A system and method for enhanced data-driven integrated prediction of building carbon emissions

Abstract

The application discloses a kind of enhanced data-driven building carbon emission integrated prediction system and method, specifically including in data enhancement module, using Bootstrap resampling method to expand building carbon emission data, and the data after amplification is combined with original data set to construct new enhanced data set;In feature extraction module, the cycle characteristics of building carbon emission time series data are extracted using autoencoder, and the original features are separated to obtain random features to form building carbon emission residual dynamic data;In model prediction module, the prediction model suitable for each other is trained using cycle characteristic data set and residual dynamic data set, the prediction model of these two parts is fused for the prediction of building carbon emission.The application solves the problem of difficult and insufficient building carbon emission data acquisition;Enhance the processing capacity of random features, can improve the prediction accuracy and operation rate of prediction model, and has high precision, strong generalization ability and good scalability.

Description

Technical Field

[0001] This invention relates to the field of building carbon emission prediction technology in the energy industry, specifically to an enhanced data-driven integrated prediction system and method for building carbon emissions. Background Technology

[0002] With the increasing severity of global climate change and environmental problems, carbon emissions from the construction sector have become a key focus of carbon neutrality goals. The construction industry is one of the major contributors to global energy consumption and CO2 emissions, accounting for nearly 40% of global carbon emissions. Scientifically and accurately predicting building carbon emissions is a crucial foundation for promoting low-carbon building design and green building policies. Therefore, achieving carbon reduction targets in the construction industry is of great significance to achieving global carbon neutrality.

[0003] Building carbon emissions primarily originate from the following sources: carbon emissions from the combustion of fuels used in building operation (such as natural gas and coal); carbon emissions from electricity consumed during building operation; and carbon emissions from the production, transportation, construction, operation, and demolition of building materials. These emissions are long-term and complex, especially throughout the building's entire life cycle. Emission characteristics vary significantly across different regions, climates, energy structures, and building types. Data collection and calculation are challenging, and traditional methods cannot fully cover the dynamic characteristics of building carbon emissions.

[0004] Traditional carbon emission prediction methods are mainly based on linear models and statistical analysis, relying on simple calculations of building area, energy consumption data, and carbon emission factors. These methods have the following shortcomings: they lack dynamic response capabilities and cannot effectively capture carbon emission changes during building operation; they lack comprehensive consideration of multiple variables (such as climate, usage patterns, energy prices, etc.), resulting in limited prediction accuracy; they are difficult to adapt to the needs of different regions, building types, or policy change scenarios, resulting in poor model applicability; and they are highly data-dependent, with weak capabilities for large-scale, real-time data analysis. Summary of the Invention

[0005] This invention provides an enhanced data-driven integrated prediction system and method for building carbon emissions, aiming to improve the accuracy and applicability of prediction models and support the low-carbon development of the construction industry.

[0006] This invention provides the following technical solution: an enhanced data-driven integrated prediction method for building carbon emissions, comprising the following steps:

[0007] S1. The Bootstrap resampling method is used to augment the original building carbon emission data to obtain virtual data, and the virtual data is combined with the original data to construct a new augmented dataset.

[0008] S2. Use an autoencoder to extract cyclic features from the building carbon emission enhancement data samples to construct a stable training set for building carbon emissions; separate the features of the original data to obtain random features that form the dynamic data of building carbon emission residuals, and convert the dynamic data of building carbon emission residuals into a training set of building carbon emission residuals; the specific steps are as follows:

[0009] The first step is to read in the building carbon emission enhancement data and perform data standardization processing;

[0010] The second step is to train the autoencoder model, including forward and backward propagation and parameter updates for the encoding and decoding parts. The forward propagation process for the encoding part is shown below:

[0011] ;

[0012] In the formula: These are the cyclic features extracted from the encoding part; and These are the weights and biases of the encoding part, respectively; The activation function is used; the forward propagation process of the decoding part is as follows:

[0013] ;

[0014] In the formula: The decoding part is for loop characteristics The reconstructed features of the eigenvalues; and These are the weights and biases of the decoding part, respectively; For activation functions;

[0015] The third step is to use an autoencoder to extract the cyclic features of building carbon emission data. Construct a stable training set for building carbon emissions, where the cyclical feature is a stable component of the original time series data;

[0016] The fourth step is to remove the stable portion of the original data to extract the dynamic data of building carbon emission residuals, and then convert the dynamic data of building carbon emission residuals into a building carbon emission residual training set.

[0017] S3. Use the building carbon emission stability training set and the building carbon emission residual training set as two datasets to train the prediction model;

[0018] Among them, the building carbon emission stability training set is used to train the cyclic feature model, which is constructed by the ELM model;

[0019] The building carbon emission residual training set is used to train the dynamic residual model, which is composed of a T-LSTM model. The specific training process of the dynamic residual model is as follows:

[0020] The first step is input data preprocessing. The input data of the dynamic residual model is first standardized through the Layer Norm layer, and then the processed data is first fed into the gated MLP for up projection to expand the data dimension and obtain richer data features.

[0021] The second step is to encode the input data. By utilizing the self-attention mechanism in Transformer, the query, key, and value of the input data are calculated to capture global dependencies, identify the correlation between positions in the sequence, and perform nonlinear transformation on the features of each position to enhance the feature representation ability.

[0022] The third step is to perform cell processing for EN-LSTM. EN-LSTM uses multiple gating mechanisms to control the flow of information. The data input to the EN-LSTM cell processing passes through the forget gate and the input gate in sequence to update the cell state in EN-LSTM. Finally, the output gate is used to obtain the input for the next time step or the output for the final result. An exponential gate is added before the input gate and the forget gate to introduce the standard state branch during the cell processing.

[0023] The fourth step is to feed the timing data processed by the LSTM unit into the gated MLP for downprojection.

[0024] The fifth step involves calculating the error between the model output and the actual target, adjusting the model parameters until the model converges; finally, a linear transformation is performed on the results in the output stage to obtain the final output.

[0025] S4. Merge the two prediction models, the cyclic feature model and the dynamic residual model, to form an integrated prediction model;

[0026] S5. Use the enhanced data as the input for the integrated prediction model, use the integrated prediction model to predict building carbon emissions, and output the prediction results.

[0027] Preferably, the cyclic feature model captures periodic changes in the time series, while the dynamic residual model is used to adjust the model prediction error.

[0028] Preferably, the output prediction result is a combination of the prediction results of the cyclic feature model and the dynamic residual model. The outputs of the cyclic feature model and the dynamic residual model are fused together by a weighted average method to construct a complete time series prediction data curve for building carbon emissions.

[0029] Preferably, the process of acquiring virtual data is as follows:

[0030] The first step is to analyze the original samples. Constructing subsamples Randomly draw samples from the original sample and generate multiple subsamples with replacement, set to 1000 times, with the size of each subsample equal to the size of the original sample;

[0031] The second step is to calculate the target statistic or model parameters for each subsample. ;

[0032] The third step is to construct an approximate Bootstrap distribution for the target statistic of each subsample;

[0033] The fourth step is to estimate the uncertainty of the parameters or construct confidence intervals based on the Bootstrap distribution, and then extract the samples within the confidence intervals as virtual data.

[0034] Preferably, the specific steps for training the ELM model are as follows:

[0035] The first step involves random feature mapping from the input layer to the hidden layer, with model weights and biases being passed and updated as follows:

[0036] ;

[0037] In the formula, The output of each neuron in the hidden layer; , These are the model's weights and biases, respectively. For activation functions;

[0038] The second step involves solving for the linear parameters from the hidden layer to the output layer by setting weights for the hidden layers. The specific process for obtaining the optimal solution is as follows:

[0039] ;

[0040] In the formula, For the final output, It is the hidden layer weight allocation.

[0041] Preferably, in step S3, backpropagation and the Adam optimizer are used to optimize parameters and adjust model parameters.

[0042] A prediction system for an enhanced data-driven integrated prediction method for building carbon emissions includes a data augmentation module, a feature extraction module, and a model prediction module.

[0043] The data augmentation module is used to address the problem of insufficient building carbon emission data. The prediction system uses the Bootstrap resampling method in the data augmentation module to augment the original building carbon emission data, and combines the augmented virtual data with the original data to construct a new augmented dataset.

[0044] The feature extraction module is used to extract cyclic features from the enhanced data samples. The feature extraction module uses an autoencoder network model to extract cyclic features from the time series data of building carbon emissions, constructs a stable training set for building carbon emissions, separates the features of the original data, deletes the stable part in the original data, and obtains random feature data to form the dynamic data of building carbon emissions residuals.

[0045] The model prediction module uses the building carbon emission stability training set and building carbon emission residual training set generated by the feature extraction module to train the corresponding prediction model, and then integrates the prediction models to predict building carbon emissions.

[0046] The model prediction module includes a cyclic feature model and a dynamic residual model; the building carbon emission stability training set is used to train the cyclic feature model, which is constructed by an ELM model; the building carbon emission residual training set is used to train the dynamic residual model, which is constructed by a T-LSTM model.

[0047] Preferably, the T-LSTM model integrates gated MLP and EN-LSTM, and incorporates an attention mechanism.

[0048] Preferably, the prediction data input for the integrated prediction model is augmented data.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] 1. This enhanced data-driven integrated prediction system and method for building carbon emissions has high accuracy, strong generalization ability and good scalability, and can be widely used in green building design and energy management.

[0051] 2. This enhanced data-driven integrated prediction system and method for building carbon emissions uses the Bootstrap data resampling method to augment the original data, combining the original data and the augmented data into a new dataset, which greatly solves the problems of difficulty and insufficiency in obtaining building carbon emission data.

[0052] 3. This enhanced data-driven integrated prediction system and method for building carbon emissions divides the time series data of building carbon emissions into cyclic feature data and random feature data through an autoencoder network model. The cyclic feature data and dynamic residual data in the original data are trained using ELM and T-LSTM models. The prediction results of the two models are combined to form the final prediction curve, which not only helps to improve the prediction accuracy of the model, but also helps to improve the computation speed of the model.

[0053] 4. This enhanced data-driven integrated prediction system and method for building carbon emissions designs a dynamic residual prediction model, T-LSTM, which integrates gated MLP and EN-LSTM. By incorporating an attention mechanism, it enhances the ability to process stochastic features, thereby accurately predicting the time series data of dynamic residuals for building carbon emissions. Attached Figure Description

[0054] Figure 1 A schematic diagram of the prediction method of an enhanced data-driven building carbon emission integrated prediction system provided by the present invention;

[0055] Figure 2 This is a diagram of the ELM backbone network architecture of the present invention;

[0056] Figure 3 This is a diagram of the EN-LSTM network structure of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] See Figures 1 to 3 This invention proposes an enhanced data-driven integrated prediction system and method for building carbon emissions, aiming to improve the accuracy and applicability of prediction models and support the low-carbon development of the construction industry. The enhanced data-driven integrated prediction method for building carbon emissions proposed in this invention includes the following steps:

[0059] S1. The Bootstrap resampling method is used to augment the original building carbon emission data to obtain virtual data. The virtual data is then combined with the original data to construct a new augmented dataset, which forms the samples for subsequent prediction model training. The process of obtaining virtual data is as follows:

[0060] The first step is to analyze the original samples. Constructing subsamples Randomly draw samples from the original sample and generate multiple subsamples with replacement, set to 1000 times, with the size of each subsample equal to the size of the original sample;

[0061] The second step is to calculate the target statistic or model parameters for each subsample. ;

[0062] The third step is to construct an approximate Bootstrap distribution for the target statistic of each subsample;

[0063] The fourth step is to estimate the uncertainty of the parameters or construct confidence intervals based on the Bootstrap distribution, and then extract the samples within the confidence intervals as virtual data.

[0064] S2. Use an autoencoder to extract cyclic features from the building carbon emission enhancement data samples to construct a stable training set for building carbon emissions; separate the features of the original data to obtain random features that form the dynamic data of building carbon emission residuals, and convert the dynamic data of building carbon emission residuals into a training set of building carbon emission residuals; the specific steps are as follows:

[0065] The first step is to read in the building carbon emission enhancement data and perform data standardization processing;

[0066] The second step is to train the autoencoder model, including forward and backward propagation and parameter updates for the encoding and decoding parts. The forward propagation process for the encoding part is shown below:

[0067] ;

[0068] In the formula: These are the cyclic features extracted from the encoding part; and These are the weights and biases of the encoding part, respectively; The activation function is used; the forward propagation process of the decoding part is as follows:

[0069] ;

[0070] In the formula: The decoding part is for loop characteristics The reconstructed features of the eigenvalues; and These are the weights and biases of the decoding part, respectively; For activation functions;

[0071] The third step is to use an autoencoder to extract the cyclic features of building carbon emission data. Construct a stable training set for building carbon emissions, where the cyclical feature is a stable component of the original time series data;

[0072] The fourth step is to remove the stable portion of the original data to extract the dynamic data of building carbon emission residuals, and then convert the dynamic data of building carbon emission residuals into a building carbon emission residual training set.

[0073] S3. Train the prediction model using two datasets: the stable training set and the residual training set for building carbon emissions. The stable training set is used for training the EML model, while the residual training set, due to its higher randomness, is used for the more robust Transformer model. This includes the following steps:

[0074] The building carbon emission stability training set is used to train the cyclic feature model, which is constructed using an ELM model; such as Figure 2 As shown, the specific steps for training the ELM model are as follows:

[0075] The first step involves random feature mapping from the input layer to the hidden layer, with model weights and biases being passed and updated as follows:

[0076] ;

[0077] In the formula, The output of each neuron in the hidden layer; , These are the model's weights and biases, respectively. For activation functions;

[0078] The second step involves solving for the linear parameters from the hidden layer to the output layer by setting weights for the hidden layers. The specific process for obtaining the optimal solution is as follows:

[0079] ;

[0080] In the formula, For the final output, It is the hidden layer weight allocation;

[0081] The building carbon emission residual training set is used to train the dynamic residual model, which is composed of a T-LSTM model. The specific training process of the dynamic residual model is as follows:

[0082] The first step is input data preprocessing. The input data of the dynamic residual model is first standardized through the Layer Norm layer, and then the processed data is first fed into the gated MLP for up projection to expand the data dimension and obtain richer data features.

[0083] The second step is to encode the input data. By utilizing the self-attention mechanism in Transformer, the query, key, and value of the input data are calculated to capture global dependencies, identify the correlation between positions in the sequence, and perform nonlinear transformation on the features of each position to enhance the feature representation ability.

[0084] The third step is to perform EN-LSTM cell processing. EN-LSTM introduces multiple gating mechanisms to control the flow of information, thereby effectively capturing long-term data dependencies. The data input to the EN-LSTM cell processing passes through the forget gate and the input gate in sequence to update the cell state in EN-LSTM. Finally, the output gate is used to obtain the input for the next time step or the output for the final result. In order to solve the strong randomness in the residual dynamic data, exponential gates are added before the input gate and the forget gate respectively. In order to enhance the stability of the model, a standard state branch is introduced during the cell processing.

[0085] The fourth step is to feed the timing data processed by the LSTM unit into the gated MLP for downprojection, which not only aligns the data format but also filters out noise and interference in the output data.

[0086] The fifth step involves calculating the error between the model output and the actual target, adjusting the model parameters, and optimizing these parameters using backpropagation and the Adam optimizer until the model converges. Finally, a linear transformation is performed on the results in the output stage to obtain the final output. Figure 3 As shown.

[0087] Cyclic feature models capture periodic changes in time series, while dynamic residual models are used to adjust model prediction errors, thereby improving the overall accuracy and stability of predictions.

[0088] S4. Merge the two prediction models, the cyclic feature model and the dynamic residual model, to form an integrated prediction model;

[0089] S5. Use the enhanced data as the input for the integrated prediction model, and use the integrated prediction model to predict building carbon emissions, outputting the prediction results. The output prediction results are a combination of the prediction results of the cyclic feature model and the dynamic residual model. The outputs of the cyclic feature model and the dynamic residual model are fused together using a weighted average method to construct a complete time-series prediction data curve for building carbon emissions.

[0090] As can be seen from the above description, when using this invention, the Bootstrap data resampling method is used to amplify the original data, and the original data and the amplified data are combined to form a new dataset, which greatly solves the problem of difficulty and insufficiency in obtaining building carbon emission data.

[0091] This invention combines the prediction results of two prediction models to form the final prediction curve, which not only helps to improve the prediction accuracy of the model, but also helps to improve the model's computing speed.

[0092] This invention designs a dynamic residual prediction model T-LSTM, which integrates gated MLP and EN-LSTM. By combining an attention mechanism, it enhances the ability to process random features, so as to accurately predict the time series data of dynamic residuals of building carbon emissions.

[0093] The above prediction method is implemented through a prediction system, such as... Figure 1 As shown, the prediction system used in an enhanced data-driven integrated prediction method for building carbon emissions adopts a modular architecture, including a data augmentation module, a feature extraction module, and a model prediction module. The data augmentation module, feature extraction module, and model prediction module operate independently, which makes the prediction system highly scalable and improves its adaptability.

[0094] The data augmentation module is used to address the problem of insufficient building carbon emission data. The prediction system uses the Bootstrap resampling method in the data augmentation module to augment the original building carbon emission data, and combines the augmented virtual data with the original data to construct a new augmented dataset for subsequent training of the prediction model.

[0095] The feature extraction module is used to extract cyclic features from the augmented data samples. The feature extraction module uses an autoencoder network model to extract cyclic features from the time series data of building carbon emissions, constructs a stable training set for building carbon emissions, and decomposes the features of the original data, deleting the stable part in the original data and obtaining random feature data to form the dynamic residual data of building carbon emissions. The feature extraction module includes the extraction of time series cyclic features from the autoencoder and the acquisition of the dynamic residual data, and constructs these two parts of data into a dataset that can be used for subsequent training.

[0096] The model prediction module uses the building carbon emission stability training set and building carbon emission residual training set generated by the feature extraction module to train the corresponding prediction model. Since the randomness of the two data features is different, different prediction models are selected and the prediction models are fused and integrated into a prediction model. The integrated prediction model is used to predict building carbon emissions, and the prediction data input of the integrated prediction model is the augmented data.

[0097] The model prediction module includes a cyclic feature model and a dynamic residual model. The stable training set for building carbon emissions is used to train the cyclic feature model, which is constructed using an ELM model. The residual training set for building carbon emissions is used to train the dynamic residual model, which is constructed using a T-LSTM model. The T-LSTM model integrates gated MLP and EN-LSTM and incorporates an attention mechanism to enhance the ability to process random features, thereby accurately predicting the time series data of dynamic residuals for building carbon emissions.

[0098] In summary, this method has high accuracy, strong generalization ability and good scalability, and can be widely used in fields such as green building design and energy management.

[0099] The contents not described in detail in this specification are prior art known to those skilled in the art. Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An enhanced data-driven integrated prediction method for building carbon emissions, characterized in that, Includes the following steps: S1. The Bootstrap resampling method is used to augment the original building carbon emission data to obtain virtual data, and the virtual data is combined with the original data to construct a new augmented dataset. S2. Use an autoencoder to extract cyclic features from the building carbon emission enhancement data samples to construct a stable training set for building carbon emissions; separate the features of the original data to obtain random features that form the dynamic data of building carbon emission residuals, and convert the dynamic data of building carbon emission residuals into a training set of building carbon emission residuals; the specific steps are as follows: The first step is to read in the building carbon emission enhancement data and perform data standardization processing; The second step is to train the autoencoder model, including forward and backward propagation and parameter updates for the encoding and decoding parts. The forward propagation process for the encoding part is shown below: ； In the formula: This refers to a sample of data on building carbon emission enhancement. These are the cyclic features extracted from the encoding part; and These are the weights and biases of the encoding part, respectively; The activation function is used; the forward propagation process of the decoding part is as follows: ； In the formula: The decoding part addresses the loop feature. The reconstructed features performed; and These are the weights and biases of the decoding part, respectively; For activation functions; The third step is to use an autoencoder to extract the cyclic features of building carbon emission data. Construct a stable training set for building carbon emissions, where the cyclical feature is a stable component of the original time series data; The fourth step is to remove the stable portion of the original data to extract the dynamic data of building carbon emission residuals, and then convert the dynamic data of building carbon emission residuals into a building carbon emission residual training set. S3. Use the building carbon emission stability training set and the building carbon emission residual training set as two datasets to train the prediction model; Among them, the building carbon emission stability training set is used to train the cyclic feature model, which is constructed by the ELM model; The building carbon emission residual training set is used to train the dynamic residual model, which is composed of a T-LSTM model. The specific training process of the dynamic residual model is as follows: The first step is input data preprocessing. The input data of the dynamic residual model is first standardized through the Layer Norm layer, and then the processed data is first fed into the gated MLP for up projection to expand the data dimension and obtain richer data features. The second step is to encode the input data. By utilizing the self-attention mechanism in Transformer, the query, key, and value of the input data are calculated to capture global dependencies, identify the correlation between positions in the sequence, and perform nonlinear transformation on the features of each position to enhance the feature representation ability. The third step is to perform cell processing for EN-LSTM. EN-LSTM uses multiple gating mechanisms to control the flow of information. The data input to the EN-LSTM cell processing passes through the forget gate and the input gate in sequence to update the cell state in EN-LSTM. Finally, the output gate is used to obtain the input for the next time step or the output for the final result. An exponential gate is added before the input gate and the forget gate to introduce the standard state branch during the cell processing. The fourth step is to feed the timing data processed by the LSTM unit into the gated MLP for downprojection. The fifth step involves calculating the error between the model output and the actual target, adjusting the model parameters until the model converges; finally, a linear transformation is performed on the results in the output stage to obtain the final output. S4. Merge the two prediction models, the cyclic feature model and the dynamic residual model, to form an integrated prediction model; S5. Use the enhanced data as the input for the integrated prediction model, use the integrated prediction model to predict building carbon emissions, and output the prediction results.

2. The enhanced data-driven integrated prediction method for building carbon emissions according to claim 1, characterized in that: The cyclic feature model captures periodic changes in the time series, while the dynamic residual model is used to adjust the model's prediction error.

3. The enhanced data-driven integrated prediction method for building carbon emissions according to claim 2, characterized in that: The output prediction result is a combination of the prediction results of the cyclic feature model and the dynamic residual model. The outputs of the cyclic feature model and the dynamic residual model are fused together by a weighted average method to construct a complete time series prediction data curve for building carbon emissions.

4. The enhanced data-driven integrated prediction method for building carbon emissions according to claim 1, characterized in that: The process of acquiring virtual data is as follows: The first step is to analyze the original samples. Constructing subsamples Samples are randomly drawn from the original sample, and multiple subsamples are generated with replacement, for a total of 1000 times. The size of each subsample is equal to the size of the original sample. Refers to the original sample The nth data sample in; It refers to the i-th subsample The nth data sample in the dataset; The second step is to calculate the target statistic or model parameters for each subsample. ; The third step is to construct an approximate Bootstrap distribution for the target statistic of each subsample; The fourth step is to estimate the uncertainty of the parameters or construct confidence intervals based on the Bootstrap distribution, and then extract the samples within the confidence intervals as virtual data.

5. The enhanced data-driven integrated prediction method for building carbon emissions according to claim 4, characterized in that: The specific steps for training the ELM model are as follows: The first step involves random feature mapping from the input layer to the hidden layer, with model weights and biases being passed and updated as follows: ； In the formula, This refers to a sample of data on building carbon emission enhancement. The output of each neuron in the hidden layer; , These are the model's weights and biases, respectively. For activation functions; The second step involves solving for the linear parameters from the hidden layer to the output layer by setting weights for the hidden layers. The specific process for obtaining the optimal solution is as follows: ； In the formula, This refers to a sample of data on building carbon emission enhancement. For the final output, It is the hidden layer weight allocation.

6. The enhanced data-driven integrated prediction method for building carbon emissions according to claim 1, characterized in that: In S3, backpropagation and the Adam optimizer are used to optimize parameters and adjust model parameters.

7. The prediction system used in the enhanced data-driven integrated prediction method for building carbon emissions according to any one of claims 1-6, characterized in that: It includes a data augmentation module, a feature extraction module, and a model prediction module; The data augmentation module is used to address the problem of insufficient building carbon emission data. The prediction system uses the Bootstrap resampling method in the data augmentation module to amplify the original building carbon emission data, and combines the amplified virtual data with the original data to construct a new augmented dataset. The feature extraction module is used to extract cyclic features from the enhanced data samples. The feature extraction module uses an autoencoder network model to extract cyclic features from the time series data of building carbon emissions, constructs a stable training set for building carbon emissions, separates the features of the original data, deletes the stable part in the original data, and obtains random feature data to form the dynamic data of building carbon emissions residuals. The model prediction module uses the building carbon emission stability training set and building carbon emission residual training set generated by the feature extraction module to train the corresponding prediction model, and then integrates the prediction models to predict building carbon emissions. The model prediction module includes a cyclic feature model and a dynamic residual model; the building carbon emission stability training set is used to train the cyclic feature model, which is constructed by an ELM model; the building carbon emission residual training set is used to train the dynamic residual model, which is constructed by a T-LSTM model.

8. The prediction system used in the enhanced data-driven integrated prediction method for building carbon emissions according to claim 7, characterized in that: The T-LSTM model integrates gated MLP and EN-LSTM, and incorporates an attention mechanism.

9. The prediction system used in the enhanced data-driven integrated prediction method for building carbon emissions according to claim 7, characterized in that: The prediction data input for the integrated prediction model is augmented data.

Images