A training method for a carbon emission prediction model

Abstract

The present invention relates to a training method for a carbon emission prediction model, belonging to the technical field of carbon emission prediction, and solves the problem of low prediction accuracy in the prior art. The method includes: obtaining the historical carbon emissions of a construction unit and the corresponding carbon emission-related factor data; constructing a training sample set based on the historical carbon emissions and the corresponding carbon emission-related factor data; constructing a multi-step time series prediction model for predicting the carbon emissions in multiple future time steps; the multi-step time series prediction model includes: a static feature extraction module for extracting static features based on static data; a time series feature extraction module for extracting time series features based on multi-dimensional time series data; a feature fusion module for fusing the static features and the time series features; a prediction module for predicting the carbon emissions in multiple future time steps based on the fused features; training the multi-step time series prediction model based on the training sample set to obtain a carbon emission prediction model. Accurate carbon emission prediction of construction units is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of carbon emission prediction, and particularly to a method for training a carbon emission prediction model. Background Art

[0002] Under the background of global response to climate change, the accurate prediction and effective control of carbon emissions have become the focus of attention of various countries and regions. Carbon emissions are one of the main factors leading to global warming, and have a profound impact on the sustainable development of the environment, ecosystem and human society. The construction industry is one of the important sources of carbon emissions, accounting for a relatively large proportion of the total global carbon emissions. This is mainly because a large amount of greenhouse gas emissions will be generated in the links such as material production during the construction process of buildings and energy consumption (such as heating, cooling, lighting, etc.) during the operation of buildings.

[0003] With the development of machine learning, deep learning, and reinforcement learning technologies, it has become possible to use intelligent algorithms to train models to predict carbon emissions. In existing carbon emission prediction models, there are certain limitations in the prediction models for construction units. Many general carbon emission prediction models do not fully consider the unique characteristics of construction units, such as the influence of factors such as the structural type, usage function, and geographical location of buildings on carbon emissions. Some existing prediction methods may only focus on a single type of data, such as only considering energy consumption data, while ignoring other factors related to carbon emissions. For construction units, in addition to energy consumption data (which is a time series data), there are also many static data, such as the area of buildings, the type of building materials, etc. These data are also of great significance when predicting carbon emissions. However, existing models often cannot make good use of these static data and time series data comprehensively, so the prediction accuracy is not high. Summary of the Invention

[0004] In view of the above analysis, an embodiment of the present invention aims to provide a method for training a carbon emission prediction model to solve the problem of low prediction accuracy of existing models.

[0005] On the one hand, an embodiment of the present invention provides a method for training a carbon emission prediction model, including the following steps:

[0006] Obtain the historical carbon emissions of a construction unit and the corresponding carbon emission-related factor data; the carbon emission-related factor data includes static data and multi-dimensional time series data;

[0007] Construct a training sample set based on the historical carbon emissions of the construction unit and the corresponding carbon emission-related factor data;

[0008] Construct a multi-step time series prediction model for predicting carbon emissions at multiple future time steps. The multi-step time series prediction model includes: a static feature extraction module for extracting static features based on the static data; a time series feature extraction module for extracting time series features based on the multi-dimensional time series data; a feature fusion module for fusing the static features and the time series features; and a prediction module for predicting carbon emissions at multiple future time steps based on the fused features.

[0009] Train the multi-step time series prediction model based on a training sample set to obtain a carbon emission prediction model.

[0010] Based on a further improvement of the above method, the following formula is used to calculate the training loss of the multi-step time series prediction model:

[0011] Loss = Loss mse + Loss shape

[0012] Loss shape = β1Loss1 + β2Loss2 + β3loss3

[0013] Where Loss mse represents the mean squared error loss, Loss shape represents the shape loss, Loss1 represents the distance loss between the label sequence and the prediction sequence, Loss2 represents the principal component loss between the label sequence and the prediction sequence, Loss3 represents the correlation loss between the label sequence and the prediction sequence, and β1, β2, and β3 represent weights.

[0014] Based on a further improvement of the above method, the following formula is used to calculate the distance loss between the label sequence and the prediction sequence:

[0015]

[0016] Where N represents the number of samples in the current training batch, T′ represents the length of the label sequence, y it represents the value of the t-th element in the label sequence of the i-th sample, represents the value of the t-th element in the prediction sequence of the i-th sample, y ij represents the value of the j-th element in the label sequence of the i-th sample, represents the value of the j-th element in the prediction sequence of the i-th sample, and d(·,·) represents the signed distance function.

[0017] Based on a further improvement of the above method, the following formula is used to calculate the principal component loss between the label sequence and the prediction sequence:

[0018]

[0019] Among them, N represents the number of samples in the current training batch, and n mf represents the number of principal components, and n f represents the number of all components, and F j (y i ) represents taking the j-th principal component of y i , represents taking the j-th principal component of represents taking the k-th non-principal component of p ||·|| represents the L-p norm, and y i represents the label sequence of the i-th sample, represents the prediction sequence of the i-th sample.

[0020] Based on the further improvement of the above method, the following formula is used to calculate the correlation loss between the label sequence and the prediction sequence:

[0021]

[0022] Among them, y i represents the label sequence of the i-th sample, represents the prediction sequence of the i-th sample, R(·,·) represents the cross-correlation function, and ||·|| p represents the L-p norm, and N represents the number of samples in the current training batch.

[0023] Based on the further improvement of the above method, a training sample set is constructed based on the historical carbon emissions of construction units and the corresponding carbon emission-related factor data, including:

[0024] For each construction unit, two sliding windows with a length of T slide in parallel on the multi-dimensional time series data and the corresponding carbon emission time series data;

[0025] The multi-dimensional time series data, carbon emission time series data, and static data of the construction unit obtained at each step of sliding are used as the input data of the sample, and the carbon emission time series data of multiple time steps after the sliding window on the carbon emission time series data are used as the labels of the sample, constituting a sample.

[0026] Based on the further improvement of the above method, the time series feature extraction module includes:

[0027] A periodic feature extraction module for extracting the periodic features of the time series data based on the multi-dimensional time series data of the sample;

[0028] A deep feature extraction module for extracting the deep features of each time step based on the periodic features using a recurrent unit;

[0029] A feature weighting module, which is used to perform weighted fusion on the depth features at each time step based on the carbon emissions at each time step to obtain temporal features.

[0030] Based on a further improvement of the above method, the periodic feature extraction module extracts periodic features from the temporal data in the following manner:

[0031] For the temporal data of each dimension, calculate K periodic components of the temporal data of this dimension, and convert the temporal data of this dimension into K two-dimensional data based on the K periodic components;

[0032] Perform feature extraction on each two-dimensional data based on a two-dimensional convolutional network to obtain K two-dimensional features;

[0033] Convert each two-dimensional feature into a one-dimensional feature; perform weighted fusion on the K one-dimensional features based on the weights of the K periodic components to obtain the periodic features corresponding to the temporal data of each dimension.

[0034] Based on a further improvement of the above method, the following formula is used to calculate the periodic features of the temporal data of this dimension:

[0035]

[0036] where f ct represents the periodic feature value at the t-th time step of the c-th dimension of the temporal data, A ck represents the amplitude corresponding to the k-th periodic component of the c-th dimension of the temporal data, represents the feature value at the t-th time step of the one-dimensional feature corresponding to the k-th periodic component of the c-th dimension of the temporal data.

[0037] Based on a further improvement of the above method, the following formula is used to calculate the weight of the depth feature at each time step:

[0038] w t =αy t +(1 - α)w t-1

[0039] where w t-1 represents the weight of the depth feature at the (t - 1)-th time step, w t represents the weight of the depth feature at the t-th time step, y t represents the carbon emissions at the t-th time step, and α represents the smoothing factor.

[0040] Compared with the prior art, the present invention constructs a training sample set by obtaining the historical carbon emissions of park construction units and the corresponding carbon emission-related factor data. By constructing a multi-step time series prediction model, static data features and multi-dimensional time series data features are extracted, and the static data and multi-dimensional time series data are fused to make full use of different types of data information, so as to predict emissions more comprehensively and accurately, achieving accurate carbon emission prediction for construction units. Different from only predicting the carbon emissions at one time point in the past, the present invention predicts the carbon emissions at multiple future time steps, thereby providing support for the long-term carbon emission management of construction units.

[0041] In the present invention, the above technical solutions can also be combined with each other to achieve more preferred combination schemes. Other features and advantages of the present invention will be described in the subsequent specification, and some advantages can be made obvious from the specification, or understood by implementing the present invention. The objectives and other advantages of the present invention can be realized and obtained from the content specifically pointed out in the specification and the drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] The drawings are only for the purpose of showing specific embodiments and are not considered as limiting the present invention. Throughout the drawings, the same reference signs denote the same components;

[0043] Figure 1 It is a flowchart of the training method of the carbon emission prediction model according to the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] The following will specifically describe the preferred embodiments of the present invention with reference to the drawings, wherein the drawings form a part of the present application and are used together with the embodiments of the present invention to explain the principles of the present invention, rather than to limit the scope of the present invention.

[0045] A specific embodiment of the present invention discloses a training method for a carbon emission prediction model, as Figure 1 shown, including the following steps:

[0046] S1. Obtain the historical carbon emissions of park construction units and the corresponding carbon emission-related factor data; the carbon emission-related factor data includes static data and multi-dimensional time series data;

[0047] S2. Construct a training sample set based on the historical carbon emissions of the construction unit and the corresponding carbon emission-related factor data;

[0048] S3. Construct a multi-step time series prediction model for predicting carbon emissions in multiple future time steps. The multi-step time series prediction model includes: a static feature extraction module for extracting static features based on the static data; a time series feature extraction module for extracting time series features based on the multi-dimensional time series data; a feature fusion module for fusing the static features and the time series features; and a prediction module for predicting carbon emissions in multiple future time steps based on the fused features.

[0049] S4. Train the multi-step time series prediction model based on the training sample set to obtain a carbon emission prediction model.

[0050] It should be noted that the building units in the park include industrial buildings, office buildings, commercial buildings, residential buildings, cultural and sports buildings, and administrative management buildings.

[0051] The static data in the carbon emission related factor data includes built environment data; the multi-dimensional time series data includes energy consumption time series data and socioeconomic time series data.

[0052] The built environment data of the building unit includes the stage of the whole life cycle in which the building unit is located, land area, floor area ratio, building type, land use nature, land use mix, building area, building shape coefficient, building height, and building orientation. The energy consumption data includes the consumption data of water, electricity, gas, and chemical fuels. The socioeconomic data includes population, economy, product output, etc. Both energy consumption and socioeconomic are multi-dimensional data.

[0053] The whole life cycle includes three stages: materialization, operation, and disposal. It should be noted that within a unit of time, a building unit can only be in one stage.

[0054] Compared with the prior art, the training method of the carbon emission prediction model provided by the present invention constructs a training sample set by obtaining the historical carbon emissions of the park building units and the corresponding carbon emission related factor data, constructs a multi-step time series prediction model, extracts static data features and multi-dimensional time series data features, and fuses the static data and the multi-dimensional time series data, making full use of different types of data information to more comprehensively and accurately predict emissions, achieving accurate carbon emission prediction of building units. Different from predicting carbon emissions at only one time point in the past, the present invention predicts carbon emissions in multiple future time steps, thus providing support for the long-term carbon emission management of building units.

[0055] The historical carbon emissions of each building unit are the carbon emissions calculated within each unit of historical time.

[0056] After obtaining the historical carbon emissions of construction units and the corresponding carbon emission-related factor data, a training sample set is constructed based on the historical carbon emissions of construction units and the corresponding carbon emission-related factor data, specifically including:

[0057] For each construction unit, two sliding windows with a length of T slide in parallel on the multi-dimensional time series data and the corresponding carbon emission time series data;

[0058] The multi-dimensional time series data, carbon emission time series data, and static data of the construction unit obtained at each step of sliding are used as the input data of the sample, and the carbon emission time series data of multiple time steps after the sliding window on the carbon emission time series data is used as the label of the sample, forming a sample.

[0059] During implementation, the built environment data of the construction unit is numerical data, including energy consumption, social economy, and carbon emission time series data, that is, the time series data composed of energy consumption, social economy, and carbon emissions at each time step in the historical time. Therefore, when constructing samples, sliding windows are used to extract time series data to construct samples.

[0060] During implementation, it is assumed that the dimension of the multi-dimensional time series data (energy consumption time series data and social economy time series data) in the carbon emission-related factor data is C. For each construction unit that has obtained the historical carbon emissions and the corresponding carbon emission-related factor data, two sliding windows with a length of T slide in parallel on the multi-dimensional time series data and the corresponding carbon emission time series data respectively. The sliding step size can be set to one time step or multiple time steps, and it can be set according to the length of the time series data and the calculation accuracy requirements.

[0061] The carbon emissions in the past time may affect the carbon emissions in the future, and the carbon emissions in the past time steps have certain reference significance for the carbon emissions in the future time. Therefore, the current carbon emission data is also used as the input data of the sample. Therefore, for each step of sliding, the multi-dimensional time series data and carbon emission time series data within the current window constitute the multi-dimensional time series data of the sample, that is, the multi-dimensional time series data of the sample includes energy consumption time series data, social economy time series data, and the corresponding carbon emission time series data. The multi-dimensional time series data and the built environment data of the construction unit are used as the input data of the sample; the carbon emission time series data of multiple time steps after the current sliding window on the carbon emission time series data is used as the label of the sample, thereby obtaining a sample, and then constructing a training sample set.

[0062] Construct a multi-step time series prediction model for predicting the carbon emissions in multiple future time steps.

[0063] Specifically, the constructed multi-step time series prediction model includes:

[0064] A static feature extraction module for extracting static features based on the static data;

[0065] A timing feature extraction module for extracting timing features based on the multi-dimensional timing data;

[0066] A feature fusion module for fusing static features and timing features;

[0067] A prediction module for predicting carbon emissions at multiple future time steps based on the fused features;

[0068] During implementation, since the built environment data is numerical data and the multi-dimensional timing data is timing data, it is necessary to extract features separately.

[0069] Therefore, the constructed timing prediction network model includes a static extraction module and a timing feature extraction module to extract the features of the built environment data, energy consumption timing data, and social and economic timing data, as well as the features of the corresponding carbon emissions respectively.

[0070] During implementation, the static feature extraction module can adopt a one-dimensional convolutional network structure with a convolution kernel size of 1.

[0071] The timing feature extraction module is used to extract the features of the timing data. Specifically, the timing feature extraction module includes:

[0072] A periodic feature extraction module for extracting the periodic features of the timing data based on the multi-dimensional timing data of the sample;

[0073] A deep feature extraction module for extracting the deep features of each time step based on the periodic features using a recurrent unit;

[0074] A feature weighting module for weighted fusion of the deep features of each time step based on the carbon emissions of each time step to obtain timing features.

[0075] During implementation, since the timing data is usually a superposition of multiple processes and has multi-period attributes, the periodic features of the multi-dimensional timing data are extracted through the periodic feature extraction module to more fully mine the features and improve the prediction accuracy.

[0076] Specifically, the period feature extraction module extracts the periodic features of the timing data in the following way:

[0077] For the timing data of each dimension, calculate the K periodic components of the timing data of this dimension, and convert the timing data of this dimension into K two-dimensional data based on the K periodic components;

[0078] Extract features from each two-dimensional data based on a two-dimensional convolutional network to obtain K two-dimensional features;

[0079] Convert each two-dimensional feature into a one-dimensional feature; perform weighted fusion on the K one-dimensional features based on the weights of the K periodic components to obtain the periodic features corresponding to the time-series data of each dimension.

[0080] During implementation, perform a fast Fourier transform on the time-series data of each dimension to obtain the period of the time-series data of each dimension and the amplitude of each period.

[0081] Select the K periods with larger amplitudes as the K periodic components of the time-series data of this dimension.

[0082] During implementation, calculate the frequency components of the time-series data of each dimension in the multi-dimensional data of the sample through a fast Fourier transform, select the main K frequency components according to the amplitude sizes of the frequency components, and obtain the K periodic components according to the sampling frequency of the time-series data and the conversion relationship between frequency and period. That is, select the periods corresponding to the first K frequencies with larger amplitudes as the K periodic components of this dimension. During implementation, the size of K is determined according to the length of the time-series data and the calculation accuracy requirements.

[0083] During implementation, after obtaining the K periodic components of the time-series data of each dimension, convert the time-series data of this dimension into two-dimensional data based on each periodic component; in the two-dimensional data, the columns and rows of the two-dimensional data respectively reflect the time-series changes within and between periods.

[0084] It should be noted that if the length of the time-series data is not a multiple of a certain periodic component, zeros are filled after the two-dimensional data corresponding to this periodic component to make the two-dimensional data structure complete. When converting the two-dimensional feature into a one-dimensional feature later, first delete the filled bits and then convert it into a one-dimensional feature.

[0085] Then, based on a two-dimensional convolutional network, perform feature extraction on each two-dimensional data to obtain K two-dimensional features. During implementation, the two-dimensional convolutional network can adopt an existing network, such as using an Inception network for two-dimensional feature extraction. The size of the extracted two-dimensional feature is the same as that of the input two-dimensional data.

[0086] After extracting the two-dimensional feature, convert the two-dimensional feature into a one-dimensional feature, and the obtained one-dimensional feature has the same length as the time-series data of this dimension.

[0087] Since there are K periodic components in the time-series data of this dimension, that is, K one-dimensional features are obtained, and the contributions of each one-dimensional feature to the data are different. Therefore, perform weighted fusion on the K one-dimensional features based on the weights of the K periodic components to obtain the periodic features corresponding to the time-series data of each dimension.

[0088] During implementation, use the amplitude corresponding to each periodic component as the weight of this periodic component. The amplitude of each periodic component, that is, the amplitude of the frequency component corresponding to each periodic component, that is, the amplitude of each frequency obtained by the fast Fourier transform.

[0089] During implementation, the following formula is used to calculate the periodic characteristics of the time-series data of this dimension:

[0090]

[0091] where f ct represents the periodic characteristic value at the t-th time step of the c-th dimensional time-series data, and A ck represents the amplitude corresponding to the k-th periodic component of the c-th dimensional time-series data, represents the characteristic value at the t-th time step of the one-dimensional characteristic corresponding to the k-th periodic component of the c-th dimensional time-series data.

[0092] For example, the size of the multi-dimensional time-series data of a sample is T×C, where T is the number of time steps and C is the dimension. The periodic feature extraction module extracts features for each dimension separately, and the obtained periodic features are still time-series data, that is, still of size T×C.

[0093] After obtaining the periodic features, deep learning is performed to extract deep features. During implementation, a recurrent unit is used to extract the deep features of each time step. That is, the periodic features obtained at each time step are used as the input for one time step, and a recurrent unit is used to extract the deep features.

[0094] During implementation, the recurrent unit can adopt a multi-layer gated recurrent unit. Each layer of the gated recurrent unit is used to extract hidden features based on the input data; the input of the first layer of the gated recurrent unit is the periodic feature of each time step; the input of each layer of the gated recurrent unit except the first layer includes the periodic feature of each time step and the hidden feature of the previous time step; the deep features of each time step output by the last layer of the gated recurrent unit.

[0095] Since the influence of each time step on future moments is different, the deep features of each time step are weighted and fused based on the carbon emissions of each time step to obtain time-series features, making the extracted features more conducive to predicting carbon emissions at future moments.

[0096] During implementation, the following formula is used to calculate the weight of the deep features of each time step:

[0097] w t =αy t +(1 - α)w t-1

[0098] where w t-1 represents the weight of the deep features at the (t - 1)-th time step, w t represents the weight of the deep features at the t-th time step, y t represents the carbon emissions at the t-th time step, and α represents the smoothing factor.

[0099] The smoothing factor determines the degree of dependence on past time steps. α ∈ [0, 1]. When α is large, the newer time steps will have a greater impact. When α is small, the earlier time steps will have a greater impact. The weight of the first time step can be preset.

[0100] During implementation, the feature fusion module concatenates the static features and the temporal features to obtain the fused features. The prediction module performs carbon emission prediction based on the fused features. The prediction module can use a fully connected layer to output the predicted values for multiple time steps.

[0101] During implementation, the following formula is used to calculate the training loss of the multi-step temporal prediction model:

[0102] Loss = Loss mse + Loss shape

[0103] where Loss mse represents the mean squared error loss, and Loss shape represents the shape loss.

[0104] Specifically, the following formula is used to calculate the shape loss:

[0105] Loss shape = β1Loss1 + β2Loss2 + β3Loss3

[0106] where Loss1 represents the distance loss between the label sequence and the prediction sequence, Loss2 represents the principal component loss between the label sequence and the prediction sequence, Loss3 represents the correlation loss between the label sequence and the prediction sequence, and β1, β2, and β3 represent weights.

[0107] By considering the mean squared error loss and the shape loss, it can more effectively guide the optimization direction of the model and improve the prediction accuracy.

[0108] During implementation, if the shapes of the prediction sequence and the label sequence are the same, then the fluctuations of the distance differences at each time step are not significant. Therefore, the following formula is used to calculate the distance loss between the label sequence and the prediction sequence:

[0109]

[0110] where N represents the number of samples in the current training batch, T′ represents the length of the label sequence, y it represents the value of the t-th element in the label sequence of the i-th sample, represents the value of the t-th element in the prediction sequence of the i-th sample, y ij represents the value of the j-th element in the label sequence of the i-th sample, represents the value of the j-th element in the predicted sequence of the i-th sample, and d(·, ·) represents the signed distance function.

[0111] If the predicted sequence has the same shape as the label sequence and the principal component differences between the predicted sequence and the label sequence are not significant, then the following formula is used to calculate the principal component loss between the label sequence and the predicted sequence:

[0112]

[0113] where N represents the number of samples in the current training batch, n mf represents the number of principal components, n f represents the number of all components, F j (y i ) represents taking the j-th principal component of y i , represents taking 's j-th principal component, represents taking 's k-th non-principal component, ||·|| p represents the L-p norm, y i represents the label sequence of the i-th sample, represents the predicted sequence of the i-th sample.

[0114] During implementation, the components of the sequence are obtained by performing a fast Fourier transform on the sequence to obtain the frequencies of the sequence. Among them, the frequencies with larger amplitudes are the principal components, and the others are non-principal components. The number of principal components can be determined according to the length of the sequence.

[0115] If the predicted sequence has the same shape as the label sequence, then the difference between the correlation coefficient of the predicted sequence and the label sequence and the autocorrelation of the label sequence is very small. Therefore, the following formula is used to calculate the correlation loss between the label sequence and the predicted sequence:

[0116]

[0117] where y i represents the label sequence of the i-th sample, represents the predicted sequence of the i-th sample, R(·, ·) represents the cross-correlation function, ||·|| p represents the L-p norm, and N represents the number of samples in the current training batch.

[0118] By considering the distance fluctuations, principal component differences, and correlation differences, the similarity between the prediction results and the labels is more fully considered, enabling the model to more accurately predict carbon emissions.

[0119] When the model training reaches the number of iterations or meets the preset training loss accuracy, the training is stopped to obtain the carbon emission prediction model.

[0120] Based on the carbon emission prediction model, the carbon emissions of each construction unit can be predicted. During implementation, the built environment data of the construction unit to be predicted and the multi-dimensional time series data of the previous T time steps before the current moment (including the time series data of energy consumption, the time series data of social economy, and the time series data of carbon emissions) can be obtained as the input data of the model to predict the carbon emissions of the construction unit to be predicted in the future for multiple time steps. Determine whether carbon control measures need to be implemented for the construction unit based on the predicted carbon emissions.

[0121] Those skilled in the art can understand that all or part of the processes for implementing the methods in the above embodiments can be completed by instructing relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium. Among them, the computer-readable storage medium is a magnetic disk, an optical disk, a read-only memory, or a random access memory, etc.

[0122] As mentioned above, only the preferred specific implementation manners of the present invention are described, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed by the present invention should be covered within the protection scope of the present invention.

Claims

1. A training method for a carbon emission prediction model, characterized in that, Including the following steps: Obtain the historical carbon emissions of a construction unit and the corresponding carbon emission-related factor data; the carbon emission-related factor data includes static data and multi-dimensional time-series data; Construct a training sample set based on the historical carbon emissions of the construction unit and the corresponding carbon emission-related factor data; Construct a multi-step time-series prediction model, which is used to predict the carbon emissions in multiple future time steps; the multi-step time-series prediction model includes: a static feature extraction module, which is used to extract static features based on the static data; a time-series feature extraction module, which is used to extract time-series features based on the multi-dimensional time-series data; a feature fusion module, which is used to fuse the static features and the time-series features; a prediction module, which is used to predict the carbon emissions in multiple future time steps based on the fused features; Train the multi-step time-series prediction model based on the training sample set to obtain a carbon emission prediction model; Constructing a training sample set based on the historical carbon emissions of a construction unit and the corresponding carbon emission-related factor data includes: For each construction unit, two sliding windows with a length of T slide in parallel on the multi-dimensional time-series data and the corresponding carbon emission time-series data; The multi-dimensional time-series data, the carbon emission time-series data, and the static data of the construction unit obtained in each sliding are used as the input data of the sample, and the carbon emission time-series data in multiple time steps after the sliding window on the carbon emission time-series data is used as the label of the sample, forming a sample; The time-series feature extraction module includes: A periodic feature extraction module, which is used to extract the periodic features of the time-series data based on the multi-dimensional time-series data of the sample; A deep feature extraction module, which is used to extract the deep features of each time step based on the periodic features by using a recurrent unit; A feature weighting module, which is used to perform weighted fusion on the deep features of each time step based on the carbon emissions of each time step to obtain time-series features; The periodic feature extraction module extracts periodic features of the time-series data in the following way: For the time-series data of each dimension, calculate K periodic components of the time-series data of this dimension, and convert the time-series data of this dimension into K two-dimensional data based on the K periodic components; Extract features from each two-dimensional data based on a two-dimensional convolutional network to obtain K two-dimensional features; Convert each two-dimensional feature into a one-dimensional feature; perform weighted fusion on the K one-dimensional features based on the weights of the K periodic components to obtain the periodic features corresponding to the time-series data of each dimension.

2. The training method of the carbon emission prediction model according to claim 1, characterized in that Use the following formula to calculate the training loss of the multi-step time-series prediction model: Loss=Loss mse +Loss shape Loss shape = β1Loss1 + β2Loss2 + β3Loss3 Among them, Loss mse represents the mean square error loss, Loss shape represents the shape loss, Loss1 represents the distance loss between the label sequence and the prediction sequence, Loss2 represents the principal component loss between the label sequence and the prediction sequence, Loss3 represents the correlation loss between the label sequence and the prediction sequence, and β1, β2, and β3 represent weights.

3. The training method of the carbon emission prediction model according to claim 2, characterized in that, Use the following formula to calculate the distance loss between the label sequence and the prediction sequence: where N represents the number of samples in the current training batch, T′ represents the length of the label sequence, y it represents the value of the t-th element in the label sequence of the i-th sample, represents the value of the t-th element in the prediction sequence of the i-th sample, y ij represents the value of the j-th element in the label sequence of the i-th sample, represents the value of the j-th element in the prediction sequence of the i-th sample, and d(·,·) represents the signed distance function.

4. The training method of the carbon emission prediction model according to claim 2, characterized in that, Use the following formula to calculate the principal component loss between the label sequence and the prediction sequence: Among them, N represents the number of samples in the current training batch, n mf represents the number of principal components, n f represents the number of all components, F j (y i ) represents taking the j-th principal component of y i The represents taking The j-th principal component of represents taking The k-th non-principal component of p represents the L-p norm, y i represents the label sequence of the i-th sample, represents the prediction sequence of the i-th sample.

5. The training method of the carbon emission prediction model according to claim 2, characterized in that Use the following formula to calculate the correlation loss between the label sequence and the prediction sequence: Among them, y i represents the label sequence of the i-th sample, represents the prediction sequence of the i-th sample, R(·,·) represents the cross-correlation function, ||·|| p represents the L-p norm, and N represents the number of samples in the current training batch.

6. The training method of the carbon emission prediction model according to claim 1, characterized in that Use the following formula to calculate the periodic features of the time-series data of this dimension: Among them, f ct represents the periodic eigenvalue at the t-th time step of the c-th dimensional time series data, and A ck represents the amplitude corresponding to the k-th periodic component of the c-th dimensional time series data, represents the eigenvalue at the t-th time step of the one-dimensional feature corresponding to the k-th periodic component of the c-th dimensional time series data.

7. The training method of the carbon emission prediction model according to claim 1, characterized in that Use the following formula to calculate the weights of the deep features of each time step: w t = αy t + (1 - α)w t-1 where, w t-1 represents the weight of the depth feature at the (t-1)-th time step, w t represents the weight of the depth feature at the t-th time step, y t represents the carbon emission at the t-th time step, and α represents the smoothing factor.

Images