Power system source side carbon emission prediction method and device based on time-frequency domain analysis

Abstract

The invention provides an electric power system source side carbon emission prediction method and device based on time-frequency domain analysis, equipment and a medium, and the method comprises the steps: related multi-source data are processed; performing enhancement processing on the fused data, improving the fidelity of the enhanced data by minimizing cross entropy, improving the diversity of the enhanced data by minimizing mutual information, and finally obtaining an enhanced multi-source time sequence; performing multi-resolution expansion on the frequency domain features and the time domain features of the enhanced multi-source time sequence, performing dynamic feature screening through an information gating unit, and performing connection to generate super-resolution features; and inputting the super-resolution features into a carbon emission prediction model to obtain a predicted value of the carbon emission. According to the method, high-frequency details such as transient fluctuation of power generation equipment are captured for a power system source side, low-frequency trends such as long-term change of an energy structure can be extracted, and multi-resolution information in a time sequence can be better captured, so that the accuracy and reliability of prediction are improved.

Description

Technical Field

[0001] This invention relates to the field of carbon emission analysis technology, and in particular to a method, apparatus, equipment and medium for predicting source-side carbon emissions from power systems based on time-frequency domain analysis. Background Technology

[0002] To achieve low-carbon development goals, accurately predicting carbon emissions from the energy system is crucial for formulating effective emission reduction policies. Currently, power system source-side carbon emission prediction largely relies on traditional statistical methods, such as linear regression models based on historical carbon emission data. These models assume a linear relationship between carbon emissions and time or a few simple variables, predicting future values ​​by fitting linear trends in historical data. However, in actual operation, when power generation equipment ages, energy structure is fine-tuned, or the operational status of environmental protection facilities changes, these models fail to accurately capture fluctuations in carbon emissions, leading to a sharp increase in prediction errors and making it difficult to meet the needs of precise control.

[0003] In recent years, machine learning algorithms have been widely used in carbon emission prediction, such as using Support Vector Regression (SVR). SVR constructs a prediction function by finding an optimal hyperplane that maximizes the distance from sample points to the hyperplane. However, SVR faces challenges when processing large-scale, high-dimensional power system data. On the one hand, the selection of kernel functions and parameter tuning are complex and time-consuming, and it has poor adaptability to different power operating conditions. On the other hand, it is difficult to effectively integrate multi-source heterogeneous information from the power production process, such as the combined impact of unit operating parameters, energy consumption structure, and meteorological conditions on carbon emissions, resulting in limited prediction accuracy and an inability to cope with complex and ever-changing power system operating scenarios.

[0004] In the field of deep learning, existing time series prediction techniques have some limitations in carbon emission prediction. Traditional methods (such as RNNs) struggle to simultaneously capture high-frequency details and low-frequency trends, and single-resolution modeling easily overlooks multi-scale features. Existing contrastive learning methods (such as TS-TCC) rely on manually designed enhancement strategies, lacking adaptability and prone to generating invalid positive and negative samples. Frequency decomposition based on Fourier transform or moving average (such as FEDformer) is susceptible to the "pick-and-fence effect," failing to fully preserve key spectral information. While Transformer-type models (such as Informer) can capture long-range dependencies, their large parameter count and slow inference speed make them difficult to apply in real-time scenarios.

[0005] Therefore, existing power system source-side carbon emission prediction technologies have many problems. Traditional statistical methods are limited by linear assumptions and are difficult to adapt to nonlinear scenarios such as equipment aging and dynamic adjustments in energy structure. Although machine learning algorithms such as SVR introduce nonlinear modeling capabilities, they face bottlenecks such as difficulty in kernel function parameter tuning and insufficient ability to fuse multi-source heterogeneous data. Although deep learning has shown potential, it has problems with the adaptability of model structure to application scenarios and is insufficient in extracting multi-resolution information and considering the impact of resolution on frequency. This makes it difficult for models to capture the multi-scale dynamic characteristics of carbon emissions, and may ignore key low-frequency signals masked by high-frequency noise, or the modulation effect of low-frequency trends on high-frequency components, resulting in insufficient recognition of periodic patterns and delayed response to transient events. Summary of the Invention

[0006] To overcome the problems existing in related technologies, this disclosure provides a method, apparatus, equipment and medium for predicting carbon emissions from the source side of a power system based on time-frequency domain analysis, so as to solve the technical problems in related technologies.

[0007] This specification provides one or more embodiments of a power system source-side carbon emission prediction method based on time-frequency domain analysis, including the following steps: Collect multi-source data related to carbon emissions from the power system source side, and perform normalization, sampling frequency alignment, and feature stitching operations on the multi-source data to obtain fused data; The fused data is augmented by minimizing cross-entropy to improve the fidelity of the augmented data and minimizing mutual information to improve the diversity of the augmented data, resulting in an augmented multi-source time series, where the true label is the actual value of carbon emissions from the power system. After multi-resolution expansion of the frequency domain features and time domain features of the enhanced multi-source time series, the features are dynamically filtered through information gating units, and the filtered features are connected to generate super-resolution features. The carbon emission prediction model is obtained by inputting super-resolution features into the carbon emission prediction model, which is a fully connected neural network trained based on the model training convergence condition.

[0008] This specification provides one or more embodiments of a power system source-side carbon emission prediction device based on time-frequency domain analysis, comprising: The data acquisition and processing module is used to collect multi-source data related to carbon emissions from the power system source side, and to perform normalization, sampling frequency alignment and feature stitching operations on the multi-source data to obtain fused data. The data augmentation module is used to enhance the fused data. It improves the fidelity of the augmented data by minimizing cross-entropy and enhances the diversity of the augmented data by minimizing mutual information, thereby obtaining the augmented multi-source time series. The true label is the actual value of carbon emissions from the power system. The time-frequency domain multi-resolution extension module is used to perform multi-resolution extension on the frequency domain features and time domain features of the enhanced multi-source time series, and then dynamically filter the features through the information gating unit, and connect the filtered features to generate super-resolution features. The carbon emission prediction module is used to obtain the predicted value of carbon emissions by inputting super-resolution features into the carbon emission prediction model. The carbon emission prediction model is a fully connected neural network trained based on the model training convergence condition.

[0009] This specification provides one or more embodiments of a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the power system source-side carbon emission prediction method based on time-frequency domain analysis as described above.

[0010] This specification provides one or more embodiments of a computer-readable storage medium storing a computer program that, when executed by a processor, implements the power system source-side carbon emission prediction method based on time-frequency domain analysis as described above.

[0011] This disclosure provides a method, apparatus, equipment, and medium for predicting carbon emissions from the power system source side based on time-frequency domain analysis. Its advantages lie in the collection of multi-source data related to carbon emissions from various stages of the power system source side. The fusion of multi-source data significantly improves prediction accuracy, offering enhanced accuracy under complex power conditions compared to traditional models that rely solely on power generation. Since the time-domain and frequency-domain characteristics of power system carbon emission-related data are multi-scale, the original multi-source data is insufficient to explicitly express multi-resolution information. Therefore, data augmentation is performed on the multi-source data to expand the representation dimensions in the time and frequency domains. Sufficient dissimilarity in the augmented multi-source time series improves the generalization ability of the contrastive learning model. Furthermore, the mutual information between the augmented data and the original data is used as an evaluation criterion to select the optimal augmented data that retains the core features of the original data while introducing diversity. High-fidelity data augmentation maintains semantic invariance to transformations, while high-diversity data augmentation exhibits sufficient dissimilarity. Finally, by performing joint time- and frequency-domain multi-resolution analysis on the enhanced multi-source time series, and through multi-resolution collaborative modeling in the time and frequency domains, it is possible to capture high-frequency details such as transient fluctuations of power generation equipment, and extract low-frequency trends such as long-term changes in energy structure. This allows for better capture of multi-resolution information in the time series, thereby improving the accuracy and reliability of predictions. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in one or more embodiments of this specification or in the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating a power system source-side carbon emission prediction method based on time-frequency domain analysis, provided for one or more embodiments of this specification; Figure 2 A schematic diagram illustrating the process of frequency-time domain multi-resolution extension provided in one or more embodiments of this specification; Figure 3 A block diagram for predicting source-side carbon emissions from a power system based on time-frequency domain analysis, provided for one or more embodiments of this specification; Figure 4 This is a schematic diagram of the structure of a computer device provided for one or more embodiments of this specification. Detailed Implementation

[0014] To enable those skilled in the art to better understand the technical solutions in one or more embodiments of this specification, the technical solutions in one or more embodiments of this specification will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this specification, and not all of the embodiments. Based on one or more embodiments of this specification, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this invention.

[0015] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings.

[0016] Method Implementation Examples According to embodiments of the present invention, a method for predicting source-side carbon emissions from power systems based on time-frequency domain analysis is provided, such as... Figure 1 The diagram shown is a flowchart of the power system source-side carbon emission prediction method based on time-frequency domain analysis provided in this embodiment. The power system source-side carbon emission prediction method based on time-frequency domain analysis according to this embodiment includes: Step S1: Collect multi-source data related to carbon emissions from the power system source side, and perform normalization, sampling frequency alignment, and feature stitching operations on the multi-source data to obtain fused data; Step S2: Enhance the fused data by minimizing cross-entropy to improve the fidelity of the enhanced data and by minimizing mutual information to improve the diversity of the enhanced data, thereby obtaining the enhanced multi-source time series, where the true label is the actual value of carbon emissions from the power system.

[0017] Step S3: After multi-resolution expansion of the frequency domain features and time domain features of the enhanced multi-source time series, the features are dynamically filtered through information gating units, and the filtered features are connected to generate super-resolution features.

[0018] Step S4: Input the super-resolution features into the carbon emission prediction model to obtain the predicted value of carbon emissions. The carbon emission prediction model is a fully connected neural network trained based on the model training convergence condition.

[0019] The power system source-side carbon emission prediction method provided in this embodiment, based on time-frequency domain analysis, collects multi-source data related to carbon emissions from various links on the power system source side. The fusion of multi-source data brings a significant improvement in prediction accuracy. Compared with traditional models that rely solely on the single factor of power generation, the prediction accuracy is improved under complex power conditions. Since the time-domain and frequency-domain characteristics of power system carbon emission-related data have multi-scale features (such as hourly fluctuations and quarterly trends), the original multi-source data is difficult to explicitly express multi-resolution information. Therefore, data augmentation is performed on the multi-source data to expand the representation dimensions in the time and frequency domains. The sufficient differences in the augmented multi-source time series can improve the generalization ability of the contrastive learning model. Then, the mutual information between the augmented data and the original data is used as the evaluation criterion to select the optimal augmented data that can retain the core features of the original data (high fidelity) and introduce diversity. High-fidelity data augmentation maintains semantic characteristics that are invariant to transformations, while high-diversity data augmentation has sufficient differences. Finally, by performing joint time- and frequency-domain multi-resolution analysis on the enhanced multi-source time series, and through multi-resolution collaborative modeling in the time and frequency domains, it is possible to capture high-frequency details such as transient fluctuations of power generation equipment, and extract low-frequency trends such as long-term changes in energy structure. This allows for better capture of multi-resolution information in the time series, thereby improving the accuracy and reliability of predictions.

[0020] In this embodiment, the multi-source data related to carbon emissions from the power system source side collected include boiler and equipment parameters, unit operation and load parameters, and coal composition parameters. Boiler and equipment parameters include boiler feedwater temperature, air supply temperature, oxygen content, thermally calculated flue gas temperature, main steam temperature, main steam pressure, reheat steam temperature, low cylinder exhaust temperature, flue gas outlet dust concentration, flue gas outlet SO2 concentration, and flue gas outlet nitrogen oxide concentration. Unit operating and load parameters include average load factor, power plant power consumption rate, and active power; The composition parameters of coal include air-dried carbon content, as-received volatile matter, as-received ash content, as-received lower heating value, and total moisture.

[0021] In this embodiment, the normalization, sampling frequency alignment, and feature stitching operations performed on the multi-source data in step S1 are due to the differences in dimensionality, sampling frequency, and distribution among the multi-source data. The data can be normalized to eliminate the differences in dimensionality, and the data with different sampling frequencies can be aggregated and aligned through interpolation or time windows. Finally, the data can be fused through feature stitching.

[0022] In this embodiment, due to noise interference in multi-source power system data, traditional data cleaning methods are prone to information loss, necessitating data quality restoration through enhancement techniques. Historical carbon emission data is constrained by energy structure and equipment status during specific periods, resulting in a singular data distribution. The time-frequency domain characteristics of carbon emissions exhibit multi-scale features (such as hourly fluctuations and quarterly trends), making it difficult for the original data to explicitly express multi-resolution information. Enhancement is needed to expand the time-frequency domain representation dimensions. Therefore, this embodiment uses step S2 to accurately measure the mutual information between the enhanced data and the original data as an evaluation criterion to select the optimal enhancement method that can both retain the core features of the original data (high fidelity) and introduce rich and diverse variations (diversity) (for the input data). x (Enhancement methods); High-fidelity data augmentation maintains semantic properties that are invariant to transformations, and sufficient diversity in highly diverse data augmentation improves the generalization ability of contrastive learning models.

[0023] In one specific embodiment, a probability function of multi-source data and random variables is used as an enhancement method, as shown in the following formula: ; in, m To enhance data fusion, x For a time series instance, is a T×F matrix, where T represents the length of the time series and F represents the feature dimension of each time step. Let be a probability function.

[0024] The objective in this embodiment is to minimize mutual information. Its definition is ,because yes The (Shannon) entropy is independent of data augmentation; the objective is equivalent to minimizing conditional entropy. .

[0025] To optimize effectively, use and We approximate it using the cross-entropy between them, where Therefore, it is to enhance the integration of data For the input prediction, , For the reason A parameterized predictive projector, optimized based on the classification target. It is a time-series encoder trained through contrastive learning. Its function is to transform raw or enhanced fused data (time series) into a high-quality representation, providing a foundation for subsequent tasks. These are the learnable parameters of the encoder, which are optimized by comparing them with the learning target.

[0026] In the contrastive learning framework, the rich diversity of augmentation samples is crucial for improving the generalization performance of the model. Entropy can be used to measure the uncertainty of the outcome contained in a random variable. Since the augmentation sample m is generated by fusing the original input x, an effective strategy to maintain high diversity in the augmentation process is to increase the conditional entropy H(m|x), which is the uncertainty of the augmentation result given the input. According to the definition of conditional entropy, the unconstrained entropy H(x) of the original input x itself may be affected by irrelevant noise, so this term can be ignored, and the focus should be on controlling the diversity of the augmentation samples. Under the continuous variable setting, in order to effectively constrain the mutual information between m and x, it can be achieved by minimizing its leave-one-out upper bound, specifically as follows: ； in, It is the input data fused from other operating conditions. The enhanced instance, x' is the data in the dataset excluding the fused data. Other instances, , and These are fused data , and The expression . Indicates the input fused data Expectation. It is a natural exponential function that maps the input to the space of positive real numbers.

[0027] Based on the definition of high-fidelity and diversity information perception, a standard is proposed to filter high-quality augmented data in the absence of prior knowledge, as follows:

[0028] in, To minimize the mutual information between the enhanced multi-source time series and the multi-source data, It is a hyperparameter used to weigh the trade-off between fidelity and diversity. It is the number of tags. s Let P(y=s) be the true label or pseudo label, representing the probability that the true label y belongs to category s. The true label is the actual value of carbon emissions from the power system. Indicates predicted label The probability of belonging to category s. These are the learnable parameters of the encoder. For predicting projectors, It is a time series encoder.

[0029] This embodiment uses precise measurement of the mutual information between the augmented data and the original data as the evaluation criterion to select the optimal augmentation method that can both retain the core features of the original data (high fidelity) and introduce rich and diverse variations (diversity). High-fidelity data augmentation maintains semantic characteristics that are invariant to transformations, while the sufficient diversity in high-diversity data augmentation improves the generalization ability of the contrastive learning model.

[0030] Existing time-series prediction techniques (such as TCN, Pyraformer, and Dsformer) mainly employ a single-domain processing approach. This involves achieving multi-resolution modeling in the time domain through downsampling, segmented sampling, or pyramidal attention, or decomposing in the frequency domain using Fourier transform. However, these methods have resolution limitations—time-domain methods only expand a single scale through dilated convolution, while frequency-domain methods lose high-frequency information due to low-pass filtering. To address this issue, this embodiment proposes a joint time-frequency domain expansion technique. Through multi-resolution collaborative modeling in both the time and frequency domains, it can capture high-frequency details such as transient fluctuations in power generation equipment, while also extracting low-frequency trends such as long-term changes in the energy structure. Furthermore, by combining fine-grained expansion in both the time and frequency domains to generate super-resolution features, it effectively avoids the "pick-up fence effect" of traditional frequency-domain methods, fully preserving key features of the carbon emission spectrum. Regarding information filtering mechanisms, existing models (such as SVR and Transformer) adopt static feature selection strategies, relying on manually designed contrastive learning or fixed attention weights, which are prone to generating redundant interference and invalid samples. In contrast, this solution innovatively designs a dynamic gating unit, which adaptively filters key features in the time and frequency domain based on the real-time operating conditions of the power system, significantly suppressing noise interference and enhancing the ability to fuse multi-source data.

[0031] Therefore, refer to Figure 2 The diagram shown is a flowchart of the frequency-time domain multi-resolution extension provided in this embodiment. Therefore, step S3, which involves multi-resolution extension of the frequency and time domain features of the enhanced multi-source time series, followed by dynamic feature selection through an information gating unit, and then connecting the selected features to generate super-resolution features, specifically includes the following steps: Step S31: Zero-padded the enhanced multi-source time series, and apply Fast Fourier Transform to the supplementary sequence obtained by zero-padded to obtain frequency domain features.

[0032] Step S32: Perform sequence interpolation on the enhanced multi-source time series, and sample the time series obtained by sequence interpolation at fixed intervals to obtain time domain features.

[0033] Step S33: Dynamically filter the frequency domain features and time domain features through the information gating unit, and connect the filtered features to generate super-resolution features.

[0034] In one specific embodiment, step S31 includes the following steps: Step S311: Zero-padded multi-source time series, and Fast Fourier Transform applied to the supplementary sequence obtained by zero-padded to obtain frequency series.

[0035] Specifically, the Fourier Transform (FFT) can only capture discrete frequency points in a sequence (such as integer multiples of the fundamental frequency), leading to the omission of spectral information (i.e., the "pick-and-fence effect"). However, carbon emission data in power systems contains high-frequency transient features and low-frequency long-term trends. Zero-filling refines the spectral resolution through frequency domain interpolation, avoiding the loss of key frequency components and accurately capturing multi-scale dynamic features of carbon emissions. It can obtain denser spectral lines, enabling neural networks to observe and capture previously undiscovered spectral components, avoiding the omission of key transient features due to the "pick-and-fence effect" in traditional methods.

[0036] First, zero-padding is applied to the sequence after data augmentation and reversible instance normalization. Then, a Fast Fourier Transform is applied to the supplemented sequence in the frequency domain, assuming the original signal length is... The length of the supplemented zeros is The Fast Fourier Transform is calculated as follows: ; in, For the enhanced multi-source time series, Indicates splicing, All zeros The multiple representing the zero fill factor is denoted as . The implementation of the Fast Fourier Transform is denoted as . , frequency dimension This indicates the result after the operation.

[0037] In this embodiment, the zero-padding operation is essentially interpolation in the frequency domain: it densifies the FFT output from X discrete frequency points to Y points. Specifically, before zero-padding, the frequency domain interval is... After padding with zeros, the interval encryption is... This adds (YX) interpolation points between the original X frequency points, resulting in a smoother and denser spectrum. This allows for more precise location and analysis of frequency characteristics while ensuring the complete preservation of the original signal's frequency information.

[0038] Step S312: The amplitude of the frequency sequence is determined by calculating the square root of the sum of the squares of the real and imaginary parts of the frequency sequence. Then, the top_k frequency components are selected based on the amplitude. The top_k method selects high-frequency components in the frequency domain to obtain frequency characteristics. This makes the trend of the sequence smoother, reduces sequence noise, and prevents overfitting.

[0039] Since frequency signals contain both valid information and noise, they need to be separated to capture global dependencies. By selecting the top_k frequency, the dominant frequency components are screened based on the sum of squared amplitudes, retaining the dominant frequency (fitted curve) that represents the trend of the sequence, and eliminating high-frequency noise.

[0040] Specifically, FFT is used to process sequences in the frequency domain. As shown below: ； ； ； ； in, Indicates the converted frequency. A represents frequency The approximate value of the amplitude, and yes The real and imaginary parts, This indicates that the most significant frequencies corresponding to the top k amplitude values ​​are selected based on the A value, where k is a hyperparameter.

[0041] In another embodiment, step S32 involves performing sequence interpolation on the enhanced multi-source time series and sampling the interpolated time series at fixed intervals to obtain time-domain features. Specifically, this includes the following steps: S321. Perform linear interpolation on the enhanced multi-source time series to obtain a linear interpolation sequence; S322, Based on retention rate Separate a portion of the interpolation sequence from the linear interpolation sequence; It should be noted that, to achieve extended time resolution, this embodiment employs a sequence interpolation method. First, linear interpolation is used to expand the dimensionality of the multi-source time series variables from... Change to ,in This represents a scaling factor. This type of interpolation does not change the overall trajectory or direction of the sequence, but rather complements the continuity of the sequence.

[0042] S323, Using a random mask matrix The separated interpolated sequence is sorted according to the retention rate. Pruning is performed, and the result is merged with the remaining interpolated sequence to obtain the time series. In this embodiment, since simple interpolation via S322 may lead to redundancy in the sequence within intervals with small data interval variations, a random mask matrix is ​​further designed for the interpolation part. Based on a certain retention rate Data pruning involves first separating the interpolated portion from the original data and randomly masking it, then merging the two to complete the interpolation operation. This is achieved using the following formula: ； ; in, Indicates a separation operation. This indicates a merge operation. Indicates based on retention rate A randomly initialized zero-one matrix, Indicates filtering The time step, of which This indicates iterating through each time step to determine... The first in the matrix If the value of each time step is 1, it is retained; otherwise, it is deleted. For the separated interpolated data, For high resolution expansion, The original time series data matrix is ​​given, where L is the number of time steps and D is the feature dimension of each time step. Let L be the time series data matrix after linear interpolation, where I is the interpolation scaling factor and L·I is the number of time steps after interpolation. This is a random mask matrix used to filter the interpolated data. γ is the retention rate of the interpolation points (0≤γ≤1), (I-1)·L is the total number of points generated by interpolation, and γ·(I-1)·L is the actual number of interpolation points retained. is the final data matrix after high-resolution expansion, and [γ·(I-1)+1]·L is the total number of time steps.

[0043] S324. Sample the time series at fixed intervals to obtain time-domain features.

[0044] Specifically, interval sampling is used to enable the network to capture more coarse and global information, and the resolution is reduced by using the opposite method of interpolation, as detailed below.

[0045] For length of Each time series is at fixed intervals. Sampling is performed to obtain a coarse and global low-resolution extension, denoted as... The sampling operation is as follows: ； ; in, This is a coarse-grained global information sequence generated by interval sampling. For the enhanced multi-source time series, For interval sampling functions, For the time points of the enhanced multi-source time series, For sampling operations, S is the sampling step size. To output the sequence length, round up to ensure all data is covered (when...). (When not an integer).

[0046] As a further preferred technical solution, the method further includes the following steps: Before steps S31 and S32, the enhanced multi-source time series must first undergo reversible instance normalization, followed by zero-padding and sequence interpolation.

[0047] In this embodiment, for the input data m, it first undergoes reversible instance normalization processing, and then is divided into two paths: frequency multi-resolution interaction and time multi-resolution interaction.

[0048] In one embodiment, when a sequence shift occurs, incorrect prediction biases may be learned during the direct extraction of time-frequency dependencies after time-frequency domain resolution expansion. Therefore, this embodiment utilizes information gating mechanisms as tools for information interaction during time-frequency multi-resolution expansion to capture the dependencies between expanded time-frequency features. These dependencies are then utilized and filtered to guide the network's prediction results. This reduces the network's sensitivity to data changes and enables it to learn more general patterns.

[0049] This embodiment captures key information that may be hidden in the original data through multi-resolution expansion in the time and frequency domains, and selectively integrates the expanded data through an information gating mechanism. In terms of information filtering mechanism, existing models (such as SVR and Transformer) adopt static feature selection strategies, relying on manually designed contrastive learning or fixed attention weights, which are prone to generating redundant interference and invalid samples. In contrast, this embodiment uses a dynamic gating mechanism to adaptively filter key features in the time and frequency domains based on the real-time operating conditions of the power system, which significantly suppresses noise interference and enhances the multi-source data fusion capability.

[0050] Specifically, the dynamic filtering features through the information gating unit are as follows: By fusing multi-resolution information from the frequency and time domains respectively, a corresponding weight matrix is ​​generated. The original observations (non-extended data) are then filtered using these weights to retain effective features. The weight matrix is ​​as follows: ； ; in , , and The learnable parameters representing the linear projection. Let A represent the softmax activation function, and let A represent the sum of squares of the real and imaginary parts of the frequency.

[0051] To extract effective features from the original observed signal, Hadamard operations are performed on the original signal using the corresponding weight matrices in both the time and frequency domains. This operation corrects signal distortion and removes filtering effects, thereby selecting the desired features. The Hadamard product is calculated as follows: ； ; in, This represents the Hadamard product. This represents the inverse fast Fourier transform.

[0052] The model ultimately aggregates the output features of all gated units and feeds them into a fully connected layer to generate the prediction results, as follows: ; in, For the splicing operation, time-domain and frequency-domain features obtained through information gating units are integrated. This method minimizes error deviation when data shifts or the expansion effect is poor. Prediction Results With complementary advantages, the time domain excels at capturing local dynamics and trends, while the frequency domain excels at characterizing global periodicity and steady-state patterns. The fusion of the two enables carbon emission prediction models to comprehensively model the complex characteristics of sequences from both time and frequency perspectives.

[0053] In this embodiment, the information gating unit essentially acts as a 'smart filter' added to the time-frequency multi-resolution extension process. It calculates the importance score of each input dimension through gating mechanisms (such as the sigmoid activation function and attention weights), determines which information is relevant to the current task or operating condition, retains key information (such as load trend characteristics in a power system) based on the scores, suppresses irrelevant information (such as occasional impulse noise), and finally outputs a filtered feature vector. This achieves dynamic adjustment of information flow to adapt to real-time operating conditions, solving the prediction bias caused by sequence offsets and enhancing the model's learning ability for key features. Ultimately, it serves to improve the accuracy of power system-related tasks (such as load forecasting and fault detection). For example, in a power system, the information gating unit can prioritize retaining time-frequency domain dependencies that match the current state based on real-time operating conditions (such as peak electricity consumption or equipment failure), while filtering out erroneous associations caused by sequence offsets.

[0054] In one embodiment, a loss function optimization strategy based on contrastive learning is used to train the carbon emission prediction model, enabling the model to learn the inherent invariant features of the data and enhancing its robustness to data noise and outliers, as detailed below: A batch of time series instances is obtained through step S1. ,for Each instance in An enhanced instance is generated through adaptively selected transformations. They are considered a pair of positive samples, while others are... Combinations (in yes Enhanced instances and Positive and negative samples are considered as negative samples; the neural network is trained using a contrastive learning loss function optimization strategy based on positive and negative samples, where the global contrastive loss function and the local contrastive loss function are as follows; Global contrastive loss aims to capture instance-level relationships in time series datasets. The global contrastive loss function is as follows: ; Local contrastive loss aims to explore temporal relationships within a time series; for time series instances An enhanced instance First, divide it into lengths of Given multiple subsequences N, for each subsequence A pair of positive samples is generated by selecting another subsequence that is close to it. Using non-adjacent samples Construct negative sample pairs, then, instance The local contrast loss is: ; in, 、 、 and These are examples 、 、 and The expression .

[0055] For all instances, there are ; The ultimate goal of comparison is: ; in, It is a hyperparameter used to weigh the global and local contrast loss.

[0056] In one embodiment, multi-source data corresponding to the period to be predicted is input into the optimized carbon emission prediction model. After complex calculations, the predicted value of carbon emissions from the power system in the future period is obtained. The prediction time scale can be set as needed, covering the next few hours to the next year. In this embodiment, the carbon emission prediction value output by the carbon emission prediction module is subjected to reversible instance normalization. This not only retains the optimization effect of instance normalization on model training, but also solves the "information loss" problem of ordinary normalization by retaining key parameters, ensuring that the final result can be restored to the original scale with physical meaning, while supporting multi-stage time series analysis tasks.

[0057] The power system source-side carbon emission prediction method based on time-frequency domain analysis provided by this invention has the following advantages: (1) By collecting multi-source data related to carbon emissions from various links of the power system source side, the fusion of multi-source data brings about a significant improvement in prediction accuracy. Compared with the traditional model that only relies on the single factor of power generation, the prediction accuracy is improved under complex power conditions. Since the time-frequency domain (time and frequency domain) characteristics of the carbon emission related data of the power system have multi-scale characteristics (such as hourly fluctuations and quarterly trends), the original multi-source data is difficult to express multi-resolution information explicitly. This invention expands the time-frequency domain (time and frequency domain) representation dimension by performing data augmentation on multi-source data. The sufficient differences in the enhanced multi-source time series variables can improve the generalization ability of the contrastive learning model. Furthermore, by performing time-frequency domain (time and frequency domain) multi-resolution joint analysis on the enhanced multi-source time series variables, through multi-resolution collaborative modeling in the time domain and frequency domain, it can capture high-frequency details such as transient fluctuations of power generation equipment, and extract low-frequency trends such as long-term changes in energy structure. It can better capture multi-resolution information in the time series, thereby improving the accuracy and reliability of prediction.

[0058] (2) The mutual information between the augmented data and the original data is accurately measured as the evaluation criterion. The optimal augmented data is selected that can retain the core features of the original data (high fidelity) and introduce rich and diverse changes. High fidelity data augmentation maintains the semantic characteristics that are invariant to transformations, and high diversity data augmentation has sufficient differences.

[0059] (3) By using the real-time operating conditions of the power system and adaptively filtering key features in the time and frequency domains based on the information gating unit, noise interference is significantly suppressed and the ability to fuse multi-source data is enhanced.

[0060] Device Examples According to embodiments of the present invention, a power system source-side carbon emission prediction device based on time-frequency domain analysis is provided, such as... Figure 3The diagram shown is a block diagram of a power system source-side carbon emission prediction device based on time-frequency domain analysis provided in this embodiment. According to an embodiment of the present invention, a power system source-side carbon emission prediction device based on time-frequency domain analysis includes: The data acquisition and processing module 10 is used to acquire multi-source data related to carbon emissions from the power system source side, and to perform normalization, sampling frequency alignment and feature splicing operations on the multi-source data to obtain fused data. Data augmentation module 20 is used to enhance the fused data. It improves the fidelity of the enhanced data by minimizing cross-entropy and enhances the diversity of the enhanced data by minimizing mutual information, thereby obtaining an enhanced multi-source time series, where the true label is the actual value of carbon emissions from the power system. The time-frequency domain multi-resolution extension module 30 is used to perform multi-resolution extension on the frequency domain features and time domain features of the enhanced multi-source time series, and then dynamically filter the features through the information gating unit, and connect the filtered features to generate super-resolution features. The carbon emission prediction module 40 is used to obtain the predicted value of carbon emissions by inputting super-resolution features into the carbon emission prediction model. The carbon emission prediction model is a fully connected neural network trained based on the model training convergence condition.

[0061] The power system source-side carbon emission prediction method provided in this embodiment, based on time-frequency domain analysis, collects multi-source data related to carbon emissions from various stages of the power system source side through the data acquisition and processing module 10. The fusion of multi-source data significantly improves prediction accuracy, offering better prediction accuracy under complex power conditions compared to traditional models that rely solely on power generation. Because the time-domain and frequency-domain characteristics of power system carbon emission-related data have multi-scale features (such as hourly fluctuations and quarterly trends), the original multi-source data is difficult to explicitly express multi-resolution information. Therefore, the data augmentation module 20 enhances the multi-source data to expand the representation dimensions in the time and frequency domains, allowing for sufficient differentiation in the enhanced multi-source time series. The generalization ability of the contrastive learning model is improved; then, the mutual information between the enhanced data and the original data is used as the evaluation criterion to select the optimal enhanced data that can retain the core features of the original data and introduce diversity. The high-fidelity data enhancement maintains the semantic characteristics that are invariant to transformations, and the high-diversity data enhancement has sufficient differences. The time-frequency domain multi-resolution extension module 30 performs time-frequency domain multi-resolution joint analysis on the enhanced multi-source time series. Through multi-resolution collaborative modeling in the time and frequency domains, it can capture high-frequency details such as transient fluctuations of power generation equipment and extract low-frequency trends such as long-term changes in energy structure. It can better capture multi-resolution information in the time series, thereby improving the accuracy and reliability of prediction.

[0062] In one embodiment, the data augmentation module 20 enhances the fused data using probability functions of multi-source data and random variables as the augmentation method, and minimizes the upper bound of the leave-one-out method ( Leave-One-Out To minimize the mutual information between the enhanced fused data and the real labels, the enhanced multi-source time series is obtained. The specific enhancement method is as follows: ; in, m To enhance data fusion, x For a time series instance, is a T×F matrix, where T represents the length of the time series and F represents the feature dimension of each time step. Let be a probability function.

[0063] The objective in this embodiment is to minimize mutual information. Its definition is ,because yes The (Shannon) entropy is independent of data augmentation; the objective is equivalent to minimizing conditional entropy. .

[0064] By minimizing the upper bound of the leave-one-out method ( Leave-One-Out To minimize mutual information, the following formula is used:

[0065] in, For input data from other operating conditions The enhanced instance, x' is the data in the dataset excluding the fused data. Other instances, , and Fusion data , and The expression, Indicates the input fused data Expectation It is a natural exponential function that maps the input to the space of positive real numbers.

[0066] Based on the definitions of high-fidelity and diversity information perception, a criterion for selecting good enhancement is proposed in the absence of prior knowledge, as follows:

[0067] in, To minimize the mutual information between the enhanced multi-source time series and the multi-source data, It is a hyperparameter used to weigh the trade-off between fidelity and diversity. For the number of tags,s Let P(y=s) be the true label or pseudo label, representing the probability that the true label y belongs to category s. The true label is the actual value of carbon emissions from the power system. Indicates predicted label The probability of belonging to category s. These are the learnable parameters of the encoder. For predicting projectors, It is a time series encoder.

[0068] In one embodiment, the time-frequency domain multi-resolution extension module 30 includes a time-domain resolution extension submodule, a frequency-domain resolution extension submodule, and a super-resolution feature generation submodule: The frequency domain resolution extension submodule is used to zero-padded the enhanced multi-source time series and apply Fast Fourier Transform to the supplementary sequence obtained by zero-padded to obtain frequency domain features.

[0069] The temporal resolution extension submodule is used to perform sequence interpolation on the enhanced multi-source time series and sample the time series obtained by sequence interpolation at fixed intervals to obtain temporal features.

[0070] The super-resolution feature generation submodule is used to dynamically filter frequency domain features and time domain features through the first and second information gating units, and then connect the filtered features to generate super-resolution features.

[0071] In one specific embodiment, the frequency domain resolution extension submodule includes a frequency domain extension unit and a frequency domain feature determination unit, wherein, The frequency domain extension unit is used to zero-padded the enhanced multi-source time series and apply Fast Fourier Transform to the supplementary sequence obtained by zero-padded to obtain the frequency sequence.

[0072] The Fast Fourier Transform is calculated as follows: ; in, For the enhanced multi-source time series, Indicates splicing, All zeros The multiple representing the zero fill factor is denoted as . The implementation of the Fast Fourier Transform is denoted as . , frequency dimension This indicates the result after the operation.

[0073] The frequency domain feature determination unit is used to determine the amplitude of the frequency sequence by calculating the square root of the sum of the squares of the real and imaginary parts of the frequency sequence, and then select the top_k frequency components based on the amplitude. The top_k method selects high-frequency components in the frequency domain to make the trend of the sequence smoother, reduce sequence noise, and prevent overfitting, thereby obtaining frequency domain features.

[0074] Specifically, FFT is used to process sequences in the frequency domain. As shown below: ; ; ; ; in, Indicates the converted frequency. A represents frequency. The approximate value of the amplitude, and yes The real and imaginary parts, This indicates that the most significant frequencies corresponding to the top k amplitude values ​​are selected based on the A value, where k is a hyperparameter.

[0075] In another embodiment, the time-domain resolution extension submodule includes a sequence interpolation processing unit, a sequence separation unit, and a time series determination unit.

[0076] The sequence interpolation processing unit is used to perform linear interpolation on the enhanced multi-source time series to obtain a linear interpolated sequence; Sequence separation unit, used to determine the retention rate Separate a portion of the interpolation sequence from the linear interpolation sequence; It should be noted that, to achieve extended time resolution, this embodiment employs a sequence interpolation method. First, linear interpolation is used to expand the dimensionality of the multi-source time series variables from... Change to ,in This represents a scaling factor. This type of interpolation does not change the overall trajectory or direction of the sequence, but rather complements the continuity of the sequence.

[0077] Time series determination unit, used to employ a random mask matrix The separated interpolated sequence is sorted according to the retention rate. Pruning is performed, and the result is merged with the remaining interpolated sequence to obtain the time series. This embodiment further incorporates a random mask matrix for the interpolation part. Based on a certain retention rate Data pruning involves first separating the interpolated portion from the original data and randomly masking it, then merging the two to complete the interpolation operation. This is achieved using the following formula: ; ; in, Indicates a separation operation. This indicates a merge operation. Indicates based on retention rate A randomly initialized zero-one matrix, Indicates filtering The time step, of which This indicates iterating through each time step to determine... The first in the matrix If the value of each time step is 1, it is retained; otherwise, it is deleted. For the separated interpolated data, For high resolution expansion, The original time series data matrix is ​​given, where L is the number of time steps and D is the feature dimension of each time step. Let L be the time series data matrix after linear interpolation, where I is the interpolation scaling factor and L·I is the number of time steps after interpolation. This is a random mask matrix used to filter the interpolated data. γ is the retention rate of the interpolation points (0≤γ≤1), (I-1)·L is the total number of points generated by interpolation, and γ·(I-1)·L is the actual number of interpolation points retained. is the final data matrix after high-resolution expansion, and [γ·(I-1)+1]·L is the total number of time steps.

[0078] S324. Sample the time series at fixed intervals to obtain time-domain features.

[0079] Specifically, interval sampling is used to enable the network to capture more coarse and global information, and the resolution is reduced by using the opposite method of interpolation, as detailed below.

[0080] For length of Each time series is at fixed intervals. Sampling is performed to obtain a coarse and global low-resolution extension, denoted as... The sampling operation is as follows: ; ; in, This is a coarse-grained global information sequence generated by interval sampling. For the enhanced multi-source time series, For interval sampling functions, For the time points of the enhanced multi-source time series, For sampling operations, S is the sampling step size. To output the sequence length, round up to ensure all data is covered (when...). (When not an integer).

[0081] The dynamic feature selection through the first and second information gating units is as follows: By fusing multi-resolution information to generate a weight matrix, the original observations (non-extended data) are filtered using the weights to retain effective features. Then, by performing Hadamard product on the corresponding weight matrix W and the original observations in the time and frequency domains, and by performing de-filtering and correction processing on the extended resolution values, the selected features are obtained.

[0082] In this preferred embodiment, a model training module is also included, which uses a loss function optimization strategy based on contrastive learning to train the carbon emission prediction model, enabling the model to learn the inherent invariant features of the data and enhance its robustness to data noise and outliers.

[0083] The embodiments of the present invention are device embodiments corresponding to the above method embodiments. The specific operations of each module processing step can be understood with reference to the description of the method embodiments, and will not be repeated here.

[0084] like Figure 4 As shown, the present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the power system source-side carbon emission prediction method based on time-frequency domain analysis in the above embodiments.

[0085] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the power system source-side carbon emission prediction method based on time-frequency domain analysis in the above embodiments.

[0086] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0087] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, for apparatus or system embodiments, since they are basically similar to method embodiments, the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments. The apparatus and system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0088] Furthermore, the functional modules in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0089] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. These modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and the contents not described in detail in the specification of the present invention are well known to those skilled in the art.

Claims

1. A method for predicting source-side carbon emissions from power systems based on time-frequency domain analysis, characterized in that, Includes the following steps: Collect multi-source data related to carbon emissions from the power system source side, and perform normalization, sampling frequency alignment, and feature stitching operations on the multi-source data to obtain fused data; The fused data is enhanced by minimizing cross-entropy to improve the fidelity of the enhanced data and by minimizing mutual information to improve the diversity of the enhanced data, resulting in an enhanced multi-source time series, where the true label is the actual value of carbon emissions from the power system. After multi-resolution expansion of the frequency domain features and time domain features of the enhanced multi-source time series, the features are dynamically filtered through information gating units, and the filtered features are connected to generate super-resolution features. The carbon emission prediction model is based on super-resolution features as input to obtain the predicted carbon emission value. The carbon emission prediction model is a fully connected neural network trained based on the model training convergence condition.

2. The power system source-side carbon emission prediction method based on time-frequency domain analysis as described in claim 1, characterized in that, The enhancement processing of the fused data is achieved through probability functions of multi-source data and random variables.

3. The power system source-side carbon emission prediction method based on time-frequency domain analysis as described in claim 1, characterized in that, The specific steps for performing multi-resolution expansion on the frequency domain and time domain features of the enhanced multi-source time series, followed by dynamic feature selection through information gating units, and then concatenating the selected features to generate super-resolution features are as follows: Zero-padding is performed on the enhanced multi-source time series, and fast Fourier transform is applied to the supplementary sequence obtained by zero-padding to obtain frequency domain features. The enhanced multi-source time series is subjected to sequence interpolation, and the time series obtained by sequence interpolation is sampled at fixed intervals to obtain time domain features; The frequency domain features and time domain features are dynamically filtered through information gating units, and the filtered features are then connected to generate super-resolution features.

4. The power system source-side carbon emission prediction method based on time-frequency domain analysis as described in claim 3, characterized in that, The specific steps for zero-padding the enhanced multi-source time series and applying Fast Fourier Transform to the supplementary sequence to obtain frequency domain features are as follows: Zero-padding is performed on the enhanced multi-source time series, and a fast Fourier transform is applied to the supplementary sequence obtained by zero-padding to obtain the frequency sequence. The amplitude of the frequency sequence is determined by calculating the square root of the sum of the squares of the real and imaginary parts of the frequency sequence; then, the top_k frequency components are selected based on the amplitude, and high-frequency components are selected in the frequency domain to obtain the frequency characteristics.

5. The power system source-side carbon emission prediction method based on time-frequency domain analysis as described in claim 3, characterized in that, The specific steps for performing sequence interpolation on the enhanced multi-source time series and sampling the interpolated time series at fixed intervals to obtain time-domain features are as follows: Linear interpolation is performed on the enhanced multi-source time series to obtain a linear interpolation sequence; Based on retention rate Separate a portion of the interpolation sequence from the linear interpolation sequence; The separated interpolated sequences are pruned according to the retention rate using a random mask matrix and then merged with the remaining interpolated sequences to obtain the time series. Time-domain features are obtained by sampling the time series at fixed intervals.

6. The power system source-side carbon emission prediction method based on time-frequency domain analysis as described in claim 1, characterized in that, It also includes the step of training the carbon emission prediction model using a loss function optimization strategy based on contrastive learning.

7. A power system source-side carbon emission prediction device based on time-frequency domain analysis, characterized in that, include: The data acquisition and processing module is used to collect multi-source data related to carbon emissions from the power system source side, and to perform normalization, sampling frequency alignment and feature stitching operations on the multi-source data to obtain fused data. The data augmentation module is used to enhance the fused data. It improves the fidelity of the augmented data by minimizing cross-entropy and enhances the diversity of the augmented data by minimizing mutual information, thereby obtaining the augmented multi-source time series. The true label is the actual value of carbon emissions from the power system. The time-frequency domain multi-resolution extension module is used to perform multi-resolution extension on the frequency domain features and time domain features of the enhanced multi-source time series, and then dynamically filter the features through the information gating unit, and connect the filtered features to generate super-resolution features. The carbon emission prediction module is used to obtain the predicted value of carbon emissions by inputting super-resolution features into the carbon emission prediction model. The carbon emission prediction model is a fully connected neural network trained based on the model training convergence condition.

8. The power system source-side carbon emission prediction device based on time-frequency domain analysis as described in claim 7, characterized in that, The time-frequency domain multi-resolution extension module includes a time-domain resolution extension submodule, a frequency-domain resolution extension submodule, and a super-resolution feature generation submodule; The frequency domain resolution extension submodule is used to zero-patch the enhanced multi-source time series and apply Fast Fourier Transform to the supplementary sequence obtained by zero-patch to obtain frequency domain features. The temporal resolution extension submodule is used to perform sequence interpolation on the enhanced multi-source time series and sample the time series obtained by sequence interpolation at fixed intervals to obtain temporal features. The super-resolution feature generation submodule is used to dynamically filter frequency domain features and time domain features through the first and second information gating units, and then connect the filtered features to generate super-resolution features.

9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the power system source-side carbon emission prediction method based on time-frequency domain analysis as described in any one of claims 1 to 6.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the power system source-side carbon emission prediction method based on time-frequency domain analysis as described in any one of claims 1 to 6.

Images