A city-scale carbon emission prediction method and system based on small sample data enhancement

Abstract

The present application relates to a kind of city scale carbon emission prediction method and system based on small sample data enhancement.The present application first obtains city night light data, utilizes multiple linear regression model to correct the fitting light proportion obtained from the data, combines city night light data to calculate the carbon dioxide emission of each city using inversion calculation formula, obtains carbon emission dataset;Carbon emission dataset is enhanced using improved time series enhancement generative adversarial network model TimeTGAN, and carbon emission enhanced dataset is obtained;Then train spatio-temporal sequence prediction network model, obtain the trained spatio-temporal sequence prediction model;Finally, collect city spatial geographic location data, input it into the trained spatio-temporal sequence prediction model, obtain carbon emission prediction result.Compared with prior art, the present application has effectively solved the small sample problem in carbon emission prediction, and has the advantages such as increasing the accuracy of prediction result.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of carbon emission prediction, in particular to a city-scale carbon emission prediction method and system based on small sample data enhancement. BACKGROUND

[0002] Today, the world is actively responding to climate change and vigorously promoting carbon emission reduction. As a highly concentrated gathering place of population, industry and economic activities, cities have a large and complex carbon emission scale. Predicting city-scale carbon emissions can provide indispensable decision-making basis for city managers to develop scientific and reasonable carbon emission reduction strategies and plan green and low-carbon development paths.

[0003] However, the openness and complexity of the current urban system make it difficult to accurately define the accounting boundary of carbon emissions, and it is difficult to collect carbon emission data. The existing carbon emission monitoring technology has limitations, and it is difficult to achieve comprehensive, real-time and accurate monitoring for some small and scattered emission sources, which increases the difficulty of data collection and leads to repeated collection and inconsistency of data.

[0004] The amount of city-scale carbon emission data that can be obtained at present is small, and a small amount of data samples cannot fully reflect the complex dynamic change rule of city carbon emission, cannot capture the subtle correlation between various influencing factors and carbon emission, and cannot provide sufficient material support for building a high-precision and high-reliability prediction model, ultimately forming a typical small sample problem, which greatly hinders the effective prediction of future city carbon emissions. SUMMARY

[0005] The purpose of the present application is to overcome the defects of the prior art and provide a city-scale carbon emission prediction method and system based on small sample data enhancement.

[0006] The purpose of the present application can be achieved by the following technical solutions:

[0007] According to one aspect of the present application, a city-scale carbon emission prediction method based on small sample data enhancement is provided, comprising the following steps:

[0008] S1, obtaining city night light data, and using a multiple linear regression model to correct the data to obtain a fitted light proportion;

[0009] S2, based on the fitted light proportion, using a city night light data inversion calculation formula to calculate the carbon dioxide emission of each city to obtain a carbon emission dataset;

[0010] S3, using an improved time series enhancement generative adversarial network model TimeTGAN to perform data enhancement on the carbon emission dataset to obtain a carbon emission enhanced dataset;

[0011] S4, training the spatio-temporal sequence prediction network model using the carbon emission enhanced data set to obtain a trained spatio-temporal sequence prediction model;

[0012] S5, collecting city spatial geographic location data and inputting the city spatial geographic location data into the trained spatio-temporal sequence prediction model to obtain a carbon emission prediction result.

[0013] As a preferred technical solution, the multiple linear regression model in S1 incorporates social and economic indicators closely related to carbon dioxide emissions, including industrial power data accounting for the proportion of total provincial industrial power data, PM2.5 emissions accounting for the proportion of total provincial emissions, population data accounting for the proportion of total provincial population data, private car ownership accounting for the proportion of total provincial ownership, and retail sales data accounting for the proportion of total provincial sales data.

[0014] As a preferred technical solution, the city night light data inversion calculation formula in S2 is specifically:

[0015]

[0016] In the formula, CE1 is municipal carbon emission data; NTC i is the corrected proportion of night light, CE P is provincial carbon emission data.

[0017] As a preferred technical solution, the process of data enhancement on the carbon emission data set in S3 introduces a bidirectional temporal convolutional network (BiTCN) network and an expanded causal convolution. The specific enhancement process is as follows: after encoding the carbon emission data set, input the encoded data set into an embedding function, perform feature extraction, data compression and pass to the latent space representation in the embedding function, output random data, use the random data as the input of the generator to generate time series data, and use the time series data discriminator to judge true or false. If the discrimination is passed, the data is converted into a usable carbon emission enhanced data set through a recovery function. If the discrimination is not passed, the new data generated in the generator is fed back to the generator, and the process of generating new data in the generator continues until the required preset amount of data is generated. The data is integrated into a carbon emission enhanced data set. As a preferred technical solution, an embedding network function is applied in the embedding network, and the specific formula is as follows:

[0018] h S_TCN =TCN(h S_F ),h X_TCN =TCN(h X_F )

[0019] h S_F =FC(S),h X_F =FC(X)

[0020] where h S_F denotes the static feature h S the result after processing by the fully connected layer; h X_F denotes the temporal feature h X the result after processing by the fully connected layer; h S_TCN is the latent space; h X_TCN is the temporal feature; S is the embedded static feature; and X is the embedded temporal feature.

[0021] As a preferred technical solution, a recovery network applies a recovery network function, and the specific formula is:

[0022]

[0023] wherein, is the recovered static feature; is the recovered temporal feature, and is an intermediate step of the recovery network; is the finally generated static feature; is the finally generated temporal feature; FC' is a fully connected layer in the recovery network; and TCN' represents a temporal convolution network in the recovery network.

[0024] As a preferred technical solution, the specific formula of the generator is:

[0025]

[0026] wherein, TCN G is a temporal convolution network in the generator network; FC G is a fully connected layer in the generator network; Z S is a static noise vector, which generates a static feature through a fully connected layer network; is its latent representation; Z 1:T is a time series noise vector, which generates an entire time series through a TCN network; is its latent representation.

[0027] As a preferred technical solution, the discriminator uses a bidirectional TCN network, and the specific formula is:

[0028]

[0029] u t = TCN forward (x t , u t-1 ), v t = TCN backward (x t , v t+1 )

[0030] h tCombine(u t ,v t )

[0031] wherein, is the evaluation result of the static feature latent representation; is the evaluation result of the temporal feature latent representation;u t is the recursive calculation formula of the forward hidden state for each time point t in the time series data x t , representing the forward propagation result of the time convolution network;v t is the recursive calculation formula of the backward hidden state for each time point t in the time series data x t , representing the backward propagation result of the time convolution network;h t is the hidden state combined with the forward and backward propagation results.

[0032] As a preferred technical solution, the spatio-temporal sequence prediction network model in S4 selects a convolutional long short-term memory network model (ConvLSTM), and the model formula is:

[0033] i t = σ(W xi *X t +W hi *H t-1 +W ci ⊙c t-1 +b i )

[0034] f t = σ(W xf *X t +W hf *H t-1 +W cf ⊙c t-1 +b f )

[0035] C t = f t ⊙C t-1 +i i ⊙tanh g (W xc *x t +W hc *H t-1 +b c )

[0036] o t = σ(W xo *X t +W h0 *H t-1 +W co ⊙C t +bo )

[0037] H t =o t ⊙tanh g (c t )

[0038] In the formula, sigma is a sigmoid activation function; tanh g is a hyperbolic tangent activation function; X t is the city space geographical position data of the current time step; H t-1 is the prediction result of the last time step; C t-1 is the memory unit of the last time step; f t , i t , C t and o t are the forget gate, the input gate, the memory unit and the output gate respectively; W xi , W xf , W xc , W xo , b i , b f , b c and b o are all model parameters; and H t is the carbon emission prediction result.

[0039] According to another aspect of the present application, a city-scale carbon emission prediction system based on small sample data enhancement is provided, which works according to the city-scale carbon emission prediction method based on small sample data enhancement as described above, and comprises a carbon emission calculation module, a carbon emission data enhancement module, a model training module and a carbon emission prediction module.

[0040] The carbon emission calculation module is configured to obtain city night light data, correct the fitting light proportion obtained by using a multiple linear regression model on the data, and calculate the carbon dioxide emission of each city based on the fitting light proportion and using a city night light data inversion calculation formula, to obtain a carbon emission dataset; the carbon emission data enhancement module is configured to use an improved time series enhancement generative adversarial network model TimeTGAN to perform data enhancement on the carbon emission dataset, to obtain a carbon emission enhanced dataset; the model training module is configured to use the carbon emission enhanced dataset to train a spatiotemporal sequence prediction network model, to obtain a trained spatiotemporal sequence prediction model; and the carbon emission prediction module is configured to collect city space geographical position data, input the data into the trained spatiotemporal sequence prediction model, and obtain a carbon emission prediction result.

[0041] Compared with the prior art, the present application has the following beneficial effects:

[0042] 1. In the present application, firstly, the city night light data is obtained, the fitting light proportion is obtained by correcting the data by using a multiple linear regression model; and based on the fitting light proportion, the carbon dioxide emissions of each city are calculated by using the city night light data to calculate the formula, to obtain the carbon emission dataset; then the carbon emission dataset is enhanced by using the improved time series enhanced generative adversarial network model TimeTGAN to obtain the carbon emission enhanced dataset; then the spatio-temporal sequence prediction network model is trained by using the carbon emission enhanced dataset to obtain the trained spatio-temporal sequence prediction model; finally, the city spatial geographic location data is collected and input into the trained spatio-temporal sequence prediction model to obtain the carbon emission prediction result. The present application generates carbon emission training data based on the improved time series enhanced generative adversarial network model TimeTGAN, which enriches the carbon emission training data and effectively solves the small sample problem in carbon emission prediction.

[0043] 2. In the process of data enhancement of the carbon emission dataset, the BiTCN network and the dilated causal convolution are introduced, and the specific enhancement process is as follows: the carbon emission dataset is encoded and input into an embedding function, feature extraction, data compression and transmission to a latent space representation are performed in the embedding function, and random data is output; the random data is used as the input of the generator to generate time series data; the time series data discriminator discriminates and judges true or false; if the discrimination is passed, the recovery function is used to convert into available carbon emission enhanced dataset; if the discrimination is not passed, the generator is fed back to continue generating new data in the generator until the required preset amount of data is generated, which is integrated into the carbon emission enhanced dataset. The long-term dependence relationship in the time series data is effectively captured, so that more accurate and coherent carbon emission enhanced data is generated, and the training resources for carbon emission prediction are enriched.

[0044] 3. In the multiple linear regression model, social indicators closely related to carbon dioxide emissions are included, including the proportion of industrial power data to the total industrial power data, the proportion of PM2.5 emissions to the total provincial emissions, the proportion of population data to the total provincial population data, the proportion of private car ownership to the total provincial ownership, and the proportion of retail sales data to the total provincial sales data. The results more accurately reflect how the social and economic characteristics of each city affect the correlation between NTL data and actual carbon dioxide emissions, and improve the accuracy of carbon dioxide emission estimation.

[0045] 4. In the present application, the convolutional long short-term memory network model is used as the spatio-temporal sequence prediction network model, and the data generated by TimeTGAN is used for training, which improves the accuracy and reliability of city-scale carbon emission prediction. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 It is a city-scale carbon emission prediction method step schematic diagram in the present application;

[0047] Figure 2 The structural schematic diagram of the urban scale carbon emission prediction system in the present application. DETAILED DESCRIPTION

[0048] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor should belong to the protection scope of the present application.

[0049] Embodiment 1

[0050] In the present embodiment, a city scale carbon emission prediction method based on small sample data enhancement is applied, and the method steps are as shown in Figure 1 The method steps are as follows:

[0051] S1, acquiring city night light data, and correcting the data by using a multiple linear regression model to obtain a fitting light proportion;

[0052] S2, based on the fitting light proportion, calculating the carbon dioxide emission of each city by using a city night light data inversion calculation formula, to obtain a carbon emission data set;

[0053] S3, using an improved time series enhancement generative adversarial network model TimeTGAN to perform data enhancement on the carbon emission data set, to obtain a carbon emission enhanced data set;

[0054] S4, using the carbon emission enhanced data set to train a spatiotemporal sequence prediction network model, to obtain a trained spatiotemporal sequence prediction model;

[0055] S5, collecting city spatial geographic location data, and inputting the data into the trained spatiotemporal sequence prediction model, to obtain a carbon emission prediction result.

[0056] In the present embodiment, first, city night light data in a selected range of time is selected, and a public data set is calculated according to the IPCC carbon emission factor method. According to the carbon emission calculation index, a cluster analysis is performed on the city scale to obtain the construction guidance type of each city, and the matching coefficient of the night lighting data (NTL) of each city is corrected based on the same. Finally, the fitting coefficient obtained by using a multiple linear regression model and a night light data inversion calculation formula are used to estimate the carbon dioxide emission of each city.

[0057] In the embodiment, the multi-linear regression model incorporates socio-economic indicators closely related to carbon dioxide emissions, including the proportion of industrial power data to the total industrial power data in the province, the proportion of PM2.5 emissions to the total emissions in the province, the proportion of population data to the total population data in the province, the proportion of private car ownership to the total ownership in the province, and the proportion of retail sales data to the total sales data in the province.

[0058] In the embodiment, the urban night light data inversion calculation formula is specifically:

[0059]

[0060] In the formula, CE1 is the municipal carbon emission data; NTC i is the corrected proportion of night light, CE P is the provincial carbon emission data.

[0061] In the embodiment, the carbon emission enhancement module uses the improved time series enhancement generative adversarial network model TimeTGAN to realize carbon emission data enhancement.

[0062] The embedding network and the restoration network are used to obtain the mapping between the carbon emission time series data features and the latent space, and to reduce the dimensionality of the carbon emission data features.

[0063] The restoration network is used to reconstruct the static and temporal features of the carbon emission data, and is responsible for converting the latent representation generated by the generator back to the original data space.

[0064] The bidirectional temporal convolutional network BiTCN (Bidirectional Temporal Convolutional Network) network and the dilated causal convolution are introduced to effectively capture the long-term dependencies in the time series data, thereby generating more accurate and coherent carbon emission enhancement data and enriching the training resources for carbon emission prediction.

[0065] In the embodiment, the process of data enhancement for the carbon emission dataset is as follows: after encoding the carbon emission dataset, input the embedding function, perform feature extraction, data compression, and pass to the latent space representation in the embedding function, output random data, use the random data as the input of the generator to generate time series data, and use the time series data discriminator to judge true or false, if the discrimination is passed, use the restoration function to convert into usable carbon emission enhancement dataset, if the discrimination is not passed, feedback to the generator to continue generating new data in the generator until the required preset amount of data is generated, and integrate the data into the carbon emission enhancement dataset.

[0066] In the embodiment, the embedding network function is applied in the embedding network, and the specific formula is as follows:

[0067] h S_TCN= TCN(h S_F ), h X_TCN = TCN(h X_F )

[0068] h S_F = FC(S), h X_F = FC(X)

[0069] where h S_F denotes the result of static feature h S after processing by the fully connected layer; h X_F denotes the result of temporal feature h X after processing by the fully connected layer; h S_TCN is the latent space; h X_TCN is the temporal feature; S is the embedded static feature; and X is the embedded temporal feature.

[0070] In this embodiment, the recovery network applies a recovery network function, and the specific formula is as follows:

[0071]

[0072] wherein, is the recovered static feature; is the recovered temporal feature, which is an intermediate step of the recovery network; is the finally generated static feature; is the finally generated temporal feature; and FC' is the fully connected layer in the recovery network; and TCN' denotes the temporal convolution network in the recovery network.

[0073] In this embodiment, the specific formula of the generator is as follows:

[0074]

[0075] wherein, TCN G is the temporal convolution network in the generator network; FC G is the fully connected layer in the generator network; Z S is the static noise vector, which generates the static feature through the fully connected layer network; is the latent representation thereof; Z 1:T is the time series noise vector, which generates the entire time series through the TCN network; is the latent representation thereof.

[0076] In this embodiment, the discriminator uses a bidirectional TCN network, and the specific formula is as follows:

[0077]

[0078] u t = TCN forward (xt ,u t-1 ),v t =TCN backward (x t ,v t+1 )

[0079] h t =Combine(u t ,v t )

[0080] where, is the evaluation result of static feature latent representation; is the evaluation result of temporal feature latent representation;u t is the recursive calculation formula of the forward hidden state for each time point t in the time series data x t , which represents the forward propagation result of the time convolution network;v t is the recursive calculation formula of the backward hidden state for each time point t in the time series data x t , which represents the backward propagation result of the time convolution network;h t is the hidden state combined with the forward and backward propagation results.

[0081] In this embodiment, first, the city scale influence factor data is collected, 18 carbon emission influence factors related to carbon emission, such as industrial structure, energy structure, and energy emission, are selected, and Pearson analysis is performed to select the top 10 influence factors with the highest correlation as characteristic values.

[0082] Further, the spatial correlation analysis of city scale carbon emission is performed, and it is concluded that city scale carbon emission has the characteristics of spatial positive correlation.

[0083] Since the city carbon emission problem has the characteristics of spatial positive correlation, the city spatial geographic location data is introduced, and the carbon emission time-space sequence prediction model ConvLSTM is selected to predict the city scale carbon emission.

[0084] In this embodiment, the spatio-temporal sequence prediction network model selects the convolution long short-term memory network model, and the model formula is:

[0085] i t =σ(W xi *X t +W hi *H t-1 +W ci ⊙c t-1 +b i )

[0086] f t =σ(W xf *X t +Whf *H t-1 +W cf ⊙c t-1 +b f )

[0087] C t =f t ⊙C t-1 +i i ⊙tanh g (W xc *x t +W hc *H t-1 +b c )

[0088] o t =σ(W xo *X t +W h0 *H t-1 +W co ⊙C t +b o )

[0089] H t =o t ⊙tanh g (C t )

[0090] In the formula, sigma is a sigmoid activation function; tanh g is a hyperbolic tangent activation function; X t is the city space geographical position data of the current time step; H t-1 is the prediction result of the previous time step; C t-1 is the memory unit of the previous time step; f t , i t , C t and o t are the forget gate, input gate, memory unit and output gate respectively; W xi , W xf , W xc , W xo , b i , b f , b c and b o are all model parameters; and H t is the carbon emission prediction result.

[0091] In summary, the method proposes to generate carbon emission training data based on an improved time series enhanced generative adversarial network model TimeTGAN, which expands and enriches the carbon emission training data, and effectively solves the small sample problem in carbon emission prediction.

[0092] Embodiment 2

[0093] In this embodiment, a city-scale carbon emission prediction system based on small sample data enhancement is applied, as shown in the system Figure 2 includes a carbon emission calculation module, a carbon emission data enhancement module, a model training module, and a carbon emission prediction module.

[0094] The carbon emission calculation module is configured to obtain city night light data, correct the fitting light proportion obtained by using a multiple linear regression model on the data, and calculate the carbon dioxide emissions of each city based on the fitting light proportion and using a city night light data inversion calculation formula, to obtain a carbon emission dataset.

[0095] The carbon emission data enhancement module is configured to use an improved time series enhancement generative adversarial network model TimeTGAN to perform data enhancement on the carbon emission dataset, to obtain a carbon emission enhanced dataset.

[0096] The model training module is configured to use the carbon emission enhanced dataset to train a spatiotemporal sequence prediction network model, to obtain a trained spatiotemporal sequence prediction model.

[0097] The carbon emission prediction module is configured to collect city spatial geographic location data, input the data into the trained spatiotemporal sequence prediction model, and obtain a carbon emission prediction result.

[0098] In this embodiment, city night light data in a certain time range is first selected, and a public dataset is calculated according to the IPCC carbon emission factor method.

[0099] In this embodiment, based on city employment and industrial structure data, the cities are classified into resource support type, private economy type, industrial support type, high-tech type, and other types according to the carbon emission calculation index k-means clustering algorithm (K-meansk). The construction orientation type of each city is obtained, and is used as a matching coefficient for correcting the city night light data (NTL).

[0100] Then, a multiple linear regression model is constructed to calculate the correction model of cities of different categories, and the carbon emissions of the cities are calculated according to the corrected night light proportion of each city in the region combined with the provincial carbon emission data.

[0101]

[0102] PCE n,m = a0+ a n x n,m + ε i

[0103] PCEn+1,m = a0 + a n x n,m + a n+1 x n+1,m + ε i

[0104] PCE n+2,m = a0 + a n x n,m + a n+1 x n+1,m + a n+2 x n+2,m + ε i

[0105] In the formula, PCE i is the proportion of carbon emissions of city i to provincial carbon dioxide emissions calculated by the multiple linear regression model, which is the dependent variable. The proportion is calculated by a linear regression model; a0 is the intercept term, a n represents the regression coefficient of the independent variable x n,m added in the nth step, which indicates that after controlling the influence of other variables, x n,m changes by one unit, the expected number of units of the dependent variable PCE i changes. β j is the regression coefficient of the independent variable X j , and j is the number of independent variables. ε i is the error term. n and m represent the number of regression steps and the type of city in the hierarchical regression model, respectively. With the step increasing by 1, an additional variable will be added in step n+i.

[0106] In this embodiment, a series of multi-source and multi-dimensional social and economic indicators closely related to carbon dioxide emissions, such as the proportion of industrial electricity data to provincial industrial electricity data (IEC), the proportion of PM2.5 emissions to provincial emissions (PM), the proportion of population data to provincial population data (POP), and the proportion of private car ownership to provincial ownership (PCO), and the proportion of retail sales data to provincial sales data (RC), etc., are included in the multiple linear regression model. The matching coefficient of the night lighting data (NTL) of each city is corrected. In this paper, POP is added in the first layer, and IEC, PM, PCO, and RC are added in the second, third, fourth, and fifth layers, respectively, on the basis of the first layer. Select variables with significance and r-square change as independent variables to construct specific regression models for these cities. The results more accurately reflect how the social and economic characteristics of each city affect the correlation between NTL data and actual carbon dioxide emissions, and improve the accuracy of carbon dioxide emission estimation. The results of the multiple hierarchical regression are shown in Table 1.

[0107] Table 1 Multiple hierarchical regression results

[0108]

[0109] Note: ***p < 0.01; **p < 0.05; *p < 0.1

[0110] In this embodiment, R2 is a determination coefficient, which represents the proportion of the variation of the dependent variable explained by the independent variable in the model. Its value is between 0 and 1, and the closer the value is to 1, the better the fitting effect of the model, and the higher the proportion of the variation of the dependent variable explained by the independent variable. Adjusted R2 is a modification of R2, which can more accurately reflect the explanatory power of the model. F value is a statistical quantity used to test the overall significance of the multiple regression model, which tests whether the explanatory power of all independent variables in the model on the dependent variable is significant. If the F value is large, it means that the model has a good explanatory power on the dependent variable as a whole. When performing stepwise regression analysis, △F value represents the change of F statistic after adding or deleting a variable in the model. If the △F value is large after adding a variable, and the corresponding p value is less than the significance level (such as 0.05), it means that the variable has a significant contribution to the model and should be retained in the model.

[0111] In this embodiment, the above regression results show that the selected variables have different effects on the five types of cities, so according to the multiple regression model to which the city belongs, a new proportion of carbon emissions at city scale to provincial emissions is obtained. Taking 2017-2020 as an example, the fitting light proportion correction formula based on social and economic elements is shown in Table 2.

[0112] Table 2 Fitting light proportion correction formula based on social and economic elements

[0113]

[0114] In this embodiment, the fitting light proportion based on social and economic elements obtained by the multiple linear regression model is used to estimate the carbon dioxide emissions of each city, and the carbon dioxide emission characteristics at the city level are analyzed accordingly. The night light data inversion model formula is as follows:

[0115]

[0116] In the formula, NTC i is the corrected night light proportion, CE P is the provincial carbon emission data, and the city-scale carbon emission data CE1 can be obtained according to the estimation.

[0117] Finally, the fitting coefficient obtained by the multiple linear regression model and the night light data inversion calculation formula are used to estimate the carbon dioxide emissions of each city.

[0118] In the embodiment, the carbon emission enhancement module implements carbon emission data enhancement by using an improved time series enhancement generative adversarial network model TimeTGAN, which includes an embedding network, a recovery network, a generator, and a discriminator.

[0119] In the embodiment, the carbon emission real data is encoded, and is transmitted to a latent space representation through embedding functions, feature extraction, data compression, and the like. At this time, the real data is converted into random data. The random data is used as an input of the generator to generate new time series data. The time series data discriminator discriminates and judges the true and false. If the discrimination is passed, the recovery function is used to convert the carbon emission enhancement data set into available carbon emission enhancement data set. If the discrimination is not passed, the sequence generator is fed back to continue generating new data in the sequence generator until sufficient available carbon emission enhancement data set is generated.

[0120] In the embodiment, the embedding network is composed of a combination of a fully connected layer FC and a time convolution network TCN. The fully connected layer FC first maps the static feature to a latent space h S_TCN Then, the time convolution network TCN further processes the temporal feature h X_TCN in the intermediate latent space to capture the dynamic changes and time dependence of the time series feature.

[0121] The embedding network function is as follows:

[0122] h S_TCN =TCN(h S_F ),h X_TCN =TCN(h X_F )

[0123] h S_F =FC(S),h X_F =FC(X)

[0124] In the formula, h S_F represents the result of the static feature h S processed by the fully connected layer (Fully Connected, FC); h X_F represents the result of the temporal feature h X processed by the fully connected layer; h S_TCN is the latent space; h X_TCN is the temporal feature; S is the embedded static feature; and X is the embedded temporal feature.

[0125] In the embodiment, the recovery network reconstructs the static and temporal features of the carbon emission data, and is responsible for converting the latent representation generated by the generator back to the original data space. The recovery network function is as follows:

[0126]

[0127] Where, is the static feature of the recovery; is the temporal feature to be recovered, is the intermediate step of the recovery network, is the final generated static feature; is the final generated temporal feature, FC′ is the fully connected layer in the recovery network, and TCN′ represents the temporal convolutional network in the recovery network.

[0128] In this embodiment, the generator is constructed using a fully connected layer and a bidirectional temporal convolutional network Bi-TCN. The generator network consists of two parts. S and Z 1:T For vector space, generate H S and H X , the generator function uses static and temporal random vector groups to generate potential codes. The generator function is as follows:

[0129]

[0130] Where TCNG represents the temporal convolutional network in the generator network, FCG represents the fully connected layer in the generator network, and the static noise vector Z S The static features are generated by the fully connected layer network, and the potential representation is Time series noise vector Z 1:T The entire time series is generated by the TCN network, and the potential representation is

[0131] In this embodiment, the discriminator uses a bidirectional TCN to analyze the potential representation, and the discriminator function is as follows:

[0132]

[0133] u t =TCN forward (x t ,u t-1 ),v t =TCN backward (x t ,v t+1 )

[0134] h t =Combine(u t ,v t )

[0135] Where, represents the evaluation result of the potential representation of static features, Represents the evaluation result of the latent representation of temporal features. t For time series data x tAt each time point t in the forward hidden state, the recursive calculation formula represents the forward propagation result of the temporal convolutional network, v t For time series data x t At each time point t in the backward hidden state, the recursive calculation formula represents the backward propagation result of the time convolution network. In the bidirectional TCN, each time point fully represents h t Represents the hidden state that combines the results of forward and backward propagation.

[0136] In this example, a time series classification model, BiTCN, is introduced to distinguish between original and generated datasets. TimeTGAN is tested and compared with different enhancement methods based on the discrimination score. The results are shown in Table 3. TimeTGAN consistently generates higher-quality synthetic data, achieving a score of 0.101 for generated samples, 23% lower than the next-best baseline (TimeGAN, 0.171), demonstrating that TimeTGAN significantly outperforms other models.

[0137] Table 3 Identification test results

[0138]

[0139] In this embodiment, during the carbon emission prediction process, a spatial correlation analysis of carbon emissions is first performed:

[0140] Moran's I analysis was used to analyze the spatial correlation of carbon emissions in the Yangtze River Delta region and explore the spatial agglomeration of emissions in the Yangtze River Delta urban agglomeration. Moran's I analysis is mainly used to evaluate spatial distribution characteristics and their differences. It is divided into a global Moran's I index and a local Moran's I index (LISA). The global Moran's I index mainly reflects the agglomeration nature of the entire space, and the analysis results range from -1 to 1. When the Moran's I index is greater than 0, it indicates spatial correlation, and the closer the analysis value is to 1, the higher the correlation. When the Moran's I index is less than 0, it indicates spatial dispersion, and the closer the analysis value is to -1, the higher the dispersion. A Moran's I index of 0 indicates no correlation between spatial units.

[0141] In this embodiment, taking the carbon emissions of the Yangtze River Delta urban agglomeration in a certain year as an example, Arcgis software was used to perform a spatial correlation analysis of carbon emissions. The analysis results are shown in Table 4.

[0142] Table 4 Spatial correlation analysis of urban agglomerations in a certain year

[0143]

[0144] The Moran I index is 0.447169, indicating that the spatial distribution of the Yangtze River Delta urban agglomeration is positively correlated with carbon emissions, and the Z (standard deviation multiple) score is 4.052541, which is greater than 2.58, indicating that the confidence interval of the data is greater than 99%, so it can be concluded that the carbon emissions of the Yangtze River Delta urban agglomeration have the characteristics of spatial clustering, so the city spatial geographic location data should be introduced when predicting the carbon emissions of the Yangtze River Delta urban agglomeration, and an algorithm model suitable for spatio-temporal sequence prediction should be selected to predict the carbon emissions of the Yangtze River Delta region.

[0145] In this embodiment, then the city scale influence factor data is collected, 18 carbon emission influence factors such as carbon emission related industrial structure, energy structure, energy emission amount, etc. are selected, and Pearson correlation analysis is performed, and 10 influence factors such as population (ten thousand people), urbanization rate, per capita GDP (yuan / person), total energy consumption (ten thousand tons of standard coal), etc. with P value greater than 0.5 are extracted as characteristic values.

[0146] In this embodiment, since the urban agglomeration carbon emissions have spatial positive correlation, the city spatial geographic location data is introduced, and a spatio-temporal sequence prediction model: Convolutional Long-Short Term Memory (ConvLSTM) is selected to predict the city scale carbon emissions. ConvLSTM is a spatio-temporal sequence prediction model combining Convolutional Neural Network CNN and Long-Short Term Memory Network LSTM, and the ConvLSTM model formula is as follows:

[0147] i t =σ(W xi *X t +W hi *H t-1 +W ci ⊙c t-1 +b i )

[0148] f t =σ(W xf *X t +W hf *H t-1 +W cf ⊙c t-1 +b f )

[0149] C t =f t ⊙C t-1 +i i ⊙tanh g (W xc *x t +W hc *H t-1 +b c)

[0150] o t = σ(W xo *X t +W h0 *H t-1 +W co ⊙C t +b o )

[0151] H t = o t ⊙tanh g (C t )

[0152] In the formula, σ is the sigmoid activation function, tanh g represents the hyperbolic tangent activation function, X t is the input of the current time step, H t-1 is the hidden state of the last time step, C t-1 is the memory cell of the last time step, f t , i t , C t , o t respectively represent the forget gate, input gate, memory cell and output gate, W and b are model parameters.i t The calculation of i is the value of the input gate, which determines how much new information will be added to the cell state; f t calculates the value of the forget gate, which determines how much old information will be forgotten from the cell state; C t The calculation of is the cell state of the current time step. The cell state is the channel of information flow in the network, which passes information through time steps; o t The calculation of is the value of the output gate. The output gate determines how much information of the cell state will be output to the hidden state. The final prediction result to be output is the hidden state H t . The hidden state is the output of the ConvLSTM network at each time step, which contains the information learned from the input sequence and can be used for subsequent carbon emission prediction tasks.

[0153] In the prediction network, the parameters in the ConvLSTM model are particularly considered, and in the experimental setting, the time steps of all models are uniformly set to 16. In the prediction network, the number of layers is also uniformly set to 2, while the number of hidden layer units is set to 12, 16, 12 and 10 respectively according to the dimensions of different data characteristics, and the number of iterations (epoch) is uniformly set to 5000. In order to ensure the accuracy and reliability of the evaluation, the present application adopts three different evaluation indexes to compare the prediction results, including root mean square error RMSE, mean absolute error MAE and mean absolute percentage error MAPE, to ensure the accuracy of the results.

[0154] Table 5 Comparison of results of each prediction model

[0155]

[0156] The prediction result comparison is shown in Table 5, and the prediction result comparison shows that the prediction model trained by the data generated by TimeTGAN has a significantly higher effect than other models, indicating that the data generated by TimeTGAN model is closer to the original distribution of carbon emission data, effectively expanding the size of the city-scale carbon emission data set, so that effective data analysis and model training can be performed under small sample conditions, not only solving the small sample problem, but also helping to improve the accuracy and reliability of city-scale carbon emission prediction, which has important significance for the formulation of city-scale carbon emission reduction policies.

[0157] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art can easily think of various equivalent modifications or replacements within the technical scope disclosed by the present application, and these modifications or replacements should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A city-scale carbon emission prediction method based on small sample data enhancement, characterized in that, The method comprises the following steps: S1, acquiring city night light data, correcting the data by using a multiple linear regression model to obtain a fitted light proportion; S2, based on the fitted light proportion, calculating the carbon dioxide emissions of each city by using a city night light data inversion calculation formula to obtain a carbon emission dataset; S3, using an improved time series enhancement generative adversarial network model TimeTGAN to perform data enhancement on the carbon emission dataset, effectively capturing long-term dependencies in time series data by introducing a bidirectional time convolution network BiTCN network and a dilated causal convolution, the specific enhancement process being: inputting the carbon emission dataset into an embedding function after encoding, performing feature extraction, data compression and passing to a latent space representation in the embedding function, outputting random data, taking the random data as input of the generator to generate time series data, discriminating true and false through a time series data discriminator, if the discrimination is passed, converting into a usable carbon emission enhanced dataset through a restoration function, if the discrimination is not passed, feeding back to the generator to continue generating new data in the generator until the required preset amount of data is generated, integrating the data into a carbon emission enhanced dataset, thereby obtaining the carbon emission enhanced dataset; S4, training a spatiotemporal sequence prediction network model using the carbon emission enhanced dataset to obtain a trained spatiotemporal sequence prediction model; S5, collecting city spatial geographic location data and inputting the data into the trained spatiotemporal sequence prediction model to obtain a carbon emission prediction result.

2. The small sample data augmentation based urban scale carbon emission prediction method according to claim 1, characterized in that, The multiple linear regression model in S1 includes social indicators closely related to carbon dioxide emissions, including industrial power data proportion of total provincial industrial power data, PM2.5 emission proportion of total provincial emissions, population data proportion of total provincial population data, private car ownership proportion of total provincial ownership, and retail sales data proportion of total provincial sales data.

3. The method of claim 1, wherein, The city night light data inversion calculation formula in S2 is specifically: In the formula, is the city-level carbon emission data; is the corrected proportion of night light, is the provincial carbon emission data.

4. The method of claim 1, wherein, An embedding network function is applied in the embedding network, and the specific formula is: wherein, denote static features after processing by a fully connected layer; denote temporal features after processing by a fully connected layer; is a latent space; is a temporal feature; is an embedded static feature; is an embedded temporal feature.

5. The small sample data augmentation based urban scale carbon emission prediction method according to claim 1, characterized in that, A restoration network function is applied in the restoration network, and the specific formula is: wherein, are restored static features; are restored temporal features, intermediate steps of the restoration network; are final generated static features; are final generated temporal features; are fully connected layers in the restoration network; denotes a temporal convolutional network in the restoration network.

6. The small sample data augmentation based urban scale carbon emission prediction method according to claim 1, characterized in that, The specific formula of the generator is: wherein is a temporal convolutional network in the generator network; is a fully connected layer in the generator network; is a static noise vector, generating static features through the fully connected layer network; is its latent representation; is a temporal noise vector, generating the entire temporal sequence through the network; latent representation.

7. The small sample data augmentation based urban scale carbon emission prediction method according to claim 6, characterized in that, The discriminator uses bidirectional The network, which is specifically formulated as: wherein is the evaluation result of the static feature latent representation; is the evaluation result of the temporal feature latent representation; is the recursive computation formula in the forward hidden state for each time point t in the time series data is the forward propagation result of the time convolution network; is the recursive computation formula in the backward hidden state for each time point t in the time series data is the backward propagation result of the time convolution network; is the hidden state combining the forward and backward propagation results.

8. The small sample data augmentation based urban scale carbon emission prediction method according to claim 1, characterized in that, The spatiotemporal sequence prediction network model in S4 selects a convolutional long short-term memory network model, and the model formula is: wherein, sigmoid is a sigmoid activation function; tanh is a hyperbolic tangent activation function; is the city spatial geographical location data of the current time step; is the prediction result of the previous time step; is the memory cell of the previous time step; , , and are a forget gate, an input gate, a memory cell and an output gate, respectively; , , , , , , and are all model parameters; is the carbon emission prediction result.

9. A city-scale carbon emission prediction system based on small sample data augmentation, characterized in that, The system works according to the city scale carbon emission prediction method based on small sample data enhancement in any one of claims 1-8, and the system comprises a carbon emission calculation module, a carbon emission data enhancement module, a model training module and a carbon emission prediction module; The carbon emission calculation module is used to acquire city night light data, correct the data by using a multiple linear regression model to obtain a fitted light proportion, and calculate the carbon dioxide emissions of each city by using a city night light data inversion calculation formula based on the fitted light proportion to obtain a carbon emission dataset; The carbon emission data enhancement module is used to perform data enhancement on the carbon emission dataset by using an improved time series enhancement generative adversarial network model TimeTGAN to obtain a carbon emission enhanced dataset; The carbon emission data enhancement module is used to perform data enhancement on the carbon emission dataset by using an improved time series enhancement generative adversarial network model TimeTGAN to obtain a carbon emission enhanced dataset; The model training module is configured to train the spatiotemporal sequence prediction network model using a carbon emission enhancement dataset to obtain a trained spatiotemporal sequence prediction model. The carbon emission prediction module is configured to collect city spatial geographic location data, input the city spatial geographic location data into the trained spatiotemporal sequence prediction model, and obtain a carbon emission prediction result.

Images