Enterprise-level power consumption health state dynamic evaluation method and system based on transfer learning

Abstract

The application discloses a kind of enterprise-level power consumption health state dynamic evaluation method and system based on transfer learning.Method includes four steps: using label complete source domain power consumption data to complete cleaning normalization, train deep neural network source domain identification model;Collect target domain user load, power quality and other original data, after interpolation, filtering cleaning, reuse source domain statistical parameter normalization, generate target domain feature matrix;Freeze the bottom layer general feature layer of source domain model, build adaptive network containing feature extractor, label predictor, field discriminator to carry out adversarial transfer training, confuse source domain and target domain feature distribution, obtain the evaluation model adapted to target domain;Real-time input target domain data output power consumption health score and grade, low score triggers gradient method abnormal attribution, generates early warning report push.The application solves the problems of data sparsity, poor cross-industry model generalization and label scarcity for small and micro enterprises, newly registered enterprises, and realizes minute-level interpretable dynamic power consumption risk assessment.

Description

Technical Field

[0001] This application relates to the field of power systems and their automation technology, specifically to an enterprise-level dynamic assessment method and system for power health status based on transfer learning. Background Technology

[0002] Currently, power grid companies have accumulated massive amounts of user electricity consumption data and developed various electricity status monitoring and analysis technologies based on this data. Existing technologies typically target user groups with complete electricity consumption data and rich historical records, such as key industrial parks and large manufacturing enterprises. They utilize continuously collected data such as electricity load curves and power quality indicators to train classification or regression models to assess electricity efficiency and identify any abnormal equipment operation. These models have achieved good results in scenarios with sufficient data and relatively stable user electricity consumption patterns. For example, for large manufacturing enterprises within the same industrial park, a specialized status identification model can be trained to determine whether the enterprise has abnormal energy consumption or potential equipment hazards.

[0003] However, in the actual business scenario of serving a large number of industrial and commercial users for power grid companies, the existing technology has the following shortcomings.

[0004] First, for newly registered enterprises, users relocated from other regions, or small and micro-sized industrial and commercial users, their historical electricity consumption data is often very sparse or even completely missing. These users may only have a few months of electricity consumption records, or even no electricity consumption history available for analysis. This makes it impossible for models trained on large amounts of historical data to be applied to these users, and the models cannot provide effective evaluation results.

[0005] Second, there are significant differences in electricity consumption patterns and load characteristics among different industries. For example, the electricity load of manufacturing enterprises typically exhibits continuous and cyclical fluctuations, while the electricity consumption of service enterprises may be concentrated during the day or at specific times, with smaller fluctuations. When models trained on manufacturing data are directly applied to service enterprises, their assessment accuracy drops significantly, and the models lack cross-industry generalization ability.

[0006] Third, accurate labels for the health status of electricity usage require on-site verification or user feedback. For example, to determine whether a company's electricity usage is "abnormal," it is necessary to send staff to the site to inspect the equipment or wait for users to proactively report faults. This method of obtaining labels is costly, and the number of labels that can be obtained is very limited. This makes it difficult to effectively train supervised learning models that rely on a large amount of labeled data.

[0007] In conclusion, improvements to existing technologies are necessary. Summary of the Invention

[0008] The purpose of this invention is to solve the technical problem in the prior art that the scarcity of enterprise electricity consumption data, the significant differences in electricity consumption patterns between industries, and the difficulty in obtaining health status labels make it impossible to accurately assess the electricity health status of newly registered enterprises, users migrating across regions, or small and micro industrial and commercial users with scarce data. The invention provides an enterprise-level dynamic assessment method and system for electricity health status based on transfer learning.

[0009] To achieve the objectives of this invention, the technical solution provided by this invention is as follows: First aspect This invention provides a dynamic assessment method for enterprise-level power health status based on transfer learning, comprising the following steps: Step 101: Construct and train a source domain power health status recognition model; Step 102: Acquire and preprocess the target domain user data to obtain a standardized target domain feature matrix; Step 103: Perform transfer learning to adapt the source domain knowledge to the target domain, and obtain the adapted dynamic assessment model of the target domain's electricity health status. Step 104: Real-time dynamic assessment and early warning.

[0010] Further, step 101 includes the following steps: This step uses high-value labeled data from the source domain to train an initial recognition model, serving as the basis for knowledge transfer.

[0011] Step 1011: Obtain the power consumption characteristic data and corresponding power health status labels of source domain users from the enterprise data center to form the source domain training dataset; Step 1012: Standardize the source domain training dataset, including missing value imputation, outlier removal, and data normalization, to generate a standardized source domain feature matrix; Step 1013: Construct a deep neural network classification model. Use the standardized source domain feature matrix as input and the electricity health status label as a supervision signal for training. The training objective is to minimize the classification error and obtain the trained source domain electricity health status identification model.

[0012] Further, step 102 includes the following steps: This step collects real-time electricity consumption data from users in the target domain and performs the same preprocessing operations as the source domain data to ensure data compatibility.

[0013] Step 1021: Obtain the load curve data, power quality index data and equipment operation logs of users in the target domain in real time through the electricity information acquisition system to obtain the raw electricity consumption data of the target domain; Step 1022: Clean the original target domain electricity consumption data, fill in the missing time periods using linear interpolation, and smooth noise points using sliding window filtering to obtain the cleaned data; Step 1023: Perform the same normalization operation as in step 1012 on the cleaned data, using the statistical values ​​of the source domain data as the normalization parameters, to generate a standardized target domain feature matrix.

[0014] Further, step 103 includes the following steps: This step utilizes transfer learning technology to transfer knowledge about healthy electricity use from the source domain identification model to the target domain, thereby constructing a target domain evaluation model.

[0015] Step 1031: Load the source domain power health status identification model trained in step 1013, freeze the first and second layer network parameters of the model, and retain its underlying general feature representation capability; Step 1032: Construct a domain adaptive network, which consists of three parts: a feature extractor, a label predictor, and a domain discriminator. The feature extractor uses unfrozen model layers, the label predictor is used to predict the power health status, and the domain discriminator is used to determine whether the feature source is the source domain or the target domain. Step 1033: Input the source domain feature matrix and the target domain feature matrix into the domain adaptive network for adversarial training; Training is performed alternately in two phases: in the first phase, the neighborhood discriminator is fixed, and the feature extractor and label predictor are updated to minimize the healthy state classification loss; in the second phase, the feature extractor and label predictor are fixed, and the neighborhood discriminator is updated to maximize the source discrimination accuracy. The process is iterated repeatedly until the neighborhood discriminator can no longer distinguish the source of features. At this point, the neighborhood discriminator is removed, and the feature extractor and label predictor are combined into an adapted dynamic assessment model of the target domain's electricity health status.

[0016] Further, step 104 includes the following steps: This step uses the adapted model to perform minute-level dynamic scoring of users in the target domain and triggers anomaly attribution and risk warnings.

[0017] Step 1041: Input the target domain feature matrix, which has been collected and preprocessed in real time, into the target domain power health status dynamic assessment model line by line; Step 1042: The target domain power health status dynamic assessment model performs forward calculation on the input data and outputs a health score value; according to the preset threshold range, the health score value is mapped to the corresponding power health status level, and the assessment time and results are recorded; Step 1043: When the health score value is lower than the health level threshold, trigger the abnormal attribution analysis; use the gradient-based feature importance calculation method to calculate the partial derivative of the output of the target domain electricity health status dynamic assessment model with respect to each input feature, and identify the features with negative partial derivatives and large absolute values ​​as the key abnormal causes of the score decline. Step 1044: Based on the health score, electricity health status level, assessment time and results, and key abnormal causes, generate a risk warning report and push it through the enterprise business system. It can be selectively pushed to account managers and enterprise users.

[0018] Second aspect Corresponding to the above methods, the present invention provides an enterprise-level dynamic assessment system for power health status based on transfer learning, comprising the following modules: The model building and training module is used to build and train a source domain power health status recognition model; The target domain feature matrix acquisition module is used to acquire and preprocess target domain user data to obtain a standardized target domain feature matrix. The dynamic evaluation model acquisition module is used to perform transfer learning, adapting source domain knowledge to the target domain to obtain the adapted dynamic evaluation model of the target domain's electricity health status. The early warning module is used for real-time dynamic assessment and early warning.

[0019] Furthermore, the model construction and training module is specifically used to perform the following steps: Step 1011: Obtain the power consumption characteristic data and corresponding power health status labels of source domain users from the enterprise data center to form the source domain training dataset; Step 1012: Standardize the source domain training dataset, including missing value imputation, outlier removal, and data normalization, to generate a standardized source domain feature matrix; Step 1013: Construct a deep neural network classification model. Use the standardized source domain feature matrix as input and the electricity health status label as a supervision signal for training. The training objective is to minimize the classification error and obtain the trained source domain electricity health status identification model.

[0020] Furthermore, the target domain feature matrix acquisition module is specifically used to perform the following steps: Step 1021: Obtain the load curve data, power quality index data and equipment operation logs of users in the target domain in real time through the electricity information acquisition system to obtain the raw electricity consumption data of the target domain; Step 1022: Clean the original target domain electricity consumption data, fill in the missing time periods using linear interpolation, and smooth noise points using sliding window filtering to obtain the cleaned data; Step 1023: Perform the same normalization operation as in step 1012 on the cleaned data, using the statistical values ​​of the source domain data as the normalization parameters, to generate a standardized target domain feature matrix.

[0021] Furthermore, the dynamic evaluation model acquisition module is specifically used to perform the following steps: Step 1031: Load the source domain power health status identification model trained in step 1013, freeze the first and second layer network parameters of the model, and retain its underlying general feature representation capability; Step 1032: Construct a domain adaptive network, which consists of three parts: a feature extractor, a label predictor, and a domain discriminator. The feature extractor uses unfrozen model layers, the label predictor is used to predict the power health status, and the domain discriminator is used to determine whether the feature source is the source domain or the target domain. Step 1033: Input the source domain feature matrix and the target domain feature matrix into the domain adaptive network for adversarial training; Training is performed alternately in two phases: in the first phase, the neighborhood discriminator is fixed, and the feature extractor and label predictor are updated to minimize the healthy state classification loss; in the second phase, the feature extractor and label predictor are fixed, and the neighborhood discriminator is updated to maximize the source discrimination accuracy. The process is iterated repeatedly until the neighborhood discriminator can no longer distinguish the source of features. At this point, the neighborhood discriminator is removed, and the feature extractor and label predictor are combined into an adapted dynamic assessment model of the target domain's electricity health status.

[0022] Furthermore, the early warning module is specifically used to perform the following steps: Step 1041: Input the target domain feature matrix, which has been collected and preprocessed in real time, into the target domain power health status dynamic assessment model line by line; Step 1042: The target domain power health status dynamic assessment model performs forward calculation on the input data and outputs a health score value; according to the preset threshold range, the health score value is mapped to the corresponding power health status level, and the assessment time and results are recorded; Step 1043: When the health score value is lower than the health level threshold, trigger the abnormal attribution analysis; use the gradient-based feature importance calculation method to calculate the partial derivative of the output of the target domain electricity health status dynamic assessment model with respect to each input feature, and identify the features with negative partial derivatives and large absolute values ​​as the key abnormal causes of the score decline. Step 1044: Based on the health score, electricity health status level, assessment time and results, and key abnormal causes, generate a risk warning report and push it through the enterprise business system.

[0023] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention effectively addresses the problem that traditional models cannot accurately assess target domain users due to data sparsity, significant industry differences, and missing labels by constructing and transferring health-related electricity consumption knowledge from high-value electricity consumption data in the source domain. This method improves the generalization ability and accuracy of the assessment model in cross-industry, cross-regional, and small-sample scenarios.

[0024] Meanwhile, this invention combines real-time data stream processing technology to achieve minute-level dynamic scoring and anomaly attribution of enterprise power health status, enhancing the real-time nature and interpretability of power status assessment.

[0025] Ultimately, this invention provides power companies' customer service centers with proactive and precise customer service capabilities, helping to identify potential electricity hazards in advance, reduce the risk of equipment failure, and improve user satisfaction with electricity usage. Attached Figure Description

[0026] Figure 1 This is a schematic diagram of the enterprise-level dynamic assessment method for power health status based on transfer learning, provided in an embodiment of the present invention. Detailed Implementation

[0027] 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0028] It should be noted that the acquisition of data and collection of information in this application are legal, compliant, or obtained with the consent of the subject of the data collection.

[0029] The following are definitions of relevant concepts in this application: Source domain: refers to the high-value data sources used to train the initial model. These data are characterized by a large sample size, stable electricity consumption patterns, and clear and complete labels on electricity health status. In this invention, the source domain specifically includes the electricity load curves, power quality indicators, and equipment operation logs of key industrial park users and large manufacturing enterprise users already accumulated by the State Grid.

[0030] Target domain: refers to the user group that needs to be evaluated but has limited data. The data distribution of these users differs from that of the source domain. In this invention, the target domain specifically includes newly registered enterprises, users who have migrated from other regions, or small and micro-sized industrial and commercial users with incomplete data records.

[0031] Electricity Health Status: A comprehensive assessment indicator used to quantitatively characterize a company's compliance with electricity usage regulations, energy efficiency, equipment stability, and the magnitude of abnormal risks within a specific time period. This status level includes at least four levels: healthy, sub-healthy, alert, and severe.

[0032] Transfer learning module: A trainable component based on deep learning technology, its role is to learn how to adapt knowledge from the source domain to the target domain. This module effectively reduces the difference between the data distribution of the source domain and the target domain through techniques such as domain adaptation, feature decoupling, and adversarial training, enabling the knowledge learned in the source domain to be reused in the target domain.

[0033] Electricity consumption characteristic data: refers to multi-dimensional time-series data that has been processed and can be input into the model. This data includes at least load curves, such as active power values ​​collected every 15 minutes; power quality indicators, such as voltage deviation percentage and harmonic distortion rate percentage; and equipment start-up and shutdown event characteristics extracted from equipment operation logs.

[0034] Knowledge of healthy electricity use: refers to the feature representations and decision boundaries extracted from the source domain data by the transfer learning module that are strongly correlated with the judgment of electricity health status and are not affected by industry or regional differences.

[0035] like Figure 1 As shown, this application provides an enterprise-level dynamic assessment method for power health status based on transfer learning, including the following steps: Step 101: Construct and train a source domain power health status recognition model This step uses high-value labeled data from the source domain to train an initial recognition model, serving as the basis for knowledge transfer.

[0036] Step 1011: Obtain the electricity consumption characteristic data and corresponding electricity health status labels of source domain users from the State Grid Data Center to form the source domain training dataset.

[0037] The source domain users refer to key industrial parks and large manufacturing enterprises. Their electricity consumption characteristic data includes active power, reactive power, three-phase voltage imbalance, total harmonic distortion rate, and daily equipment start-up and shutdown frequency, collected every fifteen minutes. The electricity consumption health status labels are assigned by professionals after on-site verification, categorized into four levels: healthy, sub-healthy, attentive, or severe. The source domain training dataset is a collection of the above feature data and labels, used for subsequent model training.

[0038] In this embodiment, historical electricity consumption data of source domain users are first exported in batches through the power big data platform interface. Next, the exported data undergoes format validation to ensure that each record contains complete feature fields and timestamps. Then, the feature data is associated and aligned with corresponding tags according to user identifier and time. Finally, the associated data is integrated into a structured dataset, which is output as the source domain training dataset.

[0039] In practical applications, a large equipment manufacturing park with complete electricity usage records is used as the source domain. This park houses 100 large enterprises, and electricity usage data has been continuously collected for over three years. Professional personnel have verified the data on each enterprise's health status monthly. Electricity usage data from the past 12 months for all 100 enterprises is exported from this park, with one record every 15 minutes. This data includes five features: active power (kW), reactive power (kVAR), three-phase voltage imbalance (%), total harmonic distortion (%), and daily equipment start-up and shutdown frequency (times), along with corresponding labels: healthy, sub-healthy, alert, or severe. This data constitutes the source domain training dataset, totaling approximately 3.5 million records.

[0040] Step 1012: Standardize the source domain training dataset, including missing value imputation, outlier removal, and data normalization, to generate a standardized source domain feature matrix.

[0041] Missing value imputation refers to filling in missing feature values ​​at a given time point by averaging the values ​​at the two times before and after that point. Outlier removal involves deleting records with values ​​exceeding the normal physical range, such as negative power, and recalculating the imputed value using linear interpolation. Data normalization uses min-max normalization to compress the value range of each feature to between 0 and 1. The normalization formula is: in, These are the original eigenvalues. This is the minimum value of the feature on the training dataset. This is the maximum value of the feature on the training dataset. These are the normalized feature values. The standardized source domain feature matrix refers to the two-dimensional array after the above processing, where the number of rows represents the number of samples and the number of columns represents the feature dimension.

[0042] In this embodiment, firstly, each feature column of the source domain training dataset is traversed, and the minimum and maximum values ​​of each feature are calculated. Next, missing values ​​in each feature column are processed: the missing position is located, and the average of the two valid values ​​before and after that position is used to fill the missing value. Then, outliers in each feature column are detected: if the value is less than 0 or greater than a preset physical upper limit (e.g., an active power limit of 10000kW), it is marked as an outlier, the record is deleted, and the filling value is recalculated using linear interpolation based on the valid records before and after it. After completing the missing and outlier processing, normalization is performed on each feature column by subtracting the minimum value of that column from each original value and dividing by the difference between the maximum and minimum values ​​of that column. Finally, all normalized feature values ​​are organized into a matrix, with each row corresponding to a sample at a given time point and each column corresponding to a feature dimension.

[0043] In practical applications, the source domain training dataset obtained in step 1011 is processed. First, the minimum active power value is calculated to be 0 kW, and the maximum value is 500 kW; the minimum reactive power value is 0 kVAR, and the maximum value is 300 kVAR; the minimum three-phase voltage imbalance value is 0%, and the maximum value is 5%; the minimum total harmonic distortion rate is 0%, and the maximum value is 10%; the minimum number of equipment start-ups and shutdowns is 0, and the maximum number is 50. For a missing active power value, the values ​​of 120 kW and 130 kW at the two moments before and after it are taken, and the average value of 125 kW is calculated for filling. For an outlier record with an active power value of -5 kW, the record is deleted, and the values ​​at the moments before and after it are interpolated and recalculated. Then, normalization is performed on each feature. For example, if the active power at a certain moment is 100 kW, the normalization value is (100-0) / (500-0) = 0.2. After all samples are processed, a standardized source domain feature matrix with 1200 rows and 5 columns is generated, where 1200 corresponds to 100 companies multiplied by 12 months.

[0044] Step 1013: Construct a deep neural network classification model. Use the standardized source domain feature matrix as input and the electricity health status label as a supervision signal for training. The training objective is to minimize the classification error and obtain the trained source domain electricity health status identification model.

[0045] In this context, a deep neural network classification model refers to a feedforward neural network containing multiple fully connected hidden layers. The number of nodes in its input layer equals the feature dimension, and the number of nodes in its output layer equals the number of health state categories. The classification error is measured using the cross-entropy loss function, with the following formula: in, The total number of health status categories. The one-hot encoding of the real label in the first... Values ​​on the class For the model to predict belonging to the first The probability of a class The loss value is denoted as . A successfully trained source domain electricity health status identification model refers to a model whose classification performance converges on the validation set after iterative parameter updates via backpropagation algorithm and optimizer.

[0046] In this embodiment, a deep neural network with three fully connected hidden layers is first constructed, with 128, 64, and 32 neurons in each hidden layer, respectively, and ReLU activation function is used. The output layer uses the Softmax activation function to output the probability distribution of four health state categories. The standardized source domain feature matrix is ​​used as the input feature, and the corresponding labels are converted into one-hot encodings as supervision signals. The network parameters are initialized, and the Adam optimizer with a learning rate of 0.001 is set. In each training round, the input data is fed into the network in batches, the output of the forward propagation is calculated, and then the error between the output and the true label is calculated using the cross-entropy loss function.

[0047] Next, the gradient of the loss function with respect to the parameters of each layer is calculated using the backpropagation algorithm, and the network parameters are updated using the Adam optimizer. After each round of training on all data, the classification accuracy of the model is evaluated on an independent validation set. Training is stopped when the validation set accuracy no longer improves for 10 consecutive rounds. The network parameters at this point are saved, resulting in the trained source domain electricity health status recognition model.

[0048] In practical applications, a neural network is constructed with an input layer of 5 nodes, hidden layers of 128, 64, and 32 nodes respectively, and an output layer of 4 nodes. The 1200-row, 5-column standardized feature matrix generated in step 1012 is used as input, and the corresponding 1200 labels are converted to one-hot encodings. The Adam optimizer is used with a learning rate of 0.001 and a batch size of 32. The training process lasts for 200 epochs. The validation set accuracy reaches 92% at the 180th epoch, and training stops thereafter. The model parameters are saved, resulting in a power health status identification model suitable for large equipment manufacturing parks.

[0049] Step 102: Acquire and preprocess target domain user data This step collects real-time electricity consumption data from users in the target domain and performs the same preprocessing operations as the source domain data to ensure data compatibility.

[0050] Step 1021: Obtain the load curve data, power quality index data and equipment operation logs of users in the target domain in real time through the electricity information acquisition system to obtain the raw electricity consumption data of the target domain.

[0051] The electricity consumption information collection system refers to the system deployed by the State Grid Corporation of China for automatically collecting data from users' smart meters. Load curve data includes active and reactive power every fifteen minutes. Power quality index data includes voltage deviation percentage and harmonic distortion rate percentage. Equipment operation logs refer to equipment start-up and shutdown event records obtained from the enterprise's smart gateway, used to extract the daily number of equipment start-ups and shutdowns. Raw target domain electricity consumption data refers to the collection of the above data without any cleaning or processing.

[0052] In this embodiment, the load curves and power quality index data of users in the target domain are first retrieved through the application programming interface of the electricity information collection system according to a preset time granularity. Simultaneously, equipment operation logs are collected through a smart gateway deployed on the user side. Next, the load curve data, power quality index data, and equipment operation logs are aligned using the same timestamp to form a unified time-series data table. Finally, the raw data, without any cleaning, is saved as the original target domain electricity consumption dataset.

[0053] In practical application, a medium-sized software service company that has recently moved into a high-tech park was selected as the target domain user. Through the electricity consumption information collection system interface, load curve data for the past two months was retrieved at 15-minute intervals, totaling approximately 5760 records. Simultaneously, equipment operation logs from the server room were collected from the smart gateway to count the daily number of equipment start-ups and shutdowns. These data were then aligned by timestamps to obtain raw target domain electricity consumption data containing five fields: active power, reactive power, voltage deviation, harmonic distortion rate, and number of start-ups and shutdowns.

[0054] Step 1022: Clean the original target domain electricity consumption data, fill in the missing time periods using linear interpolation, and smooth noise points using sliding window filtering to obtain the cleaned data.

[0055] Linear interpolation, in particular, refers to a method that, for a missing time point, uses the values ​​of two known data points before and after it to calculate an estimated value according to the time ratio. Let the missing time point be... The time of the previous known point is Value The time of the next known point is Value Then the estimated value The calculation formula is: The sliding window filtering method refers to taking a window centered on a given data point and with a width of 1 / 2π. The arithmetic mean of all points within the window is used as the smoothing value for that point, using the following formula: in, For the original number The value of each point, The smoothed value. The width is half the window width. Cleaned data refers to data sequences that have had missing data and abnormal fluctuations eliminated.

[0056] In this embodiment, the process first iterates through each feature column of the original target domain electricity consumption data, detecting discontinuous timestamps and marking missing time periods. For each missing time period, the two nearest valid data points before and after the missing point are taken, and the estimated value of each missing point is calculated and filled using a linear interpolation formula. Next, noise points in each feature column are detected: the difference between each point and its preceding and following points is calculated, and if the difference exceeds a preset threshold (e.g., 3 times the standard deviation), it is marked as a noise point. For each noise point, a sliding window with a width of 5 is taken centered on it, and the arithmetic mean of all points within the window is calculated. This average value is then used to replace the original value. Finally, the cleaned data after missing data filling and noise smoothing is output.

[0057] In practical application, the original electricity consumption data of the software service company was cleaned. It was found that the active power record at 14:00 on a certain day was missing. The value at 13:45 was 120kW, and the value at 14:15 was 130kW. The time difference was 30 minutes, and the time difference between the missing point and 13:45 was 15 minutes. Therefore, the estimated value was 120 + (130 - 120) / 30 × 15 = 125kW. For a certain moment where the harmonic distortion rate was significantly higher than the noise points before and after, the original value was 8.5%, and the values ​​of the two points before and after were 3.2%, 3.5%, 3.6%, and 3.4% respectively. The sliding window average was (3.2 + 3.5 + 8.5 + 3.6 + 3.4) / 5 = 4.44%, which was used to replace the original value. After cleaning, smooth electricity consumption data without missing values ​​was obtained.

[0058] Step 1023: Perform the same normalization operation as in step 1012 on the cleaned data, using the statistical values ​​of the source domain data as the normalization parameters, to generate a standardized target domain feature matrix.

[0059] The normalization parameters refer to the minimum and maximum values ​​of each feature calculated from the source domain data in step 1012. Using these parameters ensures that the target domain data and the source domain data are on the same numerical scale, avoiding the impact of data distribution differences on transfer learning performance. The standardized target domain feature matrix is ​​a two-dimensional array obtained by applying the normalization formula to the cleaned target domain data, with the same number of columns as the source domain feature matrix.

[0060] In this embodiment, the minimum and maximum values ​​of each feature in the source domain are first read from the record in step 1012. Then, for each feature column of the cleaned target domain data, a normalization formula is applied element-by-element: Next, the normalized feature values ​​are organized into a matrix in their original order. Finally, the standardized target domain feature matrix is ​​output for subsequent transfer learning.

[0061] In practical applications, the minimum active power of the source domain is obtained from step 1012, which is 0 kW and the maximum is 500 kW. At a certain moment, the active power of the enterprise is 100 kW, which, after normalization, is (100-0) / (500-0) = 0.2. Similarly, the minimum reactive power is 0, and the maximum is 300; at a certain moment, the reactive power of 60 kVAR is normalized to 0.2. The minimum three-phase voltage imbalance is 0, and the maximum is 5; at a certain moment, 1.5% is normalized to 0.3. The minimum total harmonic distortion is 0, and the maximum is 10; at a certain moment, 3% is normalized to 0.3. The minimum number of start-stop cycles is 0, and the maximum is 50; at a certain moment, 10 cycles are normalized to 0.2. After normalizing the data at all time points, a standard target domain feature matrix of 5760 rows and 5 columns is generated.

[0062] Step 103: Perform transfer learning to adapt knowledge from the source domain to the target domain. This step utilizes transfer learning technology to transfer knowledge about healthy electricity use from the source domain identification model to the target domain, thereby constructing a target domain evaluation model.

[0063] Step 1031: Load the source domain power health status identification model trained in step 1013, freeze the first and second layer network parameters of the model, and retain its underlying general feature representation capability.

[0064] Freezing network parameters means setting these parameters to be non-updatable during subsequent training; their gradients will not be calculated or modified by the optimizer. The first and second layers are located near the input in a deep neural network. These layers learn relatively low-level and general feature representations, such as the shape of load waveforms and basic harmonic patterns, which are applicable across different industries.

[0065] In this embodiment, the source domain model file saved in step 1013 is first read using the model loading function of the deep learning framework to restore the model's structure and parameters. Next, all weight and bias parameters of the first and second layers of the model are traversed, and their trainable attributes are set to false. Then, it is verified that these parameters are indeed frozen, meaning they will not participate in backpropagation in subsequent gradient calculations. Finally, a model with frozen low-level parameters but still trainable high-level parameters is output for subsequent domain adaptation.

[0066] In practical applications, the source domain model saved in step 1013 is loaded using the TensorFlow framework. The first layer of the model is a fully connected layer with 128 neurons, and the second layer is a fully connected layer with 64 neurons. The `trainable` property of these two layers is set to `False`. The third layer with 32 neurons and the output layer retain `trainable` set to `True`. After freezing, the parameters of the first two layers of the model are no longer updated, preserving the low-level feature extraction capabilities learned from data from large equipment manufacturing enterprises.

[0067] Step 1032: Construct a domain-adaptive network, which consists of three parts: a feature extractor, a label predictor, and a domain discriminator. The feature extractor uses unfrozen model layers, the label predictor is used to predict the electricity health status, and the domain discriminator is used to determine whether the feature source is the source domain or the target domain.

[0068] The feature extractor is the neural network component that extracts high-dimensional feature representations from the input data, and it consists of the model layers that were not frozen in step 1031. The label predictor is a fully connected layer with a Softmax activation function, which takes the output of the feature extractor as input and outputs the probability distribution of four health states. The neighborhood discriminator is a fully connected layer with a Sigmoid activation function, which takes the output of the feature extractor as input and outputs a value between 0 and 1, representing the probability that the feature originates from the source domain.

[0069] In this embodiment, the unfrozen layers are first extracted from the model in step 1031 and concatenated to form a feature extractor. Next, a fully connected layer with 4 output nodes and a Softmax activation function is connected to the output of the feature extractor, serving as a label predictor. Simultaneously, another fully connected layer with 1 output node and a Sigmoid activation function is connected in parallel to the output of the feature extractor, serving as a neighborhood discriminator. These three parts are then combined into a complete neighborhood adaptive network, with the feature extractor being a shared component. Finally, the network is compiled, specifying a cross-entropy loss function for the label predictor and a binary cross-entropy loss function for the neighborhood discriminator.

[0070] In practical applications, the 32 neurons of the third layer and subsequent layers are extracted from the source domain model as a feature extractor. The feature extractor outputs a 32-dimensional feature vector. The label predictor is a fully connected layer with 32 inputs and 4 outputs, activated using Softmax. The neighborhood discriminator is a fully connected layer with 32 inputs and 1 output, activated using Sigmoid. Both are connected in parallel after the feature extractor to form a neighborhood adaptive network.

[0071] Step 1033: Input the source domain feature matrix and the target domain feature matrix into the domain adaptive network for adversarial training. The training is performed alternately in two phases: In the first phase, the domain discriminator is fixed, and the feature extractor and label predictor are updated to minimize the health status classification loss; in the second phase, the feature extractor and label predictor are fixed, and the domain discriminator is updated to maximize the source discrimination accuracy. This process is repeated until the domain discriminator can no longer distinguish the feature source. At this point, the domain discriminator is removed, and the feature extractor and label predictor are combined to form the adapted target domain power health status dynamic assessment model.

[0072] The health status classification loss uses the cross-entropy loss function. ,in One-hot encoding of the true labels of the source domain samples. This represents the probability output by the label predictor. The neighborhood discrimination loss uses the binary classification cross-entropy loss function. ,in The domain label (source domain is 1, target domain is 0). This represents the probability output by the domain discriminator. The goal of adversarial training is to minimize the feature representation learned by the feature extractor. Simultaneously maximize The domain discriminator attempts to minimize When the discrimination accuracy of the domain discriminator approaches 50%, it indicates that the feature extractor has successfully obfuscated the domain information and achieved knowledge transfer.

[0073] In this embodiment, the source domain feature matrix and the target domain feature matrix are first merged, and a domain label (1 for the source domain and 0 for the target domain) is added to each sample. Then, alternating training rounds are set. In each training round, the first stage is performed: the parameters of the domain discriminator are fixed, the merged data is input into the network, and only the label predictor's prediction for the source domain samples is calculated. Backpropagation updates the parameters of the feature extractor and label predictor. Then, the second stage is executed: the parameters of the feature extractor and label predictor are fixed, the merged data is input into the network, and the neighborhood discriminator is computed. The parameters of the domain discriminator are updated via backpropagation. These two phases are repeated, and the accuracy of the domain discriminator on the validation set is calculated after each round. Training stops when the accuracy drops to between 50% and 55% and remains stable. Finally, the domain discriminator is removed from the network, and the feature extractor and label predictor are concatenated to form an adapted dynamic assessment model of the target domain's electricity health status.

[0074] In practical applications, 200 training samples from 1200 samples in the source domain and 5760 samples in the target domain are mixed. The source domain samples are labeled with domain label 1, and the target domain samples with domain label 0. The learning rate is set to 0.001, the batch size to 32, and training is performed for 50 epochs. In each epoch, the first stage updates the feature extractor and label predictor to increase the accuracy of the source domain health status classification; the second stage updates the domain discriminator to enable it to distinguish the source of features. Initially, the domain discriminator accuracy is 90%, which gradually decreases as training progresses. After 40 epochs, the domain discriminator accuracy stabilizes at 52%. At this point, training is stopped, the domain discriminator is removed, and a power consumption health status assessment model suitable for this software service company is obtained.

[0075] Step 104: Real-time dynamic assessment and early warning This step uses the adapted model to perform minute-level dynamic scoring of users in the target domain and triggers anomaly attribution and risk warnings.

[0076] Step 1041: Input the real-time collected and preprocessed target domain user electricity consumption characteristic data into the target domain electricity health status dynamic assessment model one by one.

[0077] Real-time acquisition refers to obtaining the latest electricity consumption data every fifteen minutes through the electricity consumption information collection system. Preprocessing includes cleaning in step 1022 and normalization in step 1023. Inputting data one at a time means feeding the feature vector of a single time point into the model each time, rather than batch input.

[0078] In this embodiment, a timer is first set to trigger a data acquisition task every fifteen minutes. The acquisition task executes steps 1021 to 1023 to obtain the standardized feature vector at the current time point. Then, the feature vector is converted into a tensor format required for model input. Finally, the model's prediction method is called, using this tensor as input to initiate forward computation.

[0079] In practical applications, the software service provider is set to collect data every 15 minutes. At 14:00 on a certain day, the following data is collected: active power 100kW, reactive power 60kVAR, three-phase voltage imbalance 1.5%, total harmonic distortion 3%, and number of start-stop cycles 10. After normalization, the feature vector [0.2, 0.2, 0.3, 0.3, 0.2] is obtained and input into the target domain model obtained in step 1033.

[0080] Step 1042: The model performs forward calculations on the input data and outputs a health score. Based on a preset threshold range, the score is mapped to the corresponding electricity health status level, and the assessment time and results are recorded.

[0081] The health score is a weighted score obtained by applying softmax activation to the model's output layer, averaging the scores across different health status categories. Alternatively, the probability value of the corresponding health category in the model's output layer can be used directly, ranging from 0 to 1. The preset threshold ranges are: 0.85 to 1.00 for a healthy level, 0.70 to 0.85 for a sub-healthy level, 0.50 to 0.70 for a attentive level, and 0 to 0.50 for a severe level. Recording the assessment time and results means storing the assessment timestamp, health score, and health status level in the database.

[0082] In this embodiment, the input feature vector is first fed into the model, and forward propagation is performed to obtain the probability distribution of four health states from the output layer. Then, a health score is calculated, which can be the probability value of the health category or a weighted sum. Next, the health score is compared with four threshold intervals to determine the corresponding level. Finally, a record is generated, containing the current timestamp, health score, and health state level, and written to the evaluation log database.

[0083] In practical applications, the input feature vector [0.2, 0.2, 0.3, 0.3, 0.2] is processed by the model's forward pass, resulting in a health category probability of 0.92, a sub-healthy probability of 0.06, a attention probability of 0.02, a severe probability of 0.00, and a health score of 0.92. Since 0.92 falls between 0.85 and 1.00, it is classified as a healthy level. System record: April 3, 2025, 14:00, score 0.92, level healthy. At another moment, the normalized value of the harmonic distortion rate in the input feature vector is 0.85, while other features remain unchanged. The model outputs a health score of 0.62, falling within the 0.50-0.70 range, and is classified as an attention level.

[0084] Step 1043: When the health score falls below the health level threshold, trigger anomaly attribution analysis. A gradient-based feature importance calculation method is used to calculate the partial derivative of the model output with respect to each input feature. Features with negative partial derivatives and large absolute values ​​are identified as key anomalies causing the score decline.

[0085] The health level threshold refers to the lower limit of the health level, which is 0.85. The gradient-based feature importance calculation method refers to calculating the score output by the model. For each input feature partial derivatives The absolute value of the partial derivative indicates the degree of influence of that feature on the output score. A positive partial derivative means that increasing the feature value will increase the score, while a negative partial derivative means that increasing the feature value will decrease the score. Key anomalies refer to the features with the largest absolute values ​​of negative partial derivatives.

[0086] In this embodiment, the health score output in step 1042 is first determined to be less than 0.85. If it is less, the attribution analysis process is triggered. Using automatic differentiation, the input feature vector is used as the variable for which gradients need to be calculated. The model's forward propagation is performed once to obtain the health score. Then, the backpropagation method is called to calculate the gradient of the score with respect to each input feature, i.e., the partial derivative. Next, for each feature, the absolute value of its negative gradient is taken and sorted in descending order. The top three features are selected as candidate causes of anomalies. Finally, the names of these features and their negative gradient values ​​are output as the anomaly attribution results.

[0087] In practical application, a health score of 0.62 at 14:15 on a certain day, which is less than 0.85, triggers attribution analysis. The input feature vector is [0.2, 0.2, 0.15, 0.85, 0.1]. The partial derivatives of each feature are calculated: active power partial derivative +0.05, reactive power partial derivative +0.02, three-phase voltage imbalance partial derivative -0.08, total harmonic distortion (THD) partial derivative -0.42, and start / stop frequency partial derivative -0.01. The absolute value of the negative gradient of THD is the largest at 0.42, followed by the absolute value of the negative gradient of three-phase voltage imbalance at 0.08. Therefore, the key anomalies are determined to be high THD and three-phase voltage imbalance.

[0088] Step 1044: Generate a risk warning report based on the health score, health status level, and abnormal attribution results, and push it to account managers and enterprise users through the enterprise business system.

[0089] The risk warning report is a structured information document that includes at least the assessment time, health score, health status level, and a description of the cause of the anomaly. The enterprise business system refers to the work order processing system and message push platform used by the State Grid Customer Service Center. Push methods include generating electronic work orders and sending them to account managers, as well as sending them to the energy consumption managers of enterprise users via SMS or mobile application messages.

[0090] In this embodiment, the assessment time, health score, health status level obtained in step 1042, and the description of the cause of the anomaly obtained in step 1043 are first combined into a report text according to a preset template. The report template is: "Assessment Time: {Time}; Health Score: {Score}; Health Status: {Level}; Anomaly Attribution: {Cause Description}; Recommendation: {Recommendation Based on the Cause}". Then, the application programming interface of the enterprise business system is called to submit the report in the form of a work order, specifying the recipient as the account manager responsible for the target domain users. At the same time, the SMS gateway interface and the application push interface are called to send the report summary to the enterprise user's mobile phone number and application account.

[0091] In practical application, an early warning report was generated based on the assessment results at 14:15: "Assessment time: April 3, 2025, 14:15; Health score: 0.62; Health status: Caution; Abnormal attribution: Total harmonic distortion rate is high at 0.85, three-phase voltage imbalance is 0.15. Recommendation: Inspect the nonlinear load equipment and power distribution lines within the enterprise." This report generates an electronic work order through the State Grid customer service platform and is automatically pushed to account manager Zhang San. Simultaneously, it is sent to the mobile phone of enterprise manager Li Si via SMS and pushed to his account via mobile application. After the push is completed, the assessment and early warning process ends.

[0092] This application successfully transferred the power health assessment capabilities of a large manufacturing enterprise in the source domain to a data-scarce software service enterprise in the target domain, achieving accurate and real-time assessment and early warning of the enterprise's power status. Based on the same inventive concept, this invention also provides an enterprise-level dynamic power health status assessment system based on transfer learning. This system includes a data acquisition module, a data preprocessing module, a model training module, a transfer learning module, a dynamic assessment module, an anomaly attribution module, and an early warning output module. These modules cooperate to implement the methods described in this invention.

[0093] In addition, the present invention also provides an enterprise-level dynamic assessment system for power consumption health status based on transfer learning, comprising the following modules: The model building and training module is used to build and train a source domain power health status recognition model; The target domain feature matrix acquisition module is used to acquire and preprocess target domain user data to obtain a standardized target domain feature matrix. The dynamic evaluation model acquisition module is used to perform transfer learning, adapting source domain knowledge to the target domain to obtain the adapted dynamic evaluation model of the target domain's electricity health status. The early warning module is used for real-time dynamic assessment and early warning.

[0094] Finally, it should be noted that the above embodiments are merely illustrative and explanatory of the present invention, and are not intended to limit the present invention to the scope of the described embodiments. Furthermore, those skilled in the art will understand that the present invention is not limited to the above embodiments, and many more variations and modifications can be made based on the teachings of the present invention, all of which fall within the scope of protection claimed by the present invention.

Claims

1. A dynamic assessment method for enterprise-level electricity health status based on transfer learning, characterized in that, Includes the following steps: Step 101: Construct and train a source domain power health status recognition model; Step 102: Acquire and preprocess the target domain user data to obtain a standardized target domain feature matrix; Step 103: Perform transfer learning to adapt the source domain knowledge to the target domain, and obtain the adapted dynamic assessment model of the target domain's electricity health status. Step 104: Real-time dynamic assessment and early warning.

2. The enterprise-level dynamic assessment method for electricity health status based on transfer learning according to claim 1, characterized in that, Step 101 includes the following steps: Step 1011: Obtain the power consumption characteristic data and corresponding power health status labels of source domain users from the enterprise data center to form the source domain training dataset; Step 1012: Standardize the source domain training dataset, including missing value imputation, outlier removal, and data normalization, to generate a standardized source domain feature matrix; Step 1013: Construct a deep neural network classification model. Use the standardized source domain feature matrix as input and the electricity health status label as a supervision signal for training. The training objective is to minimize the classification error and obtain the trained source domain electricity health status identification model.

3. The enterprise-level dynamic assessment method for power health status based on transfer learning according to claim 1, characterized in that, Step 102 includes the following steps: Step 1021: Obtain the load curve data, power quality index data and equipment operation logs of users in the target domain in real time through the electricity information acquisition system to obtain the raw electricity consumption data of the target domain; Step 1022: Clean the original target domain electricity consumption data, fill in the missing time periods using linear interpolation, and smooth noise points using sliding window filtering to obtain the cleaned data; Step 1023: Perform the same normalization operation as in step 1012 on the cleaned data, using the statistical values ​​of the source domain data as the normalization parameters, to generate a standardized target domain feature matrix.

4. The enterprise-level dynamic assessment method for electricity health status based on transfer learning according to claim 1, characterized in that, Step 103 includes the following steps: Step 1031: Load the source domain power health status identification model trained in step 1013, freeze the first and second layer network parameters of the model, and retain its underlying general feature representation capability; Step 1032: Construct a domain adaptive network, which consists of three parts: a feature extractor, a label predictor, and a domain discriminator. The feature extractor uses unfrozen model layers, the label predictor is used to predict the power health status, and the domain discriminator is used to determine whether the feature source is the source domain or the target domain. Step 1033: Input the source domain feature matrix and the target domain feature matrix into the domain adaptive network for adversarial training; Training is performed alternately in two phases: in the first phase, the neighborhood discriminator is fixed, and the feature extractor and label predictor are updated to minimize the healthy state classification loss; in the second phase, the feature extractor and label predictor are fixed, and the neighborhood discriminator is updated to maximize the source discrimination accuracy. The process is iterated repeatedly until the neighborhood discriminator can no longer distinguish the source of features. At this point, the neighborhood discriminator is removed, and the feature extractor and label predictor are combined into an adapted dynamic assessment model of the target domain's electricity health status.

5. The enterprise-level dynamic assessment method for power health status based on transfer learning according to claim 1, characterized in that, Step 104 includes the following steps: Step 1041: Input the target domain feature matrix, which has been collected and preprocessed in real time, into the target domain power health status dynamic assessment model line by line; Step 1042: The target domain power health status dynamic assessment model performs forward calculation on the input data and outputs a health score value; according to the preset threshold range, the health score value is mapped to the corresponding power health status level, and the assessment time and results are recorded; Step 1043: When the health score value is lower than the health level threshold, trigger the abnormal attribution analysis; use the gradient-based feature importance calculation method to calculate the partial derivative of the output of the target domain electricity health status dynamic assessment model with respect to each input feature, and identify the features with negative partial derivatives and large absolute values ​​as the key abnormal causes of the score decline. Step 1044: Based on the health score, electricity health status level, assessment time and results, and key abnormal causes, generate a risk warning report and push it through the enterprise business system.

6. An enterprise-level dynamic assessment system for electricity health status based on transfer learning, characterized in that, Includes the following modules: The model building and training module is used to build and train a source domain power health status recognition model; The target domain feature matrix acquisition module is used to acquire and preprocess target domain user data to obtain a standardized target domain feature matrix. The dynamic evaluation model acquisition module is used to perform transfer learning, adapting source domain knowledge to the target domain to obtain the adapted dynamic evaluation model of the target domain's electricity health status. The early warning module is used for real-time dynamic assessment and early warning.

7. The enterprise-level dynamic assessment system for electricity health status based on transfer learning according to claim 6, characterized in that, The model building and training module is specifically used to perform the following steps: Step 1011: Obtain the power consumption characteristic data and corresponding power health status labels of source domain users from the enterprise data center to form the source domain training dataset; Step 1012: Standardize the source domain training dataset, including missing value imputation, outlier removal, and data normalization, to generate a standardized source domain feature matrix; Step 1013: Construct a deep neural network classification model. Use the standardized source domain feature matrix as input and the electricity health status label as a supervision signal for training. The training objective is to minimize the classification error and obtain the trained source domain electricity health status identification model.

8. The enterprise-level dynamic assessment system for electricity health status based on transfer learning according to claim 6, characterized in that, The target domain feature matrix acquisition module is specifically used to perform the following steps: Step 1021: Obtain the load curve data, power quality index data and equipment operation logs of users in the target domain in real time through the electricity information acquisition system to obtain the raw electricity consumption data of the target domain; Step 1022: Clean the original target domain electricity consumption data, fill in the missing time periods using linear interpolation, and smooth noise points using sliding window filtering to obtain the cleaned data; Step 1023: Perform the same normalization operation as in step 1012 on the cleaned data, using the statistical values ​​of the source domain data as the normalization parameters, to generate a standardized target domain feature matrix.

9. The enterprise-level dynamic assessment system for electricity health status based on transfer learning according to claim 6, characterized in that, The dynamic evaluation model acquisition module is specifically used to perform the following steps: Step 1031: Load the source domain power health status identification model trained in step 1013, freeze the first and second layer network parameters of the model, and retain its underlying general feature representation capability; Step 1032: Construct a domain adaptive network, which consists of three parts: a feature extractor, a label predictor, and a domain discriminator. The feature extractor uses unfrozen model layers, the label predictor is used to predict the power health status, and the domain discriminator is used to determine whether the feature source is the source domain or the target domain. Step 1033: Input the source domain feature matrix and the target domain feature matrix into the domain adaptive network for adversarial training; Training is performed alternately in two phases: in the first phase, the neighborhood discriminator is fixed, and the feature extractor and label predictor are updated to minimize the healthy state classification loss; in the second phase, the feature extractor and label predictor are fixed, and the neighborhood discriminator is updated to maximize the source discrimination accuracy. The process is iterated repeatedly until the neighborhood discriminator can no longer distinguish the source of features. At this point, the neighborhood discriminator is removed, and the feature extractor and label predictor are combined into an adapted dynamic assessment model of the target domain's electricity health status.

10. The enterprise-level dynamic assessment system for electricity health status based on transfer learning according to claim 6, characterized in that, The early warning module is specifically used to perform the following steps: Step 1041: Input the target domain feature matrix, which has been collected and preprocessed in real time, into the target domain power health status dynamic assessment model line by line; Step 1042: The target domain power health status dynamic assessment model performs forward calculation on the input data and outputs a health score value; according to the preset threshold range, the health score value is mapped to the corresponding power health status level, and the assessment time and results are recorded; Step 1043: When the health score value is lower than the health level threshold, trigger the abnormal attribution analysis; use the gradient-based feature importance calculation method to calculate the partial derivative of the output of the target domain electricity health status dynamic assessment model with respect to each input feature, and identify the features with negative partial derivatives and large absolute values ​​as the key abnormal causes of the score decline. Step 1044: Based on the health score, electricity health status level, assessment time and results, and key abnormal causes, generate a risk warning report and push it through the enterprise business system.

Images