Pollutant concentration prediction method based on LightGBM multi-source data fusion

Abstract

The invention belongs to the technical field of traffic pollution prediction, and particularly relates to a LightGBM-based multi-source data fusion pollutant concentration prediction method, which comprises the following steps of: 1, integrating and processing multi-source data, and constructing a plurality of feature sets; and 2, carrying out feature analysis on meteorological, pollutant and traffic index variables through correlation analysis and time sequence analysis, and providing support for modeling and model interpretation. And step 3, dividing a training test set according to a time sequence, and training and optimizing the LightGBM pollution prediction model through parameter tuning by taking minimization of RMSE as a target. And 4, verifying the performance of the model from prediction precision, spatial distribution and wind direction influence. And 5, analyzing and displaying the key driving factor and quantifying the contribution of the key driving factor. According to the invention, through feature engineering processing and model optimization of heterogeneous data such as weather and traffic, high-precision prediction and influence factor analysis of lane-level pollutant concentration are realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of traffic pollution prediction, and particularly relates to a pollution concentration prediction method based on LightGBM multi-source data fusion. BACKGROUND

[0002] With the acceleration of urbanization, motor vehicle emissions have become an important source of urban air pollution, and accurate prediction of lane-level pollution concentration is crucial for traffic pollution prevention and control. However, traditional pollution prediction models (such as linear regression and ordinary GBDT) face significant challenges in processing meteorological, traffic flow, and historical pollution concentration data. High-dimensional sparse traffic flow data (speed, flow, occupancy) has spatial and temporal heterogeneity, and traditional models are easily disturbed by redundant features, have low training efficiency and insufficient generalization ability, making it difficult to meet the needs of large-scale data processing.

[0003] The spatiotemporal dynamic coupling relationship between meteorological factors and traffic flow is complex, and the periodic jump in wind direction angle makes it difficult for traditional methods to capture its continuous influence on pollution diffusion. In addition, there is a lack of quantitative analysis of spatial and temporal heterogeneity such as morning and evening peak congestion and main road pollution aggregation. LightGBM, as a high-efficiency gradient boosting decision tree model, is suitable for high-dimensional data modeling, but its application in lane-level pollution prediction has specific technical gaps. Existing solutions do not fully utilize its gradient one-sided sampling and mutual exclusion feature bundling mechanisms to address traffic data sparsity, and lack key feature engineering optimizations such as wind direction sine and cosine conversion and multi-source data spatiotemporal alignment, resulting in limited model fitting ability for complex pollution diffusion mechanisms. Meanwhile, the feature attribution method combined with SHAP analysis has not been systematically introduced, making it difficult to provide evidence for prevention and control strategies through quantitative feature contribution.

[0004] Therefore, there is an urgent need for a lightweight method that integrates innovative feature engineering, model optimization, and explainability analysis to break through the bottlenecks of traditional models in multi-source data processing, spatiotemporal feature modeling, and machine interpretation, achieving dual improvement in high-precision prediction and pollution cause analysis, and providing technical support for fine urban traffic pollution control. SUMMARY

[0005] The purpose of the present application is to combine the advantages of LightGBM model in efficiently processing multi-source heterogeneous data, solve the technical problems of existing pollution prediction methods in high-dimensional sparse traffic data processing, spatiotemporal correlation feature modeling, and explainability analysis, and provide a pollution concentration prediction method based on LightGBM multi-source data fusion.

[0006] To achieve the above purpose, the present application adopts the following technical solutions:

[0007] A LightGBM-based multi-source data fusion pollutant concentration prediction method, comprising the following steps:

[0008] Step one: integrate and process multi-source data, and construct multiple feature sets to realize the spatio-temporal consistency integration of air quality, meteorological, and traffic three-source heterogeneous data, and eliminate the time difference and spatial misplacement problems in data collection.

[0009] Step two: through correlation analysis and time series analysis, the meteorological, pollutant and traffic index variables are analyzed, which provides support for modeling and model interpretation.

[0010] Step three: divide the training test set according to the time sequence, take the minimum RMSE as the target, and train and optimize the LightGBM pollutant prediction model through parameter tuning.

[0011] Step four: verify the model performance from the prediction accuracy, spatial distribution, and wind direction influence.

[0012] Step five: analyze and display the key driving factors and quantify their contribution.

[0013] As a preferred solution, in step one, the specific method includes:

[0014] Divide the time axis into whole point periods (such as 00:00-01:00), and use backward resampling for non-whole point data:

[0015] (1)

[0016] Wherein, is the time identifier after resampling, is the original timestamp.

[0017] Ensure that all data is aligned to the whole point time; establish a 3km radius buffer area around the monitoring site, and implement inverse distance weighted interpolation on traffic sensor data:

[0018] (2)

[0019] (3)

[0020] Wherein, is the Euclidean distance between the sensor and the station, is the original observation value collected by the traffic sensor, is the calculated weight, is the value calculated by inverse distance weighted interpolation.

[0021] Wind direction periodicity processing

[0022] For 0°-360° ring quantity characteristics, the sine and cosine decomposition is implemented:

[0023] (4)

[0024] wherein, is the characteristic value obtained by cosine transformation of the original wind direction angle, is the characteristic value obtained by sine transformation of the original wind direction angle, is the original wind direction angle.

[0025] The converted characteristic value domain is [-1, 1], eliminating the sudden error of 360°→0°;

[0026] Missing value processing strategy:

[0027] Forward filling: for time series data (such as traffic flow), the previous valid value is used to fill the current missing value; mean interpolation: for non-time sequence features (such as air temperature), the mean value of the same hour history (near 30 days) is used to fill; preferentially use forward filling, if the continuous missing is more than 3 hours, switch to mean interpolation.

[0028] As a preferred solution, in step two, before constructing the pollution prediction model based on LightGBM, the feature correlation and time sequence rule of multi-source data need to be systematically analyzed to provide scientific basis for model input, and the specific method includes:

[0029] Variable correlation analysis

[0030] The linear correlation between variables is quantified based on Pearson correlation, and the calculation formula is:

[0031] (5)

[0032] (6)

[0033] wherein, is the observation value of variable , is the observation value of variable . is the average value of the observation value of variable , is the average value of the observation value of variable . is the predicted value of variable , is the sample size, is the determination coefficient, is the Person correlation coefficient.

[0034] By calculating the correlation coefficient matrix between pollutants, meteorological and traffic variables and visualizing it as a heat map;

[0035] Time Series Pattern Mining

[0036] The time series decomposition model is used to analyze the fluctuation pattern of variables, and the sequence is split into trend term (Tt), cycle term (St) and residual term (Rt):

[0037] At the same time, the daily variation trends of pollutant concentrations and meteorological influencing factors are plotted.

[0038] As a preferred solution, in step 3, the dataset is divided into a training set (the first 80% period) and a test set (the last 20% period) in chronological order to avoid future information leakage;

[0039] A sliding window strategy was used in the training set to construct a 5-fold cross-validation subset, with each fold separated by 24 hours to preserve the daily cycle characteristics;

[0040] Root mean square error (RMSE): The main optimization objective, which measures the deviation between the predicted value and the true value:

[0041] (7)

[0042] Mean absolute error (MAE): Assists in evaluating prediction stability:

[0043] (8)

[0044] Coefficient of determination (R²): quantifies the explanatory power of the model:

[0045] (9)

[0046] in, is a variable The observed value of is a variable The predicted value of is the sample size.

[0047] The model was initialized with key parameters: number of leaves 64, learning rate 0.005, feature sampling rate 0.75, and minimum subsample size 30. A grid search system was used to optimize the learning rate (0.001–0.01), number of leaves (32–128), feature sampling rate (0.5–0.9), and early stopping rounds (30–50). Parameter stability was evaluated using 5-fold cross-validation, and a learning rate of 0.005, number of leaves 64, sampling rate 0.75, and 50 early stopping rounds were ultimately selected.

[0048] During the training process, the validation set RMSE is monitored in real time. When the indicator does not improve for 50 consecutive rounds, the training is terminated to effectively suppress overfitting.

[0049] As the preferred solution, in step 4, we first observed the results. After completing cross-validation, the model performed well on the test set. After fitting the training set, the RMSE (RMSE) on the test set was 2.25, the MAE (MAE) was 0.83, and the coefficient of determination (R²) reached 0.915. This demonstrates that the model effectively captures the nonlinear behavior of lane-level pollution diffusion and demonstrates high confidence in its predictive capabilities.

[0050] After completing the model training, in order to more intuitively demonstrate the performance of the model in actual prediction, the goodness of fit is first evaluated by comparing the predicted values ​​with the actual observed values.

[0051] As a preferred solution, in step 5, the present invention reveals the core driving factors and their mechanisms that affect pollutant concentration prediction through SHAP (SHapley Additive exPlanations) value and feature importance analysis. The feature importance of the LightGBM model is calculated based on information gain, and its formula is:

[0052] (10)

[0053] in, is the feature importance, is the split gain of feature f in the t-th tree, is the total number of trees.

[0054] Furthermore, the nonlinear effects of variables are analyzed through the SHAP value. The core formula of the Shapley value is:

[0055] (11)

[0056] in, Features SHAP value, For the complete set of features, represents the model output of subset S, represents the factorial of the size of the subset S, Indicates the change in the prediction result after adding wind speed.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] The application creatively combines explainable feature engineering with deep spatiotemporal prediction, constructs a "mechanism-driven + data-driven" dual-engine modeling paradigm, and breaks through the technical bottleneck of traditional methods in microscale pollution prediction. Through adaptive spatiotemporal feature extraction and multi-modal data fusion technology, the application realizes dynamic visual analysis and key factor quantitative evaluation of the pollution diffusion process, and provides innovative technical path and methodological support for intelligent governance of urban traffic environment. BRIEF DESCRIPTION OF DRAWINGS

[0059] Figure 1 A multi-source variable Pearson correlation heat map of the application.

[0060] Figure 2 A daily variation trend chart of pollutants and influencing factors of the application.

[0061] Figure 3 A comparison chart of RSP concentration prediction of the application.

[0062] Figure 4 A RSP concentration spatial distribution chart of the application.

[0063] Figure 5 A pollution distribution chart of the application with wind direction vector superimposed.

[0064] Figure 6 A LightGBM feature importance column chart of the application.

[0065] Figure 7 A SHAP summary plot global attribution chart of the application.

[0066] Figure 8 A flowchart of the application. DETAILED DESCRIPTION

[0067] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings, so that those skilled in the art can better understand the advantages and features of the application, and the protection scope of the application can be defined more clearly. The described embodiments of the application are only some of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the protection scope of the application.

[0068] Embodiment:

[0069] Reference Figure 8 A pollution concentration prediction method based on LightGBM multi-source data fusion, comprising the following steps:

[0070] Step 1: Integrate multi-source data and construct multiple feature sets.

[0071] The spatio-temporal consistency of air quality, meteorological, and traffic heterogeneous data is integrated to eliminate the time difference and spatial misplacement problems in data collection. The time axis is divided into whole-hour periods (e.g., 00:00-01:00), and backward resampling is used for non-whole-hour data: (1) where, is the resampled time identifier, is the original timestamp.

[0072] A 3 km radius buffer is established around the monitoring site to implement inverse distance weighted interpolation for traffic sensor data: (2)

[0073] (3)

[0074] where, is the Euclidean distance between the sensor and the station, is the original observation value collected by the traffic sensor, is the calculated weight, is the value calculated by inverse distance weighted interpolation.

[0075] Wind direction periodic processing

[0076] For the 0°-360° ring characteristic, the cosine and sine decomposition is implemented: (4) where, is the cosine-transformed feature value of the original wind direction angle, is the sine-transformed feature value of the original wind direction angle, is the original wind direction angle.

[0077] The feature value domain after conversion is [-1, 1], eliminating the sudden error of 360°→0°;

[0078] Missing value processing strategy:

[0079] Forward filling: for time series data (such as traffic flow), use the previous valid value to fill the current missing value; mean interpolation: for non-time series features (such as air temperature), take the mean value of the same hour history (near 30 days) to fill. Prefer to use forward filling, if the continuous missing is more than 3 hours, switch to mean interpolation.

[0080] Step 2: Through correlation analysis and time series analysis, characteristic analysis of meteorological, pollutant and traffic indicator variables is carried out to provide support for modeling and model interpretation.

[0081] Before building a pollution prediction model based on LightGBM, it is necessary to systematically analyze the feature correlation and time series rules of multi-source data to provide a scientific basis for model input.

[0082] Variable correlation analysis

[0083] The linear association between variables is quantified based on the Pearson correlation coefficient, and its calculation formula is:

[0084] (5)

[0085] (6)

[0086] in, is a variable The observed value of is a variable Observed values. is a variable The average of the observations, is a variable The average of the observations. is a variable The predicted value of is the sample size, is the coefficient of determination, is the Person correlation coefficient.

[0087] The correlation coefficient matrix between pollutants, meteorological and traffic variables is calculated and visualized as a heat map.

[0088] like Figure 1 As shown, the following rules are found:

[0089] Strong positive correlation between pollutants: fine particulate matter ( , PM2.5) and inhalable particulate matter ( , PM 10 ) has a correlation coefficient of 0.92, indicating that their sources or diffusion mechanisms are highly similar;

[0090] Traffic indicator interactions: vehicle speed is positively correlated with traffic volume (r=0.65) and negatively correlated with road occupancy (r=−0.58), reflecting the direct impact of traffic congestion on emissions;

[0091] Meteorological factors influence differentiation: air temperature is negatively correlated with most pollutants (r≈−0.3), possibly because high temperatures promote diffusion; wind speed has a weak impact (|r|<0.2), or because of its directional effect, further analysis is needed in combination with vectorized wind direction.

[0092] Time series pattern mining

[0093] Using a time series decomposition model to analyze the variable fluctuation law, the sequence is divided into trend item (Tt), periodic item (St) and residual item (Rt):

[0094] At the same time, draw the daily variation trend of pollutant concentration and meteorological influence factors, as shown in Figure 2 It is found that:

[0095] RSP concentration is slightly higher in the daytime (especially from 8 to 17) than at night, showing a certain diurnal accumulation characteristics, but there is no obvious morning and evening peak fluctuation; while traffic speed decreases slightly before and after the morning peak (8), and then gradually rises, reflecting the congestion effect of commuting period. Wind speed rises slightly before and after noon, and stabilizes at night. Overall, there are some differences in the time series fluctuation between pollutants, traffic and meteorological variables, suggesting that the interaction and lag effect of time factors should be considered in modeling.

[0096] Through systematic feature analysis, the key driving factors and time series dependence relationship are determined, laying a foundation for feature engineering of LightGBM model, and improving prediction accuracy and interpretability.

[0097] Step three: divide the training and test sets according to the time sequence, minimize the RMSE as the target, and train and optimize the LightGBM pollution prediction model through parameter tuning.

[0098] Divide the data set into training set (first 80% period) and test set (last 20% period) in time sequence to avoid future information leakage.

[0099] In the training set, use sliding window strategy to build 5-fold cross-validation subsets, each interval is 24 hours to retain the daily cycle characteristics.

[0100] Root mean square error (RMSE): main optimization target, measure the deviation amplitude of predicted value and true value:

[0101] (7)

[0102] Mean absolute error (MAE): auxiliary evaluation of prediction stability:

[0103] (8)

[0104] Coefficient of determination (R²): quantifies the explanatory power of the model:

[0105] (9)

[0106] in, is a variable The observed value of is a variable The predicted value of is the sample size.

[0107] Key model parameters were initialized with the following: number of leaves 64, learning rate 0.005, feature sampling rate 0.75, and minimum subsample size 30. A grid search system was used to optimize the learning rate (0.001–0.01), number of leaves (32–128), feature sampling rate (0.5–0.9), and early stopping rounds (30–50). Parameter stability was evaluated using 5-fold cross-validation, and a learning rate of 0.005, number of leaves 64, sampling rate 0.75, and early stopping of 50 rounds were ultimately selected.

[0108] During the training process, the validation set RMSE is monitored in real time. When the indicator does not improve for 50 consecutive rounds, the training is terminated to effectively suppress overfitting.

[0109] Step 4: Verify model performance from the perspective of prediction accuracy, spatial distribution, and wind direction impact.

[0110] First, let's examine the results. After cross-validation, the model demonstrated strong predictive performance on the test set. After fitting the training set, the model achieved an RMSE of 2.25, a MAE of 0.83, and a coefficient of determination (R²) of 0.915. This demonstrates that the model effectively captures the nonlinear patterns of lane-level pollution diffusion and demonstrates high confidence in its predictive capabilities.

[0111] After completing the model training, in order to more intuitively demonstrate the performance of the model in actual prediction, the goodness of fit is first evaluated by comparing the predicted values ​​with the actual observed values. Figure 3 A trend chart comparing the model's predicted values ​​and true values ​​for the first 100 samples in the test set is plotted.

[0112] from Figure 3 It can be seen that the change trends of the predicted values ​​and the true values ​​during most periods are relatively consistent, and the model also has a certain ability to capture the response to pollutant peaks. However, there are still certain deviations in some high-fluctuation periods, which may be limited by the influence of local traffic or abnormal meteorological conditions.

[0113] To further reveal the model's ability to capture spatial differences in pollutant concentrations, a spatial distribution map of predicted pollutant concentrations was drawn.

[0114] like Figure 4As shown, the concentration of pollutants RSP presents significant differences in space, with high-value areas mainly concentrated in the north-central and eastern regions, and the pollution level is relatively high in areas with dense roads or heavy traffic, indicating that the model can better capture the actual differences in pollution distribution in the spatial dimension and has certain spatial fitting ability.

[0115] In addition, to more clearly reveal the diffusion trend of pollution with wind direction, the pollution spatial distribution map is redrawn after superimposing the wind direction vector. Figure 5 In the middle, the diffusion of pollutants in the dominant direction of the wind direction can be observed.

[0116] Figure 5 In the middle, the high concentration area of pollutants shows a certain diffusion trend in the dominant direction of the wind direction, especially in the central and eastern regions, and the RSP concentration extends along the wind direction path, indicating that wind has a significant impact on pollution transmission. This phenomenon verifies the response ability of the model to changes in pollution with wind direction. It is worth noting that the model introduces the sine and cosine components of the wind direction in the input features, thereby avoiding the periodic influence of angle data and effectively improving the expression effect of wind direction information in the model, further enhancing the fitting ability of the spatial diffusion characteristics of pollutants.

[0117] This verification system ensures the operational reliability of the model from the perspectives of data-driven and physical mechanism through hierarchical statistics, spatial autocorrelation test and meteorological model fusion, providing multi-dimensional technical support for precise pollution control.

[0118] Step five: analyze and display key driving factors and quantify their contribution.

[0119] Through SHAP (SHapley Additive exPlanations) values and feature importance analysis, the present invention reveals the core driving factors affecting the prediction of pollutant concentration and their mechanism. The feature importance of the LightGBM model is based on information gain, and its formula is:

[0120] (10)

[0121] wherein, is the feature importance, is the split gain of feature f in the tth tree, is the total number of trees.

[0122] Figure 6 is the feature importance ranking chart based on the output of the LightGBM model, which intuitively shows the contribution of each input variable to the prediction of pollutant concentration. For example, Figure 6As shown in Fig. 6, the importance of pollutant concentration (RSP), meteorological conditions such as temperature and CO, and other pollution indicators is significant, indicating that these variables play a core driving role in the prediction model. In terms of traffic factors, vehicle speed, traffic volume, and road occupancy also have high weights, reflecting the direct impact of motor vehicle activity intensity on pollution levels. In addition, the characteristic contribution of wind direction is still prominent after inputting it in the form of cosine values, verifying the effectiveness of vectorization in improving the expressiveness of wind direction features and providing a solid foundation for modeling the pollution transport of wind fields.

[0123] Furthermore, by analyzing the non-linear influence of variables through SHAP values, the core formula of Shapley value is:

[0124] (11)

[0125] where, is the SHAP value of feature , is the feature set, is the model output of subset S, is the factorial of the size of subset S, is the change in prediction after adding wind speed. From a global perspective, all features can be visualized on all samples based on SHAP values. The color represents the size of the feature value, and the distribution of points on the horizontal axis reflects the positive and negative contribution and influence intensity of the model prediction value.

[0126] As shown in Fig. 6, high values of FSP, temperature, and CO significantly increase the model output, indicating that they are positively correlated with pollution levels. The influence of NO2, wind speed, and vehicle speed shows a certain bidirectionality, and the specific contribution depends on the numerical size. Although the cosine value of wind direction has a small influence, it still has stable explanatory power, verifying the rationality of its vectorization processing. Overall, each feature has certain explanatory power for model prediction, and the influence direction and degree vary with the numerical value.

[0127] Figure 7

[0128] In summary, the model driving factors can be categorized into three types: first, pollutant concentration (cumulative contribution of 51.3%): directly representing emission intensity and chemical conversion; second, meteorological conditions (35.2%): regulating pollution distribution through diffusion efficiency and transmission path; third, traffic dynamics (29.1%): reflecting the spatiotemporal heterogeneity of motor vehicle emissions.

[0129] Based on the quantitative results, it is recommended to prioritize the control of NO2 and FSP emissions, and to start emergency emission reduction in low-temperature and calm weather , and optimize pollution evacuation paths based on wind direction (such as strengthening control in downstream areas dominated by wind direction). This analysis provides data-driven decision-making basis for targeted governance. ​​

[0130] In addition, the basic symbols in the embodiments of the present application are explained as follows:

[0131]

[0132] To sum up, through the innovative modeling and optimization scheme, the prediction ability of the lane pollution can be effectively improved, the prediction precision is as high as possible under the condition of ensuring that the model complexity is reasonable, and the lane-level pollution prevention and control strategy is more scientific and reliable.

[0133] The description and practice disclosed in the present application are easy to think and understand for ordinary skilled in the art, and some improvements and refinements can be made without departing from the principles of the present application. Therefore, the modifications or improvements made without departing from the spirit of the present application should be considered as the protection scope of the present application.

Claims

1. A pollutant concentration prediction method based on multi-source data fusion of LightGBM, characterized by: The steps include: Step 1: Integrate and process multi-source data and construct multiple feature sets to achieve spatiotemporal consistency of heterogeneous data from three sources: air quality, meteorology, and traffic, eliminating the time difference and spatial misalignment issues of data collection. Step 2: Conduct characteristic analysis of meteorological, pollutant, and traffic indicator variables through correlation analysis and time series analysis to provide support for modeling and model interpretation; Step 3: Divide the training and test sets by time series, minimize RMSE as the goal, and train and optimize the LightGBM pollution prediction model through parameter tuning; Step 4: Verify model performance from the perspective of prediction accuracy, spatial distribution, and wind direction impact; Step 5: Analyze and identify key drivers and quantify their contributions.

2. The pollutant concentration prediction method based on multi-source data fusion of LightGBM according to claim 1 is characterized in that: In step one, the specific methods include: Divide the time axis into hourly periods and use backward resampling for non-integer data: （1） in, is the time stamp after resampling, is the original timestamp; Ensure that all data is aligned to the hour; establish a 3km radius buffer around the monitoring station and perform inverse distance weighted interpolation on the traffic sensor data: （2） （3） in, is the Euclidean distance between the sensor and the site, is the original observation value collected by the traffic sensor, is the calculated weight, It is the value calculated by inverse distance weighted interpolation; ① Periodic processing of wind direction According to the 0°-360° circular quantity characteristics, sine and cosine decomposition is implemented: （4） in, is the eigenvalue obtained after cosine transform of the original wind direction angle, is the eigenvalue obtained after performing sine transformation on the original wind direction angle, is the original wind direction angle; After the transformation, the eigenvalue range is [-1, 1], eliminating the mutation error from 360° to 0°; Missing value handling strategy: Forward filling: For time series data, use the valid value of the previous moment to fill the current missing value; mean interpolation: For non-time series features, take the historical mean value of the same hour to fill the missing value; forward filling is used first. If the missing value exceeds 3 consecutive hours, switch to mean interpolation.

3. The pollutant concentration prediction method based on multi-source data fusion of LightGBM according to claim 2 is characterized in that: In step 2, before building a pollution prediction model based on LightGBM, it is necessary to systematically analyze the feature correlation and time series patterns of multi-source data to provide a scientific basis for model input. Specific methods include: Variable correlation analysis The linear association between variables is quantified based on the Pearson correlation coefficient, and its calculation formula is: （5） （6） in, is a variable The observed value of is a variable Observed values ​​of is a variable The average of the observed values, is a variable The average of the observations; is a variable The predicted value of is the sample size, is the coefficient of determination, is the Person correlation coefficient; By calculating the correlation coefficient matrix between pollutants, meteorological and traffic variables and visualizing it as a heat map; Time Series Pattern Mining The time series decomposition model is used to analyze the fluctuation pattern of variables, and the sequence is split into trend term Tt, cycle term St and residual term Rt: At the same time, the daily variation trends of pollutant concentrations and meteorological influencing factors are plotted.

4. The pollutant concentration prediction method based on multi-source data fusion of LightGBM according to claim 3 is characterized in that: In step 3, the dataset is divided into training and test sets in chronological order to avoid future information leakage; A sliding window strategy was used in the training set to construct a 5-fold cross-validation subset, with each fold separated by 24 hours to preserve the daily cycle characteristics; Root mean square error (RMSE): The main optimization objective, which measures the deviation between the predicted value and the true value: （7） Mean absolute error MAE: Assisted evaluation of prediction stability: （8） Coefficient of determination R²: quantifies the explanatory power of the model: （9） in, is a variable The observed value of is a variable The predicted value of is the sample size; The model was initialized with key parameters: number of leaves 64, learning rate 0.005, feature sampling rate 0.75, and minimum subsample size 30. A grid search system was used to optimize the learning rate between 0.001 and 0.01, number of leaves 32–128, feature sampling rate 0.5–0.9, and early stopping rounds 30–50. Parameter stability was evaluated using 5-fold cross-validation, ultimately selecting a learning rate of 0.005, number of leaves 64, sampling rate 0.75, and 50 early stopping rounds. During the training process, the validation set RMSE is monitored in real time. When the indicator does not improve for 50 consecutive rounds, the training is terminated to effectively suppress overfitting.

5. The pollutant concentration prediction method based on multi-source data fusion of LightGBM according to claim 4 is characterized in that: In step 4, after fitting the training set, the RMSE and MAE on the test set were 2.25 and 0.83, respectively, with an R² coefficient of determination of 0.

915. This indicates that the model can effectively capture the nonlinear behavior of lane-level pollution diffusion and demonstrates reliable predictive capabilities. After completing the model training, the goodness of fit is first evaluated by comparing the predicted values ​​with the actual observed values.

6. The pollutant concentration prediction method based on multi-source data fusion of LightGBM according to claim 5 is characterized in that: In step 5, through SHAP value and feature importance analysis, the feature importance of the LightGBM model is calculated based on information gain, and its formula is: （10） in, is the feature importance, is the split gain of feature f in the t-th tree, is the total number of trees; Furthermore, the nonlinear effects of variables are analyzed through the SHAP value. The core formula of the Shapley value is: （11） in, Features SHAP value, For the complete set of features, represents the model output of subset S, represents the factorial of the size of the subset S, Indicates the change in the prediction result after adding wind speed.

Images