A machine learning-based spatiotemporal quantification method for multi-source anthropogenic heat

Abstract

Anthropogenic heat has a significant impact on urban climate and air quality, but there is currently a lack of accurate and efficient estimation methods for multi-source anthropogenic heat. The present invention improves the process of anthropogenic heat modeling and proposes a multi-source anthropogenic heat spatiotemporal quantification method based on machine learning. The method comprises the following steps: step 1) calculating the county-level annual average anthropogenic heat flux (AHF) based on energy consumption and socioeconomic data; step 2) using alternative data to perform time dimension downscaling for anthropogenic heat from different sources to obtain the county-level monthly average AHF; step 3) calculating the monthly county-level average of anthropogenic heat-related multi-source data as explanatory variables, and forming training samples with the corresponding AHF; step 4) training models based on two machine learning algorithms, gradient boosting regression tree and Cubist, and performing error analysis to select the optimal algorithm modeling for different heat sources; step 5) inputting specific raster data into the optimal model to calculate the multi-source anthropogenic heat flux in a specific area at a specific time.

Description

Technical Field

[0001] The present invention relates to a multi-source anthropogenic heat spatiotemporal quantification method based on machine learning. Based on a machine learning algorithm training model, the distribution of multiple types of anthropogenic heat fluxes in a specific area and a specific period of time is calculated, providing a new idea and method for more accurate and efficient anthropogenic heat modeling. Background Art

[0002] Anthropogenic heat emissions have a significant impact on urban climate and air quality, and are also an important data input for climate modeling. Accurate anthropogenic heat data can be used as surface boundary conditions for regional or global scale climate simulations, and can reasonably assess the impact of human activities on the urban environment. It is an important basis for solving problems such as climate warming, urban heat island effect, and air pollution. Anthropogenic heat flux (AHF) is the anthropogenic heat per unit time and unit area, and is the main target of anthropogenic heat estimation. Traditional AHF estimation methods include energy balance equation method, building energy efficiency modeling, and energy consumption inventory method. Among them, the energy consumption inventory method is a relatively common and reliable method for estimating AHF. Furthermore, in order to simplify the complexity of AHF calculations and reduce repetitive work, some variables that have a strong correlation with heat emissions, such as night lights and air pollutants, are used in combination with machine learning algorithms to establish a more practical empirical model.

[0003] Although a large number of methods proposed so far can meet general research and application needs, in the face of the complex spatiotemporal changes of multi-source anthropogenic heat, simple models such as simple linear regression and single variables can no longer cope well with the challenge. In addition, in previous studies, AHF samples are usually estimated by province or city (Chen et al., 2012; Chen et al., 2020; Sailor et al., 2015), but the administrative area of ​​a province or city is generally large, and the proportion of built-up area is very low, which will result in high anthropogenic heat emissions and low AHF, which will bring certain troubles to the analysis of the estimation results. For the estimation model, although the explanatory variables are also processed into the same spatial scale, the mean processing of the large spatial scale will reduce the already scarce medium and high value AHF samples; at the same time, the estimation model based on the annual average AHF often makes the samples similar and loses a lot of temporal information. In short, current research does not fully consider the diversity and variability of training samples. On the other hand, AHF modeling based on machine learning is often affected by different algorithms. Different anthropogenic heat sources may have different optimal modeling algorithms due to their differences in spatiotemporal characteristics. This influencing factor has not yet been considered and practiced. In view of the problems existing in the existing anthropogenic heat quantification methods, the present invention constructs samples with more spatiotemporal diversity, and simultaneously uses two more complex machine learning algorithms to construct the optimal multi-source AHF estimation model.

[0004] Gradient boosting decision tree (GBDT) is a very classic ensemble learning algorithm. It generates strong learners by integrating weak learners and has been widely used in data analysis and prediction in many fields. The Cubist algorithm is developed from the model tree (Quinlan, 1992; Quinlan, 1993; Quinlan, 1996) and is a rule-based algorithm. The leaf node in the algorithm is a multivariate linear regression model rather than a single value. Similar to GBDT, Cubist can also perform ensemble learning and has good performance in the fields of traffic flow prediction, air surface temperature estimation, surface cover estimation, leaf area index estimation and anthropogenic heat estimation. Summary of the invention

[0005] In view of the lack of technology for accurate and efficient estimation of multi-source anthropogenic heat, the present invention proposes a multi-source anthropogenic heat spatiotemporal quantification method based on machine learning, which is mainly implemented based on the following steps:

[0006] Step 1) Based on energy consumption data and socioeconomic data, the annual average AHF at the county level was calculated using a top-down energy consumption inventory method;

[0007] Step 2) Use the alternative data to downscale the time dimension for anthropogenic heat from different sources to obtain the county-level monthly average AHF;

[0008] Step 3) preprocess the multi-source datasets related to anthropogenic heat, calculate the monthly county-level average values ​​as explanatory variables, and form training samples with the corresponding AHFs;

[0009] Step 4) Based on the training models of two machine learning algorithms, gradient boosting regression tree and Cubist, error analysis is performed to select the optimal algorithm modeling for different heat sources. At the same time, a simple linear regression model based on night lights is used as a reference for accuracy improvement;

[0010] Step 5) Input the specific grid data into the optimal model to calculate the anthropogenic heat flux in a specific area at a specific time and output the grid results. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] Figure 1 It is a technical flow chart;

[0012] Figure 2 is the error result of each model;

[0013] Figure 3 Output the spatiotemporal distribution map of multi-source AHF for the model; DETAILED DESCRIPTION

[0014] The present invention "a multi-source anthropogenic thermal spatiotemporal quantification method based on machine learning" is further explained below in conjunction with the accompanying drawings.

[0015] 1. Estimation of Monthly AHF at County Level

[0016] Anthropogenic heat includes four sources: industry, construction, transportation, and human metabolism. Based on the energy inventory method, the annual average AHF is estimated at the provincial, municipal, and county levels in turn. When the municipal AHF is downscaled to the county level, industrial heat is calculated based on the proportion of the number of industrial POIs in the district or county to the total city, traffic heat and building heat are calculated based on the proportion of the district or county population, and metabolic heat is directly estimated by the district or county population. The monthly AHF is calculated based on the temporal changes of alternative data, which is a commonly used temporal downscaling rule in the top-down energy inventory method. In the absence of heating in the study area, monthly building heat and industrial heat are estimated based on monthly electricity consumption, monthly traffic heat is estimated based on monthly freight volume, and metabolic heat is fixed and participates in model training as a whole with building heat. The specific calculation is as follows:

[0017]

[0018] in, represents the AHF of the mth month for heat sources of type S, i represents a prefecture-level city, and j represents a district or county; is the emission proportion (%) of the corresponding heat source in month m; are the annual heat emissions of industry, transportation and buildings (J). j is the area of ​​the district or county (m 2 );T y The time period is one year (s). This process can be implemented using Python and R programming, or directly calculated using Excel.

[0019] (II) Construction of training samples

[0020] Density grids were established for roads, railways, and industrial POIs with a search radius of 1000 meters, and distance grids were calculated at the same time, which can be calculated using the point density tool and the Euclidean distance tool in ArcGIS, respectively. Here, roads and railways are special variables for traffic heat, industrial POIs and railways are special variables for industrial heat, and the proportion of building area is used as a special variable for building metabolic heat.

[0021] The common variables are involved in the estimation of the three heat sources at the same time, and the processing is as follows: remote sensing data (surface temperature, NDVI, etc.), meteorological data (temperature, humidity, wind speed) and terrain data (DEM, slope) are all screened, synthesized by month, clipped, reprojected and resampled through Google Earth Engine (GEE); finally, the above rasters are partitioned and counted in ArcGIS, and the explanatory variable table is output, and two categorical variables, region and month, are added to form the training samples together with the monthly county-level AHF.

[0022] (III) Model training and evaluation

[0023] Import the training samples into R, randomly select 80% of the samples for training and testing, and the remaining samples are used for model verification. GBDT and Cubist both contain many hyperparameters that can be adjusted. Call caret, gbm and Cubist packages in R, train GBDT and Cubist models according to different AHF sources, use 10-fold cross validation repeated 10 times for parameter adjustment, select the optimal model parameters based on the principle of minimizing training / testing errors, and construct a multi-source anthropogenic heat spatiotemporal quantification model (GBDT and Cubist). Finally, determine the algorithm (GBDT or Cubist) used for modeling each heat source based on the verification error; the simple linear regression model (SLR) based on night lights is used as a reference to test the accuracy improvement of these two complex algorithms. The above process can be completed in R.

[0024] (III) Output of results

[0025] The raster package was called in R to convert the explanatory variables in raster form into data frame format, and a simple classification regression tree model was built to fill in the missing variables. The prediction results of the trained model were called, and the raster input prediction results were built, and finally the spatiotemporal distribution of multi-source anthropogenic heat in raster form was output.

Claims

1. A multi-source anthropogenic heat spatiotemporal quantification method based on machine learning is implemented by the following steps: Step 1) Based on energy consumption data and socioeconomic data, a top-down energy consumption inventory method is used to calculate the county-level annual average anthropogenic heat flux AHF; Step 2) Use the alternative data to downscale the time dimension for anthropogenic heat from different sources to obtain the county-level monthly average AHF; Step 3) preprocess the multi-source datasets related to anthropogenic heat, calculate the monthly county-level average values ​​as explanatory variables, and form training samples with the corresponding AHFs; Step 4) Based on the training models of two machine learning algorithms, gradient boosting regression tree and Cubist, error analysis is performed to select the optimal algorithm modeling for different heat sources. At the same time, a simple linear regression model based on night lights is used as a reference for accuracy improvement; Step 5) inputting specific grid data into the optimal model to calculate the anthropogenic heat flux in a specific area at a specific time, and outputting the grid results; In step 2), the monthly AHF is calculated based on the temporal variation of the alternative data, which is the temporal downscaling rule in the top-down energy inventory method. In the absence of heating in the study area, the monthly building heat and industrial heat are estimated based on the monthly electricity consumption, the monthly traffic heat is estimated based on the monthly freight volume, and the metabolic heat is fixed and participates in the model training as a whole with the building heat. The specific calculation is as follows: in, represents the AHF of the mth month for heat sources of type S, i represents a prefecture-level city, and j represents a district or county; is the percentage of the corresponding heat source’s emissions in month m; are the annual heat emissions of industrial transportation and buildings, J and A respectively. j The area of ​​the district / county is m 2 ; T y The time period is one year; In the step 3), a density grid is established with a search radius of 1000 meters for roads, railways and industrial POIs, and a distance grid is calculated at the same time, using the point density tool and the Euclidean distance tool in ArcGIS for calculation respectively; here, roads and railways are special variables for traffic heat, industrial POIs and railways are special variables for industrial heat, and the proportion of building area is used as a special variable for building metabolic heat; public variables are simultaneously involved in the estimation of the three heat sources, and the processing is as follows: remote sensing data including surface temperature and NDVI, meteorological data including temperature, humidity, wind speed and terrain data including DEM and slope are all screened, synthesized by month, clipped, reprojected and resampled through Google Earth Engine; finally, the above grids are subjected to zoning statistics in ArcGIS, and an explanatory variable table is output, and two classification variables, namely, the region and month, are added to form a training sample together with the county-level AHF of each month; In the step 4), the training samples are imported into R, 80% of the samples are randomly selected for training and testing, and the remaining samples are used for model verification; caret, gbm and Cubist packages are called in R, GBDT and Cubist models are trained respectively according to different AHF sources, 10-fold cross validation is repeated 10 times for parameter adjustment, and the optimal model parameters are selected according to the principle of minimizing training / testing errors, and multi-source anthropogenic heat spatiotemporal quantification models GBDT and Cubist are constructed. Finally, the algorithm used for modeling each heat source, i.e. GBDT or Cubist, is determined according to the verification error; a simple linear regression model based on night lights is used as a reference to verify the accuracy improvement of these two complex algorithms.

2. The method according to claim 1, characterized in that The step 1): Anthropogenic heat includes four sources: industry, construction, transportation and human metabolism. The annual average AHF at the provincial, municipal and county levels is estimated in turn based on the energy inventory method; when the municipal AHF is downscaled to the county level, industrial heat is calculated based on the proportion of the number of industrial POIs in the district or county to the total city, traffic heat and construction heat are calculated based on the proportion of the district or county population, and metabolic heat is directly estimated based on the district or county population.

3. The method according to claim 1, characterized in that Step 5): calling the raster package in R, converting the explanatory variables in raster form into a data frame format, and building a simple classification regression tree model to fill in the missing variables; calling the trained model prediction results, and building the raster input prediction results, and finally outputting the spatiotemporal distribution of multi-source anthropogenic heat in raster form.

Images