A power system source side carbon emission prediction method and system

Abstract

The present disclosure relates to a power system source side carbon emission prediction method and system, the method comprising: collecting multi-source data; data preprocessing of the collected multi-source data, the data preprocessing comprising data alignment, outlier processing and normalization processing; based on wavelet decomposition method and compressed sensing reconstruction algorithm, time-frequency expansion is performed on the multi-source data after data preprocessing to obtain a high-resolution wavelet coefficient matrix; the high-resolution wavelet coefficient matrix is high-resolution time-frequency data; based on the multi-scale physical properties of carbon emission and the high-resolution wavelet coefficient matrix, enhanced data are generated which conform to the physical law and cover various scenarios; based on the resolution branch design, precise feature screening is performed on the enhanced data to obtain a plurality of high-resolution branch prediction sequence matrices and output; all branch prediction sequence matrices are respectively reconstructed into corresponding time-frequency features through inverse wavelet transform, and each time-frequency feature is fused based on a preset attention mechanism to obtain a carbon emission prediction value.

Description

Technical Field

[0001] This disclosure relates to the field of electricity carbon emission prediction, specifically to a method and system for predicting carbon emissions from the source side of an electricity system. Background Technology

[0002] Power system source-side carbon emission forecasting is a key technological support for achieving the "dual carbon" goals. Its core is to predict the total amount and trend of carbon emissions from power sources in the future, based on multi-source information such as historical carbon emission data, power generation capacity, and fuel consumption. This provides a basis for decision-making in grid dispatching, fuel procurement, and carbon emission quota allocation. Currently, power system source-side carbon emission forecasting faces the following technological bottlenecks: (1) Insufficient ability to capture multi-scale features: Source-side carbon emissions are affected by multiple factors and exhibit significant multi-scale characteristics—long-term scale (seasonal changes, such as increased carbon emissions due to increased thermal power output in winter), medium-term scale (weekly / daily cycles, such as full thermal power generation during peak load periods), and short-term scale (hourly fluctuations, such as fine-tuning of carbon emissions due to unit start-up and shutdown or sudden changes in wind and solar power output). Traditional prediction methods (such as ARIMA and LSTM) often only focus on time-domain features and are difficult to capture multi-resolution information in the time and frequency domains simultaneously, easily losing key frequency components (such as low-frequency carbon emission fluctuations corresponding to unit maintenance cycles).

[0003] (2) Lack of physical constraints in data augmentation: Source-side carbon emission data suffers from limited sample size and missing data for certain time periods (e.g., sensor malfunctions), requiring data augmentation to improve model generalization capabilities. However, existing general time series augmentation methods (e.g., random jittering, time shuffling) do not consider the physical laws of the power system (e.g., linear positive correlation between carbon emissions from thermal power and fuel consumption, and negative correlation between wind and solar power output and thermal power supplementation), which may lead to data distortion and reduced prediction accuracy.

[0004] (3) Imbalance between model complexity and engineering practicality: Some high-precision models (such as Transformer-type models) can capture long-range dependencies, but they have large parameter scale and high computational cost, making them difficult to deploy on the real-time monitoring platform of the power system; while lightweight models (such as DLinear) are computationally efficient, but they lack the ability to fuse multi-source heterogeneous data and are difficult to adapt to the complex carbon emission mechanism on the source side. Summary of the Invention

[0005] To address the problems existing in the prior art, this disclosure proposes a method and system for predicting carbon emissions from the source side of a power system, thereby solving at least one of the aforementioned technical problems. The technical solution adopted in this disclosure is as follows: In a first aspect, this disclosure provides a method for predicting source-side carbon emissions from a power system, the method comprising: Collect multi-source data, including historical carbon emission data, power generation data, system load data, load cycle data, boiler and equipment parameters, and fuel-related data. The collected multi-source data is preprocessed, including data alignment, outlier handling, and normalization. Based on wavelet decomposition and compressed sensing reconstruction algorithm, time-frequency extension is performed on the preprocessed multi-source data to obtain a high-resolution wavelet coefficient matrix; the high-resolution wavelet coefficient matrix is ​​high-resolution time-frequency data. Based on the multi-scale physical characteristics of carbon emissions and the high-resolution wavelet coefficient matrix, enhanced data that conforms to physical laws and covers diverse scenarios is generated. Based on the resolution branch design, the enhanced data is accurately filtered to obtain multiple sets of high-resolution branch prediction sequence matrices and output them. Each set of branch prediction sequence matrices corresponds to a resolution branch. All the branch prediction sequence matrices are reconstructed into corresponding time-frequency features through wavelet inverse transform. The time-frequency features are then fused based on a preset attention mechanism to obtain the carbon emission prediction value.

[0006] Preferably, the historical carbon emission data includes total carbon emissions from the source side and time-series variation data (hourly level).

[0007] Preferably, the power generation data includes thermal power output data (including hourly output and ramp-up deviation data) and new energy (wind and solar) output data.

[0008] Preferably, the system load data (hourly level) includes peak and valley loads and load peak and valley deviations.

[0009] Preferably, the load cycle data (daily / weekly) includes weekday / weekend load differences and seasonal load trends.

[0010] Preferably, the boiler and equipment parameters include unit operating status data (unit start-up and shutdown status, basic value of ramp rate, etc.) and equipment maintenance status data.

[0011] Preferably, the fuel-related data includes fuel consumption time-series data (synchronized and aligned with the power generation data and historical carbon emission data) and fuel characteristic parameters (fuel calorific value and carbon emission factor).

[0012] It is understandable that the fuel consumption time-series data, power generation data, and historical carbon emission data are synchronized and aligned.

[0013] Preferably, the data alignment includes: synchronizing each data in the multi-source data according to timestamps, unifying them into an hourly time series, and supplementing missing timestamps with "forward padding + linear interpolation", thereby avoiding disruption of the continuity of unit operation.

[0014] Preferably, the outlier processing includes: identifying and correcting outlier data in the data-aligned multi-source data; such as sudden increases / decreases in carbon emissions due to sensor malfunctions—if the outlier deviates from the mean by 3 The above uses a weighted average of three adjacent time periods for replacement.

[0015] Preferably, the normalization process includes: performing reversible instance normalization on the multi-source data after outlier processing to eliminate distribution bias in each time series.

[0016] Preferably, the step of performing time-frequency extension on the preprocessed multi-source data based on wavelet decomposition, wavelet domain sparsity modeling, compressed sensing, and sparse reconstruction techniques to obtain a high-resolution wavelet coefficient matrix specifically includes the following steps: Based on the wavelet decomposition method, the multi-source data after data preprocessing is decomposed into low-frequency approximation coefficients and high-frequency detail coefficients. The low-frequency approximation coefficients and high-frequency detail coefficients are combined into a wavelet coefficient matrix, and the wavelet coefficient matrix is ​​represented as a linear combination of sparse coefficients through a wavelet domain sparse basis. Design a measurement matrix that satisfies the finite isometry property, perform linear measurements on the sparse coefficients, and obtain low-dimensional measurement values; The low-dimensional measurement values ​​are sparsely reconstructed based on the orthogonal matching pursuit algorithm to obtain the reconstructed sparse coefficients. A high-resolution wavelet coefficient matrix is ​​obtained by inverse transformation of the wavelet domain sparse basis and the reconstructed sparse coefficients. The high-resolution wavelet coefficient matrix is ​​output as the high-resolution wavelet coefficient matrix.

[0017] Preferably, the generation of enhanced data based on the multi-scale physical characteristics of carbon emissions and the high-resolution wavelet coefficient matrix, which conforms to physical laws and covers diverse scenarios, specifically includes the following steps: Based on the multi-scale physical characteristics of carbon emissions, multi-scale physical driving factors are analyzed from the multi-source data. Several sets of adjustable physical parameters are designed for each scale physical driving factor, as shown below: Fine-grained physical driving factors include unit ramping and fuel combustion fluctuations, and the corresponding adjustable physical parameters include ramping rate correction coefficient and fuel combustion efficiency fluctuation coefficient. Medium-grained physical drivers include load peaks and valleys, and intraday fluctuations in wind and solar power. Corresponding adjustable physical parameters include load peak-valley ratio correction coefficient and new energy penetration rate coefficient. Coarse-grained physical drivers include meteorological trends and load cycles, and the corresponding adjustable physical parameters include daily carbon emission baseline correction coefficients and cycle offset coefficients. Based on the high-resolution wavelet coefficient matrix and each of the adjustable physical parameters, coarse-grained baseline tuning, medium-grained fluctuation superposition, and fine-grained detail supplementation are performed sequentially to generate enhanced data; Verify the validity of the augmented data and filter out invalid sample data in the augmented data.

[0018] Preferably, the multi-scale physical driving factors include fine-grained physical driving factors, medium-grained physical driving factors, and coarse-grained physical driving factors.

[0019] Preferably, the fine-grained physical driving factors include unit ramping and fuel combustion fluctuations, and the corresponding adjustable physical parameters include ramping rate correction coefficient and fuel combustion efficiency fluctuation coefficient.

[0020] Preferably, the medium-grained physical driving factors include load peaks and valleys, and intraday fluctuations in wind and solar power, and the corresponding adjustable physical parameters include the load peak-valley ratio correction coefficient and the new energy penetration rate coefficient.

[0021] Preferably, the coarse-grained physical driving factors include meteorological trends and load cycles, and the corresponding adjustable physical parameters include daily carbon emission baseline correction coefficients and cycle offset coefficients.

[0022] Preferably, the step of performing precise feature filtering on the enhanced data based on the resolution-based branch design to obtain and output multiple sets of high-resolution branch prediction sequence matrices specifically includes the following steps: Each coefficient sequence in the augmented data is passed through a corresponding resolution branch, and all resolution branches are independent of each other, thereby avoiding mutual interference between features at different scales. Each coefficient sequence is divided into corresponding overlapping patches; Each patch is embedded into a 3D feature space through a linear layer to capture multiple local carbon emission features; A multilayer perceptron is used to capture the local dependencies between different local carbon emission characteristics to obtain local dependency features; By employing a channel attention mechanism, precise feature selection is achieved by strengthening the feature channels corresponding to the key carbon emission parameters in the local dependent features and suppressing the responses of invalid channels, resulting in a set of corresponding branch approximation coefficients and branch detail coefficients output by each sub-resolution branch. The corresponding branch approximation coefficients and branch detail coefficients of each group are combined into a set of branch prediction sequence matrices and output.

[0023] Preferably, the step of reconstructing the corresponding time-frequency features of all the branch prediction sequence matrices through wavelet inverse transform, and fusing the time-frequency features based on a preset attention mechanism to obtain the carbon emission prediction value includes the following steps: Based on the wavelet inverse transform reconstruction technique, the corresponding time-frequency features are reconstructed for each group of branch prediction sequence matrices; Based on a preset attention mechanism, each of the time-frequency features is assigned a corresponding weight and then fused to obtain a fused feature. After passing the fused features through at least two layers of pre-defined fully connected networks, the predicted carbon emissions for the next T steps are output.

[0024] Preferably, T is determined according to actual needs, and is preferably between 10 and 100.

[0025] Preferably, in the method, a total loss function can be constructed using the SmoothL1 loss function and the contrastive loss function, and the method can be optimized using the total loss function and the Adam optimizer.

[0026] A second aspect of this disclosure provides a power system source-side carbon emission prediction system, the system comprising: The data acquisition module is used to collect multi-source data, including historical carbon emission data, power generation data, system load data, load cycle data, boiler and equipment parameters, and fuel-related data. The data preprocessing module is used to preprocess the collected multi-source data, including data alignment, outlier handling, and normalization. The time-frequency extension module is used to perform time-frequency extension on the pre-processed multi-source data based on wavelet decomposition and compressed sensing reconstruction algorithm to obtain a high-resolution wavelet coefficient matrix; the high-resolution wavelet coefficient matrix is ​​high-resolution time-frequency data. The data augmentation module is used to generate augmented data that conforms to physical laws and covers diverse scenarios based on the multi-scale physical characteristics of carbon emissions and the high-resolution wavelet coefficient matrix. A multi-resolution module is used to perform precise feature filtering on the enhanced data based on the sub-resolution branch design, obtain multiple sets of high-resolution branch prediction sequence matrices and output them, with each set of branch prediction sequence matrices corresponding to a sub-resolution branch. The data prediction module is used to reconstruct the corresponding time-frequency features of all the branch prediction sequence matrices through wavelet inverse transform, and to fuse the time-frequency features based on a preset attention mechanism to obtain the carbon emission prediction value.

[0027] Preferably, the system may further include: The model optimization module is used to construct a total loss function using the SmoothL1 loss function and the contrastive loss function, and to optimize the system using the total loss function and the Adam optimizer.

[0028] In a third aspect, this disclosure provides 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 as described above.

[0029] In a fourth aspect, this disclosure provides an electronic device including a processor and a memory, the processor being configured to execute a computer program stored in the memory to implement the power system source-side carbon emission prediction method as described above.

[0030] The beneficial effects of this disclosure are as follows: This disclosure proposes a method and system for predicting carbon emissions from the source side of power systems. By constructing a prediction model that integrates time-frequency extension, physically constrained information enhancement, and resolution feature optimization, it solves the problems of accuracy and practicality in predicting carbon emissions from the source side of power systems. Addressing the problems of incomplete multi-scale feature capture, data enhancement distortion, multi-resolution information loss, and high model complexity in existing power system source-side carbon emission prediction methods, this disclosure innovatively proposes a six-stage carbon emission data prediction mechanism covering "data acquisition - data preprocessing - time-frequency extension - information perception enhancement - multi-resolution feature extraction - feature fusion and prediction".

[0031] In the system provided in this disclosure, the data preprocessing module is responsible for aligning and normalizing multi-source data, eliminating data distribution offset; the time-frequency extension module, through wavelet decomposition, compressed sensing, and wavelet domain sparse reconstruction, achieves the supplementation of implicit information in the time domain, preservation of macro trends, mitigation of the frequency domain "fence effect," and capture of implicit frequencies; the information perception enhancement module generates enhanced data based on coarse-grained scene modeling, fine-grained parameter collaboration, and a self-supervised physical verification mechanism; the resolution feature extraction module, based on a resolution branch structure and combined with a channel attention mechanism, strengthens the feature channels corresponding to key carbon emission information and suppresses invalid channel responses, thereby achieving accurate feature selection; and the feature fusion and prediction module outputs the final carbon emission prediction result through wavelet inverse transform and attention fusion. Attached Figure Description

[0032] The accompanying drawings, which form part of this application, are used to provide a further understanding of this disclosure. The illustrative embodiments of this disclosure and their descriptions are used to explain this disclosure and do not constitute an undue limitation of this disclosure.

[0033] Figure 1A flowchart of a power system source-side carbon emission prediction method provided in Embodiment 1 of this disclosure.

[0034] Figure 2 This is an architecture diagram of a power system source-side carbon emission prediction system provided in Embodiment 2 of this disclosure. Detailed Implementation

[0035] The present disclosure will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in the present application can be combined with each other.

[0036] The following detailed descriptions are exemplary and intended to provide further detailed explanation of this disclosure. Unless otherwise specified, all technical terms used in this disclosure have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. The terminology used in this disclosure is for the purpose of describing particular embodiments only and is not intended to limit the exemplary embodiments according to this disclosure.

[0037] Example 1: like Figure 1 As shown, this disclosure provides a method for predicting carbon emissions from the source side of a power system, the method comprising steps S100 to S600.

[0038] Step S100: Collect multi-source data, including historical carbon emission data, power generation data, system load data, load cycle data, boiler and equipment parameters, and fuel-related data.

[0039] Specifically, the historical carbon emission data includes total carbon emissions from the source side and time-series variation data (hourly level).

[0040] Specifically, the power generation data includes thermal power output data (including hourly output and ramp-up deviation data) and new energy (wind and solar) output data.

[0041] Specifically, the system load data (hourly level) includes peak and valley loads and load peak-valley deviations.

[0042] Specifically, the load cycle data (daily / weekly level) includes weekday / weekend load differences and seasonal load trends.

[0043] Specifically, the boiler and equipment parameters include unit operating status data (unit start-up and shutdown status, basic value of ramp rate, etc.) and equipment maintenance status data.

[0044] Specifically, the fuel-related data includes fuel consumption time-series data (synchronized and aligned with the power generation data and historical carbon emission data) and fuel characteristic parameters (fuel calorific value and carbon emission factor).

[0045] It should be noted that the fuel consumption time-series data, power generation data, and historical carbon emission data are synchronized and aligned.

[0046] Step S200: Perform data preprocessing on the collected multi-source data. The data preprocessing includes data alignment, outlier handling, and normalization.

[0047] Specifically, the data alignment includes: synchronizing each data in the multi-source data according to timestamps, unifying them into an hourly time series, and supplementing missing timestamps with "forward padding + linear interpolation" to avoid disrupting the continuity of unit operation.

[0048] Specifically, the outlier handling includes: identifying and correcting outlier data in the multi-source data after data alignment; such as sudden increases / decreases in carbon emissions due to sensor malfunctions—if the outlier deviates from the mean by 3... The above uses a weighted average of three adjacent time periods for replacement.

[0049] Specifically, the normalization process includes: performing reversible instance normalization on the multi-source data after outlier processing to eliminate distribution bias in each time series.

[0050] Furthermore, the formula used for the reversible instance normalization is as follows: ; Where X represents the data after normalization of the reversible instance; x represents the data for reversible instance normalization; , Here are the mean and standard deviation of the data; , Since the parameters are learnable, the original scale can be recovered through inverse transformation during the prediction stage, thereby ensuring the physical meaning of the prediction results.

[0051] Step S300: Based on wavelet decomposition and compressed sensing reconstruction algorithm, time-frequency extension is performed on the preprocessed multi-source data to obtain a high-resolution wavelet coefficient matrix; the high-resolution wavelet coefficient matrix is ​​high-resolution time-frequency data.

[0052] Specifically, step S300, based on wavelet decomposition, wavelet domain sparsity modeling, compressed sensing, and sparse reconstruction, performs time-frequency extension on the preprocessed multi-source data to obtain a high-resolution wavelet coefficient matrix, specifically including steps S310 to S360.

[0053] Step S310: Preprocess the multi-source data based on wavelet decomposition. Multi-scale wavelet decomposition is performed to obtain low-frequency approximation coefficients. and high frequency detail coefficient .

[0054] Specifically, the low-frequency approximation coefficients Corresponding to long-term trends in carbon emissions (such as seasonal variations); The high-frequency detail factor This corresponds to short-term fluctuations (such as unit start-up and shutdown, sudden changes in wind and solar power output, etc.). Where C is the feature dimension, such as power generation, fuel consumption, etc.; L is the sequence length. This represents the length of the sequence after m layers of wavelet decomposition.

[0055] Furthermore, the multi-scale wavelet decomposition can be achieved by iteratively applying a high-pass filter and a low-pass filter, as shown in the following expression: ; in, This represents the low-frequency approximation coefficients after the k-th level decomposition. Let represent the high-frequency detail coefficients after the k-th layer decomposition, h be the low-pass filter coefficients, and g be the high-pass filter coefficients. This is a convolution operation.

[0056] Step S320: Convert the low-frequency approximation coefficients and high frequency detail coefficient Combined into wavelet coefficient matrix Through wavelet domain sparse basis The wavelet coefficient matrix Represented as sparsity coefficient The linear combinations of are shown below: ; ; in, The wavelet coefficient matrix; These are low-frequency approximation coefficients; For high-frequency detail coefficients; The sparsity coefficient is . It is a sparse coefficient matrix; , which is a sparse basis in the wavelet domain.

[0057] Step S330: Design a measurement matrix that satisfies the finite isometric property. For sparsity coefficients Perform linear measurements to obtain low-dimensional measurement values. The expression can be as follows: ; Where M represents the measurement dimension.

[0058] Furthermore, the measurement matrix A Gaussian random matrix can be selected, which satisfies the finite isometry property condition for any sparse basis, thus fitting the characteristics of carbon emission data. Through step S330, only a small number of measurements need to be retained. It is much shorter than the length of the original wavelet coefficients. This allows for subsequent high-resolution reconstruction, avoiding traditional redundant calculations.

[0059] Step S340: Based on the orthogonal matching pursuit algorithm, analyze the low-dimensional measurement values. Perform sparse reconstruction to obtain the reconstructed sparse coefficients. .

[0060] Understandably, the sparse coefficients after reconstruction The core is to minimize the sparsity coefficient. of The orthogonal matching pursuit algorithm can be defined as follows: (The algorithm must satisfy both the norm and the measurement error constraint.) ; in, The sparse coefficients are the reconstructed values. This is the measurement error threshold.

[0061] Step S350: Through the wavelet domain sparse basis , reconstructed sparsity coefficients Inverse transform yields high-resolution wavelet coefficient matrix The expression can be as follows: ; ; in, It is a high-resolution wavelet coefficient matrix. It is a sparse basis in the wavelet domain. These are the sparse coefficients after reconstruction; , This is the resolution enhancement factor. The original wavelet coefficient matrix was supplemented. Implicit information between discrete points.

[0062] Step S360: Convert the high-resolution wavelet coefficient matrix The output is the high-resolution wavelet coefficient matrix.

[0063] Step S400: Based on the multi-scale physical characteristics of carbon emissions and the high-resolution wavelet coefficient matrix, generate enhanced data that conforms to physical laws and covers diverse scenarios.

[0064] Specifically, step S400, which generates enhanced data that conforms to physical laws and covers diverse scenarios based on the multi-scale physical characteristics of carbon emissions and the high-resolution wavelet coefficient matrix, specifically includes the following steps S410 to S430.

[0065] It is understandable that carbon emission data from the power system exhibits significant multi-scale physical coupling characteristics: 1) Fine-grained (second-minute level): Unit ramp-up, fuel combustion fluctuations; 2) Medium-grained scale (hourly level): Load peak and valley variations, and intraday fluctuations in wind and solar power output; 3) Coarse-grained (daily-weekly): seasonal weather trends, weekday / weekend load cycles.

[0066] Therefore, during implementation, by coordinating and optimizing the physical parameters at the three scales, and combining them with a self-supervised physical verification mechanism, enhanced data that conforms to the physical rules at each scale and covers diverse working conditions is generated.

[0067] Step S410: Based on the multi-scale physical characteristics of carbon emissions, analyze the multi-scale physical driving factors from the multi-source data, and design several sets of corresponding adjustable physical parameters for each scale physical driving factor.

[0068] Specifically, fine-grained physical driving factors include unit ramp-up and fuel combustion fluctuations, and the corresponding adjustable physical parameters include ramp-up rate correction coefficients. Fuel combustion efficiency fluctuation coefficient .

[0069] Specifically, the medium-grained physical driving factors include load peaks and valleys, and intraday fluctuations in wind and solar power. The corresponding adjustable physical parameters include the load peak-to-valley ratio correction coefficient. New energy penetration rate coefficient .

[0070] Specifically, coarse-grained physical drivers include meteorological trends and load cycles, and the corresponding adjustable physical parameters include the daily carbon emission baseline correction factor. Period offset coefficient .

[0071] Furthermore, it can be achieved through a matrix Define the coupling relationships between multi-scale physical driving factors, such as the penetration rate of new energy sources. As the elevation increases, the rate of increase... The parameters need to be adjusted synchronously to avoid conflicts between different scales.

[0072] Furthermore, in the process of designing several sets of adjustable physical parameters for each scale, Latin hypercube sampling (LHS) can be used to ensure the randomness of each set of adjustable physical parameters, that is, to ensure... and The randomness of the combinations makes it more in line with real-world scenarios, ultimately forming an enhanced dataset that covers diverse scenarios such as "high new energy penetration rate + low peak-to-valley ratio" and "low ramp rate + high cycle offset", ensuring diversity.

[0073] Step S420: Based on the high-resolution wavelet coefficient matrix The adjustable physical parameters are sequentially subjected to coarse-grained baseline setting, medium-grained fluctuation superposition, and fine-grained detail supplementation to generate enhanced data V.

[0074] Specifically, the coarse-grained benchmark tuning involves correcting the high-resolution wavelet coefficient matrix. The daily carbon emission baseline value in the data, to preserve the cyclical trend, can be expressed as follows: ; in, To obtain from the high-resolution wavelet coefficient matrix The daily dimension benchmark value extracted from the low-frequency long-term trend components; Simulate the periodic trend to ensure that the coarse-grained periodic characteristics remain unchanged; This is the correction factor for the daily carbon emission baseline value. This is the period offset coefficient.

[0075] Specifically, the medium-grained fluctuation superposition is based on the modified baseline superposition of load peaks and valleys and wind and solar fluctuations, and the specific expression can be as follows: ; in, For the peak-valley deviation of the original data load , To address the deviation in the output of new energy sources, through and Negative correlation coupling; This is the load peak-to-valley ratio correction factor. This represents the penetration rate coefficient of new energy sources.

[0076] Specifically, the fine-grained detail supplementation involves incorporating unit ramp-up and combustion efficiency fluctuations to ensure accurate instantaneous characteristics. The specific expression can be as follows: ; in, For deviations in thermal power output ramp-up, For slight combustion noise, and The coupling matrix W is linked with the medium-granularity parameters; Correction factor for gradient rate This is the fuel combustion efficiency fluctuation coefficient.

[0077] Step S430: Verify the validity of the enhanced data V and filter out invalid sample data in the enhanced data V.

[0078] Furthermore, step S430, verifying the validity of the enhanced data V and filtering out invalid sample data in the enhanced data V, specifically includes steps S431 to S433.

[0079] Step S431: Design a physical value discriminator (PD) and a time series characteristic discriminator (TD) respectively. The physical value discriminator (PD) is used to verify whether the key carbon emission parameters in the enhanced data V are compliant. The time series characteristic discriminator (TD) is used to verify whether the time series pattern in the enhanced data V conforms to the unit operation logic.

[0080] Specifically, the expression for the physical numerical discriminator (PD) can be as follows: ; Here, HHV represents the enhanced fuel consumption; HHV represents the high calorific value of the fuel. The total time step of the time series data. To enhance the carbon emissions data at time t, EF is the carbon emission factor of the fuel, and only retains [the data]. Sample data >0.9.

[0081] Specifically, the time-series feature discriminator (TD) is based on an LSTM pre-trained model, and its expression can be as follows: ; in, Encode the temporal characteristics of the original data. It is the temporal feature encoding extracted from the enhanced data V by an LSTM encoder. Represents the sigmoid activation function, only retaining [the relevant information]. Sample data >0.8.

[0082] Step S432: Based on the physical numerical discriminator (PD) and the time series characteristic discriminator (TD), construct a bi-objective optimization function.

[0083] Specifically, the expression for the bi-objective optimization function can be as follows: ; in, This represents the bi-objective optimization function. This represents the physical numerical discriminator (PD). This represents the time-series characteristic discriminator (TD). The entropy value of the parameter combination. =0.1, which is the diversity weight, used to balance physical fidelity and diversity.

[0084] Step S433: The physical numerical discriminator (PD) and the time series characteristic discriminator (TD) are used to verify the validity of the enhanced data V. During the verification process, the bi-objective optimization function is used as the objective function to filter out invalid sample data in the enhanced data V.

[0085] Step S500: Based on the resolution branch design, perform precise feature filtering on the enhanced data to obtain multiple sets of high-resolution branch prediction sequence matrices and output them. Each set of branch prediction sequence matrices corresponds to a resolution branch.

[0086] Specifically, step S500, which involves performing precise feature filtering on the enhanced data based on the resolution branch design to obtain and output multiple sets of high-resolution branch prediction sequence matrices, specifically includes the following steps S510 to S560.

[0087] Step S510: Pass each coefficient sequence of V in the augmented data through the corresponding resolution branch. All resolution branches are independent of each other, thereby avoiding mutual interference between features of different scales.

[0088] It should be noted that the enhanced data V and the high-resolution wavelet coefficient matrix are... and wavelet coefficient matrix The number of sequences is consistent, that is, the coefficient sequences after multi-scale wavelet decomposition. Each corresponds to one independent resolution branch, and the total number of resolution branches is m+1.

[0089] Step S520: Divide each coefficient sequence into corresponding overlapping patches.

[0090] Specifically, step S520, which involves dividing each coefficient sequence into corresponding overlapping patches, is performed using the following formula: ; in, X P For patching; X For the coefficient sequence, X ; P For patch size, S The step size.

[0091] Step S530: Embed each patch into the d-dimensional feature space through a linear layer to capture multiple local carbon emission features. X d .

[0092] Specifically, step S530, embedding each of the overlapping patches into the d-dimensional feature space through a linear layer, can be expressed as follows: .

[0093] Step S540: Use a multilayer perceptron (MLP) to capture different local carbon emission characteristics. X d Local dependencies between them are used to obtain local dependency features. .

[0094] Specifically, the structure adopted by the multilayer perceptron (MLP) is "2D-BatchNorm → Dimensional Permutation → GELU Activation → MLP Expansion → MLP Compression".

[0095] Specifically, the expression for the multilayer perceptron (MLP) can be as follows:

[0096] in, Dimensional permutation ; Extend the linear layer for features; A linear layer is used to compress features and restore the original dimension. This is the activation function.

[0097] Step S550: Employ a channel attention mechanism to enhance the local dependency features. By analyzing the characteristic channel and suppressing invalid channel responses corresponding to key carbon emission parameters, accurate feature selection is achieved, resulting in a set of corresponding branch approximation coefficients output by each resolution branch. With branch detail coefficient .

[0098] Furthermore, in step S550, a channel attention mechanism is employed to enhance the local dependency features. By analyzing the characteristic channel and suppressing invalid channel responses corresponding to key carbon emission parameters, accurate feature selection is achieved, resulting in a set of corresponding branch approximation coefficients output by each resolution branch. With branch detail coefficient This includes the following steps S551 to S556.

[0099] Step S551: For the local dependency features We perform dimensional adjustment and global information aggregation to obtain the features after dimensional adjustment and global information aggregation. .

[0100] Specifically, step S551 involves processing the local dependency features. The expression used for dimensional adjustment and global information aggregation can be as follows: ; in, Dimensional permutation .

[0101] Step S552: Adjust the dimensions and aggregate the features based on the global information. The average pooling result is obtained by fusing global average pooling (GAP) and global max pooling (GMP) with channel-wide statistics. and max pooling results .

[0102] Specifically, step S552 involves adjusting the dimensions and aggregating the features based on global information. Through global average pooling ( GAP ) and global max pooling ( GMP To perform global statistics on the fusion channel, the expression used can be as follows:

[0103] ; in, For the average pooling result, This is the result of max pooling; The features are those obtained after dimensional adjustment and aggregation of global information; GAP This represents the global average pooling operation. GMP This represents the global max pooling operation.

[0104] Step S553: ​​Calculate the average pooling result. and max pooling results After concatenation, the data is input into a shared multilayer perceptron (MLP) to learn the final channel attention weights. .

[0105] Specifically, the shared multilayer perceptron (MLP) preferably adopts the structure of "segmentation → MLP compression → GELU activation → MLP recovery → sigmoid normalization".

[0106] Specifically, in step S553, the average pooling result is... and max pooling results After concatenation, the data is input into a shared multilayer perceptron (MLP) to learn the final channel attention weights. The expression used can be as follows:

[0107] in, For the average pooling result, This is the result of max pooling; Concat represents the concatenation operation. The average pooling result and max pooling results The splicing result; Compress the linear layer for channel weights. To restore a linear layer for channel weights, For the sigmoid function, This represents the final channel attention weight.

[0108] Step S554: Apply the final channel attention weights using the Hadamard product. Features applied to the dimensional adjustment and global information aggregation The features obtained after redundant filtering are obtained. This achieves channel-level redundant filtering.

[0109] Specifically, in step S554, the final channel attention weights are calculated using the Hadamard product. Features applied to the dimensional adjustment and global information aggregation The features obtained after redundant filtering are obtained. The expression used can be as follows: ; in, These are the features after redundant filtering. For the final channel attention weights, The features are those after the dimensions are adjusted and global information is aggregated.

[0110] Step S555: Capture the redundant filtered features by embedding and mixing. Global dependencies in the data yield embedded hybrid features. .

[0111] Specifically, step S555 involves capturing the redundant filtered features through embedding and mixing. Global dependencies in the data yield embedded hybrid features. The preferred model structure is "dimensional inverse permutation → 2D-BatchNorm → GELU activation → MLP expansion → MLP compression → residual connection".

[0112] Specifically, step S555 involves capturing the redundant filtered features through embedding and mixing. Global dependencies in the data yield embedded hybrid features. The expression used can be as follows: ; ; ; in, Inverse dimensional permutation , Extend the linear layer to global features. A linear layer is used to compress global features, Batch Normalization (BN) is used, and residual connections are used to avoid gradient vanishing.

[0113] Step S556: Based on the embedded hybrid features, through a preset flattening layer and a preset linear layer... Predict and output the branch approximation coefficients corresponding to each resolution branch. With branch detail coefficient .

[0114] Specifically, in step S556, based on the embedded hybrid features, a preset flattening layer and a preset linear layer are used. Predict and output the branch approximation coefficients corresponding to each resolution branch. With branch detail coefficient The expression used can be as follows:

[0115] in, The flattening operation is achieved through the preset flattening layer; T represents the predicted time step, and the original scale can be recovered through the RevIN inverse transform.

[0116] Step S560: Calculate the branch approximation coefficients for each group. With branch detail coefficient They are respectively combined into a set of branch prediction sequence matrices. And output it.

[0117] It is understood that the branch prediction sequence matrix middle, These are the branch approximation coefficients; This represents the branch detail coefficient.

[0118] Step S600: Reconstruct the corresponding time-frequency features of all the branch prediction sequence matrices by wavelet inverse transform, and fuse the time-frequency features based on a preset attention mechanism to obtain the carbon emission prediction value.

[0119] Specifically, step S600 involves reconstructing all the branch prediction sequence matrices into corresponding time-frequency features through wavelet inverse transform, fusing the time-frequency features based on a preset attention mechanism to obtain the carbon emission prediction value, including steps S610 to S630.

[0120] Step S610: Based on wavelet inverse transform reconstruction technology, generate the branch prediction sequence matrix for each group. The corresponding time-frequency features were reconstructed respectively. .

[0121] Step S620: Based on a preset attention mechanism, for each of the time-frequency features... The corresponding weights are assigned separately and then fused to obtain the fused features. .

[0122] Specifically, in step S620, based on a preset attention mechanism, corresponding weights are assigned to each of the time-frequency features and then fused to obtain fused features. The specific expression can be as follows: ; in, This represents the attention weight of the prediction result of the i-th sub-resolution branch in the final fusion; Linear layer; m represents the total number of wavelet decomposition layers, which determines the number of resolution branches. The total number of resolution branches is m+1. For the i-th time-frequency feature, the branch prediction sequence matrix output by the i-th resolution branch is given.

[0123] Step S630: The fusion feature After passing through at least two layers of pre-defined fully connected networks, the predicted carbon emissions for the next T steps are obtained. .

[0124] Furthermore, the value of T depends on the actual needs, and is preferably between 100 and 1000.

[0125] To further improve prediction performance, the method can use the SmoothL1 loss function and the contrastive loss function to construct a total loss function, and then use the total loss function and the Adam optimizer to optimize the method.

[0126] Specifically, the expression for the total loss function can be as follows:

[0127] ; in, For the total loss function, The SmoothL1 loss function is used. To compare loss functions; These are the weighting coefficients. This is a predicted value for carbon emissions. For true carbon emissions, Indicates the time step of the prediction.

[0128] Example 2: like Figure 2 As shown, in a second aspect, this disclosure provides a power system source-side carbon emission prediction system, the system comprising: The data acquisition module 100 is used to collect multi-source data, including historical carbon emission data, power generation data, system load data, load cycle data, boiler and equipment parameters, and fuel-related data. The data preprocessing module 200 is used to perform data preprocessing on the collected multi-source data. The data preprocessing includes data alignment, outlier handling, and normalization. The time-frequency extension module 300 is used to perform time-frequency extension on the multi-source data after data preprocessing based on wavelet decomposition and compressed sensing reconstruction algorithm to obtain a high-resolution wavelet coefficient matrix; the high-resolution wavelet coefficient matrix is ​​high-resolution time-frequency data. The data augmentation module 400 is used to generate augmented data that conforms to physical laws and covers diverse scenarios based on the multi-scale physical characteristics of carbon emissions and the high-resolution wavelet coefficient matrix. The multi-resolution module 500 is used to perform precise feature filtering on the enhanced data based on the sub-resolution branch design, obtain multiple sets of high-resolution branch prediction sequence matrices and output them, with each set of branch prediction sequence matrices corresponding to a sub-resolution branch. The data prediction module 600 is used to reconstruct the corresponding time-frequency features of all the branch prediction sequence matrices through wavelet inverse transform, and to fuse the time-frequency features based on a preset attention mechanism to obtain the carbon emission prediction value.

[0129] Furthermore, the system may also include: The model optimization module is used to construct a total loss function using the SmoothL1 loss function and the contrastive loss function, and to optimize the system using the total loss function and the Adam optimizer.

[0130] The data acquisition module 100, data preprocessing module 200, time-frequency extension module 300, data enhancement module 400, multi-resolution module 500, and data prediction module 600 in the system correspond to steps S100, S200, S300, S400, S500, and S600 in the method, respectively.

[0131] It is worth noting that the system described in Example 2 is only one system implementation of the power system source-side carbon emission prediction method described in Example 1, and does not limit the power system source-side carbon emission prediction method described in Example 1 to depend on the system described in Example 2.

[0132] Example 3: Embodiment 3 of this disclosure provides 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 as described in Embodiment 1.

[0133] The computer-readable storage medium includes volatile or non-volatile, removable or non-removable media implemented in any method or technology for storing information (such as computer-readable instructions, data structures, computer program modules, or other data). Computer-readable storage media include, but are not limited to, RAM (Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory or other memory technologies, CD-ROM (Compact Disc Read-Only Memory), DVD or other optical disc storage, cartridges, magnetic tapes, disk storage or other magnetic storage devices, or any other medium that can be used to store desired information and is accessible to a computer.

[0134] Example 4: Embodiment 4 of this disclosure provides an electronic device including a processor and a memory, wherein the processor is used to execute a computer program stored in the memory to implement the power system source-side carbon emission prediction method described in Embodiment 1.

[0135] Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, and read-only memory (ROM).

[0136] Those skilled in the art will understand that embodiments of this disclosure can be provided as methods, systems, or computer program products. Therefore, this disclosure can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this disclosure can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0137] This disclosure is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this disclosure. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0138] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0139] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0140] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this disclosure. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0141] In summary, the power system source-side carbon emission prediction method and system provided in Embodiments 1-4 of this disclosure solves the problems of accuracy and practicality in power system source-side carbon emission prediction by constructing a prediction model that integrates time-frequency extension, physical constraint-based information enhancement, and resolution feature optimization. This disclosure innovatively proposes a six-stage carbon emission data prediction mechanism covering "data acquisition - data preprocessing - time-frequency extension - information perception enhancement - resolution feature extraction - feature fusion and prediction," addressing existing problems in power system source-side carbon emission prediction such as incomplete multi-scale feature capture, data enhancement distortion, multi-resolution information loss, and high model complexity.

[0142] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this disclosure and not to limit them. Although this disclosure has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of this disclosure. Any modifications or equivalent substitutions that do not depart from the spirit and scope of this disclosure should be covered within the protection scope of the claims of this disclosure.

Claims

1. A method for predicting carbon emissions from the source side of a power system, characterized in that, The method includes: Collect multi-source data, including historical carbon emission data, power generation data, system load data, load cycle data, boiler and equipment parameters, and fuel-related data; The collected multi-source data is preprocessed, including data alignment, outlier handling, and normalization. Based on wavelet decomposition and compressed sensing reconstruction algorithm, time-frequency extension is performed on the preprocessed multi-source data to obtain a high-resolution wavelet coefficient matrix; the high-resolution wavelet coefficient matrix is ​​high-resolution time-frequency data. Based on the multi-scale physical characteristics of carbon emissions and the high-resolution wavelet coefficient matrix, enhanced data that conforms to physical laws and covers diverse scenarios is generated. Based on the resolution branch design, the enhanced data is accurately filtered to obtain multiple sets of high-resolution branch prediction sequence matrices and output them. Each set of branch prediction sequence matrices corresponds to a resolution branch. All the branch prediction sequence matrices are reconstructed into corresponding time-frequency features through wavelet inverse transform. The time-frequency features are then fused based on a preset attention mechanism to obtain the carbon emission prediction value.

2. The power system source-side carbon emission prediction method as described in claim 1, characterized in that, The historical carbon emission data includes total carbon emissions from the source side and time-series variation data. The power generation data includes thermal power output data and new energy power output data; The system load data includes peak and valley loads and peak and valley load deviations. The load cycle data includes weekday / weekend load differences and seasonal load trends; The boiler and equipment parameters include unit operating status data and equipment maintenance status data; The fuel-related data includes fuel consumption time-series data and fuel characteristic parameters; The fuel consumption time-series data, power generation data, and historical carbon emission data are synchronized and aligned.

3. The power system source-side carbon emission prediction method as described in claim 1, characterized in that, The data alignment includes: synchronizing each data in the multi-source data according to timestamps, unifying them into an hourly time series, and supplementing missing timestamps using "forward padding + linear interpolation"; The outlier handling includes: identifying and correcting outlier data in the data-aligned multi-source data; such as sudden increases / decreases in carbon emissions due to sensor malfunctions—if the outlier deviates from the mean by 3... The above uses a weighted average of three adjacent time periods for replacement. The normalization process includes: performing reversible instance normalization on the multi-source data after outlier processing to eliminate distribution bias in each time series.

4. The power system source-side carbon emission prediction method as described in claim 1, characterized in that, The method, based on wavelet decomposition, wavelet domain sparsity modeling, compressed sensing, and sparse reconstruction, performs time-frequency extension on the preprocessed multi-source data to obtain a high-resolution wavelet coefficient matrix. This process specifically includes the following steps: Based on the wavelet decomposition method, the multi-source data after data preprocessing is decomposed into low-frequency approximation coefficients and high-frequency detail coefficients. The low-frequency approximation coefficients and high-frequency detail coefficients are combined into a wavelet coefficient matrix, and the wavelet coefficient matrix is ​​represented as a linear combination of sparse coefficients through a wavelet domain sparse basis. Design a measurement matrix that satisfies the finite isometry property, perform linear measurements on the sparse coefficients, and obtain low-dimensional measurement values; The low-dimensional measurement values ​​are sparsely reconstructed based on the orthogonal matching pursuit algorithm to obtain the reconstructed sparse coefficients. A high-resolution wavelet coefficient matrix is ​​obtained by inverse transformation of the wavelet domain sparse basis and the reconstructed sparse coefficients. The high-resolution wavelet coefficient matrix is ​​output as the high-resolution wavelet coefficient matrix.

5. The power system source-side carbon emission prediction method as described in claim 1, characterized in that, The generation of enhanced data based on the multi-scale physical characteristics of carbon emissions and the high-resolution wavelet coefficient matrix, which conforms to physical laws and covers diverse scenarios, specifically includes the following steps: Based on the multi-scale physical characteristics of carbon emissions, multi-scale physical driving factors are analyzed from the multi-source data, and several sets of corresponding adjustable physical parameters are designed for each scale physical driving factor. Based on the high-resolution wavelet coefficient matrix and each of the adjustable physical parameters, coarse-grained baseline tuning, medium-grained fluctuation superposition, and fine-grained detail supplementation are performed sequentially to generate enhanced data; Verify the validity of the augmented data and filter out invalid sample data in the augmented data.

6. The power system source-side carbon emission prediction method as described in claim 1, characterized in that, The step of performing precise feature filtering on the enhanced data based on the resolution-based branch design to obtain and output multiple sets of high-resolution branch prediction sequence matrices specifically includes the following steps: Each coefficient sequence in the augmented data is passed through a corresponding resolution branch, and all resolution branches are independent of each other. Each coefficient sequence is divided into corresponding overlapping patches; Each patch is embedded into a 3D feature space through a linear layer to capture multiple local carbon emission features; A multilayer perceptron is used to capture the local dependencies between different local carbon emission characteristics to obtain local dependency features; By employing a channel attention mechanism, precise feature selection is achieved by strengthening the feature channels corresponding to the key carbon emission parameters in the local dependent features and suppressing the responses of invalid channels, resulting in a set of corresponding branch approximation coefficients and branch detail coefficients output by each sub-resolution branch. The corresponding branch approximation coefficients and branch detail coefficients of each group are combined into a set of branch prediction sequence matrices and output.

7. The power system source-side carbon emission prediction method as described in claim 6, characterized in that, The method employs a channel attention mechanism, which enhances the feature channels corresponding to the key carbon emission parameters in the locally dependent features and suppresses ineffective channel responses to achieve accurate feature selection. This results in a set of corresponding branch approximation coefficients and branch detail coefficients output by each resolution branch. The method includes the following steps: The local dependent features are adjusted in dimension and aggregated with global information to obtain the features after dimension adjustment and global information aggregation. The features after dimensional adjustment and global information aggregation are fused with global average pooling and global max pooling to obtain the average pooling result and the max pooling result. The average pooling result and the max pooling result are concatenated and then input into a shared multilayer perceptron to learn the final channel attention weights. The final channel attention weights are applied to the features after dimensional adjustment and global information aggregation by Hadamard product, resulting in redundant filtered features. By embedding and blending to capture global dependencies in the redundant filtered features, embedded and blended features are obtained. By using a preset flattening layer and a preset linear layer, based on the embedded hybrid features, the branch approximation coefficient and branch detail coefficient corresponding to each sub-resolution branch are predicted and output.

8. The power system source-side carbon emission prediction method as described in claim 1, characterized in that, The step of reconstructing the corresponding time-frequency features from all the branch prediction sequence matrices using inverse wavelet transform, and fusing the time-frequency features based on a preset attention mechanism to obtain the carbon emission prediction value includes the following steps: Based on the wavelet inverse transform reconstruction technique, the corresponding time-frequency features are reconstructed for each group of branch prediction sequence matrices; Based on a preset attention mechanism, each of the time-frequency features is assigned a corresponding weight and then fused to obtain a fused feature. After passing the fused features through at least two layers of pre-defined fully connected networks, the predicted carbon emissions for the next T steps are output.

9. The power system source-side carbon emission prediction method as described in claim 1, characterized in that, In the method, a total loss function is constructed using the SmoothL1 loss function and the contrastive loss function, and the method is optimized using the total loss function and the Adam optimizer.

10. A power system source-side carbon emission prediction system, characterized in that, The system includes: The data acquisition module is used to collect multi-source data, including historical carbon emission data, power generation data, system load data, load cycle data, boiler and equipment parameters, and fuel-related data. The data preprocessing module is used to preprocess the collected multi-source data, including data alignment, outlier handling, and normalization. The time-frequency extension module is used to perform time-frequency extension on the pre-processed multi-source data based on wavelet decomposition and compressed sensing reconstruction algorithm to obtain a high-resolution wavelet coefficient matrix; the high-resolution wavelet coefficient matrix is ​​high-resolution time-frequency data. The data augmentation module is used to generate augmented data that conforms to physical laws and covers diverse scenarios based on the multi-scale physical characteristics of carbon emissions and the high-resolution wavelet coefficient matrix. A multi-resolution module is used to perform precise feature filtering on the enhanced data based on the sub-resolution branch design, obtain multiple sets of high-resolution branch prediction sequence matrices and output them, with each set of branch prediction sequence matrices corresponding to a sub-resolution branch. The data prediction module is used to reconstruct the corresponding time-frequency features of all the branch prediction sequence matrices through wavelet inverse transform, and to fuse the time-frequency features based on a preset attention mechanism to obtain the carbon emission prediction value.

Images