Prediction of regional carbon emissions using deep learning and mathematical

Abstract

Detecting carbon emissions is the key to carbon peaking and carbon neutrality goals. Existing research has focused on utilizing data-driven method to study carbon emissions off a single object. This study proposes a regional carbon emissions prediction method. The area objects are divided into dynamic objects for vehicles and static objects for buildings. For the dynamic object, carbon emissions is modeled using the vehicle location provided by the BeiDou satellite navigation system (BDS). For the static object, the neural network R3det (rotation remote sensing target detection) is used to identify the buildings in remote sensing images, and then the trained ARIMA time series model is used to predict the carbon emissions. The model is tested in an industrial park in Tangshan, Hebei Province in China. The result of the regional three-dimensional emission map shows that the method provided a novel and feasible idea for carbon emissions prediction.

Keywords

Deep learning ARIMA time series model remote sensing BDS regional carbon emissions

1. Introduction

Global warming has become a key challenge for all countries and poses a serious threat to the future of mankind. Excessive carbon emissions are the main cause of global warming and have a serious impact on economic growth [18]. According to the calculation of the United Nations Intergovernmental Panel on climate change (IPCC), if the temperature control target of 1.5°C by the Paris Agreement is achieved, the world must achieve zero net carbon emissions (also known as “carbon neutralization”) in 2050. That is, the annual carbon emissions are equal to the offset of emission reduction through tree planting. Carbon neutralization is an important means to deal with the global climate crisis. As the largest contributor of carbon emissions, China’s carbon emissions account for approximately 30% of global emissions. Faced with this severe situation, China has proposed a double carbon target of carbon peaking in 2030 and carbon neutrality in 2060 [20]. Moreover, it also clearly stipulates the carbon emissions limits of the various regions.

In addition, the process of energy conservation and emission reduction also gives birth to carbon trading [19]. As a trading mechanism for emerging things, it makes people pay more attention to the value of carbon emissions.

In summary, in the process of carbon neutralization, both enterprises and individuals are paying increasing attention to carbon emissions data. In particular, regional carbon emissions data has a great impact on China’s achievement of emission reduction targets. These data can not only promote the government’s control over regional emission reduction but also enable people to better participate in and integrate into carbon trading. Therefore, the realization of regional carbon emissions prediction is a key step in the next stage. Currently, the prediction of regional carbon emissions is mostly estimated directly through the peak change in regional carbon data. This requires measurement and statistics in advance. The carbon emissions data of this method is not only difficult to obtain but also poor in universality. Therefore, there is an urgent need for a convenient and feasible general prediction method.

Therefore, with the rapid improvement of the world economy, the amount of energy needed by various countries is also increasing. However, it will also produce many carbon emissions. Many carbon emissions will cause not only social problems but also environmental problems. It poses a serious threat to human survival and development. Global climate change caused by carbon and other greenhouse gas emissions has attracted extensive attention [7]. Moreover, countries have formulated various emission reduction measures. Energy accounts for 28% of the global carbon emissions data [17], and energy consumption is one of the main reasons for the increase in carbon emissions. Therefore, China’s control of excessive carbon emissions largely depends on the regulation of enterprise energy consumption and emissions. Transportation is another major carbon emissions source that accounts for 20% of the global total emissions [26]. Among them, private cars have the highest average energy consumption and carbon emissions. This means more than 10 times that of buses and approximately 20 times that of rail transit [22].

It can be seen from the above mentioned that the main sources of regional carbon emissions are traffic emissions and building energy consumption emissions. The proportion of the two emissions in the region can reach more than 80%. Therefore, to calculate regional carbon emissions in this study, the carbon emissions of motor vehicles and buildings were counted.

After obtaining the main source of carbon, the next step is the processing and prediction of its data. The prediction methods of carbon emissions can be divided into two categories. One is the mathematical-statistical model. From the initial linear equation fitting to the later differential equation, it is now widely used. Because the traditional prediction model only considers the regularity of historical data, it rarely considers adding new information about the development trend [28]. Then, a new gray rolling mechanism based on the principle of new information priority was proposed. The average weakening buffer operator was used to process the original sequence and new information before modeling. The proposed model had a better simulation and prediction performance and higher stability. Considering the slowdown of carbon emissions growth caused by the Gompertz law, a Gompertz differential equation was established [3]. According to the principle of differential information and fractional accumulation operators, the differential equation is transformed into a fractional accumulation gray Gompertz model. The results show that the model can better adapt to the trend of industrial carbon emissions in the United States.

The second method is prediction based on big data. With the development of artificial intelligence and the improvement of prediction accuracy and universality requirements, many machine learning (ML) methods, such as support vector machines [16], genetic algorithms [8], and BP neural networks [15], have been applied in regional carbon emissions prediction. 12 highly accurate machine learning algorithms were used to make a prediction, and four indices were considered to check the prediction accuracy of the algorithm [5]. According to the two important variables of energy consumption and economic growth, an efficient multistage method was developed by using clustering, predictive machine learning technology, and dimension reduction [12]. In addition, because deep learning (DL) has great potential in environmental remote sensing [6,11,24], traditional NN and DL models have also been used for earth environment monitoring [14]. Based on the importance of the information for various urban-related applications, such as urban planning and regional management, the urban land use information extracted from very fine spatial resolution (VFSR) remote sensing images has great data value [25]. These also provide a new development direction and ideas for the universal development of regional carbon emissions prediction.

However, currently, using only one method cannot meet the requirements of prediction. Researchers are trying to integrate the two methods into a variety of perspectives to solve the problem. They explore all walks of life and use statistical and big data methods to predict carbon emissions data.

To date, there have been many studies on carbon emissions prediction. They use different factors and parameters to predict carbon emissions from different angles. Lu used particle swarm optimization (PSO) algorithm optimized back propagation neural network (BP) model to predict future carbon emissions for heavy chemical industry [10]. And the significant magnitude of each carbon emissions driving force is acquired in terms of the absolute influence coefficient method. Zhang proposed a machine learning method (Back Propagation Neural Network – BPNN) to predict different functions of urban blocks carbon emissions (UBCE) by built environment factors (BEF) in Changxing, a representative small city in China [27]. The study found that UBCE can be significantly affected by BEF such as density, function, and morphology. Zhou used a correction factor calculated from the correction factors of energy consumption and basic land use emissions to build a regression model and used the model to estimate carbon emissions from urban land use [29]. Guo used the exponential cumulative gray model to predict the carbon emissions of BRICs countries [4]. These studies provide a reference for the prediction of carbon emissions under different scenarios from different angles and reflect the diversity of future carbon emissions prediction trends. Mardani used a two-stage methodology based on an ensemble adaptive neuro-fuzzy inference system to predict carbon emissions [13]. Ye proposed an enhanced multivariable dynamic time-delay discrete gray forecasting model for predicting China’s carbon emissions [23].

Although the above literature has solved some problems from different angles, the universal applicability of existing regional prediction is often not reflected. They are limited to the use of only one kind of industry data, which brings great constraints to the practical application. The main contribution of this study was to propose a model combining statistical methods and big data methods. The model has certain advantages, including the ARIMA model, BDS, and R3det methods, to predict carbon emissions data and distribution in any region. The innovations and main contributions of this study are as follows:

Regional objects are divided into dynamic and static for calculation, in which the carbon emissions obtained in vehicles and buildings can represent the carbon emissions of the selected region.

The method relates to an experimentally measured mathematical model for calculating carbon emissions by using vehicle position information.

The deep learning neural network of rotary remote sensing detection based on r3det was established to recognize the regional RS image. Considering the seasonal difference in carbon emissions, the ARIMA time series model was used for prediction.

2. Methods

The study proposes a new method for predicting regional carbon emissions, which helped to realize regional carbon emissions supervision by integrating the BDS, R3det, and ARIMA methods. The flowchart of the proposed method is shown in Fig. 1. The following sections will introduce the methods and steps of static and dynamic data acquisition.

Fig. 1.

Flowchart of the proposed method of regional carbon emissions prediction.

2.1. Regional vehicle carbon emissions data

To calculate the carbon emissions data of dynamic objects in the target area, the boundary coordinates of the target area are taken as constraints, and the BDS is used to obtain the location information of motor vehicles in the area. Through the speed of the vehicle model and location information update, the carbon emissions data at the current time can be predicted.

According to the transformation of vehicle position information, combined with the scale, the moving distance of the vehicle is calculated by using the Euclidean distance. Because the time interval used here is short, approximately 10 s, the vehicle movement is approximately linear. The moving speed of the vehicle is calculated in combination with the time interval of information collection.

According to the experiment conducted on the straight lane with good weather and small traffic volume [1], the curve fitting is conducted on the experimental data, and the relationship between vehicle speed (v) and fuel consumption (L) per 100 km is obtained as follows: $\begin{matrix} (1) & m = 0.0025 v^{2} - 0.2554 v + 31.75 \end{matrix}$ where v is measured experimentally, m is a computational process parameter.

The curve goodness of fit $R^{2} = 0.9253 > 0.8$ , meeting the curve fitting requirements.

The above values are used to determine the carbon emissions during fuel combustion, and the formula is as follows: $\begin{matrix} (2) & L = \sum_{i = 1}^{n} m \times ρ \times N_{i} \times F_{i} \times I_{i} \end{matrix}$ where ρ, $N_{i}$ , $F_{i}$ and $I_{i}$ can be found in Tables 1, 2 and 3.

Table 1
Energy emission factors ( $N_{i}$ ) for vehicle operation mg/MJ

Fuel type Emission factor

CO₂ CH₄ N₂O

Gasoline/diesel 74100 3 0.6

Fuel type	Emission factor
Gasoline/diesel	74100	3	0.6

Table 2

Limits of net calorific value and 95% confidence interval of energy ( $F_{i}$ )

Fuel type	Net calorific value	Net calorific value limit

		Upper limit	Lower limit
Gasoline/diesel	43.0	43.3	41.1

Table 3

Characteristic factors ( $I_{i}$ )

Greenhouse gas types	Characteristic factor	Greenhouse gas types	Characteristic factor
CO₂	1	N₂O	298
CH₄	43.0

The relationship between fuel consumption (E) and vehicle speed is brought in to obtain the relationship between vehicle speed and carbon emissions. $\begin{matrix} (3) & E = \sum_{i = 1}^{n} (0.0025 v^{2} - 0.2554 v + 31.75) \times ρ \times N_{i} \times F_{i} \times I_{i} \end{matrix}$

2.2. Acquisition of carbon emissions data of regional buildings

For static objects, R3det remote sensing object detection and the ARIMA time series model are used to complete the task. These methods are introduced in this section.

2.2.1. Remote sensing object detection to obtain the type and number of buildings

Due to the flexibility of unsupervised learning in function representation and automation, deep learning has been successfully applied to the research field of remote sensing. It shows great application potential in the object classification of remote sensing [2,9]. According to the architectural characteristics of regional urbanization, R3det can play a better effect in target intensive scenes and achieve a higher recall rate with less quantity [21]. Based on the above deep learning network model, the rectangle is changed to a square when the anchor was set, which made it more suitable for urban building detection and ensures high accuracy.

In this study, R3det is used in regional optical remote sensing image data. Its spatial resolution is 30 meters, the wavelength range was 0.58 microns, and the radiation resolution is 12 bits. The data can also ensure the accuracy of target detection when it is easy to obtain. The reason is that the spatial resolution of the image can meet the size requirements of the anchors of R3det. The trained image features can effectively classify buildings.

Fig. 2.

The structure of the neural network framework for remote sensing detection of rotating targets.

The R3det neural network is an improved single-stage detector based on the RetinaNet neural network, as is shown in Fig. 2. Its AP reaches 40.8. R3det and RetinaNet have the same network structure, including the backbone network and classified regression subnet.

The backbone FPN is composed of P3 to P7, which has changed greatly compared with P2 to P6 of the original network. To reduce the amount of computation, the FPN network here do not use P2. P3, P4 and P5 were generated by C3, C4 and C5 through up sampling and horizontal connection, which is the same as the generation method in FPN. The down sampling method of P6 modifies the maximum pooled sampling layer in the original FPN to convolution with a convolution kernel of $3 \times 3$ and a step size of 2. P7 adds a ReLU based on P6 and convolution. The convolution kernel is $3 \times 3$ , and the step size is 2. In addition, for each feature layer, Scale and Ratios are modified from 1 and 3 to 3 and 3. That is, the number of anchors generated increases from 3 to 9.

Fig. 3.

(a) Rotation detection frame construction principle diagram. (b) Object rotation detection effect diagram.

Each layer of FPN is related to a classification regression subnetwork. The predictors used by each feature layer from P3 to P7 in FPN are the same, that is, their weights are shared. In the predictor, the parameters of the classification subnet and regression subnet are separated, but the structure was similar. They all use a small FCN network with a feature layer as input and then connect four $3 \times 3$ convolution layers. After each convolution layer, there is a ReLU activation function. Finally, the classification subnetwork connects a $3 \times 3$ volume layer with a channel (KA) without an activation function. K is the number of types of detection targets, and the K value selected during training in this study is 28. A is the number of anchors of each feature layer, where A is equal to 9. To reduce the training parameters, the channel selected by the regression subnetwork is 4A. Based on the above network, to adapt to the recognition of rotating images in remote sensing images, the generation method and loss calculation of anchors are improved. First, to make it suitable for detecting rotating objects, it is necessary to predict the additional angle offset in the subnetwork. Take the shape of residential building as an example, as shown in Fig. 3. Combined with the rotating anchor, the following two sets of data will be obtained. $\begin{array}{l} (4) & b_{x} = \frac{x - x_{a}}{w_{a}} \\ (5) & b_{y} = \frac{y - y_{a}}{h_{a}} \\ (6) & b_{w} = log (w - w_{a}) \\ (7) & b_{h} = log (h - h_{a}) \\ (8) & b_{θ} = (θ - θ_{a}) \\ (9) & b_{x}^{'} = \frac{x^{'} - x_{a}}{w_{a}} \\ (10) & b_{y}^{'} = \frac{y^{'} - y_{a}}{h_{a}} \\ (11) & b_{w}^{'} = log (w^{'} - w_{a}) \\ (12) & b_{h}^{'} = log (h^{'} - h_{a}) \\ (13) & b_{θ}^{'} = (θ^{'} - θ_{a}) \end{array}$ where x, y, w, h and θ indicate the manually labeled anchor center of gravity coordinates, width, height and angle, respectively; $x_{a}$ , $y_{a}$ , $w_{a}$ , $h_{a}$ and $θ_{a}$ indicate the generated anchor center of gravity coordinates, width, height and angle, respectively; and $x^{'}$ , $y^{'}$ , $w^{'}$ , $h^{'}$ and $θ^{'}$ indicate the predicted anchor center of gravity coordinates, width, height and angle, respectively. $b_{x}$ , $b_{y}$ , $b_{w}$ , $b_{h}$ and $b_{θ}$ constitute $v_{n j}$ target vector. $b_{x}^{'}$ , $b_{y}^{'}$ , $b_{w}^{'}$ , $b_{h}^{'}$ , and $b_{θ}^{'}$ constitute $v_{n j}^{'}$ prediction offset vectors were used to build a new multicategory loss function as follows: $\begin{matrix} (14) & L = \frac{λ_{1}}{N} \sum_{n = 1}^{N} b_{n}^{'} \sum_{j \in {x, y, w, h, θ}} L_{reg} (v_{n j}^{'}, v_{n j}) + \frac{λ_{2}}{N} \sum_{n = 1}^{N} L_{cls} (p_{n}, b_{n}) \end{matrix}$

The loss function used is different from the facial loss of RetinaNet, which adds a penalty coefficient before classifying and regressing losses $λ 1$ and $λ 2$ . L is used to modify the anchors, and different IOU thresholds are used for refinement. In the first stage, the thresholds of the foreground and background are 0.5 and 0.4, respectively. In the first refining stage, 0.5 and 0.4 are used, respectively. If the refining stage is repeated many times, the remaining values are 0.7 and 0.6, respectively. The total loss of refining is defined as Eq. (15): $\begin{matrix} (15) & L_{total} = \sum_{i = 1}^{N} \propto_{i} L_{i} \end{matrix}$ where $L_{i}$ indicates the loss in the ith refining stage and $\propto_{i}$ indicates the tradeoff factor (default = 1).

The position information of the refined bounding box is recoded into the corresponding feature points to reconstruct the whole feature mapping and to realize feature alignment.

The feature mapping is added by bidirectional convolution to obtain new features. In the refinement stage, only the detection frame with the highest score of each feature point is retained to improve the speed, and at the same time, it is ensured that a feature point corresponds to only one refined boundary box. For each feature point of the feature map, the corresponding feature vector is obtained on the feature map according to the five coordinates of the refining box, and a more accurate feature vector is obtained through bilinear interpolation. Then, five feature vectors are added, and the current feature vector is replaced. After traversing the feature points, the whole feature map is reconstructed. Finally, the reconstructed feature map is added to the original feature map to complete the whole process.

Based on the rotation characteristics of the anchors of R3det, the neural network is better for detecting buildings in urban areas. R3det can not only adapt to the compact arrangement of buildings but also adapt to the characteristics of compact arrangement.

2.2.2. ARIMA time series model predicts current data

The full name of the ARIMA model is called the autoregressive integrated moving average model, which is also known as ARIMA $(p, d, q)$ . It is the most common statistical model used for time series prediction.

The observation of building emission data shows instability and random differences, which means that the data has a certain volatility.

The model is suitable for short-term prediction, especially when the past change mode of the data statistical series has not fundamentally changed. Therefore, according to the data processing and prediction characteristics of the model, the ARIMA time series model is selected to calculate the carbon emissions of buildings.

The first step is to draw a line graph of data changes over time. Then, observe whether the graph is stable in time variables. If it is not stable, first carry out a d-order difference operation to improve its stability. It is converted into a stationary time series. If it is stable, the ARMA $(p, q)$ model is used to fit the data. $\begin{matrix} (16) & y_{t} = μ + \sum_{i = 1}^{p} γ_{i} y_{t - i} + ε_{t} + \sum_{i = 1}^{q} θ_{i} ε_{t - i} \end{matrix}$ where p indicates the number of autoregressive terms, q indicates the moving average number of items, $ε_{t}$ indicates the error term, μ indicates the constant term, $γ_{i}$ indicates the autocorrelation coefficient, and $θ_{i}$ indicates the error term coefficient.

The second step is to calculate the autocorrelation coefficient (ACF) and partial autocorrelation coefficient (PACF) of the stationary time series obtained above. The ACF and PACF graphics are drawn. By analyzing the boundary value, the optimal parameter p and order q are obtained.

The third step is to construct the ARIMA time series model from the above calculated d, q, and p. Then, the model is tested on the test dataset. To ensure that the characteristics of the constructed model are consistent with the observed data. If not, return to step 2 to recalculate the parameters.

Fig. 4.

Selection of study area and display of RS map. The general map of remote sensing images of the area is shown above. Source: http://satellite.nsmc.org.cn/portalsite/default.aspx.

3. Case study

3.1. Study area

An industrial park in Tangshan city, Hebei Province, is selected as the research site of this study, as shown in Fig. 4. The industrial zone is located at 118°11′E and 39°36′N, with a span of 2 km and a longitudinal span of 1.8 km. According to China’s carbon accounting database (CEADs), it ranks second among the 30 provinces and cities in China. As a large traditional industrial province, the iron and steel, cement, coal, and transportation industries account for a heavy proportion of Hebei’s industrial structure, and the task of transformation and upgrading is arduous. Therefore, Hebei Province will build a provincial carbon emissions trading platform to accelerate the establishment of a carbon emissions trading information management platform. At the same time, it will strictly control carbon emissions and promote a green and low-carbon way of production and life. The study area is a traditional industrial zone in Tangshan, and its transformation pressure is large. According to the data of the Hebei Meteorological Bureau, the average temperature in this area is 0.5°C higher than that in the surrounding area. The research situation is grim.

The data involved in this study includes two categories. One is the remote sensing image and carbon data provided by the Fengyun satellite. The other is the building emissions of various industries from 1992 to 2010. The latter data is provided by the National Bureau of Statistics.

3.2. Construction of the experimental environment

Build the training platform environment: use Python 3.7 and Windows 10 as the operating system; CPU Intel(R) Core (TM) i7-10510U, GPU NVIDIA GeForce MX350, cuda10.1 driver, pytorch1.3 + PyCharm environment. The labeled and preprocessed remote sensing image datasets are trained. The trained neural network is used to recognize the remote sensing images of the target area and classify them according to the recognized categories.

The ARIMA time series model is constructed using SPSS 24.0. Subsequent data processing is completed using MATLAB 2020.

To improve the image definition and recognition accuracy, an image size of $600 * 600$ is comprehensively considered. The remote sensing image dataset is approximately 5000 pieces. The dataset is divided into two subsets: a training set, test set and cross set, which correspond to 60%, 20% and 20% of the total data samples, respectively.

3.3. Regional carbon emissions data

In this study, the location information is provided by BDS positioning. The time period is selected from 8:00 am to 10:00 am for statistical analysis. The location update data of the BDS are obtained every 10 s, as shown in Fig. 5. A total of 720 time point data was generated. For each set of data, the speed and carbon emissions model mentioned in Section 3.2 is used to calculate the motor vehicle emission data of the region in this time period. These are the dynamic data in the statistical area. Fig 5 shows the variation diagram of the vehicle trajectory over a time interval.

Fig. 5.

BDS positioning change diagram ( $△ t = 10 s$ ).

According to statistics, the number of vehicles passing through the selected area in the selected time period is 7334. Through the calculation of the model, the carbon emissions data of dynamic objects at the current time are obtained, as shown in Fig. 6. In the change chart of carbon emissions data, the traffic flow in this area fluctuates greatly, and there are peak fluctuations in different time periods. There is an early afternoon peak in this area. The peak fluctuation of carbon emissions in the morning and afternoon peaks is higher than that in the middle period, which is consistent with the daily data.

Fig. 6.

Fitting diagram of vehicle carbon emissions data.

Next, the static data in the region and the carbon emissions data of different types of buildings in the region are calculated. First, the method described in Section 3.3 is used. Through the neural network of R3det rotating target detection to identify the building objects in the area, the building information is selected with recognition reliability greater than 0.8 for statistics. Then, the type and quantity of buildings are obtained for statistics, as shown in Table 4 and Fig. 7. Then, the ARIMA time series model is used to combine the historical data of each type of building and predict the carbon emissions data at the current time.

Table 4

Type and quantity of the regional buildings

Building type	Number identified	Recognition credibility	Unrecognized quantity
Chemical plant	1	0.88	0
Warehouse	24	0.85	3
Enterprise	4	0.92	1
Residential building	77	0.83	4
Farm	3	0.94	0
Traffic station	12	0.89	2

Fig. 7.

Effect drawing of recognized (a) and unrecognized (b) number of various types of buildings detected by remote sensing objects and recognition accuracy (c) of various types of buildings.

The average accuracy rate in the table shows that the accuracy rate of residential buildings is the lowest, with an average of only 0.83. This may be related to the complex arrangement of residential buildings in the selected area. Among them, the recognition accuracy of rural areas and enterprises has reached more than 0.9. This may be related to their obvious characteristics. Although the recognition accuracy of the selected area is acceptable, it is generally low. This may have a direct and most important relationship with the definition of remote sensing images. In general, the unidentified buildings are below 5, which is acceptable for the study. This also shows the high feasibility of this method.

Fig. 8.

ACF and PACF diagrams of six types of buildings in the selected area in ARIMA time series model fitting.

To ensure the accuracy and timeliness of the data, the parameters of the new annual prediction model for each type of building were obtained by training the carbon emissions data for the previous 12 months. By this method, the carbon emissions dataset for each type of building from 1992 to 2010 is continuously trained and the model parameters are corrected, and the final obtained corrected model parameters are used for the prediction of carbon emissions data. The fitting results of the final set of carbon emissions data is presented.

According to the data fitting of the above process, the fitting residual ACF and PACF are basically within the confidence interval, resulting in a good fitting effect, as shown in Fig. 8.

However, according to the fitting situation shown in the ACF and PACF diagrams, there are certain fluctuations in chemical plants, enterprises, residential buildings, and farms. As far as chemical plants and enterprises are concerned, there are several large deviation fluctuations in the first and middle stages. From their scenario analysis, this is because they have the greatest fitting uncertainty in production. This explains the inevitable reason for its error. For the latter two, they fluctuate randomly in a cycle. This is not the same as the first two. From the scenario analysis, this is because such buildings are greatly affected by personnel mobility. This is also inevitable because of the uncertainty of personnel flow in a year. For the remaining two types of buildings, warehouse and transportation station, their errors are within the stable range of 0.1. Compared with the transportation station, there are still some small fluctuations in the warehouse data.

Fig. 9.

ARIMA time series model fitting diagram of six types of buildings in the selected area.

By fitting the residual ACF and PACF diagrams, the P parameter and Q parameter of each ARIMA time series model can be obtained. Then, the data of each building in the region in the past 12 months is fitted to predict the current data according to the above information, as shown in Fig. 9.

The fluctuation difference of fitting data of different types of buildings can be seen more clearly. This is more clearly reflected in the upper and lower limits of the confidence interval of the data. The six prediction models in Table 5 also show the high fit of the ARIMA time series model to such data.

The carbon emissions data of all buildings in the region at the current time are summarized to obtain the dynamic object data in the region.

According to the regional dynamic target data and static target data obtained above, the data is summarized to obtain the regional carbon emissions data. To display the regional carbon emissions distribution and emission source summary, the regional carbon emissions data is visualized. The regional carbon emissions distribution thermodynamic diagram and different types of regional carbon emissions data pie charts are constructed, as shown in Fig. 10. In the regional carbon thermodynamic map, the carbon emissions value is normalized from 0 to 100. 100 represents 43288 tons of carbon emissions. Those with carbon emissions below 500 tons are classified as 0.

Table 5

ARIMA time series model of buildings

Building type	Model	$R^{2}$
Chemical plant	ARIMA $(0, 1, 2)$	0.59
Warehouse	ARIMA $(0, 1, 2)$	0.65
Enterprise	ARIMA $(1, 0, 0)$	0.55
Residential building	ARIMA $(0, 1, 2)$	0.73
Farm	ARIMA $(0, 1, 2)$	0.63
Traffic station	ARIMA $(1, 0, 0)$	0.58

Fig. 10.

Regional thermal map (a) and carbon data proportion map (b).

Fig. 11.

Comparison between predicted data (a) and carbon satellite data (b). The abscissa and ordinate show the area of the area, and the ordinate unit is ten million tons. The gap between prediction and reality (c).

According to the thermodynamic diagram of the regional carbon emissions distribution, the carbon emissions distribution is mainly concentrated in areas B, C and F. Carbon emissions in area E are relatively small. Area A and area D have the lowest carbon emissions. Through the analysis of regional remote sensing images, region C is a residential area, so the emission is the highest. Region B and region F are industrial areas, followed by emissions. Area A and area E are suburbs, mainly distributed in farms and warehouses. Therefore, carbon emissions are relatively low. Area D was a plain with few buildings and low traffic flow, so the carbon emissions are the lowest. Through the pie chart of different types of carbon emissions data in the region, the regional daily life emissions account for the majority. This is because in the selected area, the number of residential buildings accounts for most of the number of regional buildings. The second is enterprise emissions, which has the largest daily energy consumption. Similarly, due to the large traffic flow in the industrial zone, traffic emissions cannot be ignored, accounting for 10%.

The data collected from the above carbon emissions sources accounts for most of the regional carbon emissions data, so we can use it as the carbon emissions data of our region.

3.4. Prediction model effect test

To verify the prediction effect of the model, the carbon satellite data in the region are selected. The two into three-dimensional graphics are drawn for comparison, as shown in Fig. 11.

To observe the difference between the predicted data and the actual data, a thermal map of the difference between the predicted data and the actual data in this area is constructed, as shown in Fig. 10(c). In the thermodynamic diagram, the carbon emissions difference is normalized from $- 1$ to 5. 5 represented a maximum difference of 6440 tons. If the difference is less than 500 tons, it is calculated as $- 1$ . This thermodynamic diagram shows that there is still a gap between the proposed method and the actual data. The forecast data of region C and region B are quite different. After analysis, this aspect may be caused by the inaccuracy of remote sensing image recognition. Because there are many dense layout patterns in residential buildings, it will lead to identification errors or unrecognizable problems. This should be the main reason for the problem. However, this may be due to the large fluctuation difference between the experimental model of dynamic object data and the ARIMA time series model of static object data. This will lead to fluctuations in the predicted value and have a certain impact.

Basically, the method of carbon satellite data acquisition is essentially different from the method used here in principle. In short, the data obtained by carbon satellites is the cumulative result of regional carbon emissions. The method used in the study is the state amount of carbon emissions at the current time. However, there is a certain connection between the distribution trend of the two, and they can reflect the distribution of regional carbon emissions. In other words, the two principles are different, and there is a large gap in the carbon emissions data. However, the distribution trend of carbon emissions data should be roughly the same. In the three-dimensional diagram, the distribution trend of carbon content is roughly the same, which shows that our prediction model shows a good fitting effect.

4. Conclusion and future direction

This study develops an effective method for predicting regional carbon emissions. The regional carbon emissions sources are calculated by dividing them into dynamic and static objects. First, for dynamic objects, the BDS is used to obtain the update of the motor vehicle location and apply the speed and emission model for calculation. Then, for static objects, the R3det target detection deep-learning neural network based on remote sensing is used for building recognition. The ARIMA time series model is used to fit the historical data of buildings to predict the current carbon emissions value. Finally, the regional carbon emissions data is summarized. The prediction data of an industrial park in Tangshan, Hebei Province, is compared with the prediction data of a carbon satellite. The results show that our method can effectively predict carbon emissions. In addition, compared with the existing machine learning methods, the method has higher universality and accuracy.

The research results adopt remote sensing detection and ARIMA time series prediction technology. Therefore, future projects can use other types of machine learning technologies, such as SVM and ANN, to analyze and predict regional carbon emissions. This study presents a dynamic and static prediction method for regional carbon emissions. In this sense, it is suggested to further integrate multidisciplinary methods, such as geology, hydrology and remote sensing methods. The integration of multidisciplinary methods will lead to new methods that are accurate and generally applicable in different fields. This study predicts the regional carbon emissions of an industrial park in Tangshan, Hebei Province. Therefore, it is suggested that national carbon emissions control and improvement be conducted based on this study in the future. According to these prediction data, the improvement measures of national environmental governance can be further studied. This study provides a thermal diagram of carbon emissions data, so future projects can study the improvement of building and traffic layouts according to this distribution. The carbon emissions in the region are evenly distributed and not concentrated, which is conducive to better absorption of the emitted carbon. This method can reduce the urban heat island effect and help the country’s development in environmental governance. Urban clean route navigation can also be studied according to the regional carbon emissions thermal distribution of this study. It can use the high emission area obtained by this method as the input parameter and improve the route according to this parameter in route navigation to avoid the high emission area. This can achieve the same effect as the above layout improvement and contribute to the realization of human energy conservation and emission reduction goals.

Conflict of interest

None to report.

References

Cai,

C.-I.

Sun,

Peng and

Yu, Study on the relationship between driving speed and carbon emissions, Roads and Air Transport 3 (2015), 4.

Dou,

Shen,

Li and

Guan, Time series remote sensing image classification framework using combination of deep learning and multiple classifiers system, International Journal of Applied Earth Observation and Geoinformation 103 (2021), 102477. doi:10.1016/j.jag.2021.102477.

Gao,

Yang,

Xiao and

Goh, A novel method for carbon emission forecasting based on Gompertz’s law and fractional grey model: Evidence from American industrial sector, Renewable Energy 181 (2022), 803–819. doi:10.1016/j.renene.2021.09.072.

Guo,

Liu,

Tu and

Chen, Forecasting carbon dioxide emissions in BRICS countries by exponential cumulative grey model, Energy Reports 7 (2021), 7238–7250. doi:10.1016/j.egyr.2021.10.075.

M.E.

Javanmard,

S.F.

Ghaderi and

Hoseinzadeh, Data mining with 12 machine learning algorithms for predict costs and carbon dioxide emission in integrated energy-water optimization model in buildings, Energy Conversion and Management 238 (2021), 114153. doi:10.1016/j.enconman.2021.114153.

Jing,

Cheng,

Ji,

Mao,

Li,

Duan,

Li and

Li, Detecting unknown dams from high-resolution remote sensing images: A deep learning and spatial analysis approach, International Journal of Applied Earth Observation and Geoinformation 104 (2021), 102576. doi:10.1016/j.jag.2021.102576.

Leemans and

Eickhout, Another reason for concern: Regional and global impacts on ecosystems for different levels of climate change, Global environmental change 14(3) (2004), 219–228. doi:10.1016/j.gloenvcha.2004.04.009.

Li,

Huang,

Wang and

Ding, Development and application of hybrid teaching-learning genetic algorithm in fuel reloading optimization, Progress in Nuclear Energy 139 (2021), 103856. doi:10.1016/j.pnucene.2021.103856.

Liu and

Shi, Local climate zone mapping as remote sensing scene classification using deep learning: A case study of metropolitan China, ISPRS Journal of Photogrammetry and Remote Sensing 164 (2020), 229–242. doi:10.1016/j.isprsjprs.2020.04.008.

10.

Lu,

Li and

Gao, Driving determinants and prospective prediction simulations on carbon emissions peak for China’s heavy chemical industry, Journal of Cleaner Production 251 (2020), 119642. doi:10.1016/j.jclepro.2019.119642.

11.

Ma,

Liu,

Zhang,

Ye,

Yin and

B.A.

Johnson, Deep learning in remote sensing applications: A meta-analysis and review, ISPRS journal of photogrammetry and remote sensing 152 (2019), 166–177. doi:10.1016/j.isprsjprs.2019.04.015.

12.

Mardani,

Liao,

Nilashi,

Alrasheedi and

Cavallaro, A multi-stage method to predict carbon dioxide emissions using dimensionality reduction, clustering, and machine learning techniques, Journal of Cleaner Production 275 (2020), 122942. doi:10.1016/j.jclepro.2020.122942.

13.

Mardani,

Van Fan,

Nilashi,

R.E.

Hooker,

Ozkul,

Streimikiene and

Loganathan, A two-stage methodology based on ensemble adaptive neuro-fuzzy inference system to predict carbon dioxide emissions, Journal of cleaner production 231 (2019), 446–461. doi:10.1016/j.jclepro.2019.05.153.

14.

Rodrigues,

Á.

Gomes,

A.R.

Gaspar and

C.H.

Antunes, Estimation of renewable energy and built environment-related variables using neural networks – A review, Renewable and Sustainable Energy Reviews 94 (2018), 959–988. doi:10.1016/j.rser.2018.05.060.

15.

B.O.P.

Soepangkat,

Norcahyo,

M.K.

Effendi and

Pramujati, Multi-response optimization of carbon fiber reinforced polymer (CFRP) drilling using back propagation neural network-particle swarm optimization (BPNN-PSO), Engineering Science and Technology, an International Journal 23(3) (2020), 700–713. doi:10.1016/j.jestch.2019.10.002.

16.

Suleiman,

M.R.

Tight and

A.D.

Quinn, Applying machine learning methods in managing urban concentrations of traffic-related particulate matter (PM10 and PM2.5), Atmospheric Pollution Research 10(1) (2019), 134–144. doi:10.1016/j.apr.2018.07.001.

17.

Sun and

Ren, The impact of energy consumption structure on China’s carbon emissions: Taking the Shannon–Wiener index as a new indicator, Energy Reports 7 (2021), 2605–2614. doi:10.1016/j.egyr.2021.04.061.

18.

Wang,

Fang,

Wang,

Huang and

Ma, Quantifying the relationship between urban development intensity and carbon dioxide emissions using a panel data analysis, Ecological Indicators 49 (2015), 121–131. doi:10.1016/j.ecolind.2014.10.004.

19.

X.Q.

Wang,

C.W.

Su,

O.R.

Lobonţ,

Li and

Nicoleta-Claudia, Is China’s carbon trading market efficient? Evidence from emissions trading scheme pilots, Energy 245 (2022), 123240.

20.

Wei,

Bai,

Zhao,

Zhang and

Xue, The prospect of natural gas hydrate (NGH) under the vision of peak carbon dioxide emissions in China. Petroleum 7 (2021), 357–363.

21.

Yang,

Liu,

Yan,

Li,

Zhang and

Yu, R3det: Refined single-stage detector with feature refinement for rotating object, arXiv preprint, 2019. arXiv:1908.05612.

22.

Yang,

Wang,

Liu and

Zhou, Microsimulation of low carbon urban transport policies in Beijing, Energy Policy 107 (2017), 561–572. doi:10.1016/j.enpol.2017.05.021.

23.

Ye,

Yang,

Dang and

Wang, An enhanced multivariable dynamic time-delay discrete grey forecasting model for predicting China’s carbon emissions, Energy 249 (2022), 123681.

24.

Youme,

Bayet,

J.M.

Dembele and

Cambier, Deep learning and remote sensing: Detection of dumping waste using UAV, Procedia Computer Science 185 (2021), 361–369. doi:10.1016/j.procs.2021.05.037.

25.

Zhang,

Sargent,

Pan,

Li,

Gardiner,

Hare and

P.M.

Atkinson, An object-based convolutional neural network (OCNN) for urban land use classification, Remote sensing of environment 216 (2018), 57–70. doi:10.1016/j.rse.2018.06.034.

26.

Zhang,

Long,

Li and

Wei, Potential for reducing carbon emissions from urban traffic based on the carbon emission satisfaction: Case study in Shanghai, Journal of Transport Geography 85 (2020), 102733. doi:10.1016/j.jtrangeo.2020.102733.

27.

Zhang,

Yan,

Liu and

Qiao, Towards low carbon cities: A machine learning method for predicting urban blocks carbon emissions (UBCE) based on built environment factors (BEF) in Changxing City, China, Sustainable Cities and Society 69 (2021), 102875. doi:10.1016/j.scs.2021.102875.

28.

Zhou,

Zeng,

Wang,

Luo and

Liu, Forecasting Chinese carbon emissions using a novel grey rolling prediction model, Chaos, Solitons & Fractals 147 (2021), 110968. doi:10.1016/j.chaos.2021.110968.

29.

Zhou,

Chen,

Tang and

Mei, Urbanization, land use change, and carbon emissions: Quantitative assessments for city-level carbon emissions in Beijing–Tianjin–Hebei region, Sustainable Cities and Society 66 (2021), 102701. doi:10.1016/j.scs.2020.102701.

Prediction of regional carbon emissions using deep learning and mathematical–statistical model

Abstract

Keywords

1. Introduction

2. Methods

Table 1 Energy emission factors ( N i ) for vehicle operation mg/MJ Fuel type Emission factor CO2 CH4 N2O Gasoline/diesel 74100 3 0.6

2.2.1. Remote sensing object detection to obtain the type and number of buildings

3.1. Study area

3.2. Construction of the experimental environment

3.3. Regional carbon emissions data

4. Conclusion and future direction

Conflict of interest

References

Table 1
Energy emission factors ( $N_{i}$ ) for vehicle operation mg/MJ

Fuel type Emission factor

CO₂ CH₄ N₂O

Gasoline/diesel 74100 3 0.6