Edge-cloud collaborative non-intrusive load real-time identification method and system

Abstract

The invention relates to the technical field of intelligent power grids, and discloses an edge-cloud collaborative non-intrusive load real-time identification method and system, and the method comprises the steps: collecting the power data of a total load at a power utilization entrance of a user, carrying out the processing of an edge end, extracting the characteristics of a power utilization event, recognizing the type, decomposing the total load, and obtaining the power utilization information of all equipment, an identification result is transmitted to a cloud end for analysis, a model library is dynamically updated, and then updated model library information is fed back to an edge end optimization algorithm. According to the system, an improved dynamic time warping algorithm is adopted for event matching, load decomposition is realized by using an unsupervised clustering and semi-supervised learning model, and electrical characteristics of equipment are obtained through a high-frequency sampling technology. According to the scheme, a sensor does not need to be installed on each electric device, the deployment cost is greatly reduced, meanwhile, high-precision recognition is achieved by means of an edge-cloud collaborative architecture, the accuracy rate reaches 95% or above, and an economical and efficient technical solution is provided for intelligent power consumption management.

Description

Technical Field

[0001] The present invention relates to the technical field of smart grids, and particularly to a method and system for real-time non-intrusive load identification with edge-cloud collaboration. Background Art

[0002] Smart grid is an inevitable trend in the development of power grid technology. The wide application of technologies such as communication, computer, and automation in the power grid has greatly improved traditional power technologies. As an important part of the smart power consumption technology system, non-intrusive load monitoring (NILM) can deeply analyze the load components inside users, obtain detailed power consumption information, which is of great significance to both users and power companies.

[0003] Traditional load monitoring adopts an intrusive method, that is, sensors are installed on each electrical device of users and their usage conditions are recorded. The advantage of this method is that the monitoring data is accurate and reliable, while the disadvantages are high economic cost, great implementation difficulty, and low acceptance degree by users. In contrast, non-intrusive load monitoring only needs to collect data at the entrance of the user's total circuit, and through load decomposition technology, the total load information of users is decomposed into the information of each electrical device, so as to obtain the energy consumption of electrical devices and the power consumption rules of users.

[0004] The existing non-intrusive load monitoring technologies mainly face the following problems: First, with the increasing diversification of electrical devices and the multi-state continuous change of electrical loads, the self-adaptability and robustness of event detection algorithms are insufficient, and it is difficult to effectively identify the state changes of devices in complex power consumption scenarios; second, the effective extraction technology of load characteristics needs to rely on a large amount of data support of various types of devices, and it is difficult to obtain data; finally, most of the load identification technologies based on deep learning adopt a supervised learning mode, which requires a large amount of calibrated data for training, and the types of loads and scenarios involved are limited, and the actual application performance is not good.

[0005] In addition, in the existing technologies, the edge side and the cloud system are often separated, lacking an effective collaboration mechanism. The data collection and processing capabilities of the edge side are limited and it is difficult to cope with complex and changeable power consumption scenarios; while the cloud system has powerful data analysis capabilities, but due to the lack of a real-time data feedback mechanism, it is difficult to optimize the edge-side algorithms in a timely manner. This fragmented architecture limits the overall performance of the non-intrusive load identification system and cannot achieve high-precision and high-stability load identification and decomposition.

[0006] The complexity of the user terminal load and the diversity of user electricity consumption habits pose unprecedented challenges to load modeling and the realization of intelligent demand-side optimization. Existing non-intrusive load monitoring technologies have limitations in aspects such as data acquisition frequency, algorithm robustness, adaptability, and learning mode, and cannot effectively adapt to various complex situations in actual application scenarios, thus affecting the accuracy and reliability of load identification. Summary of the Invention

[0007] The purpose of the present invention is to provide a method and system for real-time non-intrusive load identification with edge-cloud collaboration, so as to solve the technical problems of high cost and low accuracy in monitoring the electricity consumption status of existing devices.

[0008] To achieve the above technical purpose, the present invention provides a method for real-time non-intrusive load identification with edge-cloud collaboration, including:

[0009] Collect total load power data;

[0010] Process the total load power data at the edge end, extract the characteristics of electricity consumption events, and identify the types of electricity consumption events;

[0011] Decompose the total load to obtain the electricity consumption information of each electrical device;

[0012] Transmit the identified electricity consumption events and decomposition results to the cloud;

[0013] Analyze the electricity consumption event data in the cloud and dynamically update the device model library;

[0014] Feed back the updated model library information to the edge end to optimize the event recognition and load decomposition algorithms at the edge end.

[0015] Optionally, the step of collecting total load power data includes: collecting total load power data at the user electricity consumption entrance;

[0016] The acquisition frequency of the total load power data is not less than 10 Hz, and the total load power data includes at least one of active power, reactive power, and apparent power.

[0017] Optionally, the decomposition of the total load further includes:

[0018] When the total load power data includes active power and reactive power, utilize the linear superposition characteristic of the active power and the reactive power. In the case where multiple devices are working simultaneously, subtract the power part of the identified background devices to obtain the actual power curve of the newly started devices.

[0019] Optionally, the processing of the total load power data at the edge end includes using an improved dynamic time warping algorithm to match the characteristics of electricity consumption events;

[0020] The improved dynamic time warping algorithm includes:

[0021] Using the FastDTW algorithm to reduce the computational time complexity;

[0022] Adopting a segmented weighting technique to improve the matching accuracy of local morphological features.

[0023] Optionally, the segmented weighting technique includes:

[0024] Dividing the time series into a front segment, a middle segment, and a back segment, and allocating each segment according to a preset ratio;

[0025] Applying the dynamic time warping algorithm to each segment for comparison;

[0026] Performing weighted processing on the comparison results of each segment.

[0027] Optionally, the decomposition of the total load includes using an unsupervised clustering and semi-supervised learning model for load decomposition;

[0028] The unsupervised clustering and semi-supervised learning model includes:

[0029] Clustering power consumption events;

[0030] Judging the type of electrical equipment according to the clustering results;

[0031] For devices with regular features, using generative adversarial network and variational autoencoder technologies to expand the dataset for training.

[0032] Optionally, the analysis of power consumption event data in the cloud includes:

[0033] Storing the typical waveforms of devices in the cloud database;

[0034] Clustering and feature analysis of multi-user power consumption data;

[0035] Updating the device model library based on the analysis results.

[0036] The present invention also provides an edge-cloud collaborative non-intrusive load real-time identification system, including:

[0037] A load data acquisition module, configured to acquire total load power data at the user power consumption entrance;

[0038] An edge computing module, configured to process the total load power data, extract power consumption event features and identify the types of power consumption events, and decompose the total load;

[0039] A communication module, configured to transmit the identified power consumption events and decomposition results to the cloud;

[0040] A cloud analysis module for analyzing power consumption event data and dynamically updating the device model library;

[0041] A feedback module for feeding back the updated model library information to the edge computing module to optimize the event recognition and load decomposition algorithms.

[0042] Optionally, the load data acquisition module includes a processor and a sampling chip, and its sampling frequency is not lower than 8 kHz.

[0043] Optionally, the edge computing module is used to store the events to be determined based on a sliding window, compare the events within the sliding window with the events in the model library, and when the similarity reaches a preset threshold, mark the current event as the corresponding model event.

[0044] Optionally, the cloud analysis module is further used to generate a user power consumption analysis report based on the identified power consumption events and decomposition results.

[0045] Compared with the prior art, the present invention has at least the following beneficial effects:

[0046] Through the edge-cloud collaborative non-intrusive load real-time identification method and system provided by the present invention, there is no need to install sensors on each power-consuming device separately. Only by collecting the total load data at the user power consumption entrance, the identification and energy consumption analysis of each power-consuming device can be realized, greatly reducing the device monitoring cost. At the same time, through the collaborative working mechanism of edge-side processing and cloud analysis, a closed-loop optimization system is formed, so that the identification accuracy continuously improves with the accumulation of data, solving the problem of low accuracy of traditional non-intrusive load monitoring technology in complex environments.

[0047] Furthermore, the present invention adopts an improved dynamic time warping algorithm and a segmented weighting technique, effectively improving the accuracy and efficiency of power consumption event feature matching; uses unsupervised clustering and semi-supervised learning models for load decomposition, reducing the dependence on calibration data and improving the adaptive ability of the system; utilizes the linear superposition characteristics of active power and reactive power to achieve accurate identification in the case of multiple devices working simultaneously, solving the problem of difficult identification in the scenario of multiple devices working in parallel in traditional methods. Through cloud big data analysis technology, clustering and feature analysis of multi-user data are carried out, and the device model library is continuously optimized, further improving the identification accuracy and system adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 It is a flowchart of the steps of a non-intrusive load real-time identification method with edge-cloud collaboration in an embodiment of the present invention;

[0049] Figure 2 It is a schematic diagram of the power curve of the refrigerator compressor starting in an embodiment of the present invention;

[0050] Figure 3 Schematic diagram of the power curve when the microwave oven starts in the embodiment of the present invention;

[0051] Figure 4 Curve of the first startup transient process of the refrigerator compressor in the embodiment of the present invention;

[0052] Figure 5 Curve of the second startup transient process of the refrigerator compressor in the embodiment of the present invention;

[0053] Figure 6 Schematic diagram of the misjudgment situation of the DTW algorithm in the prior art;

[0054] Figure 7 Power curve of the refrigerator compressor startup when there are other working devices in the embodiment of the present invention;

[0055] Figure 8 Module diagram of an edge-cloud collaborative non-intrusive load real-time identification system in the embodiment of the present invention. Detailed implementation manners

[0056] The following will describe in more detail an edge-cloud collaborative non-intrusive load real-time identification method and system of the present invention with reference to the accompanying drawings, in which the preferred embodiments of the present invention are shown. It should be understood that those skilled in the art can modify the present invention described herein while still achieving the advantageous effects of the present invention. Therefore, the following description should be understood as a broad guidance for those skilled in the art and not as a limitation to the present invention.

[0057] The present invention will be described more specifically by way of example with reference to the accompanying drawings in the following paragraphs. According to the following description, the advantages and features of the present invention will be clearer. It should be noted that the accompanying drawings are all in a very simplified form and use non-precise scales, only for the purpose of facilitating and clearly assisting in explaining the objectives of the embodiments of the present invention.

[0058] Embodiment 1

[0059] In this embodiment, an edge-cloud collaborative non-intrusive load real-time identification method is provided. Please refer to Figure 1 , including the following steps:

[0060] S1: Collect the total load power data.

[0061] In this embodiment, a load acquisition device is installed at the power inlet of the user's home or building to collect the total load power data of the user.

[0062] The acquisition frequency of the total load power data is not less than 10 Hz to ensure that the transient characteristics during the startup and shutdown processes of the equipment can be captured.

[0063] Furthermore, the total load power data includes at least one of active power, reactive power, and apparent power.

[0064] In a specific example, the acquisition of the total load power data can be achieved through an intelligent electricity meter or a dedicated power detector. The sampling frequency of the sampling chip can reach 8 kHz or higher to capture richer electrical characteristics. These high-frequency sampling data can be used to analyze the energy distribution of the device at different frequencies, enhancing the identification ability of different devices, especially when the powers of multiple devices are similar.

[0065] Each type of electrical equipment has unique electrical structures and characteristics: equipment such as electric heaters exhibits obvious resistive circuit characteristics; refrigerators exhibit compressor circuit characteristics; washing machines exhibit motor circuit characteristics. These electrical characteristics are particularly obvious during the transient process of equipment start and stop. At the same time, there are also differences in component selection and circuit design among similar equipment of different brands. Therefore, the electrical characteristics during the start and stop process of the equipment can be used to distinguish and identify different electricity consumption events. Please refer to Figures 2 - 3 , which respectively show the power curves when the refrigerator compressor and the microwave oven start. The data sampling frequency is 10 Hz ( Figures 2 - 3 The data sampling frequency in [] is 10 Hz). It can be seen from Figures 2 - 3 that the transient process of equipment start is usually within 5 s, and during this period, the refrigerator and the microwave oven have completely different active power and reactive power curves, thus enabling the distinction of electricity consumption events.

[0066] Furthermore, on the premise that the start and stop transient processes of different equipment have obvious distinguishability, the start and stop transient process of the same equipment also needs to be reproducible to achieve the identification of electricity consumption events. Please refer to Figures 4 - 5 , Figures 4 - 5 , which respectively show the power curves of two start processes of the refrigerator compressor. It can be seen from the figure that although the curves of the two start transient processes have slight differences in specific power values, they are highly similar in shape and can be aligned through local stretching and translation. Therefore, the reproducibility of the equipment start and stop process can be used to identify the repeated start and stop of the same equipment.

[0067] S2: Process the total load power data at the edge side, extract the electricity consumption event features, and identify the electricity consumption event type.

[0068] Process the collected total load power data at the edge side (such as intelligent terminal equipment installed on the user side), extract the electricity consumption event features, and identify the electricity consumption event type.

[0069] The electricity consumption event features mainly include the transient electrical characteristics during the start and stop process of the equipment, which have good distinguishability and reproducibility.

[0070] In this embodiment, an improved dynamic time warping algorithm is used to match the characteristics of power consumption events with respect to the traditional DTW algorithm. The improved dynamic time warping algorithm includes:

[0071] Using the FastDTW algorithm (Fast Dynamic Time Warping algorithm) to reduce the computational time complexity;

[0072] Adopting a segmented weighting technique to improve the matching accuracy of local morphological features.

[0073] Specifically, the segmented weighting technique includes:

[0074] Dividing the time series into a front segment, a middle segment, and a rear segment, and allocating each segment according to a preset ratio (for example, 20% for the front segment, 60% for the middle segment, and 20% for the rear segment; it can be dynamically adjusted according to the specific device type);

[0075] Applying the dynamic time warping algorithm to each segment for comparison;

[0076] Performing weighted processing on the comparison results of each segment.

[0077] This segmented weighting technique enhances the attention to local features and improves the matching accuracy while ensuring the computational efficiency.

[0078] Furthermore, during the event recognition process at the edge side, the events to be determined are stored in the form of a sliding window, and the size of the sliding window is the same as the longest event model in the model library.

[0079] The specific recognition steps are as follows:

[0080] Read the active / reactive / apparent power values at the current moment and update the sliding window;

[0081] Traverse the event model library. First, compare the events in the sliding window with the events in the model library one by one using the FastDTW algorithm to obtain the corresponding distance values, and then compare the distance values with the judgment threshold parameter. If the distance value is less than the judgment threshold, mark the current event as this type of model event;

[0082] Verify the types of events that have been marked and clear the misjudged marks;

[0083] Wait for a fixed time interval (usually 0.5s or 1s) and repeat the above steps.

[0084] Among them, the FastDTW algorithm (Fast Dynamic Time Warping algorithm) is an optimization and improvement of the traditional DTW algorithm (Dynamic Time Warping algorithm). By reducing the search space strategy, the computational time complexity is optimized from O(n2) to O(n), greatly improving the computational efficiency and ensuring the timeliness of long time series matching.

[0085] However, the FastDTW algorithm only optimizes the computational efficiency of the DTW algorithm and is thus also limited by the limitations of the DTW algorithm itself. In practical applications, the applicant has found that there are also certain misjudgment situations, and some of the misjudgments are caused by the excessive stretching and scaling of the DTW algorithm in seeking the global optimal solution.

[0086] As Figure 6 shown, the DTW algorithm may produce misjudgments in some cases. In the figure, the red curve represents the power curve of the template, and the blue curve represents the real-time power curve. The solid blue part represents the sliding window to be matched, and the dashed blue part represents the historical power values. According to the matching principle of the DTW algorithm, the first half of the points on the red curve are all mapped to the first point of the solid blue part (as shown by the gray connections in the figure), making the algorithm think that the solid blue part has a high similarity with the template. However, judging from the historical power values (the dashed blue part), this is obviously a misjudgment situation, which may be due to the pulse signal caused by the device power perturbation rather than a real device startup event.

[0087] To solve such misjudgment problems, the present invention adopts a segmented weighting technique, and the design of this technique draws on the ideas of two solutions: Solution 1 introduces the shapeDTW algorithm. Compared with the standard DTW algorithm that only focuses on the global optimal solution, shapeDTW pays more attention to the matching degree of the local morphological features of the time series. shapeDTW first extracts the local morphological features of each time series data point using the partial data before and after it, then replaces the original data point with this local morphological feature as the input dimension, and finally applies the standard DTW algorithm to complete the similarity comparison. It can be seen that shapeDTW can use local morphological features to largely avoid misjudgment problems caused by excessive stretching and scaling. However, due to the addition of the extraction of local morphological features, shapeDTW will significantly increase the computational time complexity (shapeDTW can still be accelerated using FastDTW).

[0088] Solution 2 draws on the idea of shapeDTW to a certain extent. First, the time series is segmented, for example, divided into three segments: front, middle, and back, with each segment accounting for 20%, 60%, and 20% respectively. When comparing two time series, first apply the DTW (FastDTW) algorithm to compare them segment by segment according to the front, middle, and back segments, and then weight the comparison results of each segment. Solution 2 improves the attention to local features while ensuring that the computational time complexity remains unchanged and does not increase the system overhead, so it is more conducive to practical applications.

[0089] The segmented weighting technique in the present invention mainly adopts the implementation idea of Solution 2, effectively improving the matching accuracy and reducing the misjudgment rate without increasing the computational complexity.

[0090] S3: Decompose the total load to obtain the power consumption information of each electrical device.

[0091] In step S3, after obtaining the power consumption event type, decompose the total load to obtain the power consumption information of each electrical device.

[0092] The step of decomposing the total load specifically includes: using an unsupervised clustering and semi-supervised learning model for load decomposition.

[0093] Most deep learning load analysis algorithms adopt a supervised learning method. Supervised learning requires a large amount of calibrated data to complete modeling, that is, it is necessary to collect the individual power consumption of each device. This to some extent violates the original intention of non-intrusive measurement technology, so it is difficult to complete deployment and implementation in practical applications.

[0094] The method provided in this embodiment adopts a different technical route and realizes load analysis based on unsupervised learning and semi-supervised learning. Different from the supervised learning scheme, the single electrical device can be directly identified from the total load curve and the load decomposition can be completed to obtain the power consumption information of the device; the unsupervised learning scheme first clusters the events, and then judges which electrical device according to the unified characteristics of this type of event, and then completes the load decomposition. This scheme decouples event recognition and classification from load decomposition. While avoiding the need for a large amount of training data, it also enhances the robustness of the algorithm to a certain extent, so that it can adapt to more types of devices.

[0095] Among them, the unsupervised clustering and semi-supervised learning model includes:

[0096] Cluster the power consumption events; judge the type of electrical device according to the clustering result;

[0097] For devices with regular characteristics, use generative adversarial network and variational autoencoder technologies to expand the dataset for training.

[0098] For example, for devices with obvious regular characteristics such as refrigerators, use the semi-supervised learning method, use technologies such as generative adversarial network (GAN) and variational autoencoder (VAE) to expand the dataset and then train, so as to improve the overall recognition and decomposition accuracy of this type of device.

[0099] In a specific example, when the total load power data includes active power and reactive power, the linear superposition characteristic of the active power and the reactive power can also be used to subtract the power part of the identified background device under the condition that multiple devices are working simultaneously to obtain the actual power curve of the newly started device.

[0100] On the premise that harmonic interference is limited, the active power and reactive power approximately have the characteristic of linear superposition, that is, the total power when multiple devices work simultaneously is equal to the sum of the powers when each device works alone. Therefore, in the presence of background equipment noise, the transient process of equipment startup and shutdown can still be used to complete event recognition.

[0101] Please refer to Figure 7 , Figure 7 shows the power curve of the refrigerator starting when the kettle is working. As can be seen from Figure 7 , although the power values vary greatly, the linear superposition characteristic can be used to subtract the power part of the background equipment, so as to obtain the actual power curve of the refrigerator starting. In a specific example, a certain moment can be used as a reference, and all values at subsequent moments are subtracted from the value corresponding to this moment, so as to obtain the power curve of a single device.

[0102] Therefore, in the presence of background equipment noise, the transient process of equipment startup and shutdown can still be used to complete event recognition. It should be noted that in extreme cases, there may be an overlap in the transient processes of two devices. For example, if two devices are turned on successively within 1 s, it will affect the recognition result to a certain extent. In actual tests, it is found that the probability of this kind of situation occurring is relatively low. Even if there is an overlap, there is still a certain ability to accurately identify.

[0103] S4: Transmit the identified electricity consumption events and decomposition results to the cloud.

[0104] Specifically, transmit the identified electricity consumption events and decomposition results to the cloud through a communication network.

[0105] In step S4, the transmission can be achieved through a wired network (such as Ethernet) or a wireless network (such as Wi-Fi, 4G / 5G, etc.), and it can be flexibly selected according to the actual application scenario.

[0106] S5: Analyze the electricity consumption event data in the cloud and dynamically update the device model library.

[0107] In step S5, the specific analysis of the electricity consumption event data in the cloud includes:

[0108] Store the typical waveforms of the devices in the cloud database;

[0109] Cluster and perform feature analysis on the electricity consumption data of multiple users;

[0110] Update the device model library based on the analysis results.

[0111] S6: Feedback the updated model library information to the edge side to optimize the event recognition and load decomposition algorithms on the edge side

[0112] In step S6, the cloud feeds back the updated model library information to the edge side to optimize the event recognition and load decomposition algorithms on the edge side. This edge-cloud collaboration mechanism forms a closed loop, continuously improving the recognition accuracy and adaptability of the system.

[0113] Specifically, the cloud can push targeted model parameters to the edge side according to the characteristics and usage habits of different users, such as event matching thresholds optimized for specific users, segment weighting ratios, etc., so as to improve the recognition accuracy in specific user environments.

[0114] In addition, the cloud can also generate user electricity consumption analysis reports and provide energy-saving suggestions to users based on the recognized electricity consumption events and decomposition results, further enhancing the application value of the system.

[0115] Furthermore, no more than 10 typical electrical devices are selected, and a total of 9 electricity consumption behaviors are included for verification. The device types include: refrigerator, microwave oven, washing machine, rice cooker, hot water kettle, oven, air conditioner, coffee machine, etc. The start and stop of the devices are all for daily use, without any manual intervention. The specific electrical device parameters and startup information are shown in Table 1 below.

[0116] Table 1

[0117] Device event Maximum instantaneous power (w) Start-up times Refrigerator 1100 412 Water dispenser (heating) 950 43 Water dispenser (cooling) 90 12 Microwave oven 1100 22 Electric door 40 3782 Coffee machine 1200 89 Washing machine 200 11 Rice cooker 900 8 Oven 2000 8

[0118] In the experimental verification, event recognition verification was carried out, including single-device event verification and verification of events with more than 3 devices turned on simultaneously. The verification period was one week under typical usage environments.

[0119] The verification results are shown in Table 2. The recognition accuracy of all events is relatively high, and good recognition effects can still be maintained when there are at most 4 devices working simultaneously during the operation process.

[0120] Specifically, no more than 10 typical electrical devices were selected for verification, including refrigerator, water dispenser (heating and cooling), microwave oven, electric door, coffee machine, washing machine, rice cooker, and oven, etc. According to the analysis of the verification results, except for the rice cooker, the precision rate of event recognition has reached more than 95%, while the recall rate is slightly lower than the precision rate as a whole, with an average of more than 90%. This is mainly because relatively strict judgment threshold parameters are set, which reduces the probability of misjudgment while increasing the probability of missed judgment.

[0121] Since refrigerators, water dispenser refrigeration, microwave ovens, electric doors, washing machines, and ovens all have transient power curves with high distinguishability, a high recognition rate can be achieved. Water dispenser heating, coffee machines, and rice cookers are all resistive heating devices from the perspective of circuit design, and their powers are relatively close, so there are certain misjudgments and omissions between them, resulting in a relatively low overall recognition rate. The recall rate of microwave ovens is lower than the average level, mainly because after the microwave oven runs continuously for a long time (more than 10 minutes), the internal temperature of the device is relatively high, causing a certain degree of change in the starting transient power curve, resulting in missed judgments. The relatively low recall rate of rice cookers is due to the interference of devices such as water dispensers and coffee machines on the one hand, and the low sample size and statistical errors on the other hand.

[0122] In summary, through test verification, the average recognition accuracy of 9 electrical behaviors reaches 98.1%. The recognition accuracy of most devices reaches more than 95%, and the overall recall rate reaches more than 90%. Even when multiple devices work simultaneously, the electrical usage of each device can be effectively identified.

[0123] Table 2

[0124]

[0125]

[0126] In summary, the edge-cloud collaborative non-intrusive load real-time recognition method provided in this embodiment realizes the accurate recognition and analysis of user electrical devices by collecting total load data at the user's power consumption entrance and combining edge computing and cloud analysis. Compared with traditional intrusive monitoring methods, this method has the following beneficial effects:

[0127] It reduces the deployment cost, eliminates the need to install separate sensors on each device, and improves user acceptance; it improves the data processing efficiency and real-time performance through edge computing, reducing the network transmission pressure; it uses an improved dynamic time warping algorithm to improve the accuracy of event recognition, especially for devices with short transient processes and low-power devices; it reduces the dependence on calibration data by using unsupervised clustering and semi-supervised learning models, improving the adaptability and scalability of the system; it forms a closed-loop optimization mechanism through cloud analysis and model updates to continuously improve the system performance.

[0128] Embodiment 2

[0129] The embodiment of the present invention provides an edge-cloud collaborative non-intrusive load real-time recognition system. Please refer to Figure 8 , which includes: a load data acquisition module, an edge computing module, a communication module, a cloud analysis module, and a feedback module.

[0130] Among them, the load data acquisition module is used to collect the total load power data at the user's power consumption entrance.

[0131] The load data acquisition module includes a processor and a sampling chip, and its sampling frequency is not less than 8 kHz.

[0132] In this embodiment, the load data acquisition module uses an ARM Cortex-M3 processor, and the sampling chip uses the ADE7878 / 7880 series metering chips of ADI Corporation, with an ADC sampling frequency of 8 kHz. This module supports high-frequency data (above 10 kHz) sampling, so as to analyze the energy distribution of equipment at different frequencies and enhance the recognition of different equipment.

[0133] Among them, the edge computing module is connected to the load data acquisition module and is used to process the total load power data, extract the characteristics of power consumption events and identify the types of power consumption events, and decompose the total load.

[0134] In a specific example, the edge computing module is used to store the event to be determined based on a sliding window, compare the events within the sliding window with the events in the model library, and when the similarity reaches the preset threshold, mark the current event as the corresponding model event.

[0135] Among them, the communication module is connected to the edge computing module and is used to transmit the identified power consumption events and decomposition results to the cloud. The communication module can support multiple communication protocols, such as MQTT, HTTP, etc., to ensure the reliability and security of data transmission.

[0136] Among them, the cloud analysis module is communicatively connected to the communication module and is used to analyze the power consumption event data and dynamically update the device model library.

[0137] The cloud analysis module is also used to generate a user power consumption analysis report and provide suggestions to the user based on the identified power consumption events and decomposition results.

[0138] The cloud analysis module is deployed on a cloud server and has powerful computing and storage capabilities, capable of processing the power consumption data of a large number of users.

[0139] In a specific example, this module is used to discover the commonalities of power consumption patterns and device characteristics among different users, so as to continuously optimize the device model library.

[0140] Among them, the feedback module is connected to the cloud analysis module and the edge computing module and is used to feedback the updated model library information to the edge computing module to optimize the event recognition and load decomposition algorithms.

[0141] The feedback module is responsible for timely transmitting the analysis results and optimization parameters from the cloud to the edge computing module, forming a closed-loop optimization mechanism. At the same time, the feedback module can also push the electricity consumption analysis report and energy-saving suggestions to the user terminal to improve the user experience.

[0142] The edge-cloud collaborative non-intrusive load real-time identification system provided in this embodiment combines the advantages of edge computing and cloud computing, and realizes efficient and accurate identification and analysis of electrical equipment. The system has a clear structure, and the responsibilities of each module are clear, which is convenient for implementation and maintenance; the high-frequency sampling ability of the load data acquisition module provides a data basis for accurate identification; the edge computing module reduces the burden on the cloud and improves the system response speed and real-time performance; the cloud analysis module has powerful data processing capabilities, can discover rules from massive data, and continuously optimize the model; the feedback mechanism forms a closed loop, enabling the system to continuously improve itself and improve the identification accuracy.

[0143] In summary, the edge-cloud collaborative non-intrusive load real-time identification method and system provided by the present invention collect the total load data at the user's electricity consumption entrance, and combine edge computing and cloud analysis to realize the real-time identification and analysis of the user's electrical equipment, which can meet various load monitoring requirements in home, commercial and industrial environments, and provide strong support for energy management, electrical fault diagnosis and electricity consumption behavior analysis.

[0144] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these changes and modifications.

Claims

1. A non-intrusive load real-time identification method for edge-cloud collaboration, characterized in that Including: Collecting total load power data; Processing the total load power data at the edge side, extracting power consumption event features and identifying power consumption event types; Decomposing the total load to obtain power consumption information of each electrical device; Transmitting the identified power consumption events and decomposition results to the cloud; Analyzing the power consumption event data in the cloud and dynamically updating the device model library; Feeding back the updated model library information to the edge side to optimize the event recognition and load decomposition algorithms at the edge side.

2. The edge-cloud collaborative non-intrusive load real-time identification method according to claim 1, characterized in that The step of collecting total load power data includes: collecting total load power data at the user power consumption entrance; The collection frequency of the total load power data is not less than 10 Hz, and the total load power data includes at least one of active power, reactive power and apparent power.

3. The edge-cloud collaborative non-intrusive load real-time identification method according to claim 2, wherein The decomposition of the total load further includes: When the total load power data includes active power and reactive power, using the linear superposition characteristic of the active power and the reactive power, subtracting the power part of the identified background devices in the case of multiple devices working simultaneously to obtain the actual power curve of the newly started devices.

4. The edge-cloud collaborative non-intrusive load real-time identification method according to claim 1, characterized in that, The processing of the total load power data at the edge side includes matching the power consumption event features by using an improved dynamic time warping algorithm; The improved dynamic time warping algorithm includes: Using the FastDTW algorithm to reduce the computational time complexity; Adopting a segmented weighting technique to improve the matching accuracy of local morphological features.

5. The edge-cloud collaborative non-intrusive load real-time identification method according to claim 4, wherein, The segmented weighting technique includes: Dividing the time series into a front segment, a middle segment and a back segment, and allocating each segment according to a preset ratio; Applying the dynamic time warping algorithm to each segment for comparison; Performing weighted processing on the comparison results of each segment.

6. The edge-cloud collaborative non-intrusive load real-time identification method according to claim 1, wherein, The decomposition of the total load includes adopting an unsupervised clustering and semi-supervised learning model for load decomposition; The unsupervised clustering and semi-supervised learning model includes: Clustering power consumption events; Judging the types of electrical devices according to the clustering results; For devices with regular features, using generative adversarial network and variational autoencoder technologies to expand the dataset for training.

7. The edge-cloud collaborative non-intrusive load real-time identification method according to claim 1, characterized in that The analysis of the power consumption event data in the cloud includes: Storing the typical waveforms of devices in the cloud database; Clustering and feature analyzing the power consumption data of multiple users; Updating the device model library based on the analysis results.

8. A non-intrusive load real-time identification system with edge-cloud collaboration, characterized in that Including: A load data collection module for collecting total load power data at the user power consumption entrance; An edge computing module for processing the total load power data, extracting power consumption event features and identifying power consumption event types, and decomposing the total load; A communication module for transmitting the identified power consumption events and decomposition results to the cloud; A cloud analysis module for analyzing the power consumption event data and dynamically updating the device model library; A feedback module for feeding back the updated model library information to the edge computing module to optimize the event recognition and load decomposition algorithms.

9. The edge-cloud collaborative non-intrusive load real-time identification system according to claim 8, wherein The load data collection module includes a processor and a sampling chip, and its sampling frequency is not less than 8 kHz.

10. The edge-cloud collaborative non-intrusive load real-time identification system according to claim 8, characterized in that, The edge computing module is used to store the events to be determined based on a sliding window, compare the events within the sliding window with the events in the model library, and when the similarity reaches a preset threshold, mark the current event as the corresponding model event.

11. The edge-cloud collaborative non-intrusive load real-time identification system according to claim 8, wherein, The cloud analysis module is also used to generate a user power consumption analysis report based on the identified power consumption events and decomposition results.

Images