PM2.5 prediction method based on space-time convolution random forest model

Through the spatial and temporal convolution random forest model, combined with spatial and temporal feature extraction, the problem of nonlinear relationship and spatial and temporal feature extraction in PM2.5 concentration prediction in the existing technology is solved, and more accurate PM2.5 concentration prediction is achieved, which improves the Support for environmental monitoring.

CN118941121AInactive Publication Date: 2024-11-12QINGDAO UNIV OF TECH
View PDF 4 Cites 0 Cited by

Patent Information

Application Number
CN202411396772.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-10-09
Publication Date
2024-11-12
Estimated Expiration
Not applicable · inactive patent

AI Technical Summary

Technical Problem

The prior art is difficult to accurately capture the complex nonlinear relationships in PM2.5 concentration, especially in the extraction of spatiotemporal features, and the changes in spatiotemporal heterogeneity between geographical features and influencing factors are rarely considered.

Method used

The spatial and temporal convolution random forest model is adopted to extract the spatial and temporal characteristics of the data through the spatial convolution layer and the temporal convolution layer, and combined with the random forest algorithm, a spatiotemporal convolution random forest model is constructed for PM2.5 concentration prediction.

Benefits of technology

It effectively captures the complex spatio-temporal relationships in PM2.5 concentration, improves prediction and interpretation capabilities, provides more accurate PM2.5 concentration prediction results, and supports environmental monitoring and governance.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118941121A_ABST
    Figure CN118941121A_ABST
Patent Text Reader

Abstract

The invention discloses a PM2.5 prediction method based on a space-time convolution random forest model, relates to the technical field of environmental monitoring, and can effectively capture correlation and modes among different geographic positions through space convolution features, and space convolution can identify the distribution condition of features in a local area when remote sensing data is processed. Dynamic changes in time sequence data are concerned through time convolution features, features related to time are extracted through convolution operation, and the trend and periodicity of the data are better captured. Through combination of the spatial convolution features and the time convolution features, space and time information can be fully utilized in multi-dimensional data analysis, the prediction capability and the interpretation capability of the model are improved, and the prediction capability of PM2.5 concentration is improved through combination of the random forest and the time-space convolution features.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of environmental monitoring technology, and in particular to a PM2.5 prediction method based on a spatiotemporal convolution random forest model. Background Art

[0002] PM2.5 refers to particulate matter in the atmosphere with an aerodynamic diameter of less than or equal to 2.5 microns. Its small size allows it to remain suspended in the air for a long time without settling or dissipating easily, which poses a serious threat to human health and the environment. Therefore, quantitative and accurate estimation of PM2.5 concentrations in different spatial and temporal dimensions of a region provides important information for environmental monitoring, governance assessment, and public health.

[0003] At present, the research on PM2.5 concentration prediction mainly focuses on the following methods: Early studies mainly predicted by establishing a linear regression model or a multivariate linear regression model between AOD (aerosol optical depth) and PM2.5. However, these models cannot be applied to nonlinear relationships. Advanced statistical models are applied to the prediction of spatial PM2.5 concentration, but these models still cannot accurately capture the complex nonlinear relationships. With the rise of machine learning, support vector machines, decision trees, neural networks, and random forests have good prediction effects in large spatial ranges, but they still have certain limitations, especially in capturing spatiotemporal features. Current machine learning methods rarely consider the spatiotemporal heterogeneity between the geographical characteristics of PM2.5 itself and its influencing factors. Summary of the invention

[0004] In order to overcome the above problems existing in the prior art, the present invention proposes a PM2.5 prediction method based on a spatiotemporal convolutional random forest model.

[0005] The technical solution adopted by the present invention to solve the technical problem is: a PM2.5 prediction method based on a spatiotemporal convolution random forest model, which specifically includes the following steps: Step 1: Collect historical meteorological data of the area to be analyzed, historical PM2.5 concentration data monitored by PM2.5 monitoring stations, vegetation data and night light data obtained through satellite remote sensing technology; Step 2, preprocessing the data obtained in step 1, and dividing the preprocessed data into a training set, a test set, and a validation set; Step 3, constructing a spatiotemporal convolution random forest model including a spatiotemporal convolution module and a random forest module; Step 4: Use the data obtained in step 2 to train the spatiotemporal convolutional random forest model obtained in step 3, and use the test set and validation set to test and validate the trained model; Step 5, the pre-processed meteorological data of the area to be analyzed, the vegetation data and night light data obtained by satellite remote sensing technology are subjected to spatiotemporal convolution processing through the spatiotemporal convolution module, and the historical PM2.5 concentration data monitored by the PM2.5 monitoring station, the meteorological data of the area to be analyzed after spatiotemporal convolution processing, and the vegetation data and night light data obtained by satellite remote sensing technology after spatiotemporal convolution processing are input into the random forest module to obtain the prediction results.

[0006] Preferably, the meteorological data in step 1 includes air temperature, air pressure and relative humidity, and the meteorological data in step 1 is processed using a Kriging interpolation method.

[0007] Preferably, the spatiotemporal convolution module of the spatiotemporal convolution random forest model in step 3 includes a spatial convolution layer and a temporal convolution layer. The spatial convolution layer is used to extract features from the spatial dimension of the remote sensing data, select a given pixel point j, and use the measurement values ​​of the pixels around the pixel point j to reflect the spatial dependency; the time lag feature is introduced in the temporal convolution layer to capture the correlation between the data at different time points.

[0008] Preferably, the specific processing process of the spatial convolution layer includes: the spatial feature variable is represented by the inverse distance weighted measurement value of the position point around the given pixel point j, that is, the closer the pixel point is to the pixel point j, the greater the influence, and the weight decreases with the increase of the distance. A filter is slid on the image, and the global spatial feature is extracted by calculating the weighted sum of all pixels in the neighboring area. The specific calculation formula is: ; ; Among them, z j is the data obtained after convolution of a given pixel point j, z i is the data of each point in the convolution range near pixel j, w d is the convolution weight matrix, d ij is the distance between grid cell j and given point i.

[0009] Preferably, the specific processing process of the temporal convolution layer includes: by using the current, past and future samples, calculating the temporal correlation between the current data and the previous and subsequent data, the current and past samples are used to calculate its output at time t, and the output signal y[n] is composed of a one-dimensional input vector z t With the convolution kernel w t After convolution, we get the input vector z t The convolution calculation formula at the nth element is: ; Among them, n represents the output index, k represents the sum index, and wt Represents the convolution kernel or weight sequence, input z t represents the input signal to be processed; K represents the length of the entire sequence; since the input signal is a discrete sequence, there are differences in the degree of discreteness in adjacent data, and its weight coefficient is related to the number of days between intervals. Define d kn is the number of days between convolutions, and the convolution kernel can be expressed as: .

[0010] The beneficial effect of the present invention is that the correlation and pattern between different geographical locations can be effectively captured through spatial convolution features. When processing remote sensing data (AOD, meteorological information, etc.), spatial convolution can identify the distribution of features in a local area.

[0011] Through time convolution features, time convolution features focus on the dynamic changes in time series data, extract time-related features through convolution operations, and better capture the trends and periodicity of data.

[0012] The combination of spatial convolution features and temporal convolution features can make full use of spatial and temporal information in multidimensional data analysis and improve the model's predictive and explanatory capabilities. This dual feature extraction method enables the model to more fully understand the complex relationships in the data.

[0013] Random forest combined with spatiotemporal convolution features improves the prediction ability of PM2.5 concentration.

[0014] In summary, this patent proposes an innovative method based on spatiotemporal convolution features and random forest models for predicting PM2.5 concentration. By combining spatial convolution features and temporal convolution features, this method can fully capture the complex spatiotemporal relationships in environmental data. These beneficial effects will provide strong support for environmental monitoring and other applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 It is a flow chart of the present invention; Figure 2 This is a schematic diagram of the spatiotemporal convolution principle of the present invention; Figure 3 The relationship between the accuracy of the SCRF model in predicting PM2.5 mass concentration in the training set and the number of decision trees and the test set and the number of decision trees, where (a) is the R of the training set 2 As the decision tree changes, (b) is the RMSE of the training set, and (c) is the R of the test set. 2 As the decision tree changes, (d) is the RMSE of the test set as the decision tree changes; Figure 4 This is a feature importance ranking diagram based on the model of the present invention; Figure 5 This is a heat map of the correlation analysis between PM2.5 concentration and other characteristic variables of the present invention; Figure 6 It is a Pearson correlation coefficient bar graph of the present invention; Figure 7 The figures are the PM2.5 concentration fitting diagrams predicted by the SCRF model in the study area of ​​the embodiment of the present invention from 2016 to 2019, wherein (a) is the PM2.5 concentration fitting diagram in 2016, (b) is the PM2.5 concentration fitting diagram in 2017, (c) is the PM2.5 concentration fitting diagram in 2018, and (d) is the PM2.5 concentration fitting diagram in 2019; Figure 8 The SCRF model is used to predict seasonal PM2.5 concentration fitting graphs in the study area of ​​the embodiment of the present invention, wherein (a) is the PM2.5 concentration fitting graph in spring, (b) is the PM2.5 concentration fitting graph in summer, (c) is the PM2.5 concentration fitting graph in autumn, and (d) is the PM2.5 concentration fitting graph in winter. DETAILED DESCRIPTION

[0016] In order to enable those skilled in the art to better understand the technical solution of the present invention, the present invention is described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0017] This embodiment discloses a PM2.5 prediction method based on a spatiotemporal convolutional random forest model, which specifically includes the following steps: Step 1: Collect historical meteorological data of the area to be analyzed, historical PM2.5 concentration data monitored by PM2.5 monitoring stations, vegetation data and night light data obtained through satellite remote sensing technology; Step 2, preprocessing the data obtained in step 1, and dividing the preprocessed data into a training set, a test set, and a validation set; Step 3: construct a spatiotemporal convolutional random forest model (named SCRF model in this embodiment) including a spatiotemporal convolution module and a random forest module.

[0018] The spatiotemporal convolution module of the spatiotemporal convolutional random forest model includes a spatial convolution layer and a temporal convolution layer. The spatial convolution layer is used to extract features from the spatial dimension of remote sensing data, select a given pixel point j, and use the measurement values ​​of the pixels around pixel point j to reflect the spatial dependency. The time lag feature is introduced in the temporal convolution layer to capture the correlation between data at different time points.

[0019] Spatial convolution features are mainly used to extract features from the spatial dimension of remote sensing data. In the specific operation, we select a given pixel point j and use the measured values ​​of its surrounding pixels to reflect the spatial dependency. The spatial feature variable can be represented by the inverse distance weighted (IDW) of the measured values ​​of the points around the point, that is, the closer the pixel is to point j, the greater the influence, and the weight decreases with the increase of distance. A filter is slid on the image, and the global spatial features are extracted by calculating the weighted sum of all pixels in the neighboring area. The convolution process not only considers the direct measurement value of point j, but also combines it with the values ​​of the surrounding pixels to extract richer spatial dependency features. These spatial features can effectively describe the spatial distribution law of PM2.5 concentration and enhance the model's ability to capture spatial heterogeneity changes. The specific expression is: ; In the above expression, z j is the data obtained after convolution of a given point or grid unit j, z i is the data of each point in the convolution range near the given point, w d is the weight matrix of the convolution; ; where d ij is the distance between grid cell j and given point i.

[0020] When the time series data is established by the time factor of the time convolution feature, the data can be divided into different time periods. When building the model, the data can be divided into different time periods, such as by month, quarter or year, so that the seasonal and cyclical changes of the data can be effectively captured and analyzed. This segmentation method enables the model to identify the regular fluctuations of the data within a specific time period, thereby better reflecting the changing trends of environmental variables (such as PM2.5 concentration).

[0021] In order to enhance the model's understanding of time series, time lag features are introduced. The use of time lag features allows the model to capture the correlation between data at different time points. For example, the PM2.5 concentration at the current time point may be affected by the concentration at several past time points, especially when considering changes in environmental factors, past data can often provide important information. By incorporating these lag features into the model, random forests can be more accurate in making predictions, fully utilizing the value of historical data, and thus improving the ability to predict future changes.

[0022] Specifically, by using current, past, and future samples, the model can calculate the temporal correlation between the current data and the previous and future data. This approach can reveal the dynamic characteristics of time series and help identify trends and patterns. The current and past samples are used to calculate its output at time t. The output signal y[n] is composed of a one-dimensional input vector zt With the convolution kernel w t After convolution, we get the input vector z t The convolution calculation formula at the nth element is: ; Among them, n represents the output index, k represents the sum index, and w t represents the convolution kernel or weight sequence, and the input zt represents the input signal to be processed. Because the input signal is a discrete sequence, there are differences in the degree of discreteness in adjacent data, and its weight coefficient is related to the number of days between intervals. Define d kn is the number of days between convolutions, and the convolution kernel can be expressed as: .

[0023] The random forest module is an ensemble learning machine learning algorithm consisting of multiple decision trees. The core idea is to improve the overall accuracy and robustness by combining the prediction results of multiple models, thereby overcoming the overfitting problem that may exist in a single model. Each decision tree uses the bootstrap sampling method during the training process to randomly extract samples from the original data set so that each tree can be trained on different samples. At the same time, at each node split, a feature subset is randomly selected for optimal feature selection to ensure the diversity of each tree. This method not only improves the generalization ability of the model, but also enhances the stability of the model when processing complex data.

[0024] The SCRF model constructs a spatiotemporal convolutional random forest (SCRF) model by introducing weight matrices in spatial and temporal dimensions. The principle diagram of spatiotemporal convolution is shown in the figure. Figure 2 , where h i Indicates the current time node. Indicates the next time node, Represents the last time node, and captures the characteristic relationship of data in time through convolution. This embodiment mainly uses AOD, meteorological variables, land-related parameters, population-related indexes and spatiotemporal information obtained by satellite as input to explore the relationship between various influencing factors and PM2.5 after the model is improved. The model effectively captures and analyzes the potential spatiotemporal characteristics in the data to improve the prediction ability of complex environmental variables (such as PM2.5 concentration).

[0025] Step 4: Use the data obtained in step 2 to train the spatiotemporal convolutional random forest model obtained in step 3, and use the test set and validation set to test and validate the trained model; Step 5, the pre-processed meteorological data of the area to be analyzed, the vegetation data and night light data obtained by satellite remote sensing technology are subjected to spatiotemporal convolution processing through the spatiotemporal convolution module, and the historical PM2.5 concentration data monitored by the PM2.5 monitoring station, the meteorological data of the area to be analyzed after spatiotemporal convolution processing, and the vegetation data and night light data obtained by satellite remote sensing technology after spatiotemporal convolution processing are input into the random forest module to obtain the prediction results.

[0026] This embodiment performs PM2.5 prediction in a certain area. The flow chart is as follows Figure 1 As shown, the specific process is: 1. Data acquisition, preprocessing and normalization (1) Ground PM2.5 concentration data: PM2.5 concentration data comes from the China National Environmental Monitoring Center (http: / / www.cnemc.cn), which collected data from PM2.5 stations in Shandong Province from 2016 to 2019. First, the PM2.5 concentration data from 10:00 to 12:00 every day at each PM2.5 station were arithmetic and the average was calculated, which was used as the daily PM2.5 concentration value of the corresponding station, and finally the daily PM2.5 concentration data from 2016 to 2019 were obtained.

[0027] (2) Meteorological data: The meteorological data used in this study include temperature (TEM), air pressure (PRS), and relative humidity (RHU). The meteorological data are from the China Meteorological Science Data Center (http: / / data.cma.cn / ). Kriging is used to interpolate TEM, PRS, and RHU meteorological data, and finally obtains continuous coverage of meteorological data in the study area.

[0028] (3) Other data: Vegetation data (VEG) and night light data (NPP) are obtained through satellite remote sensing technology, mainly through the monitoring and measurement of surface vegetation cover by remote sensing satellites. These data provide an important basis for analyzing and evaluating the status of surface vegetation. Annual night light data are obtained by remote sensing satellites NPP~VIRS to observe and record the brightness of surface lights at night. These data provide us with information about urbanization levels and population distribution (i.e., population activity index). This information is provided by the Resources and Environmental Science and Data Center of the Chinese Academy of Sciences (https: / / www.resdc.cn / ).

[0029] (4) After obtaining the data from (1) to (3), the data is preprocessed, including conventional processing such as projection, cropping, and resampling. During the model training process, since the data ranges of various variables vary greatly, in order to ensure the stability of gradient changes during network training, the feature data must be normalized. The normalization calculation formula is as follows: ; In the formula: Y represents the normalized value; X is the original variable value; X min is the minimum value; X max is the maximum value. After normalization, the dataset is divided into a training set (80% of the total dataset), a test set (10% of the total dataset), and a validation set (10% of the total dataset) according to the needs of this study.

[0030] 2. Evaluation Metrics In this embodiment, the following three metrics are used to evaluate the performance of the model. First, R 2 is a statistic that measures the goodness of fit of a regression model, and its value ranges from 0 to 1. The closer the value is to 1, the better the fitting effect of the model. Second, the mean absolute error (MAE) and the root mean square error (RMSE) are used to evaluate the gap between the predicted values and the actual values of the model. The lower the values of these metrics, the better the prediction effect of the model. The following are the calculation formulas for each evaluation metric: ; Among them, n is the number of samples, is the i-th actual observation value, is the corresponding model predicted value, is the mean of the actual observation values.

[0031] 3. Experimental Parameter Settings In this embodiment, the model parameters are set as shown in Table 1, and the ten-fold cross-validation algorithm (10~CV) will be used to evaluate the performance of the model, as Figure 3 shown. In the training stage of the random forest model, 1000 decision trees are constructed. The selection of this parameter aims to ensure a good balance between the stability and accuracy of the model. The maximum depth is set to 25 to limit the growth depth of each tree and prevent overfitting. The minimum number of leaf nodes is set to 1 to ensure that each leaf node can contain at least 1 sample. And the bootstrap sampling method is used to construct each tree, thereby improving the generalization ability of the model. Through the above settings, the aim is to construct a PM2.5 prediction model with stable performance and high accuracy to improve the accuracy and reliability of air quality monitoring and prediction. At the same time, the model shows the importance ranking of the input variables, as Figure 4 shown, where Figure 4 the "I" symbol on the right in is the error bar symbol, which is obtained by calculating the standard error of the feature importance to represent the uncertainty of the feature importance. Among the 6 input features of the SCRF model, the importance ranking from high to low is AOD, TEM, PRS, RHU, NPP, VEG.

[0032] Table 1 Model Parameter Settings

[0033] 4. Analysis of PM2.5 and characteristic variables Pearson correlation analysis is a widely used statistical method, which is mainly used to measure the degree of linear correlation between two continuous variables. When studying the correlation between PM2.5 concentration and multiple characteristic variables, the Pearson correlation coefficient between PM2.5 and each characteristic variable is calculated to reveal the linear relationship between them. If the value of the correlation coefficient is close to 1 or -1, it means that there is a strong linear correlation between the two variables; conversely, if the correlation coefficient is close to 0, it means that the linear relationship between them is weak or almost non-existent.

[0034] The change of PM2.5 concentration is not only affected by human activities, but also by meteorological factors and population activities. In order to gain a deeper understanding of the correlation between PM2.5 pollutants and characteristic variables in Shandong Province, the Pearson correlation analysis was used to generate the correlation between PM2.5 and characteristic variables. Figure 5 As shown, Figure 5 ** indicates significant correlation at the 0.01 level; * indicates significant correlation at the 0.05 level; no mark indicates insignificant correlation. Figure 6 The Pearson correlation coefficient bar chart helps identify which characteristic variables have the most significant impact on PM2.5 concentration. Through this analysis method, we can have a more comprehensive understanding of the various factors that affect PM2.5 concentration, and thus provide a basis for subsequent research and governance measures.

[0035] 5. SCRF model performance In order to verify the performance of the SCRF model in predicting PM2.5 concentration in a certain area (the study area in this example), a detailed analysis was conducted. Figure 7 The scatter plot of the annual average PM2.5 site prediction of the SCRF model from 2016 to 2019 is shown. From the prediction results, the SCRF model performed quite well in these four years, with R² values ​​of 0.81, 0.70, 0.81 and 0.86, respectively, indicating that the model has high accuracy in explaining variability. In addition, the root mean square error (RMSE) values ​​of the model were 14.26, 12.33, 8.26 and 7.57 μg / m³, respectively, showing that the gap between the prediction results and the actual observations is getting smaller year by year, which means that the prediction accuracy of the SCRF model is constantly improving.

[0036] Compared with the traditional random forest model, the SCRF model has significantly better prediction results. This advantage mainly stems from the comprehensive consideration of multiple key factors by the SCRF model during the modeling process. The SCRF model not only focuses on the spatial heterogeneity of pollutant emissions, but also incorporates the temporal changes of meteorological variables. Specifically, the model analyzes the pollution sources and meteorological conditions in different regions, such as temperature and humidity, to identify their impact on PM2.5 concentrations, thereby more comprehensively capturing the impact of environmental changes on air quality. In addition, the SCRF model uses spatiotemporal convolution to effectively integrate spatiotemporal features, so that the model can better reflect the relationship between the two when processing complex meteorological and pollutant data. This innovative approach not only improves the prediction accuracy of the model, but also enhances its operability in practical applications, and can provide policymakers with more reliable air quality prediction data.

[0037] In summary, the excellent performance of the SCRF model in predicting PM2.5 concentration in the study area fully demonstrates its advantages in comprehensively considering environmental factors and provides an important reference for future air quality research and governance.

[0038] Figure 8 The scatter plots of spring, summer, autumn and winter in the study area from 2016 to 2019 based on the spatiotemporal convolutional random forest (SCRF) model are shown. Analysis of these charts shows that due to the influence of clouds in remote sensing images, especially in summer, the number of available images is relatively small, resulting in differences in the number of samples in each season. Specifically, the coefficient of determination (R²) of the SCRF model in spring, summer, autumn and winter is 0.76, 0.72, 0.82 and 0.81, respectively, indicating that the model has better prediction effect in spring and autumn, while the prediction ability in summer is relatively weak. From the perspective of root mean square error (RMSE), the RMSE values ​​of the four seasons are 11.02 μg / m³, 6.06 μg / m³, 9.48 μg / m³ and 13.70 μg / m³, respectively. The summer model with higher RMSE value has lower R² and RMSE values, which may be related to insufficient sample size and rapid changes in meteorological conditions in summer. Summer rainfall is frequent and humidity changes rapidly, which may make it difficult for the model to accurately capture and describe humidity fluctuations in a short period of time. Especially when humidity accounts for a large weight, the model fails to fully consider these rapidly changing meteorological conditions, thus affecting the prediction effect.

[0039] In contrast, PM2.5 concentrations are generally higher in spring, autumn, and winter, and there are more PM2.5 particles in the atmosphere. The concentration trend during this period may be affected by a variety of complex factors, such as meteorological factors and changes in air pollution sources. In these seasons, although the model's prediction effect is relatively good, it still faces many challenges. The fluctuation of PM2.5 concentration is not only affected by meteorological conditions, but also closely related to local industrial activities, traffic emissions, natural sources (such as sandstorms) and other factors. Therefore, even if the model can capture the trend of seasonal changes as a whole, it may still be interfered by multiple factors in specific predictions. It is worth noting that the humidity change in summer has a significant impact on PM2.5, especially after a rainfall event, the rapid change in humidity may cause instantaneous fluctuations in PM2.5 concentration, and these changes are often not easily captured by the model. In addition, the meteorological conditions in summer are more complex, with many influencing factors, and the difficulty of model establishment and training increases accordingly. Therefore, in order to improve the prediction accuracy in summer, future research can consider introducing more meteorological factors and more complex modeling methods, and combining multi-source data for model optimization.

[0040] In general, the performance differences of the SCRF model in different seasons reflect the complexity of remote sensing data processing and environmental factor analysis. By continuously optimizing the model and combining more comprehensive data, the changes in PM2.5 concentration can be more accurately evaluated and predicted, providing a scientific basis for improving air quality.

[0041] 6. Model efficiency study This example systematically compares the performance of the SCRF (Spatiotemporal Convolutional Random Forest) model with several popular machine learning models. The analysis covers the period from 2016 to 2019, aiming to evaluate the differences in prediction capabilities of the models. The compared models include back propagation neural network (BPNN), linear regression model (LME), classification and regression tree (CART), gradient boosting decision tree (GBDT), and random forest (RF). In the overall performance evaluation, the prediction results of each model are shown in Table 2, showing the differences in the capabilities of different models when handling similar tasks.

[0042] Table 2 Performance indicators of each model

[0043] Specifically, the prediction coefficient of determination (R²) values ​​of each model ranged from 0.35 to 0.79, indicating that they were able to capture the changing trends in the data to a certain extent. However, compared with these traditional models, the SCRF model significantly outperformed the other models, with an R² value of up to 0.83, showing its strong ability to capture complex data features. This result not only shows the advantages of the SCRF model, but also reflects its effectiveness in dealing with nonlinear relationships and multivariate interactions. In terms of prediction error, the root mean square error (RMSE) values ​​of each model ranged from 7.87μg / m³ to 30.16μg / m³. The SCRF model performed best with the lowest RMSE value of 7.87μg / m³, showing its excellent performance in improving prediction accuracy. It was followed by the RF model with an RMSE of 8.11μg / m³, which also reflects the advantage of random forest in feature processing. In contrast, the BPNN model had the highest RMSE value of 30.16μg / m³, showing the shortcomings of this model in this task.

[0044] Through the analysis of Table 2, it can be clearly seen that the SCRF model has significant advantages, especially in the context of being able to consider spatial heterogeneity and temporal correlation at the same time. The design concept of SCRF is to integrate spatiotemporal features, which can better adapt to the changes in actual data and thus improve the prediction effect. This feature makes SCRF more flexible and adaptable in practical applications, and can effectively cope with complex environmental changes. In addition to the SCRF and RF models, the performance of other models is GBDT, CART, LME and BPNN. This ranking not only reflects the differences in data fitting and prediction capabilities of each model, but also points out the outstanding performance of SCRF in reducing prediction errors. Especially in practical applications, the improvement of prediction accuracy is of great significance to decision support, policy making and environmental monitoring. Therefore, the introduction and application of the SCRF model provides a new perspective and direction for future research and practice in related fields.

[0045] In summary, the SCRF model shows more significant superiority in this comparison, especially in reducing prediction errors, showing its potential in the fields of machine learning and data science. Future research can further explore the performance of SCRF in other application scenarios and provide reference for a wider range of practical applications.

[0046] This embodiment solves many problems existing in the prior art by constructing the SCRF model, including: (1) Complex spatiotemporal feature extraction: PM2.5 concentration is affected by many factors, including AOD, meteorological data, etc. The traditional random forest model is difficult to effectively capture these complex spatiotemporal relationships.

[0047] Solution: By introducing spatiotemporal convolution features, features are extracted from the spatial dimension of remote sensing data, and the temporal correlation of data is calculated by using current, past, and future samples, so as to better capture the spatiotemporal changes of PM2.5 concentrations.

[0048] (2) Feature extraction of multiple environmental factors: The change of PM2.5 concentration is affected by multiple environmental factors at the same time, such as meteorological conditions, geographical factors and distribution of pollution sources. These factors work together in space and time to affect the spatiotemporal distribution of PM2.5. Traditional models are difficult to handle these complex environmental factors at the same time, which may lead to inaccurate predictions.

[0049] Solution: Extract features from multiple environmental factors (such as AOD and meteorological data) to capture the impact of different environmental factors on PM2.5 in spatial and temporal dimensions. This feature extraction method under multiple environmental factors can effectively integrate various types of information, enabling the model to fully consider the comprehensive effects of different environmental factors on PM2.5 concentration and improve the accuracy and robustness of prediction.

[0050] (3) Model robustness and generalization ability: Under different environments and conditions, the model’s robustness and generalization ability may be insufficient, resulting in unstable prediction results.

[0051] Solution: By integrating spatiotemporal convolutional features with the random forest model, the robustness and generalization ability of the model under different conditions are enhanced, ensuring the prediction stability in diverse environments.

[0052] (4) Data missing and noise processing: In actual monitoring, PM2.5, AOD and meteorological data may be missing and noisy, which will affect the accuracy of the prediction.

[0053] Solution: Solve the problem of missing data through data interpolation technology, smooth the noise data, and adopt outlier detection and elimination mechanism to ensure data quality and prediction accuracy.

[0054] The above embodiments are only exemplary embodiments of the present invention and are not intended to limit the present invention. Those skilled in the art may make various modifications or equivalent substitutions to the present invention within the essence and protection scope of the present invention, and such modifications or equivalent substitutions shall also be deemed to fall within the protection scope of the present invention.

Claims

1. PM2.5 prediction method based on spatiotemporal convolution random forest model, characterized in that: The specific steps include: Step 1: Collect historical meteorological data of the area to be analyzed, historical PM2.5 concentration data monitored by PM2.5 monitoring stations, vegetation data and night light data obtained through satellite remote sensing technology; Step 2, preprocessing the data obtained in step 1, and dividing the preprocessed data into a training set, a test set, and a validation set; Step 3, constructing a spatiotemporal convolution random forest model including a spatiotemporal convolution module and a random forest module; Step 4: Use the data obtained in step 2 to train the spatiotemporal convolutional random forest model obtained in step 3, and use the test set and validation set to test and validate the trained model; Step 5, the pre-processed meteorological data of the area to be analyzed, the vegetation data and night light data obtained by satellite remote sensing technology are subjected to spatiotemporal convolution processing through the spatiotemporal convolution module, and the historical PM2.5 concentration data monitored by the PM2.5 monitoring station, the meteorological data of the area to be analyzed after spatiotemporal convolution processing, and the vegetation data and night light data obtained by satellite remote sensing technology after spatiotemporal convolution processing are input into the random forest module to obtain the prediction results.

2. The PM2.5 prediction method based on the spatiotemporal convolution random forest model according to claim 1 is characterized in that: The meteorological data in step 1 include air temperature, air pressure and relative humidity, and the meteorological data in step 1 is processed using the Kriging interpolation method.

3. The PM2.5 prediction method based on the spatiotemporal convolution random forest model according to claim 1, characterized in that: The spatiotemporal convolution module of the spatiotemporal convolution random forest model in step 3 includes a spatial convolution layer and a temporal convolution layer. The spatial convolution layer is used to extract features from the spatial dimension of the remote sensing data, select a given pixel point j, and use the measurement values ​​of the pixels around the pixel point j to reflect the spatial dependency; the temporal convolution layer introduces a time lag feature to capture the correlation between data at different time points.

4. The PM2.5 prediction method based on the spatiotemporal convolution random forest model according to claim 3 is characterized in that: The specific processing process of the spatial convolution layer includes: the spatial feature variable is represented by the inverse distance weighted measurement value of the position points around the given pixel point j, that is, the closer the pixel point is to the pixel point j, the greater the influence, and the weight decreases with the increase of the distance. A filter is slid on the image, and the global spatial feature is extracted by calculating the weighted sum of all pixels in the neighboring area. The specific calculation formula is: ; ; Among them, z j is the data obtained after convolution of a given pixel point j, z i is the data of each point in the convolution range near pixel j, w d is the convolution weight matrix, d ij is the distance between grid cell j and given point i.

5. The PM2.5 prediction method based on the spatiotemporal convolution random forest model according to claim 3 is characterized in that: The specific processing process of the temporal convolution layer includes: by using the current, past and future samples, the temporal correlation between the current data and the previous and subsequent data is calculated, and the current and past samples are used to calculate its output at time t. The output signal y[n] is composed of a one-dimensional input vector z t With the convolution kernel w t After convolution, we get the input vector z t The convolution calculation formula at the nth element is: ; Among them, n represents the output index, k represents the sum index, and w t Represents the convolution kernel or weight sequence, input z t Represents the input signal to be processed; since the input signal is a discrete sequence, there are differences in the degree of discreteness in adjacent data, and its weight coefficient is related to the number of days between them. Define d kn is the number of days between convolutions, K represents the length of the entire sequence, and the convolution kernel can be expressed as: 。

Citation Information

Patent Citations

  • PM2.5 concentration prediction method and device considering spatio-temporal correlation and meteorological factors

    CN113610286A

  • Air quality prediction method based on space-time diagram convolutional network

    CN114240000A

  • Aerosol optical thickness space-time evolution and potential source analysis method

    CN118444408A

  • Optical system and camera module

    KR1020240023969A