A method for hourly simulation and prediction of urban road CO2 concentration based on multi-source data fusion and ensemble algorithm

Abstract

This invention discloses a method for hourly simulation and prediction of urban road CO2 concentration driven by multi-source data fusion and ensemble algorithms. It involves integrating heterogeneous data from multiple sources, including road mobile monitoring CO2 concentration data, road information and traffic data, meteorological data, and human geography data, to construct a CO2 concentration simulation model based on an ensemble algorithm combining a Random Forest (RF) model and a Long Short-Term Memory (LSTM) network model. Considering the potential inaccuracy of a single model in simulating urban road CO2 concentration, this invention employs an ensemble algorithm, combining the Random Forest model and the LSTM network model to select the optimal model for simulating road CO2 concentration, thereby effectively improving the accuracy of the simulation.

Description

Technical Field

[0001] This invention relates to the field of CO2 concentration monitoring technology, specifically to a method for hourly simulation and prediction of urban road CO2 concentration based on multi-source data fusion and integration algorithm. Background Technology

[0002] In recent years, significant technological advancements have been made in urban environmental carbon monitoring. One important technological approach is mobile monitoring based on vehicle-mounted platforms, used for real-time monitoring of atmospheric CO2 concentrations on urban roads. This solution, by installing an atmospheric CO2 concentration monitoring system on a vehicle, can acquire crucial data on CO2 concentrations and vehicle emissions while the vehicle is in motion. This method can obtain accurate CO2 concentration data in real time.

[0003] While this technology can acquire high-precision CO2 concentration data in real time, it also has some limitations, such as the huge cost. In addition, a mobile monitoring vehicle can usually only monitor a single road for a short period of time, which results in spatiotemporal discontinuity in data collection and makes it impossible to comprehensively and completely obtain continuous CO2 concentration information for all roads in the entire city at all times.

[0004] To address the aforementioned problems, this invention provides a method for hourly simulation and prediction of CO2 concentration on urban roads, driven by a multi-source data fusion and integration algorithm, which enables continuous dynamic simulation and real-time prediction of CO2 concentration on urban road networks. Summary of the Invention

[0005] This invention provides a method for hourly simulation and prediction of CO2 concentration on urban roads, driven by multi-source data fusion and integration algorithms, which enables continuous dynamic simulation and real-time prediction of CO2 concentration on urban road networks.

[0006] The purpose of this invention is to provide a method for hourly simulation and prediction of urban road CO2 concentration based on multi-source data fusion and integration algorithms. The specific steps of the method are as follows:

[0007] Step 1: Construct a road mobile monitoring system, design a mobile monitoring scheme, and acquire data on CO2 concentration, vehicle speed, road scene images, wind speed, wind direction, temperature, humidity, and the corresponding time and geographic coordinates of monitoring data for some urban roads.

[0008] Step 2: Based on CO2 concentration data from some road mobile monitoring systems, perform data cleaning and preprocessing;

[0009] Step 3: Combine the corresponding road scene images and use image recognition algorithms such as YOLO-v8 to extract information on road vehicle driving conditions and vehicle types;

[0010] Step 4: Collect road attribute information, traffic conditions, meteorological data, atmospheric background CO2 concentration observation data, land use data, POI data, and population distribution data. Match these data with the road CO2 concentration monitoring data and vehicle information extraction results using a spatiotemporal matching algorithm to construct a multi-source input feature variable dataset.

[0011] Step 5: Using road CO2 concentration data as the true value and other data as input feature variables, standardize the data and divide the data to form a training set and a validation set;

[0012] Step 6: Based on the collected multi-source heterogeneous data, construct an hourly simulation and prediction model of urban road CO2 concentration based on the random forest model and the long short-term memory network integrated algorithm, and use the training set to train the model.

[0013] Step 7: Use the validation set to verify and evaluate the accuracy of the model;

[0014] Step 8: Based on the trained urban road CO2 concentration simulation and prediction model, simulate the carbon concentration of each road in the city hourly.

[0015] Furthermore, the specific steps of step 1 are as follows:

[0016] Step 11: By installing greenhouse gas monitoring equipment, GPS system, panoramic camera and meteorological sensor system on vehicles, a mobile monitoring system for atmospheric CO2 on urban roads was constructed. Based on the mobile monitoring system, the geographical coordinates of urban road CO2 concentration, road scene images, vehicle speed, wind speed, wind direction, temperature, humidity and monitoring data were obtained.

[0017] Step 12: Subsequently, taking into account the overall urban road network planning information, road attribute characteristics, time attribute characteristics, distribution information of key urban pollution sources, topography and land use information, design a mobile monitoring route plan for CO2 on typical urban roads. The mobile monitoring route plan includes the mobile monitoring time range, mobile monitoring route, key monitoring sections, and approximate time information.

[0018] Furthermore, the specific steps of step 2 are as follows:

[0019] Step 21: First, process the monitoring data of the road CO2 concentration mobile monitoring system. Since different sensors collect different information, perform spatiotemporal matching of CO2 concentration monitoring data, GPS data, meteorological data, and panoramic image data based on the data collection time information.

[0020] Step 22: The mobile monitoring CO2 concentration data has a resolution of 2-10 seconds. Due to the influence of vehicle and road conditions, the data fluctuates significantly. Furthermore, considering spatiotemporal matching with data from multiple sensors, the original monitoring data is resampled over time, averaging the original data to the minute. Specifically, a window extending 30 seconds before and after each minute is set. Within this window, all original time-resolution monitoring data are extracted, and their arithmetic mean is calculated to obtain the current time-resampled result. The mathematical expression for data resampling is:

[0021] (1)

[0022] in, This represents the CO2 concentration value after resampling at time t. , and This represents the original temporal resolution CO2 concentration monitoring value within a 30-second time window before and after time t, where w is the number of monitoring data points within the time window.

[0023] Furthermore, the specific steps of step 3 are as follows:

[0024] Step 31: Based on the collected road scene images, use deep learning model technology to identify vehicles around the mobile monitoring vehicle. Randomly select a certain number of image frames as training samples from the road scene image sequence collected by the mobile monitoring. Manually annotate these sample images to mark the vehicle information objects in each frame, and save the annotation results as a YOLO format label file.

[0025] Step 32: Train the YOLOv8 model using labeled sample data. Learn the feature representation of vehicle objects in the image through convolutional neural network, and predict the position and category of the vehicle based on prior knowledge. During the training process, optimize the model parameters through backpropagation algorithm so that it can accurately identify vehicle objects in the image.

[0026] Step 33: Apply the trained YOLOv8 model to the panoramic image sequence collected by mobile monitoring, and perform vehicle detection on each frame of the image. The model performs dense sampling on the image by sliding window, and performs feature extraction and vehicle judgment on each window area. By setting a confidence threshold, most of the false detection results are filtered out to obtain the vehicle detection results.

[0027] Step 34: Based on the detection results of the YOLOv8 model, count the vehicles identified in each frame of the image to obtain the number of vehicles at the current time. Match the number of vehicles corresponding to each frame of the image with the CO2 concentration monitoring data according to the monitoring time to obtain the matching data pair of traffic flow and CO2 concentration.

[0028] Furthermore, the specific steps of step 4 are as follows:

[0029] Step 41: For urban POI data and high spatial resolution population distribution data, set a buffer zone with a radius of 500m for each CO2 concentration monitoring point, and count the total number of POIs and the total population of monitoring points within the buffer zone as the POI and population data values ​​corresponding to that monitoring point.

[0030] Step 42: For high spatial resolution land use data, set up a buffer zone with a radius of 500m for each CO2 concentration monitoring point, and count the area proportion of industrial land, commercial service land, residential land, forest land, grassland and water land within the buffer zone. The proportion of each type of land use corresponding to the monitoring point is used as the land use type data corresponding to the monitoring point.

[0031] Step 43: Match the collected data on the number of road lanes and road width with the road segment information based on the latitude and longitude coordinates of the road monitoring points;

[0032] Step 44: Match the road congestion data with the average vehicle speed data one by one according to the latitude and longitude coordinates of the road monitoring points and the monitoring time.

[0033] Step 45: Based on the high spatiotemporal resolution surface meteorological element data and atmospheric background CO2 concentration data, define a grid with the same spatial resolution for the road monitoring data, and determine the grid location of the road monitoring point through the latitude and longitude coordinates. Combined with the road CO2 concentration data collection time, realize the spatial and temporal matching between the monitoring point data and meteorological elements, and complete the construction of the multi-source input feature dataset.

[0034] Furthermore, the specific steps of step 5 are as follows:

[0035] Step 51: Standardize the data by selecting the maximum and minimum values: Subtract the minimum value from the data, and then divide by the range of data values ​​to ensure that the data values ​​are between [0,1]. The formula is as follows:

[0036]

[0037] in, This represents the minimum value in the original dataset. This represents the maximum value in the original dataset. These are the original values. These are the standardized values;

[0038] Step 52: Randomly divide the standardized feature dataset and the corresponding true CO2 concentration values ​​into two parts: a training set and a validation set, in an 8:2 ratio. The training set is used for model parameter learning and optimization, while the validation set is used to evaluate the model's generalization performance and prediction accuracy.

[0039] Furthermore, the specific steps of step 6 are as follows:

[0040] Step 61: Input training set data and use grid search to fine-tune the key hyperparameters of the random forest model. Determine the hyperparameters that need to be optimized: number of decision trees, maximum number of features, maximum tree depth, minimum number of split samples, and minimum number of samples per node.

[0041] Step 62: Set the value range of the hyperparameters to be optimized: For each hyperparameter to be optimized, set a set of discrete values ​​as the search space. Set the number of decision trees in the range of [300, 800] with a step size of 50 to generate a set of discrete values; set the maximum number of features to the two options of [sqrt, auto]; set the maximum tree depth in the range of [10, 100] with a step size of 10 to generate a set of discrete values; set the minimum number of split samples to three discrete values ​​of [2, 5, 10]; set the minimum number of samples per node to three discrete values ​​of [1, 2, 4].

[0042] Step 63: Generate hyperparameter combinations: Generate all possible hyperparameter combinations based on the set hyperparameter value range;

[0043] Step 64: Train and evaluate the model: For each combination of hyperparameters, use the training set data to train the random forest model using the five-fold cross-validation method. The mean squared error (MSE) is used as the performance metric for the model. The hyperparameter combination with the smallest MSE is considered the optimal hyperparameter combination. The formula for calculating the MSE is:

[0044]

[0045] in, This represents the observed CO2 concentration. This represents the predicted CO2 concentration, where n represents the number of matching data pairs.

[0046] Step 65: Construct a Long Short-Term Memory (LSTM) network model, including an input layer, hidden layers, and an output layer. The hidden layers consist of two LSTM layers and one fully connected layer. The input data of the LSTM model is in a three-dimensional format. The number of samples represents the number of samples in the training dataset. The time step is set to a 1-hour sliding time window, representing the prediction of the CO2 concentration for the current hour based on the CO2 concentration feature sequence of the previous hour. The number of features represents the input feature vector. For parameter settings, each hidden layer is trained in 20 rounds, the batch size is 64, the optimizer is adaptive moment estimation, the initial learning rate is 0.001, and the loss function is mean absolute error. After the model is deployed, the training set data is input into the model for training.

[0047] Furthermore, the specific steps of step 7 are as follows:

[0048] Step 71: Input validation set data. Select the coefficient of determination and root mean square error as evaluation metrics for the model. The coefficient of determination measures the model's ability to explain the variance of the data, and its value ranges from 0 to 1. R0 2 The closer the RSE is to 1, the better the model fits. Root error (RMSE) is the square root of the mean of the squares of the differences between predicted and true values. A smaller RMSE indicates better predictive ability. 2 The formula for calculating RMSE is:

[0049]

[0050] in, This represents the observed CO2 concentration. This represents the predicted CO2 concentration. denoted as the arithmetic mean of CO2 concentration observations, and n represents the number of matched data pairs.

[0051] Furthermore, the specific steps of step 8 are as follows:

[0052] Step 81: Based on the trained and validated random forest model and long short-term memory network model for simulating CO2 concentration on urban roads, select the model with the largest coefficient of determination and the smallest root mean square error as the final model, input the feature dataset into the model, and realize the hourly simulation of CO2 concentration on major urban roads.

[0053] The present invention has the following advantages: By integrating multi-source heterogeneous data such as CO2 concentration data from road mobile monitoring, road information and traffic data, meteorological data and human geography data, the present invention constructs a CO2 concentration simulation model based on an integrated algorithm of random forest (RF) model and long short-term memory network (LSTM) model, realizing continuous dynamic simulation and real-time prediction of CO2 concentration in urban road networks.

[0054] This invention fully leverages the value of monitoring data, developing an integrated algorithm-driven method for simulating and predicting urban road CO2 concentrations based on multi-source data. This method is scalable in both temporal and spatial dimensions, enabling a comprehensive simulation of changes in urban road CO2 concentrations and a deeper exploration of the dynamic trends in urban road carbon emissions. This innovation not only significantly reduces the cost of urban road carbon emission monitoring but also provides substantial data support for low-carbon urban development, contributing to sustainable development goals.

[0055] In simulating CO2 concentration on urban roads, this invention addresses the issue that a single model may not be able to accurately simulate the concentration. Instead, it employs an ensemble algorithm that combines a random forest model with a long short-term memory network model to select the optimal model for simulating road CO2 concentration, thereby effectively improving the accuracy of the simulation. Attached Figure Description

[0056] Figure 1 This is a technical roadmap of the present invention;

[0057] Figure 2 The diagram shows the simulation results of this invention. Detailed Implementation

[0058] This invention provides a method for hourly simulation and prediction of urban road CO2 concentration based on multi-source data fusion and integration algorithms. The specific steps of the method are as follows:

[0059] Step 1: Construct a road mobile monitoring system, design a mobile monitoring scheme, and acquire data on CO2 concentration, vehicle speed, road scene images, wind speed, wind direction, temperature, humidity, and the corresponding time and geographic coordinates of monitoring data for some urban roads.

[0060] Step 2: Based on CO2 concentration data from some road mobile monitoring systems, perform data cleaning and preprocessing;

[0061] Step 3: Combine the corresponding road scene images and use image recognition algorithms such as YOLO-v8 to extract information on road vehicle driving conditions and vehicle types;

[0062] Step 4: Collect road attribute information, traffic conditions, meteorological data, atmospheric background CO2 concentration observation data, land use data, POI data, and population distribution data. Match these data with the road CO2 concentration monitoring data and vehicle information extraction results using a spatiotemporal matching algorithm to construct a multi-source input feature variable dataset.

[0063] Step 5: Using road CO2 concentration data as the true value and other data as input feature variables, standardize the data and divide the data to form a training set and a validation set;

[0064] Step 6: Based on the collected multi-source heterogeneous data, construct an hourly simulation and prediction model of urban road CO2 concentration based on the random forest model and the long short-term memory network integrated algorithm, and use the training set to train the model.

[0065] Step 7: Use the validation set to verify and evaluate the accuracy of the model;

[0066] Step 8: Based on the trained urban road CO2 concentration simulation and prediction model, simulate the carbon concentration of each road in the city hourly.

[0067] In this embodiment, the specific steps of step 1 are as follows:

[0068] Step 11: By installing greenhouse gas monitoring equipment, GPS system, panoramic camera and meteorological sensor system on vehicles, a mobile monitoring system for atmospheric CO2 on urban roads was constructed. Based on the mobile monitoring system, the geographical coordinates of urban road CO2 concentration, road scene images, vehicle speed, wind speed, wind direction, temperature, humidity and monitoring data were obtained.

[0069] Step 12: Subsequently, taking into account the overall urban road network planning information, road attribute characteristics, time attribute characteristics, distribution information of key urban pollution sources, topography and land use information, design a mobile monitoring route plan for CO2 on typical urban roads. The mobile monitoring route plan includes the mobile monitoring time range, mobile monitoring route, key monitoring sections, and approximate time information.

[0070] In this embodiment, the specific steps of step 2 are as follows:

[0071] Step 21: First, process the monitoring data of the road CO2 concentration mobile monitoring system. Since different sensors collect different information, perform spatiotemporal matching of CO2 concentration monitoring data, GPS data, meteorological data, and panoramic image data based on the data collection time information.

[0072] Step 22: The mobile monitoring CO2 concentration data has a resolution of 2-10 seconds. Due to the influence of vehicle and road conditions, the data fluctuates significantly. Furthermore, considering spatiotemporal matching with data from multiple sensors, the original monitoring data is resampled over time, averaging the original data to the minute. Specifically, a window extending 30 seconds before and after each minute is set. Within this window, all original time-resolution monitoring data are extracted, and their arithmetic mean is calculated to obtain the current time-resampled result. The mathematical expression for data resampling is:

[0073] (1)

[0074] in, This represents the CO2 concentration value after resampling at time t. , and This represents the original temporal resolution CO2 concentration monitoring value within a 30-second time window before and after time t, where w is the number of monitoring data points within the time window.

[0075] In this embodiment, step 3 is as follows:

[0076] Step 31: Based on the collected road scene images, use deep learning model technology to identify vehicles around the mobile monitoring vehicle. Randomly select a certain number of image frames as training samples from the road scene image sequence collected by the mobile monitoring. Manually annotate these sample images to mark the vehicle information objects in each frame, and save the annotation results as a YOLO format label file.

[0077] Step 32: Train the YOLOv8 model using labeled sample data. Learn the feature representation of vehicle objects in the image through convolutional neural network, and predict the position and category of the vehicle based on prior knowledge. During the training process, optimize the model parameters through backpropagation algorithm so that it can accurately identify vehicle objects in the image.

[0078] Step 33: Apply the trained YOLOv8 model to the panoramic image sequence collected by mobile monitoring, and perform vehicle detection on each frame of the image. The model performs dense sampling on the image by sliding window, and performs feature extraction and vehicle judgment on each window area. By setting a confidence threshold, most of the false detection results are filtered out to obtain the vehicle detection results.

[0079] Step 34: Based on the detection results of the YOLOv8 model, count the vehicles identified in each frame of the image to obtain the number of vehicles at the current time. Match the number of vehicles corresponding to each frame of the image with the CO2 concentration monitoring data according to the monitoring time to obtain the matching data pair of traffic flow and CO2 concentration.

[0080] In this embodiment, step 4 is as follows:

[0081] Step 41: For urban POI data and high spatial resolution population distribution data, set a buffer zone with a radius of 500m for each CO2 concentration monitoring point, and count the total number of POIs and the total population of monitoring points within the buffer zone as the POI and population data values ​​corresponding to that monitoring point.

[0082] Step 42: For high spatial resolution land use data, set up a buffer zone with a radius of 500m for each CO2 concentration monitoring point, and count the area proportion of industrial land, commercial service land, residential land, forest land, grassland and water land within the buffer zone. The proportion of each type of land use corresponding to the monitoring point is used as the land use type data corresponding to the monitoring point.

[0083] Step 43: Match the collected data on the number of road lanes and road width with the road segment information based on the latitude and longitude coordinates of the road monitoring points;

[0084] Step 44: Match the road congestion data with the average vehicle speed data one by one according to the latitude and longitude coordinates of the road monitoring points and the monitoring time.

[0085] Step 45: Based on the high spatiotemporal resolution surface meteorological element data and atmospheric background CO2 concentration data, define a grid with the same spatial resolution for the road monitoring data, and determine the grid location of the road monitoring point through the latitude and longitude coordinates. Combined with the road CO2 concentration data collection time, realize the spatial and temporal matching between the monitoring point data and meteorological elements, and complete the construction of the multi-source input feature dataset.

[0086] In this embodiment, step 5 is as follows:

[0087] Step 51: Standardize the data by selecting the maximum and minimum values: Subtract the minimum value from the data, and then divide by the range of data values ​​to ensure that the data values ​​are between [0,1]. The formula is as follows:

[0088]

[0089] in, This represents the minimum value in the original dataset. This represents the maximum value in the original dataset. These are the original values. These are the standardized values;

[0090] Step 52: Randomly divide the standardized feature dataset and the corresponding true CO2 concentration values ​​into two parts: a training set and a validation set, in an 8:2 ratio. The training set is used for model parameter learning and optimization, while the validation set is used to evaluate the model's generalization performance and prediction accuracy.

[0091] In this embodiment, step 6 is as follows:

[0092] Step 61: Input training set data and use grid search to fine-tune the key hyperparameters of the random forest model. Determine the hyperparameters that need to be optimized: number of decision trees, maximum number of features, maximum tree depth, minimum number of split samples, and minimum number of samples per node.

[0093] Step 62: Set the value range of the hyperparameters to be optimized: For each hyperparameter to be optimized, set a set of discrete values ​​as the search space. Set the number of decision trees in the range of [300, 800] with a step size of 50 to generate a set of discrete values; set the maximum number of features to the two options of [sqrt, auto]; set the maximum tree depth in the range of [10, 100] with a step size of 10 to generate a set of discrete values; set the minimum number of split samples to three discrete values ​​of [2, 5, 10]; set the minimum number of samples per node to three discrete values ​​of [1, 2, 4].

[0094] Step 63: Generate hyperparameter combinations: Generate all possible hyperparameter combinations based on the set hyperparameter value range;

[0095] Step 64: Train and evaluate the model: For each combination of hyperparameters, use the training set data to train the random forest model using the five-fold cross-validation method. The mean squared error (MSE) is used as the performance metric for the model. The hyperparameter combination with the smallest MSE is considered the optimal hyperparameter combination. The formula for calculating the MSE is:

[0096]

[0097] in, This represents the observed CO2 concentration. This represents the predicted CO2 concentration, where n represents the number of matching data pairs.

[0098] Step 65: Construct a Long Short-Term Memory (LSTM) network model, including an input layer, hidden layers, and an output layer. The hidden layers consist of two LSTM layers and one fully connected layer. The input data of the LSTM model is in a three-dimensional format. The number of samples represents the number of samples in the training dataset. The time step is set to a 1-hour sliding time window, representing the prediction of the CO2 concentration for the current hour based on the CO2 concentration feature sequence of the previous hour. The number of features represents the input feature vector. For parameter settings, each hidden layer is trained in 20 rounds, the batch size is 64, the optimizer is adaptive moment estimation, the initial learning rate is 0.001, and the loss function is mean absolute error. After the model is deployed, the training set data is input into the model for training.

[0099] In this embodiment, step 7 is as follows:

[0100] Step 71: Input validation set data. Select the coefficient of determination and root mean square error as evaluation metrics for the model. The coefficient of determination measures the model's ability to explain the variance of the data, and its value ranges from 0 to 1. R0 2 The closer the RSE is to 1, the better the model fits. Root error (RMSE) is the square root of the mean of the squares of the differences between predicted and true values. A smaller RMSE indicates better predictive ability. 2The formula for calculating RMSE is:

[0101]

[0102] in, This represents the observed CO2 concentration. This represents the predicted CO2 concentration. denoted as the arithmetic mean of CO2 concentration observations, and n represents the number of matched data pairs.

[0103] In this embodiment, step 8 is as follows:

[0104] Step 81: Based on the trained and validated random forest model and long short-term memory network model for simulating CO2 concentration on urban roads, select the model with the largest coefficient of determination and the smallest root mean square error as the final model, input the feature dataset into the model, and realize the hourly simulation of CO2 concentration on major urban roads.

[0105] Although specific embodiments of the present invention have been described in detail with reference to the accompanying drawings, this should not be construed as limiting the scope of protection of the present invention. Various modifications and variations that can be made by those skilled in the art without inventive effort within the scope described in the claims are still within the scope of protection of the present invention.

Claims

1. A method for hourly simulation and prediction of urban road CO2 concentration based on multi-source data fusion and integration algorithms, characterized in that: The specific steps of the method are as follows: Step 1: Construct a road mobile monitoring system, design a mobile monitoring scheme, and acquire data on CO2 concentration, vehicle speed, road scene images, wind speed, wind direction, temperature, humidity, and the corresponding time and geographic coordinates of monitoring data for some urban roads. Step 2: Based on CO2 concentration data from some road mobile monitoring systems, perform data cleaning and preprocessing; Step 3: Combine the corresponding road scene images and use image recognition algorithms such as YOLO-v8 to extract information on road vehicle driving conditions and vehicle types; Step 4: Collect road attribute information, traffic conditions, meteorological data, atmospheric background CO2 concentration observation data, land use data, POI data, and population distribution data. Match these data with the road CO2 concentration monitoring data and vehicle information extraction results using a spatiotemporal matching algorithm to construct a multi-source input feature variable dataset. The specific steps of step 4 are as follows: Step 41: For urban POI data and high spatial resolution population distribution data, set a buffer zone with a radius of 500m for each CO2 concentration monitoring point, and count the total number of POIs and the total population of monitoring points within the buffer zone as the POI and population data values ​​corresponding to that monitoring point. Step 42: For high spatial resolution land use data, set up a buffer zone with a radius of 500m for each CO2 concentration monitoring point, and count the area proportion of industrial land, commercial service land, residential land, forest land, grassland and water land within the buffer zone. The proportion of each type of land use corresponding to the monitoring point is used as the land use type data corresponding to the monitoring point. Step 43: Match the collected data on the number of road lanes and road width with the road segment information based on the latitude and longitude coordinates of the road monitoring points; Step 44: Match the road congestion data with the average vehicle speed data one by one according to the latitude and longitude coordinates of the road monitoring points and the monitoring time. Step 45: Based on the high spatiotemporal resolution surface meteorological element data and atmospheric background CO2 concentration data, define a grid with the same spatial resolution for the road monitoring data, and determine the grid location of the road monitoring point through the latitude and longitude coordinates. Combined with the road CO2 concentration data collection time, realize the spatial and temporal matching between the monitoring point data and meteorological elements, and complete the construction of the multi-source input feature dataset. Step 5: Using road CO2 concentration data as the true value and other data as input feature variables, standardize the data and divide the data to form a training set and a validation set; Step 6: Based on the collected multi-source heterogeneous data, construct an hourly simulation and prediction model of urban road CO2 concentration based on the random forest model and the long short-term memory network integrated algorithm, and use the training set to train the model. Step 7: Use the validation set to verify and evaluate the accuracy of the model; Step 8: Based on the trained urban road CO2 concentration simulation and prediction model, simulate the carbon concentration of each road in the city hourly.

2. The method for hourly simulation and prediction of urban road CO2 concentration based on multi-source data fusion and integration algorithm as described in claim 1, characterized in that: The specific steps of step 1 are as follows: Step 11: By installing greenhouse gas monitoring equipment, GPS system, panoramic camera and meteorological sensor system on vehicles, a mobile monitoring system for atmospheric CO2 on urban roads was constructed. Based on the mobile monitoring system, the geographical coordinates of urban road CO2 concentration, road scene images, vehicle speed, wind speed, wind direction, temperature, humidity and monitoring data were obtained. Step 12: Subsequently, taking into account the overall urban road network planning information, road attribute characteristics, time attribute characteristics, distribution information of key urban pollution sources, topography and land use information, design a mobile monitoring route scheme for CO2 on typical urban roads. The mobile monitoring route scheme includes the mobile monitoring time range, mobile monitoring route, key monitoring sections, and time information.

3. The hourly simulation and prediction method for urban road CO2 concentration based on multi-source data fusion and integration algorithm as described in claim 1, characterized in that: The specific steps of step 2 are as follows: Step 21: First, process the monitoring data of the road CO2 concentration mobile monitoring system. Since different sensors collect different information, perform spatiotemporal matching of CO2 concentration monitoring data, GPS data, meteorological data, and panoramic image data based on the data collection time information. Step 22: The mobile monitoring CO2 concentration data has a resolution of 2-10 seconds. Due to the influence of vehicle and road conditions, the data fluctuates significantly. Furthermore, considering spatiotemporal matching with data from multiple sensors, the original monitoring data is resampled over time, averaging the original data to the minute. Specifically, a window extending 30 seconds before and after each minute is set. Within this window, all original time-resolution monitoring data are extracted, and their arithmetic mean is calculated to obtain the current time-resampled result. The mathematical expression for data resampling is: （1）； in, This represents the CO2 concentration value after resampling at time t. , and This represents the original temporal resolution CO2 concentration monitoring value within a 30-second time window before and after time t, where w is the number of monitoring data points within the time window.

4. The hourly simulation and prediction method for urban road CO2 concentration based on multi-source data fusion and integration algorithm as described in claim 1, characterized in that: The specific steps of step 3 are as follows: Step 31: Based on the collected road scene images, use deep learning model technology to identify vehicles around the mobile monitoring vehicle. Randomly select a certain number of image frames as training samples from the road scene image sequence collected by the mobile monitoring. Manually annotate these sample images to mark the vehicle information objects in each frame, and save the annotation results as a YOLO format label file. Step 32: Train the YOLOv8 model using labeled sample data. Learn the feature representation of vehicle objects in the image through convolutional neural network, and predict the position and category of the vehicle based on prior knowledge. During the training process, optimize the model parameters through backpropagation algorithm so that it can accurately identify vehicle objects in the image. Step 33: Apply the trained YOLOv8 model to the panoramic image sequence collected by mobile monitoring, and perform vehicle detection on each frame of the image. The model performs dense sampling on the image by sliding window, and performs feature extraction and vehicle judgment on each window area. By setting a confidence threshold, most of the false detection results are filtered out to obtain the vehicle detection results. Step 34: Based on the detection results of the YOLOv8 model, count the vehicles identified in each frame of the image to obtain the number of vehicles at the current time. Match the number of vehicles corresponding to each frame of the image with the CO2 concentration monitoring data according to the monitoring time to obtain the matching data pair of traffic flow and CO2 concentration.

5. The hourly simulation and prediction method for urban road CO2 concentration based on multi-source data fusion and integration algorithm as described in claim 1, characterized in that: The specific steps of step 5 are as follows: Step 51: Standardize the data by selecting the maximum and minimum values: Subtract the minimum value from the data, and then divide by the range of data values ​​to ensure that the data values ​​are between [0,1]. The formula is as follows: ； in, This represents the minimum value in the original dataset. This represents the maximum value in the original dataset. These are the original values. These are the standardized values; Step 52: Randomly divide the standardized feature dataset and the corresponding true CO2 concentration values ​​into two parts: a training set and a validation set, in an 8:2 ratio. The training set is used for model parameter learning and optimization, while the validation set is used to evaluate the model's generalization performance and prediction accuracy.

6. The hourly simulation and prediction method for urban road CO2 concentration based on multi-source data fusion and integration algorithm as described in claim 1, characterized in that: The specific steps of step 6 are as follows: Step 61: Input training set data and use grid search to fine-tune the key hyperparameters of the random forest model. Determine the hyperparameters that need to be optimized: number of decision trees, maximum number of features, maximum tree depth, minimum number of split samples, and minimum number of samples per node. Step 62: Set the value range of the hyperparameters to be optimized: For each hyperparameter to be optimized, set a set of discrete values ​​as the search space. Set the number of decision trees in the range of [300, 800] with a step size of 50 to generate a set of discrete values; set the maximum number of features to the two options of [sqrt, auto]; set the maximum tree depth in the range of [10, 100] with a step size of 10 to generate a set of discrete values; set the minimum number of split samples to three discrete values ​​of [2, 5, 10]; set the minimum number of samples per node to three discrete values ​​of [1, 2, 4]. Step 63: Generate hyperparameter combinations: Generate all possible hyperparameter combinations based on the set hyperparameter value range; Step 64: Train and evaluate the model: For each combination of hyperparameters, use the training set data to train the random forest model using the five-fold cross-validation method. The mean squared error (MSE) is used as the performance metric for the model. The hyperparameter combination with the smallest MSE is considered the optimal hyperparameter combination. The formula for calculating the MSE is: ； in, This represents the observed CO2 concentration. This represents the predicted CO2 concentration, where n represents the number of matching data pairs. Step 65: Construct a Long Short-Term Memory (LSTM) network model, including an input layer, hidden layers, and an output layer. The hidden layers consist of two LSTM layers and one fully connected layer. The input data of the LSTM model is in a three-dimensional format. The number of samples represents the number of samples in the training dataset. The time step is set to a 1-hour sliding time window, representing the prediction of the CO2 concentration for the current hour based on the CO2 concentration feature sequence of the previous hour. The number of features represents the input feature vector. In terms of parameter settings, the number of training rounds for each hidden layer is set to 20, the batch size is 64, the optimizer is the adaptive moment estimator, the initial learning rate is 0.001, and the loss function is the mean absolute error. After the model is deployed, the training set data is input into the model for training.

7. The hourly simulation and prediction method for urban road CO2 concentration based on multi-source data fusion and integration algorithm as described in claim 1, characterized in that: The specific steps of step 7 are as follows: Step 71: Input validation set data. Select the coefficient of determination and root mean square error (RMSE) as evaluation metrics for the model. The coefficient of determination measures the model's ability to explain the variance of the data, and its value ranges from 0 to 1. R0 2 The closer the RSE is to 1, the better the model fits. Root error (RMSE) is the square root of the mean of the squares of the differences between predicted and true values. A smaller RMSE indicates better predictive ability. 2 The formula for calculating RMSE is: ； in, This represents the observed CO2 concentration. This represents the predicted CO2 concentration. denoted as the arithmetic mean of CO2 concentration observations, and n represents the number of matched data pairs.

8. The hourly simulation and prediction method for urban road CO2 concentration based on multi-source data fusion and integration algorithm as described in claim 1, characterized in that: The specific steps of step 8 are as follows: Step 81: Based on the trained and validated random forest model and long short-term memory network model for simulating CO2 concentration on urban roads, select the model with the largest coefficient of determination and the smallest root mean square error as the final model, input the feature dataset into the model, and realize the hourly simulation of CO2 concentration on major urban roads.

Images