g., proximity to residential areas, commercial areas, or other industrial areas), and distance and relative position to major transportation arteries and water systems. This geographic information can be obtained through geographic information system (GIS) data, satellite remote sensing imagery, or field surveys.

[0094] Collect meteorological data: Compile meteorological data for the target industry's location, including year-round prevailing wind direction, wind speed, temperature and humidity patterns, and precipitation distribution. This data can be sourced from local meteorological observatories, climate data released by meteorological authorities, or professional meteorological data service platforms. This data can be used to analyze the transmission and diffusion of pollutants in the atmosphere.

[0095] Prediction process:

[0096] Data preprocessing and format conversion: Collected target industry information is preprocessed and formatted according to the model input requirements. For example, emission data is standardized or normalized to align with the data range and format used during model training. Geographic information and meteorological condition data are converted into numerical forms or feature vectors that the model can recognize and process, such as encoding geographic locations into longitude and latitude coordinates and converting terrain types into corresponding numerical codes.

[0097] Model Input and Prediction: The pre-processed target industry information is fed into a trained multimodal optimization deep learning model. Based on the patterns and patterns learned from previous training, the model comprehensively considers factors such as the target industry's emission characteristics, location, and meteorological conditions to predict pollution emissions for the target industry if it were located in that area.

[0098] Output prediction results and interpretation: The model outputs predictions of the target industry's impact on each PM2.5 component in the surrounding environment, including detailed information such as the concentration contribution of each component and the impact range. These predictions can be used to further analyze the potential impact of the target industry's pollution emissions on local air quality, determine whether appropriate pollution prevention and control measures are necessary, and provide a scientific basis for environmental management authorities to formulate environmental access policies and regulatory measures for the industry.

[0099] Exemplary devices

[0100] The device embodiments of this application can be used to execute the method embodiments of this application. For details not disclosed in the device embodiments of this application, please refer to the method embodiments of this application.

[0101] Figure 3The figure shows a block diagram of an industry pollution emission contribution prediction device based on a multimodal model provided by an embodiment of the present application. Figure 3 As shown, the device includes:

[0102] A collection module 31 is used to collect raw data, wherein the raw data includes PM2.5 related data;

[0103] Preprocessing module 32 is used to use the SwinLSTM model, integrating the Swin Transformer multi-scale feature extraction capability and the LSTM time series processing capability to deeply mine various types of data at different scales, extract key spatial features, and simultaneously capture the evolution of data over time and its impact on PM2.5 concentration. It preprocesses the raw data to obtain training data;

[0104] Module 33 is constructed to construct a multimodal optimized deep learning model consisting of a convolutional neural network and a graph attention network. The convolutional neural network, through a combination of convolutional layers and pooling layers, efficiently extracts spatial features of data, accurately captures the correlation between emission sources of various industries and the distribution of PM2.5 concentrations, and identifies the impact patterns of industry emissions on PM2.5 pollution in different regions. The graph attention network utilizes an attention mechanism to weight the features of emission source nodes of various industries, deeply analyzing the relationships between emission sources of various industries and their impact weights on various PM2.5 pollution components. The multimodal optimized deep learning model is used to predict the contribution ratios of different industries to various PM2.5 pollution components in different regions, different time periods, and different meteorological conditions.

[0105] A tuning module 34, configured to train and tune the multimodal optimization deep learning model based on the training data;

[0106] An acquisition module 35 is used to acquire target industry information to be predicted;

[0107] The prediction module 36 is used to use the trained multimodal optimization deep learning model to make predictions based on the target industry information to obtain a prediction of pollution emissions when the target industry is located in a certain area.

[0108] In some embodiments, the raw data includes meteorological data, geographic information data, industrial emission data, traffic emission data, energy emission data, and PM2.5 component concentration monitoring data.

[0109] In some embodiments, tuning the multimodal optimization deep learning model based on the training data includes:

[0110] Inputting part of the training data into the multimodal optimization deep learning model to perform prediction to obtain preliminary prediction results;

[0111] The preliminary prediction results are compared with the true values ​​to obtain the prediction error. The back propagation algorithm is used to feed the prediction error back to each layer of the model layer by layer, and the model weights and parameters are dynamically adjusted. After calculating the loss function, the model parameters are continuously adjusted in the direction of minimizing the loss function according to the gradient descent method, so as to gradually narrow the difference between the prediction results and the true values.

[0112] In some embodiments, the target industry information includes: the industry category, the industry's emission data, the geographical information of the location, and meteorological conditions.

[0113] Exemplary electronic devices

[0114] Below, reference Figure 4 To describe the electronic device according to the embodiment of the present application. Figure 4 The figure shows a block diagram of an electronic device according to an embodiment of the present application.

[0115] like Figure 4 As shown, electronic device 400 includes one or more processors 410 and memory 420 .

[0116] The processor 410 may be a central processing unit (CPU) or other forms of processing units having data processing capabilities and / or instruction execution capabilities, and may control other components in the electronic device 400 to perform desired functions.

[0117] The memory 420 may include one or more computer program products, which may include various forms of computer-readable storage media, such as volatile memory and / or non-volatile memory. The volatile memory may include, for example, random access memory (RAM) and / or cache memory (cache), etc. The non-volatile memory may include, for example, read-only memory (ROM), a hard disk, a flash memory, etc. One or more computer program instructions may be stored on the computer-readable storage medium, and the processor 410 may run the program instructions to implement the industry pollution emission contribution prediction method based on the multimodal model of each embodiment of the present application described above and / or other desired functions. Various contents such as category correspondences may also be stored in the computer-readable storage medium.

[0118] In one example, the electronic device 400 may further include an input device 430 and an output device 440 , and these components are interconnected via a bus system and / or other forms of connection mechanisms (not shown).

[0119] In addition, the input device 430 may also include, for example, a keyboard, a mouse, an interface, etc. The output device 440 may output various information to the outside, including analysis results, etc. The output device 440 may include, for example, a display, a speaker, a printer, a communication network and its connected remote output device, etc.

[0120] Of course, to simplify, Figure 4 Only some of the components in the electronic device related to the present application are shown, and components such as a bus, an input / output interface, etc. are omitted. In addition, the electronic device may further include any other appropriate components according to specific application scenarios.

[0121] Exemplary computer program products and computer-readable storage media

[0122] In addition to the above-mentioned methods and devices, an embodiment of the present application may also be a computer program product, which includes computer program instructions, which, when executed by a processor, enable the processor to execute the steps of the method for predicting industry pollution emission contributions based on a multimodal model according to various embodiments of the present application described in the above "Exemplary Method" section of this specification.

[0123] The computer program product may be written in any combination of one or more programming languages ​​to implement the program code for performing the operations of the embodiments of the present application, including object-oriented programming languages ​​such as Java, C++, and conventional procedural programming languages ​​such as "C" or similar programming languages. The program code may be executed entirely on the user's computing device, partially on the user's computing device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server.

[0124] In addition, an embodiment of the present application may also be a computer-readable storage medium having computer program instructions stored thereon, which, when executed by a processor, enables the processor to execute the steps of the method for predicting industry pollution emission contributions based on a multimodal model according to various embodiments of the present application described in the above "Exemplary Method" section of this specification.

[0125] The computer-readable storage medium can adopt any combination of one or more readable media. The readable medium can be a readable signal medium or a readable storage medium. The readable storage medium can, for example, include but is not limited to a system, device or component of electricity, magnetism, light, electromagnetic, infrared, or semiconductor, or any combination thereof. More specific examples (non-exhaustive list) of readable storage media include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination thereof.

[0126] The above description has been provided for the purpose of illustration and description. Furthermore, this description is not intended to limit the embodiments of the present application to the forms disclosed herein. Although a number of example aspects and embodiments have been discussed above, those skilled in the art will recognize certain variations, modifications, alterations, additions, and sub-combinations thereof.

Claims

1. A method for predicting industry pollution emission contributions based on a multimodal model, characterized by: include: Collecting raw data; wherein the raw data includes PM2.5 related data; the raw data includes meteorological data, geographic information data, industrial emission data, traffic emission data, energy emission data, and PM2.5 component concentration monitoring data; Using the SwinLSTM model, which integrates the multi-scale feature extraction capabilities of the Swin Transformer and the time series processing capabilities of the LSTM, we can deeply mine various data at different scales, extract key spatial features, and simultaneously capture the temporal evolution of data and its impact on PM2.5 concentrations. We then preprocess the raw data to generate training data. A multimodal optimization deep learning model consisting of a convolutional neural network and a graph attention network was constructed. The convolutional neural network, through a combination of convolutional and pooling layers, efficiently extracted spatial features of the data, accurately captured the correlation between emission sources across various industries and the distribution of PM2.5 concentrations, and identified the impact patterns of industry emissions on PM2.5 pollution in different regions. The graph attention network, using an attention mechanism, weighted the features of emission source nodes across various industries, deeply analyzing the relationships between emission sources across various industries and their impact weights on various PM2.5 pollution components. The multimodal optimization deep learning model was used to predict the contribution of different industries to various PM2.5 pollution components in different regions, at different time periods, and under different meteorological conditions. Training and tuning the multimodal optimization deep learning model based on the training data; Obtaining target industry information to be predicted; wherein the target industry information includes: industry category, emission data of the industry, geographical information of the location, and meteorological conditions; Using the trained multimodal optimization deep learning model, predictions are made based on the target industry information to obtain a prediction of pollution emissions when the target industry is located in a certain area.

2. The method for predicting industry pollution emission contributions based on a multimodal model according to claim 1 is characterized in that: Tuning the multimodal optimization deep learning model based on the training data includes: Inputting part of the training data into the multimodal optimization deep learning model to perform prediction to obtain preliminary prediction results; The preliminary prediction results are compared with the true values ​​to obtain the prediction error. The back propagation algorithm is used to feed the prediction error back to each layer of the model layer by layer, and the model weights and parameters are dynamically adjusted. After calculating the loss function, the model parameters are continuously adjusted in the direction of minimizing the loss function according to the gradient descent method, so as to gradually narrow the difference between the prediction results and the true values.

3. A device for predicting industry pollution emission contributions based on a multimodal model, characterized in that: include: A collection module for collecting raw data; wherein the raw data includes PM2.5 related data; The preprocessing module uses the SwinLSTM model, integrating the Swin Transformer's multi-scale feature extraction capabilities with the LSTM time series processing capabilities. This module deeply mines various data at different scales, extracts key spatial features, and captures the temporal evolution of data and its impact on PM2.5 concentration. It also preprocesses the raw data to generate training data. A construction module is used to build a multimodal optimized deep learning model composed of a convolutional neural network and a graph attention network. The convolutional neural network uses a combination of convolutional layers and pooling layers to efficiently extract spatial features of data, accurately capture the correlation between emission sources of various industries and the distribution of PM2.5 concentrations, and identify the impact patterns of industry emissions on PM2.5 pollution in different regions. The graph attention network uses the attention mechanism to weight the characteristics of emission source nodes in various industries, deeply analyzing the mutual relationship between emission sources of various industries and their influence weight on various PM2.5 pollution components. The multimodal optimized deep learning model is used to predict the contribution ratio of different industries to various PM2.5 pollution components in different regions, different time periods, and different meteorological conditions. A tuning module, configured to train and tune the multimodal optimization deep learning model based on the training data; An acquisition module is used to obtain target industry information to be predicted; A prediction module is used to use the trained multimodal optimization deep learning model to make predictions based on the target industry information to obtain a prediction of pollution emissions when the target industry is located in a certain area; The raw data include meteorological data, geographic information data, industrial emission data, traffic emission data, energy emission data and PM2.5 component concentration monitoring data; The target industry information includes: the industry category, the industry's emission data, the geographical information of the location, and meteorological conditions.

4. The device for predicting industry pollution emission contributions based on a multimodal model according to claim 3 is characterized in that: Tuning the multimodal optimization deep learning model based on the training data includes: Inputting part of the training data into the multimodal optimization deep learning model to perform prediction to obtain preliminary prediction results; The preliminary prediction results are compared with the true values ​​to obtain the prediction error. The back propagation algorithm is used to feed the prediction error back to each layer of the model layer by layer, and the model weights and parameters are dynamically adjusted. After calculating the loss function, the model parameters are continuously adjusted in the direction of minimizing the loss function according to the gradient descent method, so as to gradually narrow the difference between the prediction results and the true values.

5. An electronic device, characterized in that: include: A processor, and a memory for storing a program executable by the processor; The processor is configured to implement the method for predicting industry pollution emission contributions based on a multimodal model as described in any one of claims 1 to 2 by running the program in the memory.

6. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, enables the processor to execute the method for predicting industrial pollution emission contributions based on a multimodal model as described in any one of claims 1 to 2.

Images