Carbon emission prediction method and device based on data enhancement and time-frequency fusion comparison

Abstract

The application provides a carbon emission prediction method and device based on data enhancement and time-frequency fusion comparison, and belongs to the technical field of carbon emission prediction of power systems. First, power data P , weather data WT , line state data WS and energy storage element data SE under different power generation modes are acquired; the power data P is subjected to a data enhancement operation by using a generative adversarial network to obtain power generation data; the power data and the power generation data obtained after enhancement are subjected to data representation learning by time domain data representation modules and frequency domain data representation modules respectively to obtain time domain sample features, time domain enhanced sample features, frequency domain sample features and frequency domain enhanced sample features; after feature fusion processing, time-frequency domain comparison learning is performed to obtain a feature representation vector C i ; the splicing result of WT , WS and SE is combined with C i and P to predict the carbon emission intensity of the next stage.

Description

Technical Field

[0001] This invention relates to the field of carbon emission prediction technology for power systems, and in particular to a carbon emission prediction method and apparatus based on data augmentation and time-frequency fusion comparison. Background Technology

[0002] With the increasing severity of global climate change, reducing carbon emissions has become a key research focus. The power system is one of the major sources of carbon emissions, making carbon emission forecasting crucial for achieving low-carbon development in the power sector.

[0003] In the context of low-carbon power development, accurate prediction of the carbon emission intensity of the power system can provide data support for carbon reduction optimization strategies, thereby contributing to reducing carbon emissions from the power system. Currently, carbon emission prediction methods mainly employ traditional methods and artificial intelligence-based methods. Traditional methods are prone to prediction bias; therefore, artificial intelligence-based methods are widely used and have indeed improved prediction speed. However, in the power system, weather factors, line conditions during transmission, and the charging and discharging processes of energy storage components in the distribution section all affect carbon emission prediction. Existing artificial intelligence-based carbon emission prediction methods do not consider the actual impact of these external factors on carbon emissions, resulting in low accuracy of existing prediction methods. Summary of the Invention

[0004] In view of this, the present invention provides a carbon emission prediction method and apparatus based on data augmentation and time-frequency fusion comparison. It utilizes power data for time-frequency fusion features and time-frequency fusion comparison, and introduces an attention mechanism to combine features such as weather information, line status information and energy storage element status information to predict carbon emission intensity, which can effectively improve the accuracy of carbon emission prediction.

[0005] The technical solution adopted by the embodiments of the present invention to solve its technical problem is as follows:

[0006] A carbon emission prediction method based on data augmentation and time-frequency fusion comparison is proposed, implemented using the TF-CEP model. The TF-CEP model consists of a data acquisition module, a data augmentation module, a feature representation module, a contrastive learning module, and a downstream task module, respectively. The prediction implementation steps include:

[0007] Step S1: The data acquisition module acquires power data P under different power generation modes, as well as weather data WT, line status data WS, and energy storage element data SE; the power data includes thermal power data, hydropower data, wind power data, photovoltaic power generation data, and biomass power generation data, where P∈R L×d L represents the data length, and d represents the feature dimension;

[0008] Step S2, the data augmentation module uses a generative adversarial network (GAN) to augment the set R.L×d Each power data point P undergoes data augmentation to obtain power generation data;

[0009] Step S3: In the feature representation module, the power data and the enhanced power generation data are respectively processed by the time-domain data representation module and the frequency-domain data representation module for data representation learning. The time-domain data representation module outputs time-domain sample features. Temporal domain enhanced sample features The frequency domain data characterization module outputs frequency domain sample features f. i Frequency domain enhancement of sample features ;

[0010] Step S4, the contrastive learning module performs a comparison on t i and f i Perform feature fusion operations to obtain time-frequency fusion features. i ,right and Feature fusion operations are performed to obtain the time-frequency fusion features of the enhanced samples. Further comparative learning in the time-frequency domain yields the feature representation vector C. i ;

[0011] Step S5, the downstream task module uses the weather data WT, the line status data WS, and the energy storage element data SE in combination with C i Attention learning is performed on the concatenation result with P to obtain attention weights for the three data types, and C is further adjusted using these attention weights. i The results of splicing P and the resulting adjusted data are used as the carbon emission intensity prediction results for the next stage.

[0012] Preferably, step S2 is to perform data augmentation based on the generative model to create power generation data with the same time and space dependence as the real data; the generative model consists of a generator G and a discriminator D. The generator G generates new power generation data through linear layers and LSTM, and the discriminator D verifies the new power generation data based on one-dimensional convolution and LSTM.

[0013] The actual data distribution of electricity data is represented by P. r The data generated by GAN has a distribution of P. g The loss from adversarial training is expressed as:

[0014]

[0015] in, Represents the actual data p r and generate data p g Linear interpolation between them.

[0016] Preferably, in the feature representation module:

[0017] The time-domain data representation module consists of L stacked convolutional layers, and a single convolutional layer is represented as follows:

[0018] ;

[0019] ;

[0020] ;

[0021] ;

[0022] ;

[0023] Where ⊛ represents a one-dimensional convolution operation, || represents a concatenation operation, and the convolution kernel R k×d ={ , ,..., }, R={b (1) ,b (2) ,...,b (h)}, W fc ∈R d×h b fc ∈R 1×h ; FC(·) represents a single fully connected neural network layer; Norm() represents the layer normalization operation, μ and σ 2 Here, denoted as mean and variance, respectively; γ and β represent the parameter vectors for scaling and translation, respectively. The layer normalization process is represented as:

[0024]

[0025] The frequency domain data representation module introduces the FECAM mechanism to perform representation learning on power data and enhanced generation data: First, the power data P∈R L×d After the segmentation operation, the power generation data in each channel is split into variables {v0, v1, ..., v L-1 Then, a Discrete Cosine Transform (DCT) is introduced to obtain frequency information. Each channel variable is element-wise multiplied with all DCT components, and the complete channel tensor Freq is obtained through stacking and realigning.

[0026] ;

[0027] ;

[0028] ;

[0029] Then, frequency channel attention is learned using a fully connected neural network structure, as follows:

[0030] ;

[0031] ;

[0032] f i =FECAM(P);

[0033] =GAN(FECAM(P));

[0034] Preferably, in the contrastive learning module, the feature fusion process is represented as follows:

[0035] ;

[0036] ;

[0037] The same batch of X data points in the power data sample corresponds to sample o i and As positive sample pairs, other samples are treated as negative samples. The goal is to maximize the value of the corresponding sample o in the same batch of data. i and Time-frequency contrastive learning is performed by considering the similarity between feature representations and minimizing the similarity with other samples to obtain a high-quality feature representation vector C. i During the training of the TF-CEP model, the loss function is expressed as:

[0038] ;

[0039] ;

[0040] Among them, sim(o i , ) indicates calculation of o i and similarity, This is a hyperparameter.

[0041] Preferably, the steps of the downstream task module include:

[0042] Step S51, splicing C i The concatenation result with P is passed through an FC layer to obtain matrix D:

[0043]

[0044] In step S52, the weather data WT, line status data WS, and energy storage element data SE are processed through the FC layer and used as query vectors to perform attention learning with D to obtain the attention weight α.WT α WS α SE :

[0045] ;

[0046] ;

[0047] ;

[0048] ;

[0049] ;

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] In the formula, h represents the vector dimension. and Represents the weight parameter matrix;

[0055] Calculate D WT D WS and D SE :

[0056] ;

[0057] ;

[0058] ;

[0059] splicing D WT D WS and D SE The spliced ​​structure, after passing through the Norm() layer and the FC layer, yields the output Y, which serves as the carbon emission intensity prediction result for the next stage.

[0060]

[0061] In the TF-CEP model training process, the mean squared error is used as the loss function. :

[0062]

[0063] In the formula, let Y (z) Y' represents the prediction result. (z)This represents the true result, where L is the total number of samples.

[0064] A carbon emission prediction device based on data augmentation and time-frequency fusion comparison implements the aforementioned method based on the TF-CEP model. The TF-CEP model consists of a data acquisition module, a data augmentation module, a feature representation module, a contrastive learning module, and a downstream task module, in sequence.

[0065] The data acquisition module is used to acquire power data P under different power generation modes, as well as weather data WT, line status data WS, and energy storage element data SE; the power data includes thermal power data, hydropower data, wind power data, photovoltaic power generation data, and biomass power generation data, where P∈R L×d L represents the data length, and d represents the feature dimension;

[0066] The data augmentation module is used to apply Generative Adversarial Network (GAN) to the set R. L×d Each power data point P undergoes data augmentation to obtain power generation data;

[0067] The feature representation module includes a time-domain data representation module and a frequency-domain data representation module. These modules are used to perform data representation learning on power data and enhanced generation data, respectively, by passing the power data and the enhanced generation data through the time-domain and frequency-domain data representation modules. The time-domain data representation module outputs time-domain sample features. Temporal domain enhanced sample features The frequency domain data characterization module outputs frequency domain sample features f. i Frequency domain enhancement of sample features ;

[0068] The contrastive learning module is used to analyze t i and f i Perform feature fusion operations to obtain time-frequency fusion features. i ,right and Feature fusion operations are performed to obtain the time-frequency fusion features of the enhanced samples. Further comparative learning in the time-frequency domain yields the feature representation vector C. i ;

[0069] The downstream task module is used to combine the weather data WT, the line status data WS, and the energy storage element data SE with C. i Attention learning is performed on the concatenation result with P to obtain attention weights for the three data types, and C is further adjusted using these attention weights. i The results of splicing P and the resulting adjusted data are used as the carbon emission intensity prediction results for the next stage.

[0070] As can be seen from the above technical solution, the carbon emission prediction method and apparatus based on data augmentation and time-frequency fusion comparison provided in this invention first learns the potential knowledge of time and frequency domain data based on power data under different power generation modes. Then, based on time-frequency fusion features, it performs time-frequency fusion comparison by maximizing the similarity between positive sample feature representations and minimizing the similarity between negative sample pairs. Finally, an attention mechanism is introduced to combine features such as weather information, line status information, and energy storage component status information to predict carbon emission intensity. This invention utilizes power data for time-frequency fusion features and comparison, and introduces an attention mechanism to combine features such as weather information, line status information, and energy storage component status information to predict carbon emission intensity, which can effectively improve the accuracy of carbon emission prediction. Attached Figure Description

[0071] Figure 1 This is a diagram of the TF-CEP model architecture for carbon emission prediction based on data augmentation and time-frequency fusion comparison.

[0072] Figure 2 This is a schematic diagram of the frequency domain data representation module. Detailed Implementation

[0073] The technical solution and effects of the present invention will be further described in detail below with reference to the accompanying drawings.

[0074] To fully extract temporal and frequency information from electricity data for carbon emission intensity prediction, this paper proposes the TF-CEP method. The model architecture of TF-CEP is as follows: Figure 1 As shown, the model consists of a data module, a data augmentation and feature representation module, a contrastive learning module, and a downstream task module. Power data information under different power generation modes is of great significance for predicting carbon emission intensity. In this paper, the power data under different power generation modes mainly include coal power data (based on carbon energy power generation data), hydroelectricity power data (based on hydropower data), wind power data (based on wind power data), photovoltaic power data (based on photovoltaic power generation data), and biomass power data (based on biomass power generation data). To fully explore the power data information under different power generation modes, firstly, data augmentation operations are performed on the power data through the data augmentation and feature representation module, and feature representation learning is performed in the time and frequency domains, allowing the model to fully explore the potential knowledge in the power data and obtain time-domain sample features. Temporal domain enhanced sample features Frequency domain sample features f i Frequency domain enhancement of sample features Then, in the contrastive learning module, the time-frequency domain sample features and the time-frequency domain enhanced sample features are fused together, and contrastive learning is performed in the time-frequency domain. In the downstream task module, information such as weather conditions is introduced to predict carbon emission intensity, and the model's predictive performance is improved through an attention mechanism. The following details the specific implementation process of the data augmentation module, feature representation module, contrastive learning module, and downstream task module in the TF-CEP model.

[0075] Step S1: The data acquisition module acquires power data P under different power generation modes, as well as weather data WT, line status data WS, and energy storage element data SE; the power data includes thermal power data, hydropower data, wind power data, photovoltaic power generation data, and biomass power generation data, where P∈R L×d L represents the data length, and d represents the feature dimension;

[0076] Step S2, the data augmentation module uses a generative adversarial network (GAN) to augment the set R. L×d Each power data point P undergoes data augmentation to obtain power generation data;

[0077] Step S3: In the feature representation module, the power data and the enhanced power generation data are respectively processed by the time-domain data representation module and the frequency-domain data representation module for data representation learning. The time-domain data representation module outputs the time-domain sample features. Temporal domain enhanced sample features The frequency domain data representation module outputs frequency domain sample features f. i Frequency domain enhancement of sample features ;

[0078] Step S4, compare the learning module on t i and f i Perform feature fusion operations to obtain time-frequency fusion features. i ,right and Feature fusion operations are performed to obtain the time-frequency fusion features of the enhanced samples. Further comparative learning in the time-frequency domain yields the feature representation vector C. i ;

[0079] Step S5, the downstream task module combines weather data WT, line status data WS, and energy storage element data SE with C i Attention learning is performed on the concatenation result with P to obtain attention weights for the three data types, and C is further adjusted using these attention weights. i The results of splicing P and the resulting adjusted data are used as the carbon emission intensity prediction results for the next stage.

[0080] To enhance robustness and improve the model's generalization ability, we first process the power data P∈R under different power generation modes as input. L×d Data augmentation is performed. This invention uses generative adversarial networks (GANs) for data augmentation. This module aims to learn a generative model that can create power generation data with the same temporal and spatial dependencies as real data. The data augmentation module mainly consists of a generator G and a discriminator D. The generator generates new power generation data through linear layers and LSTM, while the discriminator validates the new power generation data based on one-dimensional convolution and LSTM.

[0081] The actual data distribution of electricity data is represented by P. r The data generated by GAN has a distribution of P. g The loss from adversarial training is expressed as:

[0082] (1)

[0083] in, Represents the actual data p r and generate data p g Linear interpolation between them.

[0084] The feature representation module is used to learn the representations of power data and augmented generation data through the time-domain data representation module and the frequency-domain data representation module, respectively.

[0085] The temporal data representation module consists of L stacked convolutional layers, and a single convolutional layer is represented as:

[0086] (2)

[0087] (3)

[0088] (4)

[0089] (5)

[0090] (6)

[0091] Where ⊛ represents a one-dimensional convolution operation, || represents a concatenation operation, and the convolution kernel R k×d ={ , ,..., }, R={b (1) ,b (2) ,...,b (h)}, W fc ∈R d×h b fc ∈R 1×h; FC(·) represents a single fully connected neural network layer; Norm() represents the layer normalization operation, μ and σ 2 Here, denoted as mean and variance, respectively; γ and β represent the parameter vectors for scaling and translation, respectively. The layer normalization process is represented as:

[0092] (7)

[0093] The frequency domain data representation module introduces the FECAM mechanism to perform representation learning on power data and enhanced power generation data. The specific process is as follows: Figure 2 As shown: First, the power data P∈R L×d After the segmentation operation, the power generation data in each channel is split into variables {v0, v1, ..., v L-1 Then, a Discrete Cosine Transform (DCT) is introduced to obtain frequency information. Each channel variable is element-wise multiplied with all DCT components, and the complete channel tensor Freq is obtained through stacking and realigning.

[0094] (8)

[0095] (9)

[0096] (10)

[0097] Then, frequency channel attention is learned using a fully connected neural network structure, as follows:

[0098] (11)

[0099] (12)

[0100] f i =FECAM(P) (13)

[0101] =GAN(FECAM(P)) (14)

[0102] In the contrastive learning module, the feature fusion process is represented as follows:

[0103] (15)

[0104] (16)

[0105] The same batch of X data points in the power data sample corresponds to sample o i and As positive sample pairs, other samples are treated as negative samples. The goal is to maximize the value of the corresponding sample o in the same batch of data. i and Time-frequency contrastive learning is performed by considering the similarity between feature representations and minimizing the similarity with other samples to obtain a high-quality feature representation vector C. i During the training of the TF-CEP model, the loss function is expressed as:

[0106] (17)

[0107] (18)

[0108] Among them, sim(o i , ) indicates calculation of o i and similarity, This is a hyperparameter.

[0109] The steps of the downstream task module include:

[0110] Step S51, splicing C i The concatenation result with the original data P is passed through an FC layer to obtain matrix D:

[0111] (19)

[0112] Step S52: Since carbon emission intensity in the power system is closely related to weather data, line status, and energy storage component status, we incorporate the acquired weather data, line status data, and energy storage component data into the attention learning process based on power data features. Weather data (WT), line status data (WS), and energy storage component data (SE) are each processed through an FC layer and then used as query vectors to perform attention learning with D, thereby obtaining the attention weight α. WT α WS α SE :

[0113] (20)

[0114] (twenty one)

[0115] (twenty two)

[0116] (twenty three)

[0117] (twenty four)

[0118] (25)

[0119] (26)

[0120] (27)

[0121] (28)

[0122] In the formula, h represents the vector dimension. and Represents the weight parameter matrix;

[0123] Calculate D WT D WS and D SE :

[0124] (29)

[0125] (30)

[0126] (31)

[0127] splicing D WT D WS and D SE The spliced ​​structure, after passing through the Norm() layer and the FC layer, yields the output Y, which serves as the carbon emission intensity prediction result for the next stage.

[0128] (32)

[0129] In the TF-CEP model training process, the mean squared error is used as the loss function. :

[0130] (33)

[0131] In the formula, let Y (z) Y' represents the prediction result. (z) This represents the true result, where L is the total number of samples.

[0132] A carbon emission prediction device based on data augmentation and time-frequency fusion comparison implements the aforementioned method based on the TF-CEP model. The TF-CEP model consists of a data acquisition module, a data augmentation module, a feature representation module, a contrastive learning module, and a downstream task module, in sequence.

[0133] The data acquisition module is used to acquire power data P under different power generation modes, as well as weather data WT, line status data WS, and energy storage element data SE; the power data includes thermal power data, hydropower data, wind power data, photovoltaic power generation data, and biomass power generation data, where P∈R L×d L represents the data length, and d represents the feature dimension;

[0134] The data augmentation module is used to augment the set R using a generative adversarial network (GAN).L×d Each power data point P undergoes data augmentation to obtain power generation data;

[0135] The feature representation module includes a time-domain data representation module and a frequency-domain data representation module. These modules are used to learn the representations of power data and enhanced generation data, respectively, and the time-domain data representation module outputs time-domain sample features. Temporal domain enhanced sample features The frequency domain data representation module outputs frequency domain sample features f. i Frequency domain enhancement of sample features ;

[0136] The contrastive learning module is used to analyze t i and f i Perform feature fusion operations to obtain time-frequency fusion features. i ,right and Feature fusion operations are performed to obtain the time-frequency fusion features of the enhanced samples. Further comparative learning in the time-frequency domain yields the feature representation vector C. i ;

[0137] The downstream task module is used to combine weather data (WT), line status data (WS), and energy storage element data (SE) with C. i Attention learning is performed on the concatenation result with P to obtain attention weights for the three data types, and C is further adjusted using these attention weights. i The results of splicing P and the resulting adjusted data are used as the carbon emission intensity prediction results for the next stage.

[0138] Data and Experimental Description

[0139] To verify the effectiveness of the TF-CEP model method, this paper uses data from a power service system, containing 627,263 sequence samples. The ratio of the training set, validation set, and test set is set to 0.7:0.15:0.15. The experimental environment is shown in Table 1.

[0140] Table 1 Experimental Environment

[0141]

[0142] In this invention, MSE and MAE are selected as evaluation indicators.

[0143] (34)

[0144] (35)

[0145] Where N represents the number of samples, For the true value, These are the predicted values. The smaller the values ​​of MSE and MAE, the smaller the difference and average difference between the model's predicted values ​​and the actual values, meaning the better the model's predictive performance.

[0146] Comparison of prediction performance (model performance)

[0147] (1) To verify the performance of the model, we compared the proposed method TF-CEP with eight baseline methods, and the prediction results are shown in Table 2. The experimental results show that the TF-CEP method outperforms the other baseline methods. The performance improvement of the TF-CEP method is due to its thorough learning of the time and frequency domain characteristics of power data, and its ability to mine the potential knowledge in the features through comparative learning.

[0148] Table 2 Comparison of results from different prediction methods

[0149]

[0150] (2) In this paper, we compare the TF-CEP method with four state-of-the-art methods: Ts2vec, BTSF, SimCLR and TS-TCC. As shown in Table 3, the TF-CEP method has advantages over the other four methods.

[0151] Ts2vec[1]: A representation learning model that hierarchically distinguishes between positive and negative samples from the perspectives of instance and time.

[0152] BTSF[2]: This method proposes a Bilinear Temporal-SpectralFusion framework and improves model performance by learning feature representations in the time domain and frequency domain respectively.

[0153] SimCLR[3]: A self-supervised contextual contrastive learning framework.

[0154] TS-TCC[4]: A contrastive learning framework based on time contrast module and context contrast module.

[0155] Table 3 Comparison of Advanced Methods

[0156]

[0157] Ablation experiment:

[0158] To further verify the contribution of each module in the TF-CEP method, we conducted ablation experiments. Variants of the method presented in this paper are shown in Table 4.

[0159] Table 4. Variants of TF-CEP

[0160]

[0161] Table 5 shows the performance of different variants of the TF-CEP method. Experiments demonstrate that the TF-CEP method outperforms the other four variants. Furthermore, the experiments show that the data augmentation module, time-frequency domain representation learning module, and contrastive learning module of the TF-CEP method all have a significant impact on the performance of the carbon emission prediction model.

[0162] Table 5 Performance comparison of different variants

[0163]

[0164] Parameter effects:

[0165] To improve carbon emission intensity prediction and reduce the impact of model parameters on prediction performance, this paper conducts parameter experiments to evaluate the model's prediction performance under different embedding dimensions. The parameter h is set to 8, 16, 32, 64, and 128. Experimental results show that larger h values ​​result in better model prediction performance, but when h > 16, the improvement in prediction performance is not significant. Therefore, based on the parameter experiments, this paper sets h to 16.

[0166] The above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the invention. Those skilled in the art will understand that implementing all or part of the above-described embodiments and making equivalent changes in accordance with the claims of the present invention are still within the scope of the invention.

[0167] [1]Choi S, Kang D, Cho M. Contrastive Mean-Shift Learning for Generalized Category Discovery[C]. Proceedings of the IEEE / CVF Conference onComputer Vision and Pattern Recognition. 2024: 23094-23104.

[0168] [2]Wu J, Sun W, Gan T, et al. Neighbor-guided consistent and contrastive learning for semi-supervised action recognition[J]. IEEETransactions on Image Processing, 2023, 32: 2215-2227.

[0169] [3]Shu X, Xu B, Zhang L, et al. Multi-granularity anchor-contrastiverepresentation learning for semi-supervised skeleton-based action recognition[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2022, 45(6): 7559-7576.

[0170] [4]Basak H, Yin Z. Pseudo-label guided contrastive learning for semi-supervised medical image segmentation[C]. Proceedings of the IEEE / CVFconference on computer vision and pattern recognition. 2023: 19786-19797.

Claims

1. A carbon emission prediction method based on data enhancement and time-frequency fusion contrast, characterized in that, Based on the implementation of the TF-CEP model, the TF-CEP model is composed of a data acquisition module, a data enhancement module, a feature representation module, a contrast learning module and a downstream task module in turn, and the prediction implementation steps include: In step S1, the data acquisition module acquires power data P in different power generation modes, weather data WT, line state data WS and energy storage element data SE; the power data includes thermal power data, hydropower data, wind power data, photovoltaic power generation data and bio-power data, P ∈ R L×d L represents the data length, and d represents the feature dimension. In step S2, the data enhancement module utilizes a generative adversarial network (GAN) to generate a set R L×d Each power data P is subjected to a data enhancement operation to obtain power generation data. Step S3: In the feature representation module, the power data and the enhanced power generation data are respectively processed by the time-domain data representation module and the frequency-domain data representation module for data representation learning. The time-domain data representation module outputs time-domain sample features. Temporal domain enhanced sample features The frequency domain data characterization module outputs frequency domain sample features f. i Frequency domain enhancement of sample features ; Step S4, the contrastive learning module performs a comparison on t i and f i Perform feature fusion operations to obtain time-frequency fusion features. i ,right and Feature fusion operations are performed to obtain the time-frequency fusion features of the enhanced samples. Further comparative learning in the time-frequency domain yields the feature representation vector C. i ; Step S5, the downstream task module utilizes the weather data WT, the line state data WS and the energy storage element data SE in combination with C i Attention learning is performed on the spliced result of P, and attention weights about the three types of data are obtained, and the attention weights are further utilized to adjust C i The spliced result of P, and the obtained adjustment data is used as the carbon emission intensity prediction result of the next stage.

2. The carbon emission prediction method based on data augmentation and time- frequency fusion contrast of claim 1, wherein, The step S2 is data enhancement based on a generation model, and power generation data with the same time and space dependence as the real data is created; the generation model is composed of a generator G and a discriminator D, the generator G generates new power generation data through a linear layer and an LSTM, and the discriminator D verifies the new power generation data based on one-dimensional convolution and LSTM; The actual data distribution of the power data is represented as P r The data distribution generated by the GAN is represented as P g The loss of the adversarial training is represented as: ； wherein, represents a linear interpolation between the real data p r and the generated data p g .

3. The carbon emission prediction method based on data augmentation and time- frequency fusion contrast of claim 2, wherein, In the feature representation module: The time domain data representation module is stacked by L' convolution layers, and a single convolution layer is represented as: ； ； ； ； ； wherein, represents a one-dimensional convolution operation, represents a concatenation operation, the convolution kernel R k×d = , ,..., }, R = {b (1) ,b (2) ,...,b (h)}, W fc ∈ R d×h , b fc ∈ R 1×h ; FC(·) represents a single fully connected neural network layer; Norm() represents a layer normalization operation, μ and σ 2 are the mean and variance, respectively; γ and β represent the scaling and translation parameter vectors respectively, and the layer normalization process is represented as: ； The frequency domain data representation module introduces the FECAM mechanism to perform representation learning on the power data and the enhanced power generation data: first, the power data P ∈ R L×d After the segmentation operation, the power generation data in each channel is split to obtain variables {v0, v1,..., v L’-1}; Then, the discrete cosine transform DCT is introduced to obtain frequency information, each channel variable is multiplied with all DCT components in elements, and a complete channel tensor Freq is obtained through stacking and realignment: ； ； ； Then, the frequency channel attention is learned using a fully connected neural network structure, and the process is: ； ； f i = FECAM(P); = GAN(FECAM(P)).

4. The carbon emission prediction method based on data augmentation and time- frequency fusion contrast of claim 3, wherein, In the contrast learning module, the feature fusion process is represented as: ； ； X data pairs corresponding to sample o in the same batch of power data samples i And As positive sample pairs, other samples as negative samples, by maximizing the similarity between the corresponding samples o in the same batch of data i And The similarity between the feature representations is minimized, and the similarity between other samples is minimized to perform time-frequency contrast learning, and a high-quality feature representation vector C is obtained i ; In the training process of the TF-CEP model, the loss function is represented as: ； ； where sim(o i , ) denotes the similarity of o i and , is a hyperparameter.

5. The carbon emission prediction method based on data augmentation and time- frequency fusion contrast of claim 4, wherein, The steps of the downstream task module include: Step S51, concatenating C i and P, the concatenation result passes through the FC layer to obtain matrix D: ； Step S52, weather data WT, line state data WS and energy storage element data SE respectively after FC layer as query vector and D attention learning to get attention weight α WT 、α WS 、α SE : ； ； ； ； ； ； ； ； ； where h represents the vector dimension, and denotes a weight parameter matrix; D is calculated WT D WS and D SE : ； ； ； Splicing D WT , D WS and D SE , the output Y obtained after the Norm() layer and the FC layer is spliced is taken as the carbon emission intensity prediction result of the next stage: ； The loss in the training process of the TF-CEP model adopts a mean square error as a loss function : ； where Y (z) represents the prediction result, Yˊ (z) represents the true result, and L is the total number of samples.

6. A carbon emission prediction device based on data augmentation and time-frequency fusion contrast, characterized in that, Based on the implementation of the TF-CEP model, the TF-CEP model is composed of a data acquisition module, a data enhancement module, a feature representation module, a contrast learning module and a downstream task module in turn: The data acquisition module is used for acquiring power data P in different power generation modes, weather data WT, line state data WS and energy storage element data SE; the power data includes thermal power data, hydropower data, wind power data, photovoltaic power generation data and bio-power generation data, and P is in R L×d L represents the data length, and d represents the feature dimension. The data enhancement module is configured to utilize a generative adversarial network (GAN) on the set R L×d Each power data P is subjected to a data enhancement operation to obtain power generation data; The feature representation module comprises a time domain data representation module and a frequency domain data representation module, which are used for performing data representation learning on the power data and the enhanced power generation data respectively through the time domain data representation module and the frequency domain data representation module , the time domain data representation module outputs time domain sample features , the frequency domain data representation module outputs frequency domain sample features f i and frequency domain enhanced sample features ; The contrast learning module is configured to perform feature fusion operation on t i and f i to obtain time-frequency fusion features o i . and to obtain time-frequency fusion features of the enhanced sample , and further perform contrast learning in the time-frequency domain to obtain a feature representation vector C i . The downstream task module is configured to combine the weather data WT, the line state data WS, and the energy storage element data SE to obtain adjusted data i The attention learning is performed on the spliced result of P to obtain attention weights for the three types of data, and the attention weights are further utilized to adjust C i The spliced result of P, and the obtained adjusted data is used as the carbon emission intensity prediction result of the next stage.

Images