A dual-dimensional green electricity identification and dynamic carbon accounting method based on deep learning

Abstract

The present application relates to the technical field of electric power carbon metering, and discloses a dual-dimension green electricity identification and dynamic carbon accounting method based on deep learning, which solves the technical problems of insufficient reliability, rough processing of electrical quantity characteristics, and poor granularity and timeliness caused by static lag in carbon emission accounting in the prior art due to the fact that green electricity identification relies on single-dimension information. The present application overcomes the defect of unreliable single information source identification through protocol and physical dual-dimension fusion identification, improves the distinguishing accuracy of different green electricity types through a physical feature grouping interactive convolution model, solves the problems of static accounting lag and coarse granularity through real-time calculation of dynamic emission factors and carbon quota balance, and realizes significant improvement in green electricity identification accuracy, real-time dynamic updating of carbon emission accounting, and terminal closed-loop calculation of carbon quota.

Description

Technical Field

[0001] This invention relates to the field of electricity carbon metering technology, and in particular to a two-dimensional green electricity identification and dynamic carbon accounting method based on deep learning. Background Technology

[0002] Currently, the common practice in existing technologies for green electricity consumption and carbon emission management for industrial and commercial users is to receive green electricity identification frames issued by the power grid through power line carrier communication, parse them to obtain the power source type and power supply period, collect the user's electricity consumption and multiply it by the regional average emission factor, calculate the total carbon emissions offline, and then compare it with the user's carbon quota afterward. On this basis, some solutions will also perform statistical analysis on the electricity consumption side to help determine the source of electricity.

[0003] However, the aforementioned existing technologies have the following shortcomings in practical applications:

[0004] I. The green electricity identification process mainly relies on single-dimensional information. When relying solely on protocol parsing, the identification frame may be distorted due to signal interference, data tampering, or mismatch of time period information; while analyzing only electrical quantity characteristics makes it difficult to reliably distinguish different power supply types under load fluctuations or noise interference.

[0005] Second, the processing of electrical quantity features is relatively crude. They are usually simply stacked and then fed into the classification model without considering the differences in the ability of different physical quantities to represent power types. Simple splicing makes it difficult to fully explore the local patterns and controllable interactions between features, which limits the model's accuracy in distinguishing different types of green electricity.

[0006] Third, a fixed regional average emission factor is usually used, which does not take into account the actual dynamic factors such as the difference in power supply structure during peak and valley periods of the power grid and the transmission loss at different nodes. As a result, the calculation results are coarse-grained and have poor timeliness. Summary of the Invention

[0007] The purpose of this invention is to solve the technical problems in the prior art, such as insufficient reliability due to reliance on single-dimensional information for green electricity identification, coarse processing of electrical quantity characteristics, and poor granularity and timeliness due to static lag in carbon emission accounting. The invention provides a two-dimensional green electricity identification and dynamic carbon accounting method based on deep learning.

[0008] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0009] A deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method includes the following steps:

[0010] The electricity meter parses the green electricity identification signal to obtain the green electricity identification frame, verifies the green electricity identification frame, and extracts the core protocol feature parameters of the green electricity identification frame when the verification is successful.

[0011] The meter obtains voltage and current sampling sequences through sampling, extracts multi-dimensional physical features based on the voltage and current sampling sequences, normalizes the multi-dimensional physical features, and obtains a physical feature vector.

[0012] The physical feature grouping interactive convolutional model divides physical feature vectors into related and unrelated classes to obtain the physical matching degree.

[0013] By using dynamic correction factors to integrate physical matching degree and protocol reliability, a green electricity purity index is obtained. By combining the green electricity purity index with the power type code, it is determined whether the current electricity consumption is green electricity.

[0014] If it is green electricity, the carbon accounting model obtains the green electricity consumption and smoothed green electricity purity index within the current electricity supply period, and obtains the dynamic emission factor through the basic emission factor, the period correction coefficient and the node correction coefficient.

[0015] The carbon accounting model obtains the accumulated carbon emissions through dynamic emission factors, obtains the corrected carbon emission reduction using green electricity consumption and a smoothed green electricity purity index, and obtains the carbon quota balance through the accumulated carbon emissions and the corrected carbon emission reduction.

[0016] To address the technical problem of low reliability in existing green electricity identification methods that rely on single-dimensional information (simple protocol parsing is easily interfered with or tampered with, and analyzing only electrical quantity characteristics is difficult to reliably distinguish power types), this invention constructs a two-dimensional identification mechanism of "protocol layer identification + physical feature extraction". It uses a physical feature grouping interactive convolution model to process multi-dimensional physical features, and uses a dynamic correction factor to fuse protocol credibility and physical matching degree to obtain a green electricity purity index. This achieves complementary verification and adaptive fusion of two-dimensional information, significantly improving the accuracy and anti-interference ability of green electricity determination.

[0017] To address the technical problem that existing technologies often handle electrical quantity features coarsely, simply stacking them before feeding them into the model, and failing to fully explore the local patterns and controllable interactions of different physical quantities, thus limiting the accuracy of distinguishing different types of green electricity, this invention introduces a physical feature grouping interactive convolution model. This model groups the normalized physical feature vectors into frequency domain, power fluctuation, and waveform morphology categories. Local patterns are extracted through intra-group convolution, and controllable interactions are achieved through inter-group physical mask fully connected layers. This effectively enhances the model's ability to express and distinguish the features of different power sources such as photovoltaic, wind power, and hydropower.

[0018] To address the technical problems of existing carbon emission accounting methods that use fixed regional average emission factors and fail to consider dynamic factors such as differences in power supply structure during peak and off-peak periods and transmission losses at nodes, resulting in coarse calculation results, poor timeliness, and the inability to obtain carbon allowance balances in real time during electricity consumption, this invention constructs a dynamic emission factor by using a basic emission factor, a time period correction coefficient, and a node correction coefficient. It then combines green electricity consumption with a smoothed green electricity purity index to obtain cumulative carbon emissions and corrected carbon emission reductions, thereby calculating carbon allowance balances in real time and achieving a dynamic and accurate closed-loop carbon accounting system on the user side.

[0019] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention overcomes the unreliability of identification from a single information source by fusing protocol and physical dimensions for identification; it improves the differentiation accuracy of different green energy types by using a physical feature grouping interactive convolution model; it solves the problems of static accounting lag and coarse granularity by using real-time calculation of dynamic emission factors and carbon quota balances; and it achieves a significant improvement in the accuracy of green energy identification, real-time dynamic updates of carbon emission accounting, and closed-loop calculation of carbon quotas at the terminal.

[0020] Furthermore, a deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method is proposed, in which the electricity meter parses the green electricity identification signal to obtain a green electricity identification frame, verifies the green electricity identification frame, and extracts the core protocol feature parameters of the green electricity identification frame when the verification passes, including the following sub-steps:

[0021] The electricity meter uses power line carrier to receive the green electricity identification signal sent by the power grid side, and uses Manchester encoding to parse the green electricity identification signal to obtain the green electricity identification frame;

[0022] The electricity meter performs triple verification on the green electricity identification frame by weighted summation to obtain the protocol's credibility.

[0023] When the reliability of the protocol When the credibility threshold is reached, the core protocol feature parameters of the green electricity identification frame are extracted; the core protocol feature parameters include power type encoding, power supply period, and power supply priority.

[0024] In the above scheme, this invention receives the green electricity identification signal sent by the power grid side via power line carrier, parses it using Manchester encoding to obtain the green electricity identification frame, and performs a triple weighted summation of frame format verification, signature verification, and time period validity verification on the frame to calculate the protocol credibility. When the protocol credibility is not lower than the credibility threshold, core protocol feature parameters such as power type code, power supply period, and power supply priority are extracted. Addressing the technical problem of green electricity identification information distortion caused by carrier signal interference, tampering of identification frames, or mismatch of time period information in existing protocol parsing methods, this invention adopts a triple verification and weighted summation mechanism, setting frame format verification weight, signature verification weight, and time period validity verification weight respectively. By comprehensively calculating the protocol credibility and setting a threshold judgment, unreliable identification frames are effectively eliminated, avoiding misjudgment of power type or power supply period due to protocol information errors; at the same time, the edge transition characteristics of Manchester encoding are used to enhance the signal's anti-interference capability. This invention provides accurate, reliable, and quantifiable protocol layer input information for subsequent fusion identification of protocol features and physical features, ensuring the basic data quality of two-dimensional green electricity identification.

[0025] Furthermore, a deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method is proposed. The method involves obtaining voltage and current sampling sequences from the electricity meter, extracting multi-dimensional physical features based on these sequences, and normalizing these features to obtain a physical feature vector. The method includes the following sub-steps:

[0026] The electricity meter samples the voltage and current on the electricity consumption side at a fixed sampling period to obtain voltage sampling sequences and current sampling sequences;

[0027] Based on voltage and current sampling sequences, multi-dimensional physical features are extracted; these multi-dimensional physical features include fundamental frequency, total harmonic distortion, phase offset, voltage fluctuation amplitude, current peak factor, active power fluctuation coefficient, reactive power ratio, voltage harmonic order, current harmonic amplitude, and waveform stability.

[0028] Normalize each dimension of the multi-dimensional physical features to obtain the normalized value of each dimension.

[0029] The normalized values ​​of each dimension feature are integrated to obtain the physical feature vector.

[0030] In the above scheme, the present invention synchronously samples the voltage and current on the electricity consumption side using an electricity meter at a fixed sampling period to obtain a voltage sampling sequence and a current sampling sequence; based on the sequence, it extracts ten-dimensional multi-dimensional physical features including fundamental frequency, total harmonic distortion rate, phase offset, voltage fluctuation amplitude, current peak factor, active power fluctuation coefficient, reactive power ratio, voltage harmonic order, current harmonic amplitude, and waveform stability; each feature is min-max normalized to eliminate the influence of dimensions, and the normalized feature values ​​are integrated into a physical feature vector. To address the technical problems of existing electrical quantity feature extraction methods that directly perform simple statistics on the original sampled sequences or lack sufficient preprocessing, making them susceptible to sensor noise, power grid harmonic interference, and sampling quantization errors, resulting in large fluctuations and poor stability of feature values, and making it difficult to directly use features of different dimensions for subsequent model training, this invention employs fixed-period synchronous sampling to ensure data consistency. It comprehensively characterizes the differentiated characteristics of power supply types in the frequency, time, and power domains through ten-dimensional features, and uses normalization operations to map each feature to the [0,1] interval, effectively suppressing the influence of noise and interference on feature values ​​and eliminating numerical deviations caused by differences in dimensions. This invention provides stable and standardized input feature vectors for physical feature grouping interactive convolution models, significantly improving the convergence speed of model training and the recognition accuracy of different green electricity types.

[0031] Furthermore, a deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method, wherein the extraction of multi-dimensional physical features based on voltage sampling sequences and current sampling sequences includes the following sub-steps:

[0032] Perform a fast Fourier transform on the voltage sampling sequence to extract the fundamental frequency and voltage amplitude of each spectral line of the grid voltage signal;

[0033] The fundamental voltage amplitude and harmonic voltage amplitude are extracted from the voltage amplitude of each spectral line, and the total harmonic distortion rate is obtained using the fundamental voltage amplitude and harmonic voltage amplitude.

[0034] Perform a fast Fourier transform on the current sampling sequence to obtain the current amplitude of each spectral line;

[0035] Take the phase angle of the fundamental voltage complex number and the fundamental current complex number, and use the phase angle to obtain the phase offset;

[0036] The voltage fluctuation amplitude is obtained by using the maximum and minimum values ​​of the voltage sampling sequence.

[0037] The current peak factor is obtained by using the maximum value of the effective current value and the absolute value of the current sampling sequence.

[0038] The active power fluctuation coefficient is obtained by using active power and the average active power.

[0039] The reactive power ratio is obtained by using the total active power and the total reactive power.

[0040] The highest harmonic order, whose harmonic voltage amplitude is greater than 5% of the fundamental voltage amplitude, is used to obtain the voltage harmonic order.

[0041] The amplitude of the current harmonics is obtained by analyzing the voltage harmonic order.

[0042] In the above scheme, the present invention extracts the fundamental frequency and voltage amplitude of each spectral line by performing a fast Fourier transform on the voltage sampling sequence, and then calculates the total harmonic distortion rate; performs a fast Fourier transform on the current sampling sequence to obtain the current amplitude of each spectral line; obtains the phase offset by using the phase angle difference between the complex fundamental voltage and the complex fundamental current; obtains the voltage fluctuation amplitude by using the difference between the maximum and minimum values ​​of the voltage sequence; obtains the current peak factor by combining the effective value and the maximum absolute value of the current; obtains the active power fluctuation coefficient based on the piecewise active power and its average value; calculates the total active power and total reactive power from the fundamental voltage, current and phase, and then obtains the reactive power ratio; extracts the highest harmonic order whose harmonic voltage amplitude exceeds 5% of the fundamental amplitude as the voltage harmonic order, and obtains the current harmonic amplitude accordingly. Existing technologies typically extract only a few electrical quantities (such as voltage, current RMS values, and power), failing to systematically construct a multi-dimensional physical feature set covering the frequency domain (fundamental frequency, harmonic distortion rate, voltage harmonic order), time domain (voltage fluctuation, current peak factor, waveform stability), and power domain (phase shift, active power fluctuation, reactive power ratio). This results in insufficient description and limited distinguishing ability of the electrical characteristics of different power source types (wind power, photovoltaic, hydropower, and thermal power). This invention, through the collaborative extraction of the above 10 dimensions of physical features, comprehensively characterizes the differentiated features of different power sources in terms of spectral structure, power transmission dynamics, and waveform morphology. It not only captures the essential differences between green electricity and thermal power in harmonic distribution and power fluctuation, but also provides information-rich and highly discriminative input for subsequent physical feature grouping and interactive convolutional models, thereby significantly improving the accurate identification capability of different green electricity types.

[0043] Furthermore, a deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method, wherein the physical feature grouping interactive convolution model divides the physical feature vectors into associated and unassociated classes for processing, and the physical matching degree includes the following sub-steps:

[0044] The physical feature vectors are input into the feature grouping layer, which divides the physical feature vectors into different physical feature groups in width according to their inherent correlation, thus obtaining frequency domain feature groups, power fluctuation feature groups, and waveform morphology feature groups.

[0045] The first convolutional layer within the group performs one-dimensional convolution and activation on the frequency domain feature group to obtain the frequency domain feature map; the second convolutional layer within the group performs one-dimensional convolution and activation on the power fluctuation feature group to obtain the power fluctuation feature map; and the third convolutional layer within the group performs one-dimensional convolution and activation on the waveform morphology feature group to obtain the waveform morphology feature map.

[0046] Flattening layer 1 flattens the frequency domain feature map to obtain the frequency domain vector; flattening layer 2 flattens the power fluctuation feature map to obtain the power fluctuation vector; flattening layer 3 flattens the waveform morphology feature map to obtain the waveform morphology vector.

[0047] The splicing layer splices the frequency domain vector, power fluctuation vector, and waveform morphology vector to obtain the total feature vector;

[0048] Based on the correlation between frequency domain vector, power fluctuation vector, and waveform morphology vector, a block mask matrix is ​​constructed;

[0049] The first feature mapping layer maps and activates the interactive feature vectors. The activated vectors are then randomly deactivated to obtain the mapped vectors.

[0050] The two-layer feature mapping process maps the mapping vector and then activates it to obtain the mapped features.

[0051] The output layer is activated after mapping the features to obtain the physical matching degree.

[0052] In the above scheme, the present invention inputs the physical feature vector into the feature grouping layer, which is divided into frequency domain feature group, power fluctuation feature group and waveform morphology feature group according to the inherent correlation. One-dimensional convolution and activation are performed through intra-group convolution layer 1, intra-group convolution layer 2 and intra-group convolution layer 3 respectively to obtain frequency domain feature map, power fluctuation feature map and waveform morphology feature map. After flattening, frequency domain vector, power fluctuation vector and waveform morphology vector are obtained, and then fused into a total feature vector by the splicing layer. Based on the physical correlation between the three groups of vectors, a block mask matrix is ​​constructed, which only allows reasonable inter-group connections (frequency domain and power fluctuation, power fluctuation and waveform morphology) and prohibits unreasonable connections (frequency domain and waveform morphology). After passing through feature mapping layer 1 (including random deactivation), feature mapping layer 2 and output layer, the physical matching degree is finally obtained. To address the technical problems of existing technologies that simply process electrical quantity features and directly stack them before feeding them into the classification model without considering the inherent correlation and controllable interaction between different physical quantities, resulting in redundant feature representation, low discrimination accuracy, and poor model generalization ability, this invention mines local patterns within groups through grouped convolution and achieves sparse inter-group connections based on physical priors through a block mask matrix. This preserves the information flow of strongly correlated features while cutting off the noise coupling of weakly correlated features, effectively reducing the number of model parameters and the risk of overfitting, and significantly improving the recognition accuracy and robustness of different green energy types such as wind power, photovoltaics, and hydropower.

[0053] Furthermore, a deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method involves element-wise multiplication of the block mask matrix and the trainable weight matrix, followed by matrix multiplication of the product with the total feature vector. After activation, the resulting interactive feature vector is obtained, as shown in the formula:

[0054] ;

[0055] in, This is a block mask matrix, where 1 represents a matrix of all 1s and 0 represents a matrix of all 0s. For trainable weight matrix, The feature vector after interaction, For element-wise multiplication, This is the total eigenvector. This is the ReLU activation function.

[0056] In the above scheme, this invention constructs a block mask matrix, performs element-wise multiplication with the trainable weight matrix, then performs matrix multiplication with the total feature vector, and finally processes it through the ReLU activation function to obtain the interactive feature vector. The block mask matrix divides the input and output features into three blocks according to the order of frequency domain vector, power fluctuation vector, and waveform morphology vector. The block sizes are 32, 32, and 12 respectively, and the following masking rules are set: blocks within each vector are all 1s (connection allowed); bidirectional connections between the frequency domain vector and the power fluctuation vector are all 1s (connection allowed); bidirectional connections between the power fluctuation vector and the waveform morphology vector are all 1s (connection allowed); bidirectional connections between the frequency domain vector and the waveform morphology vector are all 0s (connection prohibited). To address the technical problem of existing fully connected layers in power systems that apply full connectivity to all feature channels without considering the real causal relationships between different physical quantity groups, resulting in a large number of invalid connections participating in training and causing redundant model parameters, high risk of overfitting, and poor physical interpretability, this invention employs a fixed block mask matrix to achieve sparse connectivity. This permanently sets unreasonable connections to zero, preventing them from participating in training, while retaining necessary inter-group interactions (such as the impact of harmonics on power fluctuations and the impact of power fluctuations on waveform morphology). This design significantly reduces the number of trainable parameters, lowers model complexity, effectively avoids overfitting with limited samples, and enhances the fit between the network output and the physical laws of the power system, thereby improving the model's generalization ability, interpretability, and anti-interference performance.

[0057] Furthermore, a deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method, which utilizes a dynamic correction factor to fuse physical matching degree and protocol credibility degree to obtain a green electricity purity index, and then uses the green electricity purity index combined with power type coding to determine whether the current electricity consumption is green electricity, includes the following sub-steps:

[0058] By introducing a dynamic correction factor and integrating physical matching degree and protocol credibility degree, a green electricity purity index is obtained. ;

[0059] Combined with power type coding And green electricity determination threshold Determine whether the current electricity consumption is green electricity:

[0060] like ,and If the current electricity consumption is green electricity, the power supply period, power type code, and green electricity purity index of the current electricity consumption will be output to the carbon accounting model.

[0061] like ,and If the current electricity consumption is false green electricity, an early warning will be triggered indicating an abnormal power type code.

[0062] like If the current electricity consumption is non-green electricity, then the electricity consumption data should be recorded.

[0063] in, At that time, the power source type was wind power. At that time, the power source type was photovoltaic. At that time, the power source was hydroelectric. At that time, the power source was thermal power.

[0064] Furthermore, a deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method is provided, wherein the formula for obtaining the green electricity purity index is:

[0065] ;

[0066] in, The green electricity purity index, Let be the dynamic correction factor at time t. Let be the power grid signal interference intensity at time t. Let be the signal-to-noise ratio of the voltage waveform at time t. This is the preset maximum signal-to-noise ratio. , For signal power, For noise power, For the credibility of the protocol, This represents the physical matching degree.

[0067] In the above scheme, this invention introduces a dynamic correction factor to integrate physical matching degree and protocol credibility degree to obtain a green electricity purity index. The green electricity purity index is compared with a preset green electricity judgment threshold, and combined with the power source type code (wind power, photovoltaic, and hydropower correspond to codes 1, 2, and 3, and thermal power corresponds to code 4) to determine whether the current electricity consumption is green electricity, fake green electricity, or non-green electricity: if the green electricity purity index is not lower than the threshold and the power source type code is wind power, photovoltaic, or hydropower, it is judged as valid green electricity, and the power supply period, power source type code, and green electricity purity index are output to the carbon accounting model; if the index is not lower than the threshold but the power source type code is thermal power, it is judged as fake green electricity and an early warning is triggered; if the index is lower than the threshold, it is judged as non-green electricity, and only regular electricity consumption data is recorded. To address the technical problem that existing green electricity identification methods use fixed weights when fusing protocol information and physical features, failing to adaptively adjust according to operating conditions such as signal quality and interference intensity, leading to deviations from reality in the fusion results when protocol information is interfered with or physical features fluctuate, this invention utilizes a dynamic correction factor to sense the grid signal interference intensity in real time and dynamically adjust the weights of protocol reliability and physical matching in the fusion process. This allows the green electricity purity index to adaptively reflect the relative reliability of the two-dimensional information. Simultaneously, by combining power type coding and threshold-based hard judgment logic, it effectively distinguishes between valid green electricity, false green electricity, and non-green electricity, solving the problem of misjudgment under harsh conditions caused by traditional fixed-weight fusion methods. This significantly improves the robustness and accuracy of green electricity identification, providing a reliable basis for subsequent carbon accounting.

[0068] Furthermore, a deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method is proposed. If the electricity is green, the carbon accounting model obtains the green electricity consumption and smoothed green electricity purity index within the current electricity supply period. The dynamic emission factor is obtained through the basic emission factor, period correction coefficient, and node correction coefficient, including the following sub-steps:

[0069] Collect the real-time power during the power supply period corresponding to the current power consumption, and obtain the green electricity consumption by discretizing the real-time power.

[0070] The green electricity purity index of the continuous identification period is smoothed by using slip mode to obtain the smoothed green electricity purity index.

[0071] By querying the peak and valley time periods to which the current time belongs, the time period correction coefficient for the current time is obtained from the preset time period coefficient table;

[0072] Read the transmission loss power and total power supply of the power grid node where the electricity meter is located, and obtain the node correction coefficient;

[0073] The dynamic emission factor is obtained by using the regional baseline emission factor, the time period correction factor, and the nodal correction factor;

[0074] Dynamic emission factors are acquired through a dual-mode approach of local storage and cloud synchronization, creating a dynamic spatiotemporal emission factor library.

[0075] In the above scheme, after determining that the current electricity consumption is green electricity, the present invention obtains the real-time power during the power supply period through a carbon accounting model, and obtains the green electricity consumption through discretization integration; it uses a slip mode to smooth the green electricity purity index of the continuous identification period through a sliding window to obtain a smoothed green electricity purity index; it queries a preset time period coefficient table to obtain a time period correction coefficient according to the peak and valley time period to which the current time belongs; it reads the transmission loss power and total power supply of the power grid node where the electricity meter is located to calculate the node correction coefficient; it combines the regional benchmark emission factor, the time period correction coefficient and the node correction coefficient to obtain a dynamic emission factor; and it updates the parameters every 15 minutes through local storage and cloud synchronization to build a dynamic spatiotemporal emission factor library. To address the technical problems in existing carbon emission accounting, such as the failure to consider purity reduction in green electricity consumption, the use of fixed regional averages for emission factors, and the neglect of peak-valley power structure differences and node transmission losses, resulting in coarse carbon accounting results that fail to reflect the actual emission reduction effect of green electricity, this invention introduces a green electricity purity index to reduce the purity of green electricity consumption. It dynamically adjusts the emission factor using time-period correction coefficients and node correction coefficients, while employing a slippage model to smooth instantaneous fluctuations in the purity index. Combined with cloud synchronization and local caching, it ensures continuous parameter updates, achieving refined green electricity consumption accounting, spatiotemporal dynamics of emission factors, and reliable accounting parameters. This provides a real-time, accurate, and traceable data foundation for subsequent carbon quota balance calculations.

[0076] Furthermore, a deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method is proposed. The carbon accounting model obtains accumulated carbon emissions through a dynamic emission factor, and obtains corrected carbon emission reductions using green electricity consumption and a smoothed green electricity purity index. The carbon quota balance is then obtained through the accumulated carbon emissions and the corrected carbon emission reductions, comprising the following sub-steps:

[0077] The electricity meter collects real-time load power, and the carbon accounting model combines load type, electricity consumption period, and power fluctuation coefficient to obtain load characteristic coefficients using a weighted summation method.

[0078] Read the dynamic emission factors for the accounting period from the dynamic spatiotemporal emission factor library, and use the real-time load power, dynamic emission factors and load characteristic coefficients to obtain the accumulated carbon emissions for the accounting period through the discretization integration method.

[0079] The corrected carbon emission reduction is obtained by using green electricity consumption, smoothed green electricity purity index, regional benchmark emission factor and transmission loss correction coefficient.

[0080] Carbon allowance balance is obtained by accumulating carbon emissions, corrected carbon emission reductions, and the initial total carbon allowance.

[0081] In the above scheme, during the carbon accounting stage, the present invention collects real-time load power from electricity meters. The carbon accounting model combines load type, electricity consumption period, and power fluctuation coefficient, and uses a weighted summation method to calculate the load characteristic coefficient. The dynamic emission factors within the accounting period are read from the dynamic spatiotemporal emission factor library. Using real-time load power, dynamic emission factors, and load characteristic coefficient, the cumulative carbon emissions within the accounting period are obtained through a discretized integral method. At the same time, the corrected carbon emission reduction is calculated using green electricity consumption, a smoothed green electricity purity index, regional benchmark emission factors, and transmission loss correction coefficients. Finally, the current carbon quota balance is obtained by subtracting the cumulative carbon emissions and the corrected carbon emission reduction from the initial total carbon quota. To address the technical problems of existing carbon accounting methods that only calculate total carbon emissions without dynamically incorporating carbon reductions from green electricity consumption into quota management, and that fail to consider the differences in the impact of load characteristics (such as load type, electricity consumption period, and power fluctuations) on actual carbon emissions, resulting in carbon quota balances failing to reflect the effectiveness of green electricity consumption in real time and a disconnect between accounting and actual electricity consumption behavior, this invention introduces a load characteristic coefficient to correct carbon emission intensity. It independently calculates the carbon reductions after reducing the purity of green electricity and combines them with cumulative carbon emissions to influence the carbon quota balance, achieving a closed-loop coupling of green electricity consumption, carbon emissions, and carbon quotas at the terminal. The use of discretized integrals and dynamic emission factors ensures the real-time performance and spatiotemporal accuracy of the calculation, enabling users to dynamically monitor their carbon quota balance during electricity consumption. This invention provides an accurate and closed-loop data foundation for implementing user-side load regulation, prioritizing green electricity consumption, and providing carbon quota early warning. Attached Figure Description

[0082] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0083] Figure 1 This is a flowchart of a deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method.

[0084] Figure 2 The structure diagram of the interactive convolutional model grouped by physical features. Detailed Implementation

[0085] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0086] It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance, or suggesting any such actual relationship or order between these entities or operations. Additionally, the terms "connected," "linked," etc., can refer to a direct connection between elements or an indirect connection via other elements.

[0087] like Figure 1 As shown, a deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method includes the following steps:

[0088] S1: The electricity meter parses the green electricity identification signal, obtains the green electricity identification frame, verifies the green electricity identification frame, and extracts the core protocol feature parameters of the green electricity identification frame when the verification is successful.

[0089] S11: The electricity meter uses power line carrier to receive the green electricity identification signal sent by the power grid side, and uses Manchester encoding to parse the green electricity identification signal to obtain the green electricity identification frame;

[0090] S12: The electricity meter performs a triple check on the green electricity identification frame using a weighted summation to obtain the protocol's reliability. The formula is as follows:

[0091] ;

[0092] in, For the credibility of the protocol, For frame format verification weights, For signature verification weight, For valid time period verification weights, For frame format verification results, , For signature verification results, , For the valid time period verification results, , .

[0093] In the embodiments, , , .

[0094] It should be noted that if the frame format verification passes, then ,otherwise If the signature verification passes, then ,otherwise The valid time period check determines whether the current time t is within the valid time range T. If it is, then... ,otherwise .

[0095] In this embodiment, the frame format verification protocol includes the frame header, frame trailer, and data length. Signature verification uses an encryption algorithm to verify the signature. The public key extracted from the protocol is compared with a preset public key. Assuming the effective photovoltaic time range T1 is 9:00-17:00, and the current time t is 14:30... When the current time t is 18:00 .

[0096] S13: When Extract the core protocol feature parameters of the green electricity identification frame; the core protocol feature parameters include power type code, power supply period, and power supply priority;

[0097] The power type code is: .

[0098] in, , Power type encoding, At that time, the power source type was wind power. At that time, the power source type was photovoltaic. At that time, the power source was hydroelectric. At that time, the power source type was thermal power;

[0099] The power supply period is as follows: ;

[0100] in, During the power supply period, The power supply start time. This is the time when power supply ends.

[0101] The power supply priority is as follows: ;

[0102] in, For power supply priority, when At that time, it is powered by conventional power supply, when At that time, it supplies electricity using green electricity.

[0103] S2: The meter obtains voltage sampling sequences and current sampling sequences through sampling, extracts multi-dimensional physical features based on the voltage sampling sequences and current sampling sequences, normalizes the multi-dimensional physical features, and obtains physical feature vectors;

[0104] S21: The meter uses a fixed sampling period. The voltage and current on the power-consuming side are sampled to obtain a voltage sampling sequence. and current sampling sequence ;

[0105] in, , For the index of the sampling point, This represents the number of sampling points;

[0106] In the embodiments, , voltage sampling sequence and current sampling sequence The sliding average or median filtering is performed using slip mode (window size 5) to reduce sensor random noise and remove obvious outliers (such as voltage > 450V, current < 0), replacing the outliers with the average of the previous and next effective values.

[0107] S22: Based on voltage sampling sequence and current sampling sequence Extract multi-dimensional physical features; the multi-dimensional physical features include fundamental frequency, total harmonic distortion rate, phase offset, voltage fluctuation amplitude, current peak factor, active power fluctuation coefficient, reactive power ratio, voltage harmonic order, current harmonic amplitude, and waveform stability.

[0108] S221: Perform a Fast Fourier Transform on the voltage sampling sequence to extract the fundamental frequency and voltage amplitude of each spectral line of the grid voltage signal. The formula is as follows:

[0109] ;

[0110] in, Let be the fundamental frequency, and k be the spectral line index in the frequency domain of the Fast Fourier Transform. To return the spectral line index k where the amplitude reaches its maximum value, For Fast Fourier Transform, For voltage sampling sequence After performing a Fast Fourier Transform, the voltage amplitude of the k-th spectral line is... Let be the voltage amplitude of the k-th spectral line. ;

[0111] In the embodiments, photovoltaic and thermal power are used as examples. The fundamental frequency fluctuation of photovoltaic power is small, for example... (Standard is) Thermal power plants have large fluctuations in their fundamental frequency, for example... .

[0112] S222: Extract the fundamental voltage amplitude and harmonic voltage amplitude from the voltage amplitude of each spectral line, and obtain the total harmonic distortion using the fundamental voltage amplitude and harmonic voltage amplitude. The formula is as follows:

[0113] ;

[0114] in, The fundamental voltage amplitude (the voltage amplitude of the first spectral line is taken as the fundamental voltage amplitude). The total harmonic distortion (THD) is the harmonic distortion rate. The amplitude of the harmonic voltage. ;

[0115] In the embodiments, photovoltaic and thermal power are used as examples. The total harmonic distortion rate of photovoltaic power proves that there are fewer harmonics. For example, (Effective harmonics are only 2-5th order), the total harmonic distortion rate of thermal power plants proves that there are more high-order harmonics, for example... .

[0116] S223: Perform a Fast Fourier Transform on the current sampling sequence to obtain the current amplitude of each spectral line, using the following formula:

[0117] ;

[0118] in, For current sampling sequence After performing a fast Fourier transform, the current amplitude of the k-th spectral line is... Let k be the current amplitudes of the spectral lines;

[0119] S224: Take the phase angle of the fundamental complex voltage and fundamental complex current, and use the phase angle to obtain the phase offset. The formula is:

[0120] ;

[0121] in, The phase angle is the complex value of the fundamental voltage. Let be the phase angle of the complex fundamental current. The fundamental voltage is a complex number. For voltage sampling sequence After performing a fast Fourier transform, the complex voltage of the first spectral line is... The fundamental current is a complex number. For current sampling sequence After performing a fast Fourier transform, the complex current of the first spectral line is... This is the phase offset.

[0122] It should be noted that this application uses whole-cycle sampling, so that the fundamental frequency falls exactly on the first spectral line.

[0123] In the embodiment, the phase offset of the photovoltaic for This proves that photovoltaic power has a high power factor and small phase shift, while thermal power has a low power factor. .

[0124] S225: The voltage fluctuation amplitude is obtained by using the maximum and minimum values ​​of the voltage sampling sequence, using the following formula:

[0125] ;

[0126] in, For voltage fluctuation amplitude, The maximum value of the voltage sampling sequence. This represents the minimum value of the voltage sampling sequence. To obtain the maximum value, To obtain the minimum value;

[0127] S226: The current peak factor is obtained by using the maximum value of the effective current value and the absolute value of the current sampling sequence. The formula is as follows:

[0128] ;

[0129] in, The absolute value of the current sampling sequence. The peak current factor, The maximum value of the absolute value of the current sampling sequence. This is the effective value of the current;

[0130] S227: The active power fluctuation coefficient is obtained by using active power and the average active power, using the following formula:

[0131] ;

[0132] in, The active power fluctuation coefficient, Let m be the active power in the m-th time period, where m = 1, 2, 3...M, m is the sequence number of the time period, and M is the total number of time periods. This represents the average active power over all M time periods;

[0133] S2271: Divide the voltage sampling sequence and current sampling sequence into M consecutive time periods, each time period including L sampling points. Calculate the active power of the voltage and current sampling values ​​in the m-th time period using the following formula:

[0134] ;

[0135] Where L is the number of sampling points included in each time period, and j is the sampling sequence number within the time period. , This represents the j-th voltage sample value within the m-th time period. This represents the j-th current sample value within m time intervals;

[0136] S2272: Calculate the average active power over all M time periods to obtain the average active power value. The formula is:

[0137] ;

[0138] S228: The reactive power ratio is obtained by using the total active power and total reactive power, using the following formula:

[0139] ;

[0140] in, This represents the fundamental current amplitude (the current amplitude of the first spectral line). Total active power Total reactive power This represents the proportion of reactive power.

[0141] It should be noted that since the harmonic content in the power grid is usually small, the harmonic power can be ignored, and only the fundamental frequency is used for calculation here.

[0142] S229: Take the highest harmonic order whose harmonic voltage amplitude is greater than 5% of the fundamental voltage amplitude to obtain the voltage harmonic order. The formula is:

[0143] ;

[0144] in, This represents the voltage harmonic order.

[0145] S2210: Obtain the current harmonic amplitude by using the voltage harmonic order. ;

[0146] S2211: Based on the voltage sampling sequence, waveform stability is obtained using the following formula:

[0147] ;

[0148] in, For waveform stability, ;

[0149] S23: Normalize each dimension of the multi-dimensional physical features to obtain the normalized value of each dimension. The formula is as follows:

[0150] ;

[0151] in, The minimum value of each feature dimension. The maximum value of each feature dimension. The original values ​​for each feature dimension. Normalized values ​​for each feature dimension.

[0152] In the embodiment, the active power fluctuation coefficient , , After normalization .

[0153] S24: Integrate the normalized values ​​of each dimension's features to obtain the physical feature vector, using the following formula:

[0154] ;

[0155] in, For physical feature vectors, The normalized fundamental frequency, The normalized total harmonic distortion (THD) This is the normalized phase offset. The normalized voltage fluctuation amplitude, The normalized peak current factor, The normalized active power fluctuation coefficient, The normalized reactive power ratio, For the normalized voltage harmonic order, For the normalized current harmonic amplitude, For normalized waveform stability, This is the matrix transpose.

[0156] S3: The physical feature grouping interactive convolution model divides physical feature vectors into associated and unassociated classes to obtain the physical matching degree.

[0157] like Figure 2As shown, the physical feature grouping interactive convolutional model includes a feature grouping layer, an intra-group convolutional layer, a flattening layer, a concatenation layer, a physical interaction mask fully connected layer, a feature mapping layer, and an output layer; the intra-group convolutional layer includes intra-group convolutional layer 1, intra-group convolutional layer 2, and intra-group convolutional layer 3; the flattening layer includes flattening layer 1, flattening layer 2, and flattening layer 3; and the feature mapping layer includes feature mapping layer 1 and feature mapping layer 2.

[0158] The internal connectivity of the physical feature grouping interactive convolutional model is as follows: the first output of the feature grouping layer is connected to the input of the intra-group convolutional layer 1; the second output of the feature grouping layer is connected to the input of the intra-group convolutional layer 2; the third output of the feature grouping layer is connected to the input of the intra-group convolutional layer 3; the output of the intra-group convolutional layer 1 is connected to the input of the flattening layer 1; the output of the intra-group convolutional layer 2 is connected to the input of the flattening layer 2; the output of the intra-group convolutional layer 3 is connected to the input of the flattening layer 3; the output of the flattening layer 1 is connected to the first input of the splicing layer; the output of the flattening layer 2 is connected to the second input of the splicing layer; the output of the flattening layer 3 is connected to the third input of the splicing layer; the output of the splicing layer is connected to the input of the physical interactive mask fully connected layer; the output of the physical interactive mask fully connected layer is connected to the input of the feature mapping layer 1; the output of the feature mapping layer 1 is connected to the input of the feature mapping layer 2; and the output of the feature mapping layer 2 is connected to the output layer.

[0159] S31: Transfer physical feature vectors The input is fed into the feature grouping layer, which then converts the physical feature vectors into... Based on their inherent correlation, the physical features are divided into different groups in width, resulting in frequency domain feature groups, power fluctuation feature groups, and waveform morphology feature groups. The formula is as follows:

[0160] ;

[0161] in, R is the set of real numbers. For frequency domain feature groups, , This is a feature group for power fluctuations. , This is a waveform morphology feature group. ;

[0162] It is important to note that the feature grouping module groups features according to the inherent correlation of each dimension. Frequency domain feature groups reflect the essential spectral structure of the power source, power fluctuation features reflect the dynamic characteristics of power transmission, and waveform morphology feature groups reflect the geometric shape of the time domain waveform. Different physical feature groups are independent of each other, with high physical correlation between features within a group and weak coupling between features between groups, which facilitates the subsequent extraction of more discriminative intra-group patterns and controllable inter-group interactions.

[0163] S32: The first intra-group convolutional layer performs one-dimensional convolution and activation on the frequency domain feature group to obtain the frequency domain feature map; the second intra-group convolutional layer performs one-dimensional convolution and activation on the power fluctuation feature group to obtain the power fluctuation feature map; the third intra-group convolutional layer performs one-dimensional convolution and activation on the waveform morphology feature group to obtain the waveform morphology feature map. The formula is as follows:

[0164] ;

[0165] in, For frequency domain feature maps, , It is the ReLU activation function. The weights of the convolutional kernels in the group's convolutional layer 1. The bias weights of the convolutional layer 1 within the group. This is a one-dimensional convolution operation. This is a power fluctuation characteristic diagram. , The weights of the convolutional kernels in the two convolutional layers within the group. The bias weights for the two convolutional layers within the group. This is a waveform morphology feature diagram. , The weights of the convolutional kernels in the three convolutional layers within the group. These are the bias weights for the 3 convolutional layers within the group.

[0166] In this embodiment, the convolutional layers within a group employ one-dimensional convolution with a kernel size of 2 and a stride of 1.

[0167] S33: Flattening layer 1 flattens the frequency domain feature map to obtain the frequency domain vector; flattening layer 2 flattens the power fluctuation feature map to obtain the power fluctuation vector; flattening layer 3 flattens the waveform morphology feature map to obtain the waveform morphology vector. The formula is:

[0168] ;

[0169] in, It is a frequency domain vector. , For the flattening operation of the first layer, For power fluctuation vectors, , To flatten the two layers, A waveform shape vector. , To perform a 3-layer flattening operation;

[0170] S34: The splicing layer concatenates the frequency domain vector, power fluctuation vector, and waveform morphology vector to obtain the total feature vector, as shown in the formula:

[0171] ;

[0172] in, This is the total eigenvector. For splicing operations, .

[0173] S35: Based on the correlation between frequency domain vector, power fluctuation vector, and waveform morphology vector, a block mask matrix is ​​constructed, with the following formula:

[0174] ;

[0175] in, This is a block mask matrix, where 1 represents a matrix of all 1s and 0 represents a matrix of all 0s;

[0176] It is important to note that the block mask matrix divides the input and output features according to... Sequentially divided into blocks, each with dimensions of 32, 32, 12, and the masking rule is based on the internal frequency domain vector. Fully allowed, power fluctuation vector interior Fully allowed, waveform shape vector inside Fully allowed; frequency domain vector and power fluctuation vector during cross-block connections. or Fully permissible, power fluctuation vector and waveform shape vector or Fully allowed, frequency domain vector and waveform morphology vector or prohibit.

[0177] In the above scheme, the design of the block mask matrix ensures that there are no connections between unrelated frequency domain vectors and waveform morphology vectors, while related frequency domain vectors and power fluctuation vectors, as well as power fluctuation vectors and waveform morphology vectors, retain bidirectional full connections. (For example, the fundamental frequency and current peak factor do not have a direct physical causal relationship; forcibly connecting them would introduce noise and overfitting.) This permanently sets unreasonable connections to zero, preventing them from participating in training, reducing model complexity, avoiding overfitting with limited samples, and eliminating redundant and harmful interactions while ensuring the flow of necessary information. This improves the model's generalization ability, interpretability, training efficiency, and robustness.

[0178] The block mask matrix and the trainable weight matrix are multiplied element-wise. This product is then multiplied with the total feature vector. After activation, the resulting interactive feature vector is obtained, as shown in the formula:

[0179] ;

[0180] in, For trainable weight matrix, , The feature vector after interaction, For element-wise multiplication, .

[0181] S36: Feature mapping layer 1 maps the interactive feature vectors and then activates them. The activated vectors are then randomly deactivated to obtain the mapped vectors, as shown in the formula:

[0182] ;

[0183] in, The weight matrix of the first layer of the feature mapping is... The bias vector of the first layer of the feature mapping. The activated vector, The activated vector With probability Random inactivation, This is a random deactivation operation. For mapping vectors, .

[0184] In the embodiments, .

[0185] S37: The second layer of feature mapping activates the mapped vector to obtain the mapped features, as shown in the formula:

[0186] ;

[0187] in, The weight matrix for the two-layer feature mapping is... The bias vector for the two layers of feature mapping. For mapping features, .

[0188] S38: The output layer is activated after mapping the features to obtain the physical matching degree, as shown in the formula:

[0189] ;

[0190] in, For physical matching degree, The weight vector of the output layer. For the bias scalar of the output layer, , This is the Sigmoid activation function.

[0191] It is important to note that The closer it is to 1, the better the voltage sampling sequence collected by the meter. and current sampling sequence The higher the degree of matching between the described real-time voltage and current time-domain waveforms and the green electricity standard waveforms, the better.

[0192] The physical feature grouping interactive convolutional model is also trained using a cross-entropy loss function, the formula of which is:

[0193] ;

[0194] in, Let cross-entropy be the loss function. Where i is the batch size and i is the sample index. Let i be the true label of the i-th sample, when At that time, the sample was green electricity. At that time, the sample was non-green electricity. Let be the physical matching degree of the i-th sample.

[0195] In this embodiment, the Adam optimization method is used, with a learning rate of... 1000 iterations, batch size 32, regularization coefficient The training samples consist of 1000 samples each for photovoltaic (power range 50-200kW), wind power (power range 80-300kW), and thermal power (power range 100-400kW).

[0196] In the embodiment, photovoltaic Wind power thermal power Suspicious green electricity (a mixture of photovoltaic and thermal power) .

[0197] S4: By using dynamic correction factors to integrate physical matching degree and protocol credibility degree, a green electricity purity index is obtained. By combining the green electricity purity index with the power type code, it is determined whether the current electricity consumption is green electricity.

[0198] S41: Introducing a dynamic correction factor to fuse physical matching degree and protocol credibility degree, the green electricity purity index is obtained, with the following formula:

[0199] ;

[0200] in, The green electricity purity index, Let be the dynamic correction factor at time t. Let be the power grid signal interference intensity at time t. Let be the signal-to-noise ratio of the voltage waveform at time t. This is the preset maximum signal-to-noise ratio. , For signal power, Noise power;

[0201] In the embodiment, under normal interference, , ,but , If strong interference exists (such as lightning interference). ,but , .

[0202] In the above scheme, the existing fixed weight fusion method cannot detect changes in power grid interference. When the protocol frame is interfered with or tampered with, it is still fused according to a fixed ratio, which is prone to misjudgment. The present invention calculates the weight dynamically based on the real-time signal-to-noise ratio. The stronger the interference, the lower the protocol credibility weight and the higher the physical feature matching weight. Under severe conditions such as strong lightning and load change, it can still accurately identify green electricity based on waveform features, which significantly enhances the anti-interference ability and identification robustness.

[0203] S42: Combining the power type code and the green electricity determination threshold, determine whether the current electricity consumption is green electricity:

[0204] like ,and (Wind power, photovoltaic, hydropower), then the current electricity consumption is green electricity, and the power supply period, power type code, and green electricity purity index of the current electricity consumption are output to the carbon accounting model;

[0205] like ,and If the current electricity consumption is a false green electricity source (for thermal power plants), an early warning will be triggered indicating an abnormal power type code.

[0206] like If the current electricity consumption is non-green electricity, then the electricity consumption data should be recorded.

[0207] in, Threshold for determining green electricity;

[0208] The current electricity consumption specifically refers to the electrical energy corresponding to the real-time voltage and current time-domain waveforms collected by the meter.

[0209] The electricity consumption data includes conventional metering data such as effective voltage value, effective current value, active power, reactive power, and active electrical energy, but does not include green electricity attribute information such as power type code and power supply period in the green electricity identification frame. Fake green electricity can be continuously collected and re-identified with power type code.

[0210] In the embodiments, For example, during normal photovoltaic interference, , , If the current electricity consumption is green electricity, then it is green electricity; for example, when there is strong wind power interference, , , , , If the electricity used is green electricity, then the current electricity consumption is green electricity; for example, when using thermal power, , When it comes to non-green electricity, such as suspected green electricity, , , , If so, it is considered fake green electricity.

[0211] S5: If it is green electricity, the carbon accounting model obtains the green electricity consumption and smoothed green electricity purity index within the current power supply period. Through the basic emission factor, the time period correction coefficient and the node correction coefficient, the dynamic emission factor is obtained.

[0212] S51: Collect the real-time power during the power supply period corresponding to the current electricity consumption, and obtain the green electricity consumption by discretizing the real-time power. The formula is:

[0213] ;

[0214] in, The sampling interval is... The number of data collections during the power supply period. For the b-th sampling time, This represents the accumulated green electricity consumption during the power supply period, where b is the sampling number. The b-th sampling time The corresponding real-time power;

[0215] In the embodiment, the sampling interval is .

[0216] S52: The green electricity purity index of the continuous identification period is smoothed using a slip mode to obtain the smoothed green electricity purity index, as shown in the formula:

[0217]

[0218] in, To adjust the sliding window size, For the first Green electricity purity index for a continuous identification cycle This is the smoothed green electricity purity index.

[0219] In the embodiments, .

[0220] The green electricity purity index of the continuous identification cycle refers to the green electricity purity index output after each cycle of the complete green electricity identification process is executed in multiple consecutive identification cycles.

[0221] It should be noted that in actual continuous monitoring, due to instantaneous interference in the power grid (such as load switching and harmonic fluctuations), the green electricity purity index output in adjacent cycles may have slight jumps (e.g., 0.85, 0.82, 0.86). If the jump values ​​are used directly for carbon accounting or green electricity consumption accumulation, short-term noise will be introduced. Therefore, a slip mode is used to smooth the green electricity purity index of the continuous identification cycle.

[0222] It is also necessary to verify the green electricity consumption and the smoothed green electricity purity index: if or If the data is found to be abnormal during the power supply period, the green electricity identification and calculation will be re-executed during the next power supply period.

[0223] S53: By querying the peak / valley time period (peak / average / valley) to which the current time belongs, obtain the time period correction coefficient for the current time from the preset time period coefficient table. .

[0224] Specifically, the time period correction factor comes from the time period correction factor corresponding to the peak-valley time period division published by the power grid company or power dispatching agency.

[0225] S54: Read the transmission loss power and total power supply of the power grid node where the meter is located, and obtain the node correction coefficient. The formula is:

[0226] ;

[0227] in, Let be the transmission loss power at node s of the power grid. Let be the total power supplied to grid node s, where s is the grid node number. For node s in the power grid;

[0228] S55: The dynamic emission factor is obtained by using the regional baseline emission factor, time-period correction factor, and nodal correction factor. The formula is as follows:

[0229] ;

[0230] in, Let be the dynamic emission factor of grid node s at time t. The regional baseline emission factor (obtained from the power grid company, with a value ranging from 0.65 to 0.78 tons) is used. ).

[0231] S56: Dynamic emission factors are obtained through a dual-mode approach of local storage and cloud synchronization, creating a dynamic spatiotemporal emission factor library.

[0232] Specifically, the system retrieves the latest basic emission factors, time-period correction coefficients, and node correction coefficients from the cloud every 15 minutes and updates the dynamic spatiotemporal emission factor library. If the cloud is interrupted, the system uses the previously successfully retrieved basic emission factors, time-period correction coefficients, and node correction coefficients from the local cache (cache validity period ≤ 2 hours) to ensure that the basic emission factors, time-period correction coefficients, and node correction coefficients are not missing and that carbon accounting is continuous.

[0233] S6: The carbon accounting model obtains the accumulated carbon emissions through dynamic emission factors, obtains the corrected carbon emission reduction by using green electricity consumption and smoothed green electricity purity index, and obtains the carbon quota balance by accumulating carbon emissions and correcting carbon emission reduction.

[0234] S61: Electricity meter collects real-time load power The carbon accounting model combines load type, electricity consumption period, and power fluctuation coefficient, and uses a weighted summation method to obtain the load characteristic coefficient, as shown in the formula:

[0235] ;

[0236] in, Weights for load type, For load type coefficient, Weighted by electricity consumption time period This is the electricity consumption period coefficient. As power fluctuation weight, For power fluctuation coefficient, Let be the load characteristic coefficient at time t. .

[0237] In the embodiment, when When it is a core load (ensuring the necessities of production and daily life). ,when When it is a secondary load (not essential but high value). ,when Under normal load conditions (idle, high energy consumption). ;

[0238] When it is a trough period When it is a normal period, During peak hours, , , , .

[0239] S62: Read the dynamic emission factors for the calculation period from the dynamic spatiotemporal emission factor database. Using real-time load power, dynamic emission factors, and load characteristic coefficients, obtain the accumulated carbon emissions for the calculation period through the discretization integration method. The formula is as follows:

[0240] ;

[0241] in, N1 represents the cumulative carbon emissions during the accounting period, and N1 represents the total number of discrete samples during the accounting period. For discrete sampling point numbers, For the first Next sampling time For the first Real-time load power at the next sampling time For the first The dynamic emission factor of grid node s at the next sampling time. For the first Load characteristic coefficient at the next sampling time The discrete sampling interval.

[0242] In the embodiments, Consistent with the updates of the dynamic spatiotemporal emission factor library, .

[0243] The accounting period is the time window for carbon emission statistics (such as hour, day, month).

[0244] S63: The corrected carbon emission reduction is obtained by using green electricity consumption, smoothed green electricity purity index, regional baseline emission factor and transmission loss correction coefficient.

[0245] S631: Carbon emission reduction is obtained by using green electricity consumption, a smoothed green electricity purity index, and a regional baseline emission factor. The formula is as follows:

[0246] ;

[0247] in, For carbon emission reduction;

[0248] S632: The carbon emission reduction is corrected using a transmission loss correction factor to obtain the corrected carbon emission reduction. The formula is as follows:

[0249] ;

[0250] in, This is the transmission loss correction factor. This is the revised carbon emission reduction.

[0251] In the embodiments, .

[0252] S64: The carbon allowance balance is obtained by accumulating carbon emissions, corrected carbon emission reductions, and the initial total carbon allowance, using the following formula:

[0253] ;

[0254] in, This represents the initial total carbon allowance. This represents the remaining carbon allowance.

[0255] It should be noted that it is possible to At that time, a carbon exceedance warning was triggered.

[0256] In this embodiment, taking a certain industrial park as an example, the accounting period is 24 hours. .

[0257] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method, characterized in that, Includes the following steps: The electricity meter parses the green electricity identification signal to obtain the green electricity identification frame, verifies the green electricity identification frame, and extracts the core protocol feature parameters of the green electricity identification frame when the verification is successful. The meter obtains voltage and current sampling sequences through sampling, extracts multi-dimensional physical features based on the voltage and current sampling sequences, normalizes the multi-dimensional physical features, and obtains a physical feature vector. The physical feature grouping interactive convolutional model divides physical feature vectors into related and unrelated classes to obtain the physical matching degree. By using dynamic correction factors to integrate physical matching degree and protocol reliability, a green electricity purity index is obtained. By combining the green electricity purity index with the power type code, it is determined whether the current electricity consumption is green electricity. If it is green electricity, the carbon accounting model obtains the green electricity consumption and smoothed green electricity purity index within the current electricity supply period, and obtains the dynamic emission factor through the basic emission factor, the period correction coefficient and the node correction coefficient. The carbon accounting model obtains the accumulated carbon emissions through dynamic emission factors, obtains the corrected carbon emission reduction using green electricity consumption and a smoothed green electricity purity index, and obtains the carbon quota balance through the accumulated carbon emissions and the corrected carbon emission reduction.

2. The deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method according to claim 1, characterized in that, The electricity meter parses the green electricity identification signal to obtain a green electricity identification frame, verifies the green electricity identification frame, and extracts the core protocol feature parameters of the green electricity identification frame when the verification passes, including the following sub-steps: The electricity meter uses power line carrier to receive the green electricity identification signal sent by the power grid side, and uses Manchester encoding to parse the green electricity identification signal to obtain the green electricity identification frame; The electricity meter performs triple verification on the green electricity identification frame by weighted summation to obtain the protocol's credibility. When the reliability of the protocol When the credibility threshold is reached, the core protocol feature parameters of the green electricity identification frame are extracted; the core protocol feature parameters include power type encoding, power supply period, and power supply priority.

3. The deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method according to claim 1, characterized in that, The electricity meter obtains voltage and current sampling sequences through sampling, extracts multi-dimensional physical features based on the voltage and current sampling sequences, and normalizes the multi-dimensional physical features to obtain a physical feature vector, including the following sub-steps: The electricity meter samples the voltage and current on the electricity consumption side at a fixed sampling period to obtain voltage sampling sequences and current sampling sequences; Multi-dimensional physical features are extracted based on voltage and current sampling sequences. The multi-dimensional physical characteristics include fundamental frequency, total harmonic distortion rate, phase offset, voltage fluctuation amplitude, current peak factor, active power fluctuation coefficient, reactive power ratio, voltage harmonic order, current harmonic amplitude, and waveform stability. Normalize each dimension of the multi-dimensional physical features to obtain the normalized value of each dimension. The normalized values ​​of each dimension feature are integrated to obtain the physical feature vector.

4. The deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method according to claim 3, characterized in that, The extraction of multi-dimensional physical features based on voltage and current sampling sequences includes the following sub-steps: Perform a fast Fourier transform on the voltage sampling sequence to extract the fundamental frequency and voltage amplitude of each spectral line of the grid voltage signal; The fundamental voltage amplitude and harmonic voltage amplitude are extracted from the voltage amplitude of each spectral line, and the total harmonic distortion rate is obtained using the fundamental voltage amplitude and harmonic voltage amplitude. Perform a fast Fourier transform on the current sampling sequence to obtain the current amplitude of each spectral line; Take the phase angle of the fundamental voltage complex number and the fundamental current complex number, and use the phase angle to obtain the phase offset; The voltage fluctuation amplitude is obtained by using the maximum and minimum values ​​of the voltage sampling sequence. The current peak factor is obtained by using the maximum value of the effective current value and the absolute value of the current sampling sequence. The active power fluctuation coefficient is obtained by using active power and the average active power. The reactive power ratio is obtained by using the total active power and the total reactive power. The highest harmonic order, whose harmonic voltage amplitude is greater than 5% of the fundamental voltage amplitude, is used to obtain the voltage harmonic order. The amplitude of the current harmonics is obtained by analyzing the voltage harmonic order.

5. The deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method according to claim 1, characterized in that, The physical feature grouping interactive convolutional model divides physical feature vectors into associated and unassociated classes for processing, and obtains the physical matching degree through the following sub-steps: The physical feature vectors are input into the feature grouping layer, which divides the physical feature vectors into different physical feature groups in width according to their inherent correlation, thus obtaining frequency domain feature groups, power fluctuation feature groups, and waveform morphology feature groups. The first convolutional layer within the group performs one-dimensional convolution and activation on the frequency domain feature group to obtain the frequency domain feature map; the second convolutional layer within the group performs one-dimensional convolution and activation on the power fluctuation feature group to obtain the power fluctuation feature map; and the third convolutional layer within the group performs one-dimensional convolution and activation on the waveform morphology feature group to obtain the waveform morphology feature map. Flattening layer 1 flattens the frequency domain feature map to obtain the frequency domain vector; flattening layer 2 flattens the power fluctuation feature map to obtain the power fluctuation vector; flattening layer 3 flattens the waveform morphology feature map to obtain the waveform morphology vector. The splicing layer splices the frequency domain vector, power fluctuation vector, and waveform morphology vector to obtain the total feature vector; Based on the correlation between frequency domain vector, power fluctuation vector, and waveform morphology vector, a block mask matrix is ​​constructed; The first feature mapping layer maps and activates the interactive feature vectors. The activated vectors are then randomly deactivated to obtain the mapped vectors. The two-layer feature mapping process maps the mapping vector and then activates it to obtain the mapped features. The output layer is activated after mapping the features to obtain the physical matching degree.

6. The deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method according to claim 5, characterized in that, The step involves element-wise multiplication of the block mask matrix and the trainable weight matrix, followed by matrix multiplication with the total feature vector. After activation, the resulting interactive feature vector is obtained. The formula is as follows: ； in, This is a block mask matrix, where 1 represents a matrix of all 1s and 0 represents a matrix of all 0s. For trainable weight matrix, The feature vector after interaction, For element-wise multiplication, This is the total eigenvector. This is the ReLU activation function.

7. The deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method according to claim 1, characterized in that, The process of using a dynamic correction factor to fuse physical matching degree and protocol reliability degree to obtain a green electricity purity index, and then using the green electricity purity index in conjunction with the power type code to determine whether the current electricity consumption is green electricity, includes the following sub-steps: By introducing a dynamic correction factor and integrating physical matching degree and protocol credibility degree, a green electricity purity index is obtained. ; Combined with power type coding And green electricity determination threshold Determine whether the current electricity consumption is green electricity: like ,and If the current electricity consumption is green electricity, the power supply period, power type code, and green electricity purity index of the current electricity consumption will be output to the carbon accounting model. like ,and If the current electricity consumption is false green electricity, an early warning will be triggered indicating an abnormal power type code. like If the current electricity consumption is non-green electricity, then the electricity consumption data should be recorded. in, At that time, the power source type was wind power. At that time, the power source type was photovoltaic. At that time, the power source was hydroelectric. At that time, the power source was thermal power.

8. The deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method according to claim 1, characterized in that, The formula for obtaining the green electricity purity index is: ； in, The green electricity purity index, Let be the dynamic correction factor at time t. Let be the power grid signal interference intensity at time t. Let be the signal-to-noise ratio of the voltage waveform at time t. This is the preset maximum signal-to-noise ratio. , For signal power, For noise power, For the credibility of the protocol, This represents the physical matching degree.

9. The deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method according to claim 1, characterized in that, If the electricity is green, the carbon accounting model obtains the green electricity consumption and smoothed green electricity purity index within the current electricity supply period. The dynamic emission factor is obtained through the basic emission factor, time period correction coefficient, and nodal correction coefficient, including the following sub-steps: Collect the real-time power during the power supply period corresponding to the current power consumption, and obtain the green electricity consumption by discretizing the real-time power. The green electricity purity index of the continuous identification period is smoothed by using slip mode to obtain the smoothed green electricity purity index. By querying the peak and valley time periods to which the current time belongs, the time period correction coefficient for the current time is obtained from the preset time period coefficient table; Read the transmission loss power and total power supply of the power grid node where the electricity meter is located, and obtain the node correction coefficient; The dynamic emission factor is obtained by using the regional baseline emission factor, the time period correction factor, and the nodal correction factor; Dynamic emission factors are acquired through a dual-mode approach of local storage and cloud synchronization, creating a dynamic spatiotemporal emission factor library.

10. The deep learning-based two-dimensional green electricity identification and dynamic carbon accounting method according to claim 1, characterized in that, The carbon accounting model obtains accumulated carbon emissions through dynamic emission factors, and obtains corrected carbon emission reductions using green electricity consumption and a smoothed green electricity purity index. The carbon allowance balance is then calculated using accumulated carbon emissions and corrected carbon emission reductions, comprising the following sub-steps: The electricity meter collects real-time load power, and the carbon accounting model combines load type, electricity consumption period, and power fluctuation coefficient to obtain load characteristic coefficients using a weighted summation method. Read the dynamic emission factors for the accounting period from the dynamic spatiotemporal emission factor library, and use the real-time load power, dynamic emission factors and load characteristic coefficients to obtain the accumulated carbon emissions for the accounting period through the discretization integration method. The corrected carbon emission reduction is obtained by using green electricity consumption, smoothed green electricity purity index, regional benchmark emission factor and transmission loss correction coefficient. Carbon allowance balance is obtained by accumulating carbon emissions, corrected carbon emission reductions, and the initial total carbon allowance.

Images