Non-intrusive identification method, device and system for charging behavior of household electric vehicle

Abstract

The invention provides a non-intrusive identification method, device and system for the charging behavior of a household electric vehicle, and the method comprises the following steps: collecting the aggregated power utilization information at a power inlet of a distribution box, analyzing the charging behavior and charging load characteristics of the household electric vehicle, and constructing the sample data needed by the training of a neural network model; constructing an SAED model for effectively identifying the EV charging load, and performing model training and optimization by using the sample data set to obtain a trained SAED model; a trained SAED model is deployed, non-intrusive electric quantity subentry refined identification (English short for NILM, Non-Intrusive Load Monitor) is carried out on the charging behavior of a user in situ, and the abnormal power consumption condition is discriminated. According to the method, the charging load of the electric vehicle can be effectively identified, the electric vehicle which is charged privately can be positioned, potential safety hazards and non-standard power consumption behaviors can be found, a good power consumption environment and safe and stable operation of a power grid are guaranteed, and more accurate, efficient and flexible power consumption management service is provided for the power industry.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent monitoring of electrical loads, and particularly relates to a non-invasive identification method, device and system for household electric vehicle charging behavior. Background Technique

[0002] With the large-scale popularization of electric vehicles, the charging demand has surged. However, due to a series of factors such as policy regulations, installation environment, and application procedures, some households do not set up special electric vehicle charging piles separately inside the house, but charge electric vehicles through privately installed charging piles or flying wires, which poses a great safety hazard. At the same time, during the sampling inspection of charging pile users, it is found that there are many phenomena of privately connecting other loads to electric vehicle charging piles, evading the residential ladder electricity price and affecting the normal power supply and consumption order.

[0003] However, there is a lack of effective and intelligent supervision means for electric vehicle charging, and it is difficult to detect irregular charging behaviors in a timely manner. Therefore, it is necessary to study the identification of electric vehicle charging loads, locate the privately charged electric vehicles, which can not only actively provide the application service of charging piles for users who meet the installation conditions and improve the quality of service level, but also discover potential safety hazards and irregular electricity consumption behaviors and ensure a good power consumption environment.

[0004] Electric load monitoring has important promoting significance for solving a series of electrical engineering problems, such as safe and intelligent power consumption, power system planning and prediction, power grid stability simulation modeling, formulation of demand response policies, and energy conservation and emission reduction issues. Traditional load monitoring takes sockets as units, and sensors are equipped for the same type of electrical appliances or even single electrical appliances to collect load power consumption information, which is not only costly and difficult to maintain, but also difficult to achieve real-time transmission of high-frequency sensing data. Summary of the Invention

[0005] In order to solve the above problems, the present invention provides a non-invasive identification method, device and system for household electric vehicle charging behavior, which can effectively identify abnormal electric vehicle charging behaviors.

[0006] The technical solution adopted by the present invention to solve its technical problems is as follows:

[0007] In a first aspect, a non-invasive identification method for household electric vehicle charging behavior provided by an embodiment of the present invention includes the following steps:

[0008] Step S1, collect the aggregated power consumption information at the power inlet of the distribution box, analyze the household electric vehicle charging behavior and charging load characteristics, and construct the sample data required for training the neural network model;

[0009] Step S2, construct a SAED (Self-Attentive Energy Disaggregator) model for effectively identifying the charging load of EV (Electric Vehicle), and use the sample data set to train and optimize the model to obtain a trained SAED model;

[0010] Step S3, deploy the trained SAED model to finely identify the user's charging behavior in-situ by NILM (Non-Intrusive Load Monitoring), and distinguish abnormal electricity consumption situations.

[0011] Furthermore, the step S1 includes the following steps:

[0012] Install a smart meter at the power inlet of the household distribution box to collect aggregated power consumption information including the electric vehicle load data of the household users;

[0013] Perform data cleaning and denoising preprocessing on the collected data;

[0014] Combine the typical household electric vehicle charging load data sets published internationally and the collected household electric vehicle load data, and perform simulation amplification of the data volume according to the characteristics of the electric vehicle charging load to form a sample data set.

[0015] Furthermore, the sample data set includes the curve and waveform data of the EV charging load.

[0016] Furthermore, the SAED model includes a CNN convolutional layer, a self-attention mechanism layer, a bidirectional GRU layer, and two fully connected network layers. The CNN convolutional layer is used to extract the feature information of the identified load signal; the self-attention mechanism layer is used to focus on the most important correlation relationships in the feature information; the bidirectional GRU layer is used to identify the pattern rules of the load sequence data; the two fully connected network layers are used to regress and generate the final load sequence structure.

[0017] Furthermore, in step S2, using the sample data set to train and optimize the model to obtain a trained SAED model includes the following steps:

[0018] Use the sample data set to train the SAED model in a supervised learning manner;

[0019] During the training process, use the concise and efficient PyTorch as the support framework for the model's training and inference operations. The model training and tuning work is carried out on the kubeflow platform;

[0020] Implement the distributed training of the model on the GPU cluster through the training-operator of the kubeflow platform, and implement the model parameter tuning through katib;

[0021] Manage all metadata through metadata, and automate the deep learning workflow through ml-pipeline;

[0022] After multiple rounds of training and validation, continuously optimize the parameters and performance of the model so that the SAED model can accurately identify the charging load of electric vehicles.

[0023] Furthermore, step S3 includes the following steps:

[0024] Deploy the trained SAED model to the edge gateway intelligent device in the ONNX format;

[0025] Sample the smart meter at the power inlet of the user's distribution box from seconds to minutes to monitor the user's electricity consumption in real time;

[0026] When detecting suspicious charging load information, use the SAED model for analysis to determine whether there is abnormal charging / electricity consumption behavior;

[0027] If an abnormality is found, an alarm message will be sent in a timely manner.

[0028] Furthermore, the abnormal electricity consumption conditions include:

[0029] Low-voltage users privately install charging piles to charge or directly pull wires to charge;

[0030] Charging pile users privately add charging piles for use;

[0031] Users privately connect their electricity loads to charging piles for use to avoid the residential ladder electricity price.

[0032] In a second aspect, a non-intrusive identification device for home electric vehicle charging behavior provided by an embodiment of the present invention includes:

[0033] A sample data construction module for collecting the aggregated electricity consumption information at the power inlet of the distribution box, analyzing the home electric vehicle charging behavior and charging load characteristics, and constructing the sample data required for neural network model training;

[0034] A model training and optimization module for constructing a SAED model that can effectively identify EV charging loads, and using the sample data set for model training and optimization to obtain a trained SAED model;

[0035] An abnormal electricity consumption discrimination module for deploying the trained SAED model to locally perform NILM fine-grained identification of the user's charging behavior and discriminating abnormal electricity consumption conditions.

[0036] In a third aspect, an non-intrusive identification system for household electric vehicle charging behavior provided by an embodiment of the present invention includes a smart meter, an edge gateway intelligent device, and a cloud server; the smart meter is used to sample the power load of a suspicious user; the edge gateway intelligent device is deployed with a SAED model, which is used to analyze the power load of the suspicious user, accurately identify the charging load information in the power load, and combine the reported charging pile installation information to accurately associate and analyze abnormal power consumption situations; the cloud server is responsible for calculating task allocation and storing the power load data and its identification results.

[0037] Further, the SAED model includes a CNN convolutional layer, a self-attention mechanism layer, a bidirectional GRU layer, and two fully connected network layers. The CNN convolutional layer is used to extract the feature information of the identification load signal; the self-attention mechanism layer is used to focus on the most important correlation relationships in the feature information; the bidirectional GRU layer is used to identify the pattern rules of the load sequence data; the two fully connected network layers are used to regress and generate the final load sequence structure.

[0038] Compared with the prior art, the beneficial effects are as follows:

[0039] (1) The present invention can effectively identify the electric vehicle charging load, locate the privately charged electric vehicle, discover potential safety hazards and irregular power consumption behaviors, ensure a good power consumption environment, and at the same time ensure the safe and stable operation of the power grid, providing a more accurate, efficient, and flexible power consumption management service for the power industry.

[0040] (2) The present invention proposes a NILM artificial intelligence identification charging load model, which can accurately identify and statistically analyze the electric vehicle charging load based on a low sampling frequency. The data identification accuracy is high, and it can be deployed and run on edge intelligent devices, greatly improving the monitoring efficiency and accuracy.

[0041] (3) The present invention proposes a SAED model, which is composed of a feature extraction convolutional network layer, a self-attention processing layer, and a load curve generation bidirectional GRU network layer. It can accurately identify and statistically analyze the electric vehicle charging load based on a low sampling frequency. The data identification accuracy is high, and it can be deployed and run on edge intelligent devices, greatly improving the monitoring efficiency and accuracy.

[0042] (4) By analyzing the electrical energy data at the user's power inlet, the present invention obtains the load categories and their electricity consumption information existing within the user area. It is not only applicable to EV charging management but also enables more scientific power grid planning, power generation scheduling, and formulation of dynamic electricity price policies. It can more accurately conclude and verify demand response (DR), audit the energy efficiency information of enterprises, etc., which helps the safe and economic operation of the power grid. For household users, the transparency of electricity consumption information can help them understand their own energy consumption composition and promote the formation of energy-saving awareness. Description of the Drawings

[0043] Figure 1 is a flowchart of a non-intrusive identification method for household electric vehicle charging behavior of the present invention;

[0044] Figure 2 is a schematic structural diagram of a non-intrusive identification device for household electric vehicle charging behavior of the present invention;

[0045] Figure 3 is a schematic structural diagram of a non-intrusive identification system for household electric vehicle charging behavior of the present invention;

[0046] Figure 4 is a schematic structural diagram of the SEAD neural network model of the present invention;

[0047] Figure 5 is a specific flowchart for the present invention to identify household electric vehicle charging behavior;

[0048] Figure 6 is a schematic diagram of 8 typical charging load curves of household EVs. Detailed Embodiment

[0049] The following will elaborate in detail on the specific embodiments of a non-intrusive identification method, device, and system for household electric vehicle charging behavior of the present invention with reference to the accompanying drawings.

[0050] Non-intrusive load monitoring (NILM) technology takes rooms, user units, or even buildings and campuses as units. It only needs to collect information at the power inlet of the distribution box and can decompose and infer the real-time electricity consumption information of various loads within the area (such as the power of a certain type of electrical appliance load, historical electricity consumption data statistics, electricity consumption suggestions, etc.) by processing data such as current and power through intelligent algorithms. It can be seen that compared with traditional electricity consumption monitoring methods, the NILM system can greatly reduce the investment in sensing devices, is economical and environmentally friendly, has better controllability in data generation, and has better functional expansion and interoperability capabilities, becoming a new generation of digital and intelligent power load monitoring technology.

[0051] At the current stage, there are certain technical gaps in the intelligent supervision during the charging process of electric vehicles, which makes it difficult for regulatory agencies to effectively detect and handle irregular behaviors during the charging process in a timely manner. This invention has conducted research on monitoring methods based on NILM such as extraction of electric vehicle charging load characteristics, detection of switch events, load decomposition algorithms, and system interaction communication, proposed a novel deep learning neural network model, and demonstrated its performance and effectiveness on real intelligent meter data; furthermore, proposed an edge computing architecture of NILM in the Internet of Things scenario and designed and implemented an intelligent device for identifying abnormal charging behaviors of electric vehicles.

[0052] As Figure 1 shown, a non-intrusive identification method for the charging behavior of household electric vehicles provided by an embodiment of the present invention includes the following steps:

[0053] Step S1, collect the aggregated electricity consumption information at the power inlet of the distribution box, analyze the charging behavior of household electric vehicles and the charging load characteristics, and construct the sample data required for training the neural network model;

[0054] Step S2, construct a SAED model for effectively identifying EV charging loads, and use the sample data set for model training and optimization to obtain a trained SAED model;

[0055] Step S3, deploy the trained SAED model to conduct NILM refined identification of the user's charging behavior in-situ and distinguish abnormal electricity consumption situations.

[0056] As a possible implementation manner of this embodiment, the step S1 includes the following steps:

[0057] Install an intelligent meter at the power inlet of the household user's distribution box to collect the aggregated electricity consumption information including the load data of the household user's electric vehicle;

[0058] Perform data cleaning and denoising preprocessing on the collected data;

[0059] Combine the published typical household electric vehicle charging load data sets internationally and the collected household user electric vehicle load data, and perform simulation amplification of the data volume according to the electric vehicle charging load characteristics to form a sample data set.

[0060] As a possible implementation manner of this embodiment, the sample data set in step S2 includes the curve and waveform data of the EV charging load.

[0061] As a possible implementation of this embodiment, the SAED model includes a CNN convolutional layer, a self-attention mechanism layer, a bidirectional GRU layer, and two fully-connected network layers. The CNN convolutional layer is used to extract the feature information of the identification load signal; the self-attention mechanism layer is used to focus on the most important correlation relationships in the feature information; the bidirectional GRU layer is used to identify the pattern rules of the load sequence data; and the two fully-connected network layers are used to regress and generate the final load sequence structure.

[0062] As a possible implementation of this embodiment, using the sample data set for model training and optimization to obtain a trained SAED model includes the following steps:

[0063] Use the sample data set to train the SAED model in a supervised learning manner;

[0064] During the training process, use the concise and efficient PyTorch as the support framework for model training and inference operations, and the model training and tuning work is carried out on the kubeflow platform;

[0065] Implement the distributed training of the model on the GPU cluster through the training-operator of the kubeflow platform, and implement model parameter tuning through katib;

[0066] Manage all metadata through metadata, and implement the automation of the deep learning workflow through ml-pipeline;

[0067] After multiple rounds of training and validation, continuously optimize the model parameters and performance so that the SAED model can accurately identify the electric vehicle charging load.

[0068] As a possible implementation of this embodiment, step S3 includes the following steps:

[0069] Deploy the trained SAED model to the edge gateway intelligent device in the ONNX format;

[0070] Sample the smart meter at the power inlet of the user's distribution box from seconds to minutes to monitor the user's electricity consumption in real time;

[0071] When suspicious charging load information is detected, use the SAED model for analysis to determine whether there is abnormal charging / electricity consumption behavior;

[0072] If an abnormality is found, send an alarm message in time.

[0073] As a possible implementation of this embodiment, the abnormal electricity consumption conditions include:

[0074] 1) Low-voltage users install charging piles privately to charge or directly pull wires privately to charge;

[0075] 2) Charging pile users add charging piles privately for use;

[0076] 3) Users connect their electricity loads privately to charging piles for use to avoid residential ladder electricity prices.

[0077] As Figure 2 shown, a non-intrusive identification device for home electric vehicle charging behavior provided by an embodiment of the present invention includes:

[0078] A sample data construction module, configured to collect the aggregated electricity consumption information at the power inlet of the distribution box, analyze the home electric vehicle charging behavior and charging load characteristics, and construct the sample data required for training the neural network model;

[0079] A model training and optimization module, configured to construct a SAED model for effectively identifying EV charging loads, and use the sample data set for model training and optimization to obtain a trained SAED model;

[0080] An abnormal electricity consumption discrimination module, configured to deploy the trained SAED model to perform NILM fine-grained identification of the user's charging behavior locally, and discriminate abnormal electricity consumption situations.

[0081] As Figure 3 shown, a non-intrusive identification system for home electric vehicle charging behavior provided by an embodiment of the present invention includes a smart meter, an edge gateway intelligent device, and a cloud server; the smart meter is used to sample the electricity load of a suspicious user; the edge gateway intelligent device is deployed with a SAED model, which is used to analyze the electricity load of the suspicious user, accurately identify the charging load information in the electricity load, and combine the reported charging pile installation information to accurately associate and analyze abnormal electricity consumption situations; the cloud server is responsible for computing task allocation and storing the electricity load data and its identification results.

[0082] As Figure 4 shown, as a possible implementation manner of this embodiment, the SAED model includes a CNN convolutional layer, a self-attention mechanism layer, a bidirectional GRU layer, and two fully connected network layers. The CNN convolutional layer is used to extract the feature information of the identification load signal; the self-attention mechanism layer is used to focus on the most important correlation relationships in the feature information; the bidirectional GRU layer is used to identify the pattern rules of the load sequence data; the two fully connected network layers are used to regress and generate the final load sequence structure.

[0083] I. Analyze the home electric vehicle charging behavior and charging load characteristics, and construct high-quality sample data required for training the neural network model.

[0084] The charging load of electric vehicles is special compared with other household electrical equipment. The charging load signal has two states, on and off, and the switching frequency between the two states is very low. For example Figure 6 The charging curves of eight typical household electric vehicles are shown. It can be seen that the rated charging power range is 3kW - 20kW. The curve shape of the charging load is approximately rectangular. The characteristics of the charging load can be summarized as follows: 1) rising edge; 2) stable charging stage; 3) falling edge. The duration of the rising edge is short, and the power of the falling edge of some EV charging curves shows a gentle downward trend. Therefore, a path to identify the electric vehicle charging signal is to determine the positions of the rising edge, falling edge, and steady-state charging power.

[0085] Based on the actual electric vehicle load data of household users collected in the target area and some typical household electric vehicle charging load datasets EV-CPW (EV Charging Profiles and Waveforms, https: / / openenergyhub.ornl.gov / explore / dataset / ev-cpw-ev-charging-profiles-and-wavef orms / information / ) released internationally, combined with the load characteristics as shown in Figure 6 The figure, the simulation amplification of the data volume is carried out to support the systematic training, verification, and continuous optimization of the developed neural network model.

[0086] II. Construct an end-to-end deep learning neural network model that can effectively identify EV charging loads.

[0087] Traditional system models are often composed of many cooperating subsystems in terms of architecture. The subsystems provide the required information for each other and are indispensable, providing a complete path for the information flow, and can be called an Ensemble Model. To improve the overall accuracy and expressiveness of the model, researchers need to design a self-consistent system structure by combining professional theories and various characteristics of the thing being modeled, which requires in-depth design and tuning of each subsystem separately. Deep learning is a modeling method with heuristic learning ability. A complete model is composed of many neural network layers, and can even reach thousands of layers, which is the origin of the term "deep". The end-to-end architecture is to use only one neural network to represent the entire system by leveraging the heuristic learning ability of deep learning, abandoning the design of sub-modules in the integrated architecture. In addition, the design of the entire model is completed in one go, and the training process is also carried out synchronously for all parameters, which is quite different from the asynchronous training and parameter tuning of the integrated model. This end-to-end architecture has been widely used in speech recognition, autonomous driving, etc.

[0088] This invention focuses on researching such end-to-end neural network model solutions. The load decomposition model of NILM is designed in an End to End mode. Instead of being divided into two steps of load detection and load decomposition, after the original electrical quantities are input into the model, the decomposed sub-load electrical quantities corresponding to the time nodes are directly output. This invention constructs a hybrid model combining CNN (Convolutional Neural Network) and RNN (Recurrent Neural Network). Through supervised learning, the CNN model in the architecture "remembers" the changes in electrical quantities when the target load starts and stops, replacing the manually set strategy. Finally, the RNN model in the architecture completes the load decomposition according to the trend law of the time series and directly outputs the identified complete charging load signal.

[0089] Based on this, this invention proposes a novel model architecture - SAED (Self Attentive Energy Disaggregator). SAED is an End to End mode NILM neural network model, which is composed of a feature extraction convolutional network layer (CNN), a self-attention processing layer, and a sub-item load generation bidirectional GRU network layer (a type of RNN). Compared with general deep learning neural networks, this model has only 48k parameters while maintaining high data recognition accuracy.

[0090] Specifically, as Figure 4 shown, the SAED model is a hybrid model of four small neural network models:

[0091] 1) A CNN convolutional layer is used to extract and identify the feature information of the load signal;

[0092] 2) A self-attention mechanism layer is used to focus on the most important correlation relationships in the feature information;

[0093] 3) A bidirectional GRU layer is used to identify the pattern rules of the load sequence data;

[0094] 4) Two fully connected network layers are used to regress and generate the final load sequence structure. For this invention, it is the electric vehicle charging load curve finally obtained.

[0095] The trained model can be deployed and run on the cloud or edge intelligent devices. This model uses the sampling data of the conventional smart meter AMI (sampling frequency of 1 minute - 15 minutes) to accurately identify and statistically analyze the electric vehicle charging load.

[0096] III. The trained SAED model is deployed and run on the edge gateway intelligent device to conduct NILM fine-grained identification of users' charging behaviors locally.

[0097] The Internet of Things (IoT) is a collection of multiple technologies, including wireless sensor networks (WSNs), RFID radio frequency, etc. The IoT can also be regarded as a huge ecosystem of embedded technologies, such as various sensing devices, proximal communication devices, automation devices, etc. Naturally, the IoT has gradually penetrated into many industries, such as information collection and processing in smart grids, smart homes, and smart transportation. The emergence of the IoT can provide a perfect communication architecture and information storage solution for large-scale deployed smart meters, making large-scale NILM systems possible.

[0098] This invention focuses on constructing a technical framework with three main entities: the cloud, the edge gateway, and the terminal (smart meter). The cloud is responsible for computing task allocation, storage, decision-making, user-side interaction, etc. The edge gateway (also the intelligent device mainly developed in this invention) deployed near the smart meter plays the role of abstracting the sensor bottom layer. The smart meter terminal is combined with the communication component of this invention to abstract the communication between it and the edge gateway, including querying, receiving, executing decisions, and obtaining underlying data operations.

[0099] Deep learning models mainly function as inference on the IoT platform, while model training still relies on x86 servers equipped with GPUs. In addition, another advantage of deep learning is its compatibility with different practical problems, and it can perform automatic feature extraction on data of different specifications. When processing multimedia information, the performance of traditional machine learning models depends on the accuracy of features defined by artificial feature engineering; deep learning can gradually extract accurate features using large-scale data during the training process. At the same time, due to the significant reduction in the cost of current embedded NPUs (Neural Processing Units) and GPUs (Graphic Processing Units), and the greatly enhanced performance of ARM architecture embedded CPUs, the inference speed of deep learning models in embedded devices is already faster than that of traditional machine learning models based on the X86 platform.

[0100] The emergence of edge computing is to discretize computing power from the centralized cloud to a large number of edge nodes, but still retain the computing power of the cloud as the computing power supplement for the entire IoT, which is also the biggest difference from distributed computing. Edge computing has two improvements compared to existing cloud computing:

[0101] 1) Edge nodes can preprocess large-scale sensing data and, if necessary, send it to the cloud later.

[0102] 2) After edge nodes have their own computing power, cloud resources can be utilized more efficiently, leaving more powerful computing power for tasks that need it more.

[0103] The present invention focuses on the cooperation among smart meters, edge gateways, and the cloud. Smart meters can actually be abstracted as sensors, which may have very weak computing power, but the computing performance of the edge gateway can make up for it and can deploy and execute deep learning algorithms.

[0104] Based on this, the present invention has developed an edge gateway intelligent device based on the Raspberry Pi Zero W system, which can sample the smart meter at the power inlet of the user's distribution box from seconds to minutes. The trained SAED model can also be deployed and run on this device to locally and finely identify the user's charging behavior through NILM.

[0105] IV. Discrimination criteria for three types of abnormal charging / electricity consumption behaviors based on the developed model, edge computing architecture, and intelligent device.

[0106] Based on the above complete technical solution and intelligent device, for the charging load information analyzed by the intelligent device, combined with the relevant reported information on the installation of charging piles, abnormal electricity consumption situations can be accurately correlated and analyzed, such as:

[0107] Low-voltage users privately install charging piles to charge or directly pull wires to charge;

[0108] Charging pile users privately add charging piles for use;

[0109] Users privately connect their electricity loads to charging piles for use to avoid residential ladder electricity prices, etc.

[0110] After testing and verification in the target area and the laboratory, the above system solution and device can effectively support the project requirements:

[0111] 1) For the charging loads of various types of charging piles (guns) and typical electric vehicle types, high-precision identification can be achieved. The identification accuracy is higher than or not lower than other NILM methods in the current industry.

[0112] 2) The SAED model deployed on the cloud server can effectively distinguish the load curves of electric vehicles in normal areas and abnormal area load curves.

[0113] 3) The developed intelligent device can accurately sample and analyze the electricity load of suspicious users from seconds to minutes through the smart meter, and can accurately identify the charging load information in the electricity load by running the SAED model locally.

[0114] 4) For the charging load information analyzed by the intelligent device, combined with the reported charging pile installation information, abnormal electricity consumption situations can be accurately associated and analyzed. In the present invention, the identification and analysis problems of the following three situations are mainly solved:

[0115] a) Low-voltage users privately install charging piles for charging or directly pull wires for charging;

[0116] The system can completely automatically judge this situation. When a user privately installs a charging pile or pulls a wire for charging on a branch under the jurisdiction of a certain intelligent electric meter without application, the present invention will detect the charging load in the aggregated data of the intelligent electric meter where there should be no electric vehicle charging load component. The system will immediately give an alarm, and the relevant department can also impose an economic penalty on the user according to the accurate power (kW) and electricity (kWh) data monitored by NILM.

[0117] b) Users with existing charging piles privately add charging piles for use;

[0118] From the perspective of electricity consumption load, if a user privately adds a charging pile (for multiple vehicles to charge in parallel) on a branch under the jurisdiction of a certain intelligent electric meter, a charging load that does not conform to the charging law of general single-pile household users will be observed. Generally, the charging duration in a residence exceeds 30 minutes, while the charging behavior rarely lasts more than 200 minutes. It is unlikely that residents charge more than 3 times a day, and the interval between two adjacent charging events is often greater than 2 hours. If there are multiple piles (multiple vehicles charging), it can be detected that the interval between charging events is much smaller, and there may even be a situation where several groups of charging loads overlap. The overlap of multiple groups of charging loads will cause obvious changes in the load waveform and maximum amplitude compared with a single group of charging loads. The present invention has considered this situation in model training and added corresponding training samples. In most cases, even with load overlap, it can be effectively identified. Therefore, the system of the present invention can identify and analyze this type of situation.

[0119] c) Users privately connect their electricity consumption loads to charging piles to avoid the residential ladder electricity price, etc.

[0120] For the aggregated power data of the intelligent electric meter collected, after removing the charging load sub-components identified by the device, the remaining power components are analyzed and judged. If there is a situation where the volatility is relatively large and there is a relatively large power (8 to 10 kW) in some time periods, it can be judged that there is a situation where the electricity consumption load is privately connected to this charging pile.

[0121] In view of the characteristic curve of the electric vehicle charging load, the present invention proposes a deep learning neural network model for non-intrusive electricity sub-item measurement in an End to End mode - the SAED model, which can effectively detect the load curve of electric vehicles by non-intrusively sampling and identifying the aggregated electricity consumption information at the power inlet of the distribution box. Furthermore, from the perspective of edge computing, a NILM edge computing architecture is proposed, and an embedded device for identifying abnormal behaviors of electric vehicle charging is designed and implemented, which can accurately identify various abnormal situations for suspicious individual users or user areas.

[0122] The present invention proposes a SAED model, which consists of a feature extraction convolutional network layer, a self-attention processing layer, and a load curve generation bidirectional GRU network layer. Based on a low sampling frequency, it can achieve accurate discrimination and statistical analysis of the electric vehicle charging load, with high data recognition accuracy, and can be deployed and run on edge intelligent devices, greatly improving the efficiency and accuracy of monitoring.

[0123] In the edge computing scenario, the present invention proposes a technical architecture for implementing NILM, which can adopt a sampling frequency at the minute level for suspicious electricity users, and run the artificial intelligence model designed by the present invention to effectively identify and alarm various abnormal situations.

[0124] The present invention proposes a novel deep learning neural network model - the SAED model from aspects such as electric vehicle charging load feature extraction, switch event detection, load decomposition algorithm, and cloud-edge collaborative architecture based on the NILM monitoring method. By analyzing the electrical energy data at the user's power inlet, the load categories and their electricity consumption information existing within the user area can be obtained. It is not only for EV charging management, but also can be used for more scientific power grid planning, power generation scheduling, formulating dynamic electricity price policies; more accurately concluding and verifying demand response (DR), auditing the energy efficiency information of enterprises, etc., which is helpful for the safe and economic operation of the power grid. For household users, the transparency of electricity consumption information can help them understand their own energy consumption composition and promote the formation of energy-saving awareness.

[0125] Finally, it should be noted that the technical solutions of the present invention are only illustrated in combination with the above embodiments and are not limited thereto. Those of ordinary skill in the art should understand that those skilled in the art can modify or equivalently replace the specific embodiments of the present invention, but these modifications or changes are all within the scope of protection of the claims pending for approval.

Claims

1. A non-invasive identification method for the charging behavior of household electric vehicles, characterized in that It includes the following steps: Step S1: Collect the aggregated power consumption information at the power inlet of the distribution box, analyze the charging behavior and charging load characteristics of household electric vehicles, and construct the sample data required for training the neural network model; Step S2: Construct a SAED model for effectively identifying EV charging loads, and use the sample data set for model training and optimization to obtain a trained SAED model; Step S3: Deploy the trained SAED model to conduct NILM refined identification of user charging behavior locally and identify abnormal power consumption situations.

2. The non-invasive identification method of a household electric vehicle charging behavior according to claim 1, wherein, Step S1 includes the following steps: Install a smart meter at the power inlet of the household user's distribution box to collect the aggregated power consumption information including the load data of the household user's electric vehicle; Perform data cleaning and denoising preprocessing on the collected data; Combine the typical household electric vehicle charging load data sets published internationally and the collected load data of household user electric vehicles, and perform simulation amplification of the data volume according to the charging load characteristics of electric vehicles to form a sample data set.

3. The non-invasive identification method for the charging behavior of a household electric vehicle according to claim 2, characterized in that The sample data set includes the curve and waveform data of EV charging loads.

4. The non-intrusive identification method for a household electric vehicle charging behavior according to claim 1, characterized in that, The SAED model includes a CNN convolutional layer, a self-attention mechanism layer, a bidirectional GRU layer, and two fully connected network layers. The CNN convolutional layer is used to extract the feature information of the identification load signal; the self-attention mechanism layer is used to focus on the most important correlation relationships in the feature information; the bidirectional GRU layer is used to identify the pattern rules of the load sequence data; the two fully connected network layers are used to regress and generate the final load sequence structure.

5. The non-invasive identification method of a household electric vehicle charging behavior according to claim 4, characterized in that, In step S2, using the sample data set for model training and optimization to obtain a trained SAED model includes the following steps: Adopt the supervised learning method to train the SAED model using the sample data set; During the training process, use the simple, efficient and fast PyTorch as the support framework for model training and inference operations. The model training and tuning work are carried out on the kubeflow platform; Realize the distributed training of the model on the GPU cluster through the training-operator of the kubeflow platform, and realize the model parameter tuning through katib; Manage all metadata through metadata, and realize the automation of the deep learning workflow through ml-pipeline; After multiple rounds of training and verification, continuously optimize the model parameters and performance to enable the SAED model to accurately identify the charging loads of electric vehicles.

6. The non-invasive identification method of a household electric vehicle charging behavior according to claim 1, characterized in that, Step S3 includes the following steps: Deploy the trained SAED model to the edge gateway intelligent device in the ONNX format; Sample the smart meter at the power inlet of the user's distribution box from the second level to the minute level to monitor the user's power consumption in real time; When suspicious charging load information is detected, use the SAED model for analysis to judge whether there are abnormal charging / electricity consumption behaviors; If an abnormality is found, send an alarm message in time.

7. An non-invasive identification method for the charging behavior of a household electric vehicle according to any one of claims 1-6, characterized in that The abnormal electricity consumption situations include: Low-voltage users privately install charging piles for charging or directly pull wires for charging; Charging pile users privately add charging piles for use; Users connect their private electricity loads to charging piles to avoid the residential ladder electricity price.

8. A non-invasive identification device for household electric vehicle charging behavior, which uses a non-invasive identification method for household electric vehicle charging behavior described in any one of claims 1-7 to perform non-invasive identification of household electric vehicle charging behavior, and is characterized in that, Including: A sample data construction module, which is used to collect the aggregated electricity consumption information at the power inlet of the distribution box, analyze the charging behavior of household electric vehicles and the characteristics of charging loads, and construct the sample data required for neural network model training; A model training and optimization module, which is used to construct a SAED model for effectively identifying EV charging loads, and use the sample data set for model training and optimization to obtain a trained SAED model; An abnormal electricity consumption discrimination module, which is used to deploy the trained SAED model to conduct NILM fine-grained identification of users' charging behaviors locally and identify abnormal electricity consumption situations.

9. A non-intrusive identification system for the charging behavior of a household electric vehicle, which uses a non-intrusive identification method for the charging behavior of a household electric vehicle as described in any one of claims 1-7 to perform non-intrusive identification of the charging behavior of a household electric vehicle, characterized in that, It is characterized in that It includes a smart meter, an edge gateway intelligent device and a cloud server; the smart meter is used to sample the electricity load of suspicious users; the edge gateway intelligent device is deployed with a SAED model, which is used to analyze the electricity load of suspicious users, accurately identify the charging load information in the electricity load, and combine the reported charging pile installation information to accurately associate and analyze abnormal electricity consumption situations; the cloud server is responsible for computing task allocation and storing electricity load data and its identification results.

10. The non-intrusive identification system for the charging behavior of a household electric vehicle according to claim 9, characterized in that, The SAED model includes a CNN convolutional layer, a self-attention mechanism layer, a bidirectional GRU layer and two fully connected network layers. The CNN convolutional layer is used to extract and identify the characteristic information of the load signal; the self-attention mechanism layer is used to focus on the most important correlation relationship in the characteristic information; the bidirectional GRU layer is used to identify the pattern rules of the load sequence data; the two fully connected network layers are used to regress and generate the final load sequence structure.

Images