A method and system for monitoring electric vehicle charging

Abstract

This invention discloses an electric vehicle charging monitoring method and system. The method includes: a cloud server acquiring data to train an electric vehicle charging monitoring model until the model parameters converge or the maximum number of training iterations is reached, ultimately obtaining the model parameters of the electric vehicle charging monitoring model; the cloud server compressing the trained electric vehicle charging monitoring model to obtain a compressed electric vehicle charging monitoring model; the compressed electric vehicle charging monitoring model being transferred and deployed to multiple user edge devices in an unsupervised manner under a federated transfer learning framework; and multiple user edge devices using the transferred and deployed electric vehicle charging monitoring model to monitor the charging of their respective electric vehicles, performing filtering processing, and finally outputting the charging monitoring results. This invention deploys the model on low-computing-power edge devices, achieving accurate and low-latency detection, and enabling large-scale deployment while protecting user data privacy.

Description

Technical Field

[0001] This invention relates to the field of charging monitoring technology, and in particular to a method and system for monitoring electric vehicle charging. Background Technology

[0002] In recent years, with increased attention to carbon emissions, the development of battery energy storage technology, and policy support, the number of electric vehicles (EVs) has grown rapidly. It is projected that by 2027, the total number of EVs in society will reach 233 million, with a compound annual growth rate exceeding 20%. On the positive side, intelligent and connected EVs are developing into energy storage terminals for the energy internet, using a V2G (Vehicle-to-grid) model to help the grid with peak shaving and frequency regulation, promote the integration of renewable energy, and become an important way of comprehensive energy utilization. On the negative side, the highly dynamic energy demand brought about by large-scale EV charging will greatly affect the stability of the power grid (e.g., the EV charging process will introduce harmonics into the grid). As grid operators, if they can detect EV charging events in the grid quickly and economically and make corresponding load adjustments, they will be able to better support grid planning and operation.

[0003] Compared to existing charging stations, the number of home electric vehicle charging stations will be much larger in the future. However, due to issues such as user privacy protection, complex data collection, and differences in vehicle-to-everything (V2X) interfaces, it is difficult to directly obtain charging event information from home charging stations. Currently, for residences with smart meters, non-intrusive load monitoring (NILM) can identify or decompose the load of a single appliance from the user's total load curve. Therefore, this method can be applied to the charging detection of home electric vehicles. Because of the unique "constant current-constant voltage" charging mode of electric vehicles (with relatively small fluctuations in the charging load curve), directly identifying electric vehicle charging events from the user's load curve is economically feasible.

[0004] Existing non-intrusive electric vehicle charging monitoring methods can be mainly divided into two types: rule-based and learning-based.

[0005] (1) Rule-based methods: For example, a steady-state load curve matching method for detecting electric vehicle charging was proposed based on the analysis of the constant current-constant voltage mode of electric vehicles. However, the state pattern matching process in this method has high computational cost and requires a long period of data accumulation for closed-loop computation. At high sampling rates, the computation time is very long, resulting in a large detection time delay. For example, a model that performs threshold processing on the total load curve and filters out spikes / noise is used to detect and decompose the load curve of household electric vehicle charging. This model shows good performance at low sampling rates (1 / 60 Hz) and maintains good accuracy in cases with some high-power load appliances (such as air conditioners). However, the robustness of this model has not been fully verified in real-world power consumption data with more high-power appliances. At the same time, the model performs iterative filtering calculations on a daily basis, resulting in a high detection delay. In addition, none of the above methods consider scalability and user data privacy protection in actual large-scale deployment scenarios.

[0006] (2) Learning-based methods: Currently, there are electric vehicle charging detection methods based on deep neural networks (DNNs), such as autoencoder models to identify electric vehicle charging curves. This model consists of a one-dimensional convolutional layer and an autoencoder block, and optimizes the charging detection results by filtering the total power signal. For example, there are also deep generative models that use an embedded HMM (Hidden Markov Model) and apply neural networks to approximate the posterior distribution of the hidden units of the HMM model, thereby identifying the charging load of electric vehicles. The successful training of this generative model depends heavily on data preprocessing, and the cost of preprocessing is very high. In addition, neither of the above two methods considers factors such as model scalability, detection latency, and privacy protection in the model design.

[0007] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0008] The main objective of this invention is to provide an electric vehicle charging monitoring method and system, which aims to solve the problems of existing electric vehicle charging monitoring methods being unable to be deployed on a large scale, lacking user data privacy protection, and having inaccurate detection.

[0009] To achieve the above objectives, the present invention provides a method for monitoring electric vehicle charging, the method comprising the following steps:

[0010] The cloud server acquires aggregated household energy consumption data and individual electric vehicle charging energy consumption data, annotates the individual electric vehicle charging energy consumption data to obtain event tags, and trains the electric vehicle charging monitoring model based on the aggregated household energy consumption data and event tags until the model parameters converge or the maximum number of training loops is reached, and finally obtains the model parameters of the electric vehicle charging monitoring model.

[0011] The cloud server compresses the parameters of each layer in the trained electric vehicle charging monitoring model according to a given pruning and compression ratio to obtain the compressed electric vehicle charging monitoring model.

[0012] Multiple user edge terminals receive compressed electric vehicle charging monitoring models sent by the cloud server, and migrate and deploy the electric vehicle charging monitoring models to their respective user edge terminals in an unsupervised manner under the federated transfer learning framework.

[0013] Multiple user edge devices use the migrated and deployed electric vehicle charging monitoring model to monitor the charging of their respective electric vehicles, and filter the model output results based on the output filter to finally output the charging monitoring results.

[0014] Optionally, in the electric vehicle charging monitoring method, the cloud server acquires aggregated household energy consumption data and individual electric vehicle charging energy consumption data, labels the individual electric vehicle charging energy consumption data to obtain event tags, and trains the electric vehicle charging monitoring model based on the aggregated household energy consumption data and event tags until the model parameters converge or the maximum number of training iterations is reached, ultimately obtaining the model parameters of the electric vehicle charging monitoring model, specifically including:

[0015] The cloud server acquires aggregated household energy consumption data and individual electric vehicle charging energy consumption data, and uses the Doane frequency classification method to label the individual electric vehicle charging energy consumption data to obtain event labels;

[0016] Aggregated household energy consumption data and event labels are input into a pre-designed electric vehicle charging monitoring model for supervised training.

[0017] That is: the aggregated energy consumption reading sequence is input into the electric vehicle charging monitoring model, the predicted electric vehicle charging status label sequence is output, the model loss is calculated by comparing it with the actual charging status label sequence, the gradient information is obtained to update the model parameters, until the model parameters converge or the maximum number of training loops is reached, and finally the model parameters of the electric vehicle charging monitoring model are obtained.

[0018] Optionally, in the electric vehicle charging monitoring method, the step of labeling individual electric vehicle charging energy consumption data using the Doane frequency classification method to obtain event labels specifically includes:

[0019] When labeling individual electric vehicle charging energy consumption data, the cloud server uses the Doane frequency classification method to extract statistical features of the electric vehicle charging load.

[0020] The |mean-interquartile range| value of the load curve is used as the standard threshold for labeling charging events. If the value of a single electric vehicle charging energy consumption data is greater than the standard threshold, it is marked as a charging event, and an event label is obtained.

[0021] Optionally, in the electric vehicle charging monitoring method, the actual charging status label sequence is a sequence obtained by labeling individual electric vehicle charging energy consumption data.

[0022] Optionally, in the electric vehicle charging monitoring method, the cloud server compresses the parameters of each layer in the trained electric vehicle charging monitoring model according to a given pruning and compression ratio to obtain a compressed electric vehicle charging monitoring model, specifically including:

[0023] The cloud server sorts the L1 norm values ​​of the parameter vectors in ascending order using the L1 norm for each layer of the trained electric vehicle charging monitoring model.

[0024] Given a pruning and compression ratio, the parameters of the models at the top of the list are set to 0 proportionally to compress the electric vehicle charging monitoring model, resulting in a compressed electric vehicle charging monitoring model.

[0025] Optionally, in the electric vehicle charging monitoring method, the pre-designed electric vehicle charging monitoring model is a deep neural network model, which includes 5 one-dimensional convolutional layers, 2 bidirectional LSTM layers, 2 fully connected layers, 2 ReLU functions and 1 Sigmoid function. Except for the last fully connected layer, each layer contains a normalization operation.

[0026] Optionally, in the electric vehicle charging monitoring method, the input of the electric vehicle charging monitoring model is an aggregated energy consumption reading sequence, and the output is an electric vehicle charging status label sequence of the corresponding input length.

[0027] The aggregated energy consumption reading sequence is aggregated household energy consumption data, and the electric vehicle charging status tag sequence is a predicted value based on the aggregated household energy consumption data.

[0028] Optionally, in the electric vehicle charging monitoring method, the multiple user edge terminals receive a compressed electric vehicle charging monitoring model sent by a cloud server, and migrate and deploy the electric vehicle charging monitoring model to their respective user edge terminals in an unsupervised manner under a federated transfer learning framework, specifically including:

[0029] Multiple user edge devices receive global initial status and some tagged data sent by the cloud server;

[0030] Multiple user edge devices use transfer learning to update the model parameters to adapt to the local data distribution, and then upload the updated model parameters to the cloud server.

[0031] Optionally, the electric vehicle charging monitoring method further includes, whereby the plurality of user edge terminals use transfer learning to update the parameters of the model to adapt to the local data distribution, and upload the updated model parameters to a cloud server, and then further includes:

[0032] The cloud server performs federated averaging aggregation on the received local models to obtain a new global model state. This updated global state (i.e., global model parameters) can then be migrated and deployed to newly added user edge terminals in an unsupervised manner for electric vehicle charging monitoring.

[0033] In addition, to achieve the above objectives, the present invention also provides an electric vehicle charging monitoring system, wherein the electric vehicle charging monitoring system includes: a cloud server and multiple user edge terminals, wherein the cloud server is communicatively connected to the multiple user edge terminals respectively;

[0034] The cloud server is used to acquire aggregated household energy consumption data and individual electric vehicle charging energy consumption data, annotate the individual electric vehicle charging energy consumption data to obtain event tags, and train the electric vehicle charging monitoring model based on the aggregated household energy consumption data and event tags until the model parameters converge or the maximum number of training loops is reached and terminated, and finally obtain the model parameters of the electric vehicle charging monitoring model.

[0035] The cloud server is also used to compress the parameters of each layer in the trained electric vehicle charging monitoring model according to a given pruning compression ratio to obtain a compressed electric vehicle charging monitoring model.

[0036] Multiple user edge terminals are used to receive the compressed electric vehicle charging monitoring model sent by the cloud server. Under the federated transfer learning framework, the electric vehicle charging monitoring model is transferred and deployed to their respective user edge terminals in an unsupervised manner.

[0037] Multiple user edge devices use the migrated and deployed electric vehicle charging monitoring model to monitor the charging of their respective electric vehicles, and filter the model output results based on the output filter to finally output the charging monitoring results.

[0038] In this invention, a cloud server acquires aggregated household energy consumption data and individual electric vehicle (EV) charging energy consumption data. The individual EV charging energy consumption data is labeled to obtain event tags. Based on the aggregated household energy consumption data and event tags, an EV charging monitoring model is trained until the model parameters converge or the maximum number of training iterations is reached, ultimately obtaining the model parameters. The cloud server compresses the parameters of each layer in the trained EV charging monitoring model according to a given pruning compression ratio, resulting in a compressed EV charging monitoring model. Multiple user edge devices receive the compressed EV charging monitoring model sent by the cloud server and migrate and deploy the EV charging monitoring model to their respective user edge devices in an unsupervised manner under a federated transfer learning framework. The multiple user edge devices use the migrated and deployed EV charging monitoring model to monitor the charging of their respective EVs and filter the model output results based on an output filter, ultimately outputting the charging monitoring results. This invention can deploy the EV charging monitoring model on low-computing-power edge devices, achieving accurate and low-latency detection, and can be deployed on a large scale while protecting user data privacy. Attached Figure Description

[0039] Figure 1 This is a flowchart of a preferred embodiment of the electric vehicle charging monitoring method of the present invention;

[0040] Figure 2 This is a schematic diagram of the network structure of the electric vehicle charging monitoring model in a preferred embodiment of the electric vehicle charging monitoring method of the present invention;

[0041] Figure 3 This is a schematic diagram comparing layer normalization and batch normalization in a preferred embodiment of the electric vehicle charging monitoring method of the present invention.

[0042] Figure 4 This is a schematic diagram of the framework and process of a non-invasive EV charging detection method in a preferred embodiment of the electric vehicle charging monitoring method of the present invention.

[0043] Figure 5 This is a flowchart illustrating the algorithm process of the output filter filtering out (erroneous) predictions with short durations in a preferred embodiment of the electric vehicle charging monitoring method of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0045] To address the shortcomings of existing non-intrusive EV charging monitoring methods in terms of robustness, scalability, and privacy protection, this invention proposes a non-intrusive EV charging detection method (i.e., electric vehicle charging monitoring method) based on a federated learning framework. The model in this method can be deployed on edge devices with low computing power, achieving accurate and low-latency detection, and has good scalability, enabling large-scale deployment while protecting user data privacy.

[0046] The electric vehicle charging monitoring method described in the preferred embodiment of the present invention, such as... Figure 1 As shown, the electric vehicle charging monitoring method includes the following steps:

[0047] Step S10: The cloud server obtains aggregated household energy consumption data and individual electric vehicle charging energy consumption data, labels the individual electric vehicle charging energy consumption data to obtain event tags, and trains the electric vehicle charging monitoring model based on the aggregated household energy consumption data and event tags until the model parameters converge or the maximum number of training loops is reached, and finally obtains the model parameters of the electric vehicle charging monitoring model.

[0048] Specifically, the training and compression of the model in this invention are both completed on a cloud server. First, the cloud server obtains aggregated household energy consumption data (e.g., represented by D1) and individual electric vehicle charging energy consumption data (i.e., individual EV charging energy consumption data, e.g., represented by D2). The D1 and D2 data can be obtained from publicly available datasets or by recording power consumption readings from installed electricity meters (e.g., D1 is obtained from the main electricity meter and D2 is obtained from the sub-meters). It should be noted that the two datasets used by the cloud server are from publicly available datasets or have been anonymized and do not contain any user privacy information, thus protecting user privacy.

[0049] The individual electric vehicle charging energy consumption data (D2) needs to be labeled using the Doane frequency classification method to obtain event labels. Then, the aggregated household energy consumption data (D1) and event labels are input into the electric vehicle charging monitoring model (EVSense model) to perform supervised training on the electric vehicle charging monitoring model and obtain the model parameters of the electric vehicle charging monitoring model (the model parameters are the parameters of 5 one-dimensional convolutional layers, 2 bidirectional LSTM layers and 2 fully connected layers). That is, an electric vehicle charging monitoring model (EVSense model) has been pre-designed, but the model parameters of this model have not been determined. This is the process of determining the model parameters.

[0050] For the acquired D2 data, the Doane frequency classification method was used to extract statistical features of the electric vehicle charging load. The |mean-interquartile range| value of the load curve was used as the standard threshold for labeling charging events. When labeling individual electric vehicle charging energy consumption data, if the value of an individual electric vehicle charging energy consumption data is greater than the standard threshold, it is marked as a charging event (e.g., a value of 1, otherwise 0), thus obtaining event labels. Compared with the traditional method of using "0" as the labeling threshold, the labeling method of this invention can not only eliminate the influence of meter reading noise, but also extract more complete electric vehicle charging events, thus making the labeling of D2 data more accurate.

[0051] Furthermore, such as Figure 2 As shown, the pre-designed electric vehicle charging monitoring model is a deep neural network model. The electric vehicle charging monitoring model includes 5 one-dimensional convolutional layers, 2 bidirectional LSTM layers, 2 fully connected layers, 2 ReLU functions and 1 Sigmoid function. The activation function of the last output layer is Sigmoid, and the remaining activation functions are ReLU. In addition, except for the last fully connected layer, each layer includes a normalization operation.

[0052] Layer normalization is used between all layers. The input to the electric vehicle charging monitoring model is an aggregated energy consumption reading sequence, and the output is an electric vehicle charging status label sequence of the corresponding input length. The aggregated energy consumption reading sequence represents aggregated household energy consumption data, and the electric vehicle charging status label sequence represents predicted values ​​based on the aggregated household energy consumption data. Specifically, the input to the electric vehicle charging monitoring model is an aggregated energy consumption reading sequence of length 20 or 10; the output is an EV charging status label sequence of the corresponding input length, with values ​​of "1" or "0" ("1" represents charging status, and "0" represents not charging status).

[0053] Compared to the batch normalization method commonly used in other deep neural network models, the EVSense model proposed in this invention uses layer normalization, such as... Figure 3 As shown, this allows the standardization operation to be performed only within the sequence, and the model training (the standardization process in the model training) is not affected by the number of batch samples.

[0054] The aggregated energy consumption reading sequence is input into the electric vehicle charging monitoring model, which outputs a predicted electric vehicle charging status label sequence. This sequence is compared with the actual charging status label sequence, and the model loss is calculated. Gradient information is obtained to update the model parameters until the model parameters converge or the maximum number of training iterations is reached, thus terminating the process. Finally, the model parameters of the electric vehicle charging monitoring model are obtained.

[0055] The model parameters are updated using batch sample learning and gradient descent. Specifically, the input data is fed into the electric vehicle charging monitoring model (EVSense model) to obtain the predicted EV charging state label sequence output by the EVSense model. The model loss is calculated by comparing it with the real charging state label sequence (the real charging state label sequence is the sequence obtained by labeling the individual electric vehicle charging energy consumption data D2). Gradient information is then obtained to update the model parameters (backpropagation) until the model parameters converge or the maximum number of training iterations is reached.

[0056] Furthermore, since the duration of a home electric vehicle charging event constitutes a relatively small portion of the entire time series record of the electricity meter, and is a sparse event with imbalanced samples, this invention uses binary classification Dice loss to measure the error in model training. That is, during the model parameter training process, the loss function is used to calculate the model prediction error under the current parameters, and gradient information is obtained through the model prediction error. Based on the gradient information, the model parameters are optimized through the backpropagation algorithm (to further reduce the prediction error).

[0057] Step S20: The cloud server compresses the parameters of each layer in the trained electric vehicle charging monitoring model according to a given pruning and compression ratio to obtain the compressed electric vehicle charging monitoring model.

[0058] Specifically, the cloud server sorts the L1 norm values ​​of each layer of the electric vehicle charging monitoring model trained in step S10 in ascending order using the L1 norm (the parameters of the model can be represented by vectors, and the L1 norm of a vector is defined as the sum of the absolute values ​​of its elements). Under a given pruning and compression ratio, the model parameters at the top of the list are set to 0 proportionally. The parameters set to 0 will not play a role in the model inference process, which is equivalent to deleting the parameters from the model (hence the figurative term "pruning"). After deleting these parameters, the entire model becomes smaller, so in essence, model compression is performed. The compressed model is easier to deploy to multiple user edge terminals.

[0059] Step S30: Multiple user edge terminals receive the compressed electric vehicle charging monitoring model sent by the cloud server, and migrate and deploy the electric vehicle charging monitoring model to their respective user edge terminals in an unsupervised manner under the federated transfer learning framework.

[0060] Specifically, such as Figure 4 As shown, the EVSense model can be deployed unsupervised to edge devices of different users within a federated transfer learning framework; the specific process is as follows:

[0061] The first step is to train and compress an electric vehicle charging monitoring model (EVSense model) on a cloud server using existing labeled data, and use this as the initial global initial state (the initial state is the initial model, which will be distributed from the cloud server to each edge terminal (i.e., multiple user edge terminals) at the beginning, hence the name global initial model or global initial state). The global initial state and some labeled data (i.e., some D1 and D2 data) are distributed to the user edge side (user edge terminal). The user edge side (user edge terminal) is the user side, and "edge" is relative to the cloud server (center point).

[0062] The second step is to use transfer learning to update the model parameters at the user edge to adapt to the local data distribution, and then upload the updated model parameters to the cloud server. After the global initial model is deployed to a user, it needs to be fine-tuned according to the user's electricity load characteristics (because the global model is learned from a public dataset and may not be applicable to the specific user). This process of fine-tuning the model parameters is called transfer learning, which involves updating the original global initial model parameters and then using the updated model to perform EV charging detection locally.

[0063] Transfer learning uses a relevance alignment method to transfer a model from the source domain (the cloud server is equivalent to the source domain) to the target domain (the user edge is equivalent to the target domain). First, the covariances of the model parameters in the source domain and the target domain are calculated, denoted as C1 and C2 respectively. Then, the difference L between C1 and C2 is calculated using the following formula. Coral = , where ||A|| F Let L be the Frobenius norm of matrix A, defined as the square root of the sum of the squares of the absolute values ​​of all elements in matrix A; then, the following formula is used as the loss function to retrain and update the local model parameters on the user side: L User = L Supv + L Coral L SupvThe loss value is calculated from partially labeled data, which makes the output signal distribution of the model consistent in the source and target domains.

[0064] The third step involves performing federated average aggregation on the received local model on the cloud server to obtain a new global model state. This updated global state (i.e., global model parameters) can then be migrated and deployed to the newly added user edge in an unsupervised manner for electric vehicle charging monitoring.

[0065] In the above process, the first and second steps can be regarded as the starting stage of the entire model migration process. The third step needs to be repeated iteratively and terminated when the corresponding performance is reached locally or when a maximum number of loops is given. Another advantage of using the above federated learning framework is that the user's data does not need to be uploaded to the cloud server for model training. The public model from the cloud server can be obtained and migrated and updated locally, thus protecting user data privacy and ensuring the accuracy of the prediction model.

[0066] Step S40: Multiple user edge terminals use the migrated and deployed electric vehicle charging monitoring model to monitor the charging of their respective electric vehicles, and filter the model output results based on the output filter to finally output the charging monitoring results.

[0067] Specifically, multiple user edge terminals use the migrated and deployed electric vehicle charging monitoring model to monitor the charging of their respective electric vehicles. Each user edge terminal will adaptively fine-tune the parameters of its corresponding deployed electric vehicle charging monitoring model before monitoring the charging of its own electric vehicle, which can directly obtain electric vehicle charging events. This allows for large-scale deployment while protecting user data privacy.

[0068] Furthermore, the model output results will be filtered through an output filter to finally output the charging monitoring results. Specifically, for the electric vehicle charging state label sequence, a sliding window value and a threshold are set, and erroneous predictions are filtered out based on the output filter to eliminate frequent state jumps in the electric vehicle charging state label sequence.

[0069] Specifically, electric vehicle charging events are generally continuous, ranging from a few minutes to several hours. For the predicted charging events output by the model (i.e., the EV charging state label sequence), this invention designs an output filter to further improve the accuracy of the model's prediction results. For a model prediction sequence of a given length as input, by providing a sliding window value and a threshold, the output filter can filter out short-duration (erroneous) predictions, eliminating frequent state jumps in the predicted charging event sequence. The time complexity of this output filtering method is O(N), where N is the length of the input sequence, and no prior training or learning is required.

[0070] After the electric vehicle charging monitoring model is migrated and deployed at the user edge, it is used to detect EV charging. Since the output of the electric vehicle charging monitoring model may contain some false positives (such as frequent switching on / off cycles within a short period), an output filter is designed to smooth the results. The specific operation of the filter is as follows: Figure 5 As shown.

[0071] The entire process of this invention is as follows: global model training and compression (cloud server) - user-side model migration and deployment (edge ​​terminal device) - user-side model inference (i.e., charging monitoring and filtering); wherein after the user-side model is updated through transfer learning, the updated model can be uploaded to the cloud server to correct and update the original global initial model.

[0072] In this invention, the training and compression of the EVSense model are completed on a cloud server (with strong computing and storage capabilities), while the migration and deployment of the model are completed on edge devices (with weak computing and storage capabilities).

[0073] The EVSense model's inherent structural design contributes to its robustness; the migration and deployment of lightweight models (i.e., compressed models) ensures the method's scalability. Within this framework, users can directly use the compressed (and migrated) cloud-based model for inference without uploading local data, thus providing privacy protection. The EVSense model supports shorter time-series inputs (on the order of minutes), effectively improving the real-time performance of electric vehicle charging monitoring.

[0074] The unique model design of EVSense in this invention (i.e., 5 one-dimensional convolutional layers, 2 bidirectional LSTM layers, and 2 fully connected layers) has not been proposed or used in previous methods; EVSense uses the Dice loss function for training, and the training data labeling method is determined by calculating statistical features; layer normalization is used between the intermediate layers of the neural network in the EVSense model to enhance the stability of training.

[0075] This invention utilizes a federated learning framework and incorporates special operations such as model compression and transfer learning. The resulting solution effectively protects user privacy and data security while simultaneously addressing the deployment and migration challenges of the EVSense model on large-scale, low-computing-power terminals.

[0076] Beneficial effects:

[0077] (1) Quickly and accurately detect electric vehicle charging events from the total load signal of the meter.

[0078] (2) The detection algorithm or model can run on edge devices with low computing power at the user end and can be deployed on a large scale.

[0079] (3) It has user privacy protection attributes, user data is completely stored locally, and data is not shared between different users.

[0080] Furthermore, the non-intrusive electric vehicle charging monitoring method proposed in this invention can be implemented in various other ways. For example, the number and width of intermediate hidden layers can be increased in the main EVSense model. Besides using Dice loss, other loss functions that can address training in imbalanced sample scenarios can also be applied to model training. The methods for model compression and model transfer proposed in this invention also have other approaches. For example, in terms of model compression, there are structured pruning methods, such as using the Frobenius norm as a parameter pruning metric. Regarding model transfer, in addition to the CORAL feature alignment method used in this invention, other domain adaptation methods in unsupervised learning can also be applied to the model transfer framework.

[0081] Furthermore, such as Figure 1 As shown, based on the above-described electric vehicle charging monitoring method, this invention also provides an electric vehicle charging monitoring system. The system includes a cloud server and multiple user edge terminals, with the cloud server communicating with each user edge terminal. The cloud server acquires aggregated household energy consumption data and individual electric vehicle charging energy consumption data, labels the individual electric vehicle charging energy consumption data to obtain event tags, and trains an electric vehicle charging monitoring model based on the aggregated household energy consumption data and event tags until the model parameters converge or the maximum number of training iterations is reached, ultimately obtaining the model parameters of the electric vehicle charging monitoring model. The cloud server also compresses the parameters of each layer in the trained electric vehicle charging monitoring model according to a given pruning compression ratio to obtain a compressed electric vehicle charging monitoring model. Multiple user edge terminals receive the compressed electric vehicle charging monitoring model sent by the cloud server and migrate and deploy the electric vehicle charging monitoring model to their respective user edge terminals in an unsupervised manner under a federated transfer learning framework. The multiple user edge terminals use the migrated and deployed electric vehicle charging monitoring model to monitor the charging of their respective electric vehicles and filter the model output results based on an output filter to finally obtain the charging monitoring results.

[0082] In summary, this invention provides an electric vehicle charging monitoring method and system. The method includes: a cloud server acquiring aggregated household energy consumption data and individual electric vehicle charging energy consumption data; labeling the individual electric vehicle charging energy consumption data to obtain event tags; training an electric vehicle charging monitoring model based on the aggregated household energy consumption data and event tags until the model parameters converge or the maximum number of training iterations is reached, ultimately obtaining the model parameters of the electric vehicle charging monitoring model; the cloud server compressing the parameters of each layer in the trained electric vehicle charging monitoring model according to a given pruning compression ratio to obtain a compressed electric vehicle charging monitoring model; multiple user edge terminals receiving the compressed electric vehicle charging monitoring model sent by the cloud server, migrating and deploying the electric vehicle charging monitoring model to their respective user edge terminals in an unsupervised manner under a federated transfer learning framework; multiple user edge terminals fine-tuning the model parameters according to their respective local data characteristics, then using the fine-tuned electric vehicle charging monitoring model to monitor the charging of their respective electric vehicles, and filtering the model output results based on an output filter to finally obtain the charging monitoring results. This invention can deploy electric vehicle charging monitoring models on low-computing-power edge devices and achieve accurate and low-latency detection, enabling large-scale deployment while protecting user data privacy.

[0083] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0084] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware (such as a processor, controller, etc.). The program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The computer-readable storage medium can be a memory, magnetic disk, optical disk, etc.

[0085] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A method for monitoring electric vehicle charging, characterized in that, The electric vehicle charging monitoring method includes: The cloud server acquires aggregated household energy consumption data and individual electric vehicle charging energy consumption data, annotates the individual electric vehicle charging energy consumption data to obtain event tags, and trains the electric vehicle charging monitoring model based on the aggregated household energy consumption data and event tags until the model parameters converge or the maximum number of training loops is reached, and finally obtains the model parameters of the electric vehicle charging monitoring model. The cloud server compresses the parameters of each layer in the trained electric vehicle charging monitoring model according to a given pruning and compression ratio to obtain the compressed electric vehicle charging monitoring model. Multiple user edge terminals receive compressed electric vehicle charging monitoring models sent by the cloud server, and migrate and deploy the electric vehicle charging monitoring models to their respective user edge terminals in an unsupervised manner under the federated transfer learning framework. Multiple user edge devices use the migrated and deployed electric vehicle charging monitoring model to monitor the charging of their respective electric vehicles, and filter the model output results based on the output filter to finally output the charging monitoring results.

2. The electric vehicle charging monitoring method according to claim 1, characterized in that, The cloud server acquires aggregated household energy consumption data and individual electric vehicle charging energy consumption data. It labels the individual electric vehicle charging energy consumption data to obtain event tags. Based on the aggregated household energy consumption data and event tags, it trains the electric vehicle charging monitoring model until the model parameters converge or the maximum number of training iterations is reached, ultimately obtaining the model parameters of the electric vehicle charging monitoring model, specifically including: The cloud server acquires aggregated household energy consumption data and individual electric vehicle charging energy consumption data, and uses the Doane frequency classification method to label the individual electric vehicle charging energy consumption data to obtain event labels; Aggregated household energy consumption data and event labels are input into a pre-designed electric vehicle charging monitoring model for supervised training. The aggregated energy consumption reading sequence is input into the electric vehicle charging monitoring model, which outputs a predicted electric vehicle charging status label sequence. This sequence is compared with the actual charging status label sequence, and the model loss is calculated. Gradient information is obtained to update the model parameters until the model parameters converge or the maximum number of training iterations is reached, thus terminating the process. Finally, the model parameters of the electric vehicle charging monitoring model are obtained.

3. The electric vehicle charging monitoring method according to claim 2, characterized in that, The process of labeling individual electric vehicle charging energy consumption data using the Doane frequency classification method to obtain event labels specifically includes: When labeling individual electric vehicle charging energy consumption data, the cloud server uses the Doane frequency classification method to extract statistical features of the electric vehicle charging load. The |mean-interquartile range| value of the load curve is used as the standard threshold for labeling charging events. If the value of a single electric vehicle charging energy consumption data is greater than the standard threshold, it is marked as a charging event, and an event label is obtained.

4. The electric vehicle charging monitoring method according to claim 2, characterized in that, The actual charging status label sequence is a sequence obtained by labeling individual electric vehicle charging energy consumption data.

5. The electric vehicle charging monitoring method according to claim 3, characterized in that, The cloud server compresses the parameters of each layer in the trained electric vehicle charging monitoring model according to a given pruning and compression ratio to obtain a compressed electric vehicle charging monitoring model, specifically including: The cloud server sorts the L1 norm values ​​of the parameter vectors in ascending order using the L1 norm for each layer of the trained electric vehicle charging monitoring model. Given a pruning and compression ratio, the parameters of the models at the top of the list are set to 0 proportionally to compress the electric vehicle charging monitoring model, resulting in a compressed electric vehicle charging monitoring model.

6. The electric vehicle charging monitoring method according to claim 5, characterized in that, The pre-designed electric vehicle charging monitoring model is a deep neural network model. The electric vehicle charging monitoring model includes 5 one-dimensional convolutional layers, 2 bidirectional LSTM layers, 2 fully connected layers, 2 ReLU functions and 1 Sigmoid function. Except for the last fully connected layer, each layer contains a normalization operation.

7. The electric vehicle charging monitoring method according to claim 5, characterized in that, The input to the electric vehicle charging monitoring model is an aggregated energy consumption reading sequence, and the output is an electric vehicle charging status label sequence of the corresponding input length. The aggregated energy consumption reading sequence is aggregated household energy consumption data, and the electric vehicle charging status tag sequence is a predicted value based on the aggregated household energy consumption data.

8. The electric vehicle charging monitoring method according to claim 5, characterized in that, The multiple user edge terminals receive the compressed electric vehicle charging monitoring model sent by the cloud server, and migrate and deploy the electric vehicle charging monitoring model to their respective user edge terminals in an unsupervised manner under a federated transfer learning framework. Specifically, this includes: Multiple user edge devices receive global initial status and some tagged data sent by the cloud server; Multiple user edge devices use transfer learning to update the model parameters to adapt to the local data distribution, and then upload the updated model parameters to the cloud server.

9. The electric vehicle charging monitoring method according to claim 8, characterized in that, The multiple user edge terminals use transfer learning to update the model parameters to adapt to the local data distribution, and then upload the updated model parameters to the cloud server. This process also includes: The cloud server performs federated averaging aggregation on the received local models to obtain a new global model state. This updated global state is then migrated and deployed in an unsupervised manner to the newly added user edge for electric vehicle charging monitoring.

10. An electric vehicle charging monitoring system, characterized in that, The electric vehicle charging monitoring system includes: a cloud server and multiple user edge terminals, wherein the cloud server is communicatively connected to the multiple user edge terminals respectively; The cloud server is used to acquire aggregated household energy consumption data and individual electric vehicle charging energy consumption data, annotate the individual electric vehicle charging energy consumption data to obtain event tags, and train the electric vehicle charging monitoring model based on the aggregated household energy consumption data and event tags until the model parameters converge or the maximum number of training loops is reached and terminated, and finally obtain the model parameters of the electric vehicle charging monitoring model. The cloud server is also used to compress the parameters of each layer in the trained electric vehicle charging monitoring model according to a given pruning compression ratio to obtain a compressed electric vehicle charging monitoring model. Multiple user edge terminals are used to receive the compressed electric vehicle charging monitoring model sent by the cloud server. Under the federated transfer learning framework, the electric vehicle charging monitoring model is transferred and deployed to their respective user edge terminals in an unsupervised manner. Multiple user edge devices use the migrated and deployed electric vehicle charging monitoring model to monitor the charging of their respective electric vehicles, and filter the model output results based on the output filter to finally output the charging monitoring results.

Images