Electric carbon metering and credible traceability method and system based on deep learning

Abstract

The invention relates to the technical field of energy metering and carbon emission management, and provides an electricity carbon metering and credible traceability method and system based on deep learning, and the method comprises the steps: collecting the electric energy data, environment and state data, and external correlation data of a power system; performing rejection, denoising and abnormal data monitoring and screening on the data to obtain credible and high-resolution input; constructing a dynamic memory reconstruction network, learning typical time sequence characteristics of the system in a normal power consumption mode, and further identifying an abnormal power consumption mode; a gradient boosting decision tree model GBDT is adopted to carry out dynamic correction with boundary constraint on the carbon emission factors, and then the carbon emission is calculated; and finally, based on a block chain and cryptography technology, carrying out evidence storage and source tracing on the output carbon emission and associated data. Through the method, the electric carbon metering precision is improved.

Description

Technical Field

[0001] This invention relates to the field of energy metering and carbon emission management technology, specifically to a deep learning-based method and system for electricity carbon metering and reliable traceability. Background Technology

[0002] With the introduction of the "dual carbon" target, accurate measurement and reliable record-keeping of carbon emissions during electricity use have become fundamental tasks in areas such as carbon trading and energy conservation regulation. Currently, mainstream electricity carbon measurement methods have the following shortcomings: (1) Static calculation model: Most methods use fixed carbon emission factors and linear formulas, which cannot dynamically adapt to differences in equipment type, environmental changes and electricity consumption behavior, resulting in large errors in measurement results.

[0003] (2) Incomplete data processing: The original electricity consumption data is affected by measurement errors, communication delays and equipment failures, resulting in noise, missing data and anomalies. There is a lack of systematic cleaning and correction mechanisms.

[0004] (3) Lack of a reliable traceability mechanism: Existing systems usually store measurement results in a centralized database, which is at risk of being tampered with and makes it difficult to provide a strong legal basis for carbon trading and auditing.

[0005] (4) Limited source tracing function: It is not possible to perform precise retrieval based on time, equipment, emissions, etc., resulting in low query efficiency.

[0006] Current technologies employ artificial intelligence models to predict regional carbon emission factors, such as the Chinese invention patent application CN120470553A, "Regional Carbon Emission Factor Prediction Model Training Method, Prediction Method, and Electricity Carbon Metering Terminal." This method obtains historical power generation data from a power system database; improves the LSTM based on a constructed multi-head attention mechanism to obtain a lightweight LSTM network with low model complexity; and trains the lightweight LSTM network using a feature matrix to obtain a regional carbon emission factor prediction model. However, the raw data upon which the aforementioned patent's carbon emission factor prediction is based is susceptible to fluctuations due to errors in measurement equipment and power system equipment failures, leading to prediction errors. Furthermore, it does not consider possible abnormal electricity consumption patterns on the user side, preventing dynamic correction of carbon emission factors based on electricity consumption patterns. In summary, this amplifies the error in electricity carbon metering.

[0007] Therefore, how to avoid dynamic errors in data collection and carbon emission factors, and how to ensure the immutability of data through reliable evidence storage technology, have become urgent problems to be solved. Summary of the Invention

[0008] The technical problem to be solved by this invention is how to reduce data errors in power systems, thereby improving the accuracy of carbon metering based on reliable data.

[0009] The present invention solves the above-mentioned technical problems through the following technical means: This invention provides a deep learning-based method for electrocarbon metering and reliable traceability, comprising the following steps: S1. Collect power system data, environmental and status data, and external correlation data; S2. First, exclude significant outliers from the data collected in step S1 and normalize the data. Then, learn the wavelet basis through a neural network and perform threshold denoising on the data after excluding outliers in the wavelet coefficient domain to obtain a multidimensional time series data sequence X. S3. Perform anomaly monitoring on the multidimensional time-series data sequence X, filter out data with abnormal fluctuations, and obtain reliable and high-resolution input. , For the first Time of the first Effective values ​​of dimensional features; S4. Construct a dynamic memory reconstruction network to learn the typical timing features under the normal power consumption mode of the system, and then identify abnormal power consumption modes. S5. Based on the results of abnormal power consumption pattern identification and combined with real-time power grid data, the gradient boosting decision tree model GBDT is used to dynamically correct the carbon emission factor with boundary constraints. S6. Calculate carbon emissions based on the dynamically corrected carbon emission factor; S7. Based on blockchain and cryptography technologies, the output carbon emissions and related data are stored and traced.

[0010] Preferably, the electrical energy data includes voltage, current, active / reactive power, power factor, and frequency of the power grid, photovoltaic system, energy storage system, and load side; the environmental and status data includes ambient temperature and humidity, light intensity, and equipment operating status data that affect system operating efficiency; and the external correlation data includes carbon emission factors and time-of-use pricing strategies.

[0011] Further, step S2 includes the following steps: S21. Use the IQR baseline to identify significant outliers in the raw data. Use a mask to mark the identification results and remove data that is marked as invalid by the mask. S22. Normalize the valid data. The normalization process is as follows:

[0012] Where x represents the original data, x' is the normalized data result, and x min and x max These represent the minimum and maximum values ​​in the original data, respectively. S23, Regarding the signal Wavelet decomposition is performed as follows:

[0013] in, j The number of decomposition layers, These are the wavelet coefficients at the corresponding scale; S24. Use a differentiable threshold function in the coefficient domain. Obtain the set of coefficients after threshold suppression. As shown in the following formula:

[0014] Among them, threshold As trainable parameters; S25, will The signal is reconstructed by performing inverse wavelet transform to obtain a denoised and smoothed time-series signal. As shown in the following formula:

[0015] S26. Construct a multidimensional time-series data sequence X, as shown in the following formula:

[0016] in, Indicates a time step t The observation vector, where C is the dimension.

[0017] Furthermore, step S3 includes the following steps: S31. Divide the continuous time series data into fixed lengths. Divide into multiple sample windows, as shown in the following formula

[0018] in, Indicates the first i A window of samples; Indicates the window length; S32, For each sample window A Fast Fourier Transform is performed on each channel in the time dimension, as shown in the following equation:

[0019] in, This indicates that the c-th channel is at frequency The Fourier transform result is shown below. Calculate the average amplitude spectrum across channels:

[0020] in, Indicates frequency The average amplitude spectrum at that location; Within the preset frequency range The cycle length corresponding to each main frequency is:

[0021] in, The period length is the number of periods found by the FFT. j One main frequency The corresponding period length, The number of columns is the number of periodic segments into which a window length can be divided.

[0022] Complete the original window sequence to... Length and reshape into a two-dimensional tensor:

[0023] in, Indicates according to the number j The two-dimensional tensor after reshaping the main frequency; This represents the sequence after padding; This indicates a reshaping operation; the row direction of the two-dimensional tensor represents a sequence of phases with the same phase during the period, and the column direction represents the phase within the period; S33, convert each two-dimensional tensor Inputting a shared two-dimensional convolutional feature extraction network yields a two-dimensional feature map, as shown in the following equation:

[0024] in, This represents the feature map extracted from the two-dimensional tensor. This represents a two-dimensional convolutional feature extraction network; Flatten the two-dimensional feature map and crop it back to its original time length to obtain one-dimensional temporal features:

[0025] in, Indicates from the j A characteristic sequence of each period; This indicates a flattening and cutting operation; express The dimension, that is d represents the length of the time window, and d represents the feature dimension at each time point. The soft maximum normalized weights are calculated using the amplitude spectrum of each major frequency, as shown in the following formula:

[0026] in, Indicates adaptive fusion weights; Indicates the first j The amplitude of each main frequency; kIndicates the number of selected clock frequencies; The characteristic representation of each period of the weighted fusion is as follows:

[0027] in, This represents the temporal characteristics after fusion; Indicates adaptive fusion weights; S34. Merge the temporal features The global representation vector is obtained by time convergence:

[0028] in, Display window The global representation vector; This indicates an average pooling operation; Indicates the dimension of the global representation vector; Then, the vector of the contrastive learning space is obtained through the projection mapping function g(•), as shown in the following equation:

[0029] in, Represents the projection vector; This represents the normalized projection vector; The Euclidean norm of a vector; S35. Construct an augmented view for each sample using a random augmentation method. Repeat steps S32-S34 to obtain the positive sample projection vector. ; When the sample batch size is B, the contrastive learning loss is obtained using the InfoNCE loss function. As shown in the following formula:

[0030] in, B represents the contrastive learning loss; B represents the batch size of the samples. Indicates sample i The positive sample projection vector; Represents the projection vector of other samples within the batch; Indicates transpose; Indicates the inner product; For temperature parameters; S36. Merge the temporal features Input Reconstruction Network The reconstructed sequence is obtained as follows:

[0031] in, This represents the window data reconstructed by the decoder; This indicates the need to rebuild the network; The mean square error of reconstruction is calculated as follows:

[0032] in, Indicates the mean square error of reconstruction; Indicates the original window number t The value of channel c at time c; Indicates the reconstructed value; S37. Based on steps S35 and S36, the total loss function of the joint optimization is as follows:

[0033] in, and These are the loss weighting coefficients; S38. Calculate the reconstruction error score. As shown in the following formula:

[0034] in, Indicates the first i Window reconstruction error score; Indicates the original window number t The value of channel c at time c; Indicates the reconstructed value; The similarity score, i.e., the maximum similarity with the normal sample projection set Q, is calculated as follows:

[0035] in, Indicates the first i The maximum similarity between the window and the normal sample set; Q represents the set of projection vectors of the normal samples; This represents a specific element in the normal sample projection set Q; Then, the first i The overall anomaly score for the window is:

[0036] For the first i A comprehensive anomaly score for the window; These are the weight parameters.

[0037] S39, the first i Window's overall anomaly score and threshold For comparison, see the following formula:

[0038] Remove tags Data assigned the value of 0 yields reliable and high-resolution input. ,in, For time steps, , This represents the total time duration. For feature dimension, , For the total number of features, For the first Time of the first Effective values ​​of dimensional features.

[0039] Further, step S4 includes the following steps: S41. Construct a dynamic memory reconstruction network, as follows:

[0040] in, It represents the internal state of the network at the previous moment and is used to record the recent time-series changes. It represents the state of the long-term memory pool at the previous moment, storing critical load patterns that span a long time scale; , These are the weight matrices corresponding to the input, short-term memory, and long-term memory, respectively, which are obtained through training. It is a non-linear activation function used to introduce non-linear expressive power; It is the internal state variable at the current moment, which combines short-term and long-term memory information; S42. The reconstructed mean square error (MSE) is used as the indicator for identifying abnormal power consumption patterns, as shown in the following formula:

[0041] in, The reconstructed value of the dynamic memory reconstruction network is determined by the short-term state. and long-term memory The combined effect is generated by the encoder and is regarded as the predicted output of the system under normal power consumption mode; when the MSE is higher than the dynamic threshold set based on the mean-standard deviation MSD or the median-absolute deviation MAD, it is determined to be an abnormal power consumption mode.

[0042] Further, step S5 includes the following steps: S51. Obtain the basic amount of carbon intensity in the power grid, as follows:

[0043] in, Indicates time t The grid carbon intensity, in kgCO2 / kWh, is preferably measured using real-time monitoring data. When real-time data is missing, the grid carbon intensity is estimated based on the unit configuration, where: For the first k The percentage of internet-connected power consumption for this type of power supply. This represents the average carbon intensity of the power source.

[0044] S52, Define the fusion weights Used to dynamically adjust the baseline carbon emission factor Real-time grid carbon intensity The proportion; in, The official annual / regional average carbon emission factor for electricity consumption is used; firstly, a gradient boosting decision tree model (GBDT) is trained using historical data samples, and then the feature vector is input. Gradient Boosting Decision Tree Model (GBDT) Output As shown in the following formula:

[0045] Among them, the input feature vector It includes abnormal electricity consumption pattern identification results, time-of-use pricing, regional load, and meteorological factors; For the Sigmoid function; S53, Calculating Dynamic Carbon Emission Factors As shown in the following formula:

[0046] for t Dynamic carbon emission factors at any given time; S54. Use the exponential moving average (EMA) to filter instantaneous fluctuations for dynamic carbon emission factors. Smoothing is performed as follows:

[0047] in, for t Smoothing results at time -1; The weighting parameter controls the proportion of influence between real-time data and historical smoothing results; S55 sets upper and lower limits for dynamic carbon emission factors that meet the requirements of physical logic or business rules, respectively. and Output the final dynamic carbon emission factor As shown in the following formula: .

[0048] Furthermore, the specific method for training the gradient boosting decision tree model GBDT as described in step S52 is as follows: When real-time grid carbon intensity data is accurate and reliable, the fusion weights are set. The label is close to 1; when the primary reliance needs to be placed on the carbon emission benchmark factor, the fusion weight is set. Labels close to 0; other intermediate cases are assigned fusion weights based on data quality and scenario. Continuous values ​​between 0 and 1; trained using historical data samples, enabling the Gradient Boosting Decision Tree (GBDT) model to output fusion weights given features. .

[0049] Furthermore, the calculation of carbon emissions in step S6 is specifically as follows:

[0050] in, For a moment t Carbon emissions; The output of the dynamic memory reconstruction network during the reconstruction process t Predicted power channel values ​​at any given time; For measurement time intervals; Therefore, the cumulative carbon emissions over a certain monitoring period [1, T] are as follows: .

[0051] Furthermore, step S7 specifically includes: (1) To t Carbon emissions at any time timestamp t Equipment identification d and dynamic factors Generate a unique hash value as a data integrity credential, as follows:

[0052] Where || represents string concatenation. Stored in consortium blockchain blocks; (2) Sensitive information is encrypted in plaintext using Paillier homomorphic encryption and then uploaded to the cloud to achieve full-process privacy protection; (3) Automatically implement evidence enhancement or early warning based on emission thresholds, specifically: when When an exception is triggered on the chain, the exception flag is set. The emission threshold is set for a preset time period. For floating coefficients; exception markers and Bind and synchronously record MSE values ​​during abnormal periods; (4) Online query period emissions At that time, the authenticity of the data is verified by hash comparison, as shown in the following formula:

[0053] in To query the returned value, To correspond to the dynamic factor, consistency indicates that the data has not been tampered with; otherwise, it is marked as untrustworthy and an abnormal node is indicated. (5) Combine the results of abnormal electricity consumption pattern identification with dynamic carbon emission factor logs to achieve classification and location of abnormal causes; (6) A permission system based on the Paillier encryption mechanism distinguishes the search scope.

[0054] This invention also provides a deep learning-based system for measuring and tracing the source of carbon dioxide. The system operates using the aforementioned method and includes the following modules: The data acquisition module is used to collect electrical energy data, environmental and status data, and external correlation data of the power system; The data rejection and denoising module is used to first exclude significant outliers from the data acquired by the data acquisition module and normalize the data. Then, it learns a wavelet basis through a neural network and performs threshold denoising on the data after excluding outliers in the wavelet coefficient domain to obtain a multidimensional time series data sequence X. The data anomaly monitoring module is used to monitor for anomalies in a multidimensional time-series data sequence X, filtering out data with abnormal fluctuations to obtain reliable and high-resolution input. , For the first Time of the first Effective values ​​of dimensional features; The power consumption pattern recognition module is used to build a dynamic memory reconstruction network, learn the typical timing features of the system under normal power consumption patterns, and then identify abnormal power consumption patterns. The carbon emission factor correction module is used to dynamically correct the carbon emission factor with boundary constraints based on the results of abnormal electricity consumption pattern identification and combined with real-time power grid data, using the Gradient Boosting Decision Tree (GBDT) model. The carbon emission calculation module is used to calculate carbon emissions based on the dynamically corrected carbon emission factor. The data storage and traceability module, based on blockchain and cryptography technologies, stores and traces the output carbon emissions and related data.

[0055] The advantages of this invention are: (1) This invention uses data rejection, data denoising, data feature extraction and comparative learning to gradually clean the original data and obtain reliable and high-resolution time series data for subsequent identification of abnormal electricity consumption patterns and dynamic correction of carbon emission factors, thereby improving the accuracy of electricity carbon metering.

[0056] (2) By identifying abnormal electricity consumption patterns, the present invention uses the gradient boosting decision tree model GBDT to make a constrained dynamic correction of carbon emission factors. Compared with the existing technology, which generally uses fixed carbon emission factors or linear formulas, this invention further improves the accuracy of electricity carbon measurement.

[0057] (3) By introducing blockchain and cryptography, this invention can store data and abnormal indicators in the entire carbon metering process in a reliable manner, prevent tampering and make them verifiable, and support traceability query under multiple conditions. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating a deep learning-based method for measuring and tracing the source of carbon dioxide in an embodiment of the present invention. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] Example 1 This embodiment provides a deep learning-based method for electrocarbon metering and reliable traceability, the process of which is as follows: Figure 1 As shown, it includes the following steps: S1. Collect power system data, environmental and status data, and external correlation data; achieve three-dimensional data collection by deploying smart meters, multi-functional power monitoring instruments, IoT environmental sensors, and edge computing gateways at user sites. The power system data includes voltage, current, active / reactive power, power factor, and frequency of the power grid, photovoltaic system, energy storage system, and load side. The environmental and status data includes ambient temperature and humidity, light intensity, and equipment operating status data affecting system efficiency, such as inverter operating mode and battery SOC. The external correlation data includes carbon emission factors and time-of-use pricing strategies.

[0061] To ensure data quality and transmission reliability, the acquisition terminal incorporates high-precision time synchronization and edge computing capabilities, supporting the cleaning, aggregation, and local caching of raw data. Its robust breakpoint resumption mechanism automatically resumes data uploads after network fluctuations or interruptions, ensuring the integrity and continuity of the data stream. This provides a high-quality, highly reliable data foundation for upper-layer applications and offers excellent compatibility and scalability.

[0062] S2. First, exclude significant outliers from the data collected in step S1 and normalize the data. Then, learn the wavelet basis through a neural network and perform threshold denoising on the data after excluding outliers in the wavelet coefficient domain to obtain a multidimensional time series data sequence X. This step involves efficient and accurate preprocessing of the raw power system signal. Core tasks include signal denoising, trend extraction, outlier removal, and preliminary frequency domain or time-frequency domain conversion. Key features in the raw signal are clearly revealed, providing a reliable guarantee for the accuracy and robustness of feature extraction and subsequent detection models, and further improving the system's ability to detect and respond to abnormal behavior. The specific implementation includes the following steps: S21. Use IQR baseline to identify significant outliers in the raw data. The identification results are then masked, and data marked as invalid by the mask is removed. Specifically, a sliding window-based IQR method is used to identify significant outliers.

[0063] in, Indicates the first quartile. Indicates the third quartile. Indicates the interquartile range. This is a constant, with a value of 1.5. A mask is used. The validity of the data is marked, and data with a mask of 0 is discarded to ensure the accuracy and robustness of subsequent modeling.

[0064] S22. Normalize the valid data. The normalization process is as follows:

[0065] Where x represents the original data, x' is the normalized data result, and x min and x max These represent the minimum and maximum values ​​in the original data, respectively. Temperature-related attributes are normalized to the (0, 1) interval, while other types of feature data are uniformly scaled to the (0, 10) interval to ensure the comparability of different features and model compatibility under a unified scale.

[0066] S23, Regarding the signal Wavelet decomposition is performed as follows:

[0067] in, j The number of decomposition layers, These are the wavelet coefficients at the corresponding scale; S24. Use a differentiable threshold function in the coefficient domain. Obtain the set of coefficients after threshold suppression. As shown in the following formula:

[0068] Among them, threshold As a trainable parameter; a typical soft threshold is as follows:

[0069] S25, will The signal is reconstructed by performing inverse wavelet transform to obtain a denoised and smoothed time-series signal. As shown in the following formula:

[0070] S26. Construct a multidimensional time-series data sequence X, as shown in the following formula:

[0071] in, Indicates a time step t The observation vector, where C is the dimension.

[0072] S3. Perform anomaly monitoring on the multidimensional time-series data sequence X, filter out data with abnormal fluctuations, and obtain reliable and high-resolution input. , For the first Time of the first Effective values ​​of dimensional features; This step focuses on the in-depth mining and identification of abnormal patterns in the preprocessed time-series signal. The core tasks include anomaly detection, mutation analysis, and the identification and localization of anomalous sequences. Using contrastive learning, it can accurately capture the non-stationarity and complex dynamic changes in the signal, effectively distinguishing between normal and anomalous fluctuations. Through in-depth analysis of the time-series anomaly detection layer, it not only achieves rapid response to sudden abnormal events but also provides the system with comprehensive and fine-grained anomaly diagnosis capabilities, significantly enhancing the overall system's early warning accuracy and real-time monitoring level. The specific implementation includes the following steps: S31. Divide the continuous time series data into fixed lengths. Divide into multiple sample windows, as shown in the following formula

[0073] in, Indicates the first i A window of samples; Indicates the window length; S32, For each sample window A Fast Fourier Transform is performed on each channel in the time dimension, as shown in the following equation:

[0074] in, This indicates that the c-th channel is at frequency The Fourier transform result is shown below. Calculate the average amplitude spectrum across channels:

[0075] in, Indicates frequency The average amplitude spectrum at that location; Within the preset frequency range The cycle length corresponding to each main frequency is:

[0076] in, The period length is the number of periods found by the FFT. j One main frequency The corresponding period length, The number of columns is the number of periodic segments into which a window length can be divided.

[0077] Complete the original window sequence to... Length and reshape into a two-dimensional tensor:

[0078] in, Indicates according to the number j The two-dimensional tensor after reshaping the main frequency; This represents the sequence after padding; This indicates a reshaping operation; the row direction of the two-dimensional tensor represents a sequence of phases with the same phase during the period, and the column direction represents the phase within the period; S33, convert each two-dimensional tensor Inputting a shared two-dimensional convolutional feature extraction network yields a two-dimensional feature map, as shown in the following equation:

[0079] in, This represents the feature map extracted from the two-dimensional tensor. This represents a two-dimensional convolutional feature extraction network; Flatten the two-dimensional feature map and crop it back to its original time length to obtain one-dimensional temporal features:

[0080] in, Indicates from the j A characteristic sequence of each period; This indicates a flattening and cutting operation; express The dimension, that is d represents the length of the time window, and d represents the feature dimension at each time point. The soft maximum normalized weights are calculated using the amplitude spectrum of each major frequency, as shown in the following formula:

[0081] in, Indicates adaptive fusion weights; Indicates the first j The amplitude of each main frequency; k Indicates the number of selected clock frequencies; The characteristic representation of each period of the weighted fusion is as follows:

[0082] in, This represents the temporal characteristics after fusion; Indicates adaptive fusion weights; S34. Merge the temporal features The global representation vector is obtained by time convergence:

[0083] in, Display window The global representation vector; This indicates an average pooling operation; Indicates the dimension of the global representation vector; Then, the vector of the contrastive learning space is obtained through the projection mapping function g(•), as shown in the following equation:

[0084] in, Represents the projection vector; This represents the normalized projection vector; The Euclidean norm of a vector; S35. Construct an augmented view for each sample using a random augmentation method. Repeat steps S32-S34 to obtain the positive sample projection vector. ; When the sample batch size is B, the contrastive learning loss is obtained using the InfoNCE loss function. As shown in the following formula:

[0085] in, B represents the contrastive learning loss; B represents the batch size of the samples. Indicates sample i The positive sample projection vector; Represents the projection vector of other samples within the batch; Indicates transpose; Indicates the inner product; For temperature parameters; S36. Merge the temporal features Input Reconstruction Network The reconstructed sequence is obtained as follows:

[0086] in, This represents the window data reconstructed by the decoder; This indicates the need to rebuild the network; The mean square error of reconstruction is calculated as follows:

[0087] in, Indicates the mean square error of reconstruction; Indicates the original window number t The value of channel c at time c; Indicates the reconstructed value; S37. Based on steps S35 and S36, the total loss function of the joint optimization is as follows:

[0088] in, and These are the loss weighting coefficients; S38. Calculate the reconstruction error score. As shown in the following formula:

[0089] in, Indicates the first i Window reconstruction error score; Indicates the original window number t The value of channel c at time c; Indicates the reconstructed value; The similarity score, i.e., the maximum similarity with the normal sample projection set Q, is calculated as follows:

[0090] in, Indicates the first i The maximum similarity between the window and the normal sample set; Q represents the set of projection vectors of the normal samples; This represents a specific element in the normal sample projection set Q; Then, the first i The overall anomaly score for the window is:

[0091] For the firsti A comprehensive anomaly score for the window; These are the weight parameters.

[0092] S39, the first i Window's overall anomaly score and threshold For comparison, see the following formula:

[0093] Remove tags Data assigned the value of 0 yields reliable and high-resolution input. ,in, For time steps, , This represents the total time duration. For feature dimension, , For the total number of features, For the first Time of the first Valid values ​​for dimensional features. Threshold. Determined by percentile method of normal data score distribution.

[0094] S4. Construct a dynamic memory reconstruction network to learn the typical timing features under the normal power consumption mode of the system, and then identify abnormal power consumption modes. Traditional LSTM-Autoencoders (LSTM-AEs) suffer from two main shortcomings in processing long-sequence power data: first, insufficient memory capacity for long-term dependencies, leading to decreased prediction accuracy under seasonal or periodic loads; and second, poor reconstruction performance when faced with sudden abnormal electricity consumption behaviors (such as sudden peaks caused by electricity theft or load jumps caused by equipment aging), with abnormal signals easily masked. To address these issues, this embodiment proposes a dynamic memory reconstruction network that introduces a "dual-memory structure" within the autoencoder framework, consisting of: (1) Short-term memory unit: captures rapid changes in signals such as voltage, current, and power in real time; (2) Long-term memory pool: retain key historical patterns and avoid forgetting information that depends on distant sources.

[0095] The two elements are adaptively weighted through a dynamic gating mechanism, enabling the network to simultaneously monitor local fluctuations and global trends, thereby more accurately reconstructing the normal patterns of power load. In implementation, the network's "internal state" can be understood as a memory variable that updates over time.

[0096] The "valid value" obtained after the time-series data anomaly detection layer is basic data purification, which solves static anomalies (such as value jumps caused by sensor failures), but cannot solve dynamic anomalies (such as hidden loads caused by electricity theft, or inefficient power consumption caused by equipment aging). Therefore, this embodiment introduces a dynamic memory reconstruction network to process multi-dimensional time-series data. By encoding and decoding features (such as voltage, current, power, temperature, and humidity), the typical timing characteristics of the system under "normal power consumption mode" can be learned. If there is a significant deviation between the actual data and the reconstructed results, it indicates that a dynamic anomaly may have occurred at that moment. The specific implementation method includes the following steps: S41. Construct a dynamic memory reconstruction network, as follows:

[0097] in, It represents the internal state of the network at the previous moment and is used to record the recent time-series changes. It represents the state of the long-term memory pool at the previous moment, storing critical load patterns that span a long time scale; , These are the weight matrices corresponding to the input, short-term memory, and long-term memory, respectively, which are obtained through training. It is a non-linear activation function used to introduce non-linear expressive power; As an internal state variable at the current moment, it integrates short-term and long-term memory information; through this dynamic memory mechanism, it can stably reconstruct normal power consumption patterns, providing a reliable "benchmark profile" for subsequent anomaly detection.

[0098] S42. The reconstructed mean square error (MSE) is used as the indicator for identifying abnormal power consumption patterns, as shown in the following formula:

[0099] in, The reconstructed value of the dynamic memory reconstruction network is determined by the short-term state. and long-term memory The combined effect is generated by the encoder and is considered as the predicted output of the system under normal power consumption mode. When the MSE is higher than the dynamic threshold set based on the mean-standard deviation (MSD) or median-absolute deviation (MAD), it is judged as an abnormal power consumption mode. Moreover, the higher the reconstruction error (MSE), the more significant the deviation of the power consumption mode from the normal state.

[0100] S5. Based on the abnormal electricity consumption pattern identification results and combined with real-time power grid data, the Gradient Boosting Decision Tree (GBDT) model is used to dynamically correct the carbon emission factor with boundary constraints; the specific implementation includes the following steps: S51. Obtain the basic amount of carbon intensity in the power grid, as follows:

[0101] in, Indicates time t The grid carbon intensity, in kgCO2 / kWh, is preferably measured using real-time monitoring data. When real-time data is missing, the grid carbon intensity is estimated based on the unit configuration, where: For the first k The percentage of grid-connected electricity generated by different power sources, such as coal-fired, gas-fired, wind, and solar power. This represents the average carbon intensity of the power source.

[0102] Unit composition estimation refers to a method that indirectly estimates the overall grid carbon intensity when real-time monitoring data of grid carbon intensity is missing or unavailable. It is based on the power generation ratio of different types of generating units in the grid (such as coal-fired units, gas-fired units, wind turbines, photovoltaic units, nuclear power units, etc.) and combined with the average carbon emission intensity of each type of unit. It is a supplementary means to ensure the continuity of carbon metering.

[0103] S52, Define the fusion weights Used to dynamically adjust the baseline carbon emission factor Real-time grid carbon intensity The proportion; in, The official annual / regional average carbon emission factor for electricity consumption is used; firstly, a gradient boosting decision tree model (GBDT) is trained using historical data samples, and then the feature vector is input. Gradient Boosting Decision Tree Model (GBDT) Output As shown in the following formula:

[0104] Among them, the input feature vector It includes abnormal electricity consumption pattern identification results, time-of-use pricing, regional load, and meteorological factors; For the Sigmoid function; The specific method for training the gradient boosting decision tree model GBDT is as follows: When real-time grid carbon intensity data is accurate and reliable, the fusion weights are set. The label is close to 1; when the primary reliance needs to be placed on the carbon emission benchmark factor, the fusion weight is set. Labels close to 0; other intermediate cases are assigned fusion weights based on data quality and scenario. Continuous values ​​between 0 and 1; trained using historical data samples, enabling the Gradient Boosting Decision Tree (GBDT) model to output fusion weights given features. That is, under given characteristic conditions, how reliable should the real-time carbon intensity be?

[0105] When an abnormal power consumption pattern is identified, the feature vector The "reconstruction error" eigenvalue increases, and when input into the GBDT model, its output value... The weights will decrease accordingly after the Sigmoid function mapping. Automatic adjustment (e.g., from 0.8 to 0.3) means that when calculating the dynamic carbon emission factor, the system will reduce its reliance on real-time carbon intensity and rely more on the baseline factor to avoid distortion caused by abnormal power consumption.

[0106] S53, Calculating Dynamic Carbon Emission Factors As shown in the following formula:

[0107] for t Dynamic carbon emission factors at any given time; S54. Use the exponential moving average (EMA) to filter instantaneous fluctuations for dynamic carbon emission factors. Smoothing is performed as follows:

[0108] in, for t Smoothing results at time -1; The weighting parameter controls the proportion of influence between real-time data and historical smoothing results; S55 sets upper and lower limits for dynamic carbon emission factors that meet the requirements of physical logic or business rules, respectively. and Output the final dynamic carbon emission factor As shown in the following formula: .

[0109] S6. Calculate carbon emissions based on the dynamically corrected carbon emission factor; the specific formula is as follows:

[0110] in, For a moment t Carbon emissions; The output of the dynamic memory reconstruction network during the reconstruction process t Predicted power channel values ​​at any given time; For measurement time intervals; Therefore, the cumulative carbon emissions over a certain monitoring period [1, T] are as follows: .

[0111] S7. Based on blockchain and cryptographic technologies, the output carbon emissions and related data are stored and traced. Specifically, this includes: (1) To t Carbon emissions at any time timestamp t Equipment identification d and dynamic factors Generate a unique hash value as a data integrity credential, as follows:

[0112] Where || represents string concatenation. Stored in consortium blockchain blocks; (2) Sensitive information is encrypted in plaintext using Paillier homomorphic encryption and then uploaded to the cloud to achieve full-process privacy protection; specifically, the cloud can only perform aggregation operations on the encryption gradient and cannot reverse the original data. The ciphertext is as follows:

[0113] in, is the encrypted ciphertext, and m is the plaintext of sensitive information (such as device ID, power consumption data, local gradient parameters, etc.). and For public key parameters, The formula uses Paillier homomorphic encryption to protect the privacy of sensitive data, ensuring that the data cannot be directly parsed during transmission and storage. Only authorized nodes (such as regulatory agencies and auditors) can decrypt the data using a private key, thus protecting user privacy.

[0114] (3) Automatically implement evidence enhancement or early warning based on emission thresholds, specifically: when When an exception is triggered on the chain, the exception flag is set. The emission threshold is set for a preset time period. The floating coefficient is set to 1.2 in this embodiment based on historical normal emission distribution; anomaly markers and Binding and synchronously recording of MSE values ​​during abnormal periods provides relevant clues for tracing the source; (4) Online query period emissions At that time, the authenticity of the data is verified by hash comparison, as shown in the following formula:

[0115] in To query the returned value, To correspond with dynamic factors, consistency indicates that the data has not been tampered with; otherwise, it is marked as untrustworthy and an abnormal node is indicated. The verification result is synchronously associated with the blockchain block height and transaction ID, and users can directly view the signature of the evidence storage node, ensuring that the "measurement-evidence storage-query" link is transparent and traceable.

[0116] (5) Combining the results of abnormal electricity consumption pattern identification with dynamic carbon emission factor logs, the abnormal causes are classified and located; and a report is output including: the on-chain evidence storage location (block ID) during the abnormal period, and the associated parameter logs (such as... Weights, EMA smoothing records, and responsibility nodes (data acquisition / metering nodes).

[0117] Specifically, If an abnormal power consumption pattern is identified, then associate it with the power prediction log ( (Deviation from measured value) indicates "abnormal power consumption mode"; If dynamic carbon emission factors Exceeding Then retrieve the line loss rate and the grid carbon intensity. Time series curves were analyzed to identify "factor calculation anomalies". (6) A permission system based on Paillier encryption mechanism distinguishes the search scope. The search dimension supports the combination of "time interval + device ID + emission threshold" (e.g., "device d1 is in peak hours and..."). Records of >100kgCO2”, outputting a statistical summary (total emissions, percentage of anomalies) and a link to the original evidence.

[0118] Specifically, For enterprise users, the permissions are limited to querying only the emissions and average dynamic factor of their own device d. and anomaly markers, preventing access to encrypted raw data (such as...) (real-time value); If the entity is a regulator or auditor, the access is granted by decrypting the data using a private key, allowing them to retrieve the entire dataset (including...). The original curves and unit composition estimation logs are stored, but the operation logs are automatically uploaded to the blockchain for verification, ensuring that permissions are controllable.

[0119] Example 2 This embodiment also provides a deep learning-based system for measuring and tracing the source of carbon dioxide. The system executes the method described in Embodiment 1 during operation and includes the following modules: The data acquisition module is used to collect electrical energy data, environmental and status data, and external correlation data of the power system; The data rejection and denoising module is used to first exclude significant outliers from the data acquired by the data acquisition module and normalize the data. Then, it learns a wavelet basis through a neural network and performs threshold denoising on the data after excluding outliers in the wavelet coefficient domain to obtain a multidimensional time series data sequence X. The data anomaly monitoring module is used to monitor for anomalies in a multidimensional time-series data sequence X, filtering out data with abnormal fluctuations to obtain reliable and high-resolution input. , For the first Time of the first Effective values ​​of dimensional features; The power consumption pattern recognition module is used to build a dynamic memory reconstruction network, learn the typical timing features of the system under normal power consumption patterns, and then identify abnormal power consumption patterns. The carbon emission factor correction module is used to dynamically correct the carbon emission factor with boundary constraints based on the results of abnormal electricity consumption pattern identification and combined with real-time power grid data, using the Gradient Boosting Decision Tree (GBDT) model. The carbon emission calculation module is used to calculate carbon emissions based on the dynamically corrected carbon emission factor. The data storage and traceability module, based on blockchain and cryptography technologies, stores and traces the output carbon emissions and related data.

[0120] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A deep learning-based method for the measurement and reliable traceability of carbon dioxide, characterized in that, Includes the following steps: S1. Collect power system data, environmental and status data, and external correlation data; S2. First, exclude significant outliers from the data collected in step S1 and normalize the data. Then, learn the wavelet basis through a neural network and perform threshold denoising on the data after excluding outliers in the wavelet coefficient domain to obtain a multidimensional time series data sequence X. S3. Perform anomaly monitoring on the multidimensional time-series data sequence X, filter out data with abnormal fluctuations, and obtain reliable and high-resolution input. , For the first Time of the first Effective values ​​of dimensional features; S4. Construct a dynamic memory reconstruction network to learn the typical timing features under the normal power consumption mode of the system, and then identify abnormal power consumption modes. S5. Based on the results of abnormal power consumption pattern identification and combined with real-time power grid data, the gradient boosting decision tree model GBDT is used to dynamically correct the carbon emission factor with boundary constraints. S6. Calculate carbon emissions based on the dynamically corrected carbon emission factor; S7. Based on blockchain and cryptography technologies, the output carbon emissions and related data are stored and traced.

2. The method for electrocarbon metering and reliable traceability based on deep learning according to claim 1, characterized in that, The electrical energy data includes voltage, current, active / reactive power, power factor, and frequency of the power grid, photovoltaic system, energy storage system, and load side; the environmental and status data includes ambient temperature and humidity, light intensity, and equipment operating status data that affect system operating efficiency; and the external correlation data includes carbon emission factors and time-of-use pricing strategies.

3. The method for electrocarbon metering and reliable traceability based on deep learning according to claim 1, characterized in that, Step S2 includes the following steps: S21. Use the IQR baseline to identify significant outliers in the raw data. Use a mask to mark the identification results and remove data that is marked as invalid by the mask. S22. Normalize the valid data. The normalization process is as follows: Where x represents the original data, x' is the normalized data result, and x min and x max These represent the minimum and maximum values ​​in the original data, respectively. S23, Regarding the signal Wavelet decomposition is performed as follows: in, j The number of decomposition layers, These are the wavelet coefficients at the corresponding scale; S24. Use a differentiable threshold function in the coefficient domain. Obtain the set of coefficients after threshold suppression. As shown in the following formula: Among them, threshold As trainable parameters; S25, will The signal is reconstructed by performing inverse wavelet transform to obtain a denoised and smoothed time-series signal. As shown in the following formula: S26. Construct a multidimensional time-series data sequence X, as shown in the following formula: in, Indicates a time step t The observation vector, where C is the dimension.

4. The method for electrocarbon metering and reliable traceability based on deep learning according to claim 3, characterized in that, Step S3 includes the following steps: S31. Divide the continuous time series data into fixed lengths. Divide into multiple sample windows, as shown in the following formula in, Indicates the first i A window of samples; Indicates the window length; S32, For each sample window A Fast Fourier Transform is performed on each channel in the time dimension, as shown in the following equation: in, This indicates that the c-th channel is at frequency The Fourier transform result is shown below. Calculate the average amplitude spectrum across channels: in, Indicates frequency The average amplitude spectrum at that location; Within the preset frequency range The cycle length corresponding to each main frequency is: in, The period length is the number of periods found by the FFT. j One main frequency The corresponding period length, The number of columns is the number of periodic segments into which a window length can be divided. Complete the original window sequence to... Length and reshape into a two-dimensional tensor: in, Indicates according to the number j The two-dimensional tensor after reshaping the main frequency; This represents the sequence after padding; This indicates a reshaping operation; the row direction of the two-dimensional tensor represents a sequence of phases with the same phase during the period, and the column direction represents the phase within the period; S33, convert each two-dimensional tensor Inputting a shared two-dimensional convolutional feature extraction network yields a two-dimensional feature map, as shown in the following equation: in, This represents the feature map extracted from the two-dimensional tensor. This represents a two-dimensional convolutional feature extraction network; Flatten the two-dimensional feature map and crop it back to its original time length to obtain one-dimensional temporal features: in, Indicates from the j A characteristic sequence of each period; This indicates a flattening and cutting operation; express The dimension, that is d represents the length of the time window, and d represents the feature dimension at each time point. The soft maximum normalized weights are calculated using the amplitude spectrum of each major frequency, as shown in the following formula: in, Indicates adaptive fusion weights; Indicates the first j The amplitude of each main frequency; k Indicates the number of selected clock frequencies; The characteristic representation of each period of the weighted fusion is as follows: in, This represents the temporal characteristics after fusion; Indicates adaptive fusion weights; S34. Merge the temporal features The global representation vector is obtained by time convergence: in, Display window The global representation vector; This indicates an average pooling operation; Indicates the dimension of the global representation vector; Then through the projection mapping function g( The vectors of the contrastive learning space are obtained as follows: in, Represents the projection vector; This represents the normalized projection vector; The Euclidean norm of a vector; S35. Construct an augmented view for each sample using a random augmentation method. Repeat steps S32-S34 to obtain the positive sample projection vector. ; When the sample batch size is B, the contrastive learning loss is obtained using the InfoNCE loss function. As shown in the following formula: in, B represents the contrastive learning loss; B represents the batch size of the samples. Indicates sample i The positive sample projection vector; Represents the projection vector of other samples within the batch; Indicates transpose; Indicates the inner product; For temperature parameters; S36. Merge the temporal features Input Reconstruction Network The reconstructed sequence is obtained as follows: in, This represents the window data reconstructed by the decoder; This indicates the need to rebuild the network; The mean square error of reconstruction is calculated as follows: in, Indicates the mean square error of reconstruction; Indicates the original window number t The value of channel c at time c; Indicates the reconstructed value; S37. Based on steps S35 and S36, the total loss function of the joint optimization is as follows: in, and These are the loss weighting coefficients; S38. Calculate the reconstruction error score. As shown in the following formula: in, Indicates the first i Window reconstruction error score; Indicates the original window number t The value of channel c at time c; Indicates the reconstructed value; The similarity score, i.e., the maximum similarity with the normal sample projection set Q, is calculated as follows: in, Indicates the first i The maximum similarity between the window and the normal sample set; Q represents the set of projection vectors of the normal samples; This represents a specific element in the normal sample projection set Q; Then, the first i The overall anomaly score for the window is: For the first i A comprehensive anomaly score for the window; These are weight parameters; S39, the first i Window's overall anomaly score and threshold For comparison, see the following formula: Remove tags Data assigned the value of 0 yields reliable and high-resolution input. ,in, For time steps, , This represents the total time duration. For feature dimension, , For the total number of features, For the first Time of the first Effective values ​​of dimensional features.

5. The method for electrocarbon metering and reliable traceability based on deep learning according to claim 4, characterized in that, Step S4 includes the following steps: S41. Construct a dynamic memory reconstruction network, as follows: in, It represents the internal state of the network at the previous moment and is used to record the recent time-series changes. It represents the state of the long-term memory pool at the previous moment, storing critical load patterns that span a long time scale; , These are the weight matrices corresponding to the input, short-term memory, and long-term memory, respectively, which are obtained through training. It is a non-linear activation function used to introduce non-linear expressive power; It is the internal state variable at the current moment, which combines short-term and long-term memory information; S42. The reconstructed mean square error (MSE) is used as the indicator for identifying abnormal power consumption patterns, as shown in the following formula: in, The reconstructed value of the dynamic memory reconstruction network is determined by the short-term state. and long-term memory The combined effect is generated by the encoder and is regarded as the predicted output of the system under normal power consumption mode; when the MSE is higher than the dynamic threshold set based on the mean-standard deviation MSD or the median-absolute deviation MAD, it is determined to be an abnormal power consumption mode.

6. The method for electrocarbon metering and reliable traceability based on deep learning according to claim 5, characterized in that, Step S5 includes the following steps: S51. Obtain the basic amount of carbon intensity in the power grid, as follows: in, Indicates time t The grid carbon intensity, in kgCO2 / kWh, is preferably measured using real-time monitoring data. When real-time data is missing, the grid carbon intensity is estimated based on the unit configuration, where: For the first k The percentage of internet-connected power consumption for this type of power supply. This represents the average carbon intensity of the power source; S52, Define the fusion weights Used to dynamically adjust the baseline carbon emission factor Real-time grid carbon intensity The proportion; in, The official annual / regional average carbon emission factor for electricity consumption is used; firstly, a gradient boosting decision tree model (GBDT) is trained using historical data samples, and then the feature vector is input. Gradient Boosting Decision Tree Model (GBDT) Output As shown in the following formula: Among them, the input feature vector It includes abnormal electricity consumption pattern identification results, time-of-use pricing, regional load, and meteorological factors; For the Sigmoid function; S53, Calculating Dynamic Carbon Emission Factors As shown in the following formula: for t Dynamic carbon emission factors at any given time; S54. Use the exponential moving average (EMA) to filter instantaneous fluctuations for dynamic carbon emission factors. Smoothing is performed as follows: in, for t Smoothing results at time -1; The weighting parameter controls the proportion of influence between real-time data and historical smoothing results; S55 sets upper and lower limits for dynamic carbon emission factors that meet the requirements of physical logic or business rules, respectively. and Output the final dynamic carbon emission factor As shown in the following formula: 。 7. The method for electrocarbon metering and reliable traceability based on deep learning according to claim 6, characterized in that, The specific method for training the gradient boosting decision tree model GBDT as described in step S52 is as follows: When real-time grid carbon intensity data is accurate and reliable, the fusion weights are set. The label is close to 1; when the primary reliance needs to be placed on the carbon emission benchmark factor, the fusion weight is set. Labels close to 0; other intermediate cases are assigned fusion weights based on data quality and scenario. Continuous values ​​between 0 and 1; trained using historical data samples, enabling the Gradient Boosting Decision Tree (GBDT) model to output fusion weights given features. .

8. The method for electrocarbon metering and reliable traceability based on deep learning according to claim 6, characterized in that, The calculation of carbon emissions in step S6 is specifically as follows: in, For a moment t Carbon emissions; The output of the dynamic memory reconstruction network during the reconstruction process t Predicted power channel values ​​at any given time; For measurement time intervals; Therefore, the cumulative carbon emissions over a certain monitoring period [1, T] are as follows: 。 9. The method for electrocarbon metering and reliable traceability based on deep learning according to claim 8, characterized in that, Step S7 specifically includes: (1) To t Carbon emissions at any time timestamp t Equipment identification d and dynamic factors Generate a unique hash value as a data integrity credential, as follows: Where || represents string concatenation. Stored in consortium blockchain blocks; (2) Sensitive information is encrypted in plaintext using Paillier homomorphic encryption and then uploaded to the cloud to achieve full-process privacy protection; (3) Automatically implement evidence enhancement or early warning based on emission thresholds, specifically: when When an exception is triggered on the chain, the exception flag is set. The emission threshold is set for a preset time period. For floating coefficients; exception markers and Bind and synchronously record MSE values ​​during abnormal periods; (4) Online query period emissions At that time, the authenticity of the data is verified by hash comparison, as shown in the following formula: in To query the returned value, To correspond to the dynamic factor, consistency indicates that the data has not been tampered with; otherwise, it is marked as untrustworthy and an abnormal node is indicated. (5) Combine the results of abnormal electricity consumption pattern identification with dynamic carbon emission factor logs to achieve classification and location of abnormal causes; (6) A permission system based on the Paillier encryption mechanism distinguishes the search scope.

10. A deep learning-based system for measuring and credibly tracing the source of carbon dioxide, characterized in that, Includes the following modules: The data acquisition module is used to collect electrical energy data, environmental and status data, and external correlation data of the power system; The data rejection and denoising module is used to first exclude significant outliers from the data acquired by the data acquisition module and normalize the data. Then, it learns a wavelet basis through a neural network and performs threshold denoising on the data after excluding outliers in the wavelet coefficient domain to obtain a multidimensional time series data sequence X. The data anomaly monitoring module is used to monitor for anomalies in a multidimensional time-series data sequence X, filtering out data with abnormal fluctuations to obtain reliable and high-resolution input. , For the first Time of the first Effective values ​​of dimensional features; The power consumption pattern recognition module is used to build a dynamic memory reconstruction network, learn the typical timing features of the system under normal power consumption patterns, and then identify abnormal power consumption patterns. The carbon emission factor correction module is used to dynamically correct the carbon emission factor with boundary constraints based on the results of abnormal electricity consumption pattern identification and combined with real-time power grid data, using the Gradient Boosting Decision Tree (GBDT) model. The carbon emission calculation module is used to calculate carbon emissions based on the dynamically corrected carbon emission factor. The data storage and traceability module, based on blockchain and cryptography technologies, stores and traces the output carbon emissions and related data.

Images