Distributed photovoltaic system abnormal state detection-oriented sustainable edge computing framework construction method and system

By changing the percentage of power amplitude in a distributed photovoltaic system, evaluating machine learning model performance, and integrating the material layer, edge layer, fog layer and cloud layer, a sustainable edge computing framework is built to solve large-scale distributed The problem of abnormal state detection of photovoltaic systems in smart cities or power grids improves detection accuracy and efficiency.

CN119025849APending Publication Date: 2024-11-26STATE GRID SHANDONG ELECTRIC POWER CO
View PDF 0 Cites 3 Cited by

Patent Information

Application Number
CN202411048844.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-26

AI Technical Summary

Technical Problem

The existing technology is difficult to deploy an abnormal state detection edge computing framework for distributed photovoltaic systems on a large scale in smart cities or power grids, and most of the studies are only applicable to small-scale systems, which fail to effectively solve the resource limitation of distributed photovoltaic systems. Computing requirements on edge devices.

Method used

Create synthetic datasets by changing the percentage of power amplitude, evaluate the abnormal state detection performance of machine learning models, and determine the optimal model, integration layer, based on performance metrics such as F1 score, training time, prediction time, RAM usage, and model size, Edge layer, fog layer and cloud layer to build a sustainable edge computing framework.

Benefits of technology

The distributed photovoltaic system abnormal state detection edge computing architecture is implemented for large-scale deployment in smart cities or power grids, which improves detection accuracy and efficiency, and is suitable for resource-constrained edge devices.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119025849A_ABST
    Figure CN119025849A_ABST
Patent Text Reader

Abstract

The invention discloses a sustainable edge computing framework construction method and system for abnormal state detection of a distributed photovoltaic system, which is applied to the technical field of power system automation, and comprises the following steps: changing a power amplitude percentage to create a synthetic data set as different abnormal scenes; evaluating the abnormal state detection performance of the model on the central server under different abnormal scenes by using the training prediction time and the F1 score under the macroscopic average technology; evaluating the abnormal state detection performance of the model on the edge device under different abnormal scenes by using the prediction time, the RAM use and the model size; based on abnormal state detection performance of the model, an optimal machine learning model is determined, edge nodes, an integrated object layer, an edge layer, a fog layer and a cloud layer are embedded, and an edge calculation framework is constructed. According to the method, an abnormal state detection edge calculation framework deployed in a large scale in a smart city or a power grid is realized, and the edge calculation precision of abnormal state detection of the distributed photovoltaic system is effectively improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of power system automation technology, and in particular to a sustainable edge computing framework construction method and system for abnormal state detection of distributed photovoltaic systems. Background Art

[0002] In recent years, the rapid development of technology has brought about a digital revolution in urban areas, especially in the context of growing renewable distributed energy systems. New computing frameworks are being widely studied to solve current problems in monitoring and fault detection. Currently, with the increasing use of sensor networks, there has been a significant surge in the generation and real-time processing of large amounts of data, but sensor data may show anomalies due to sensor failures or real anomalies in the detected environment. Therefore, it is crucial to detect anomalies in a timely manner and effectively mitigate the occurrence of catastrophic situations.

[0003] In addition, due to the continuous development of digitalization, a large amount of data is transmitted to centralized servers, which puts tremendous pressure on the network. The diversity, speed, accuracy and variability of the data volume make this process more challenging. In addition, the data circulating in the network needs to be cleaned, analyzed and processed to provide effective information and useful knowledge. At the same time, the data from thousands of connected devices must be processed quickly, otherwise the network will experience severe delays and crashes.

[0004] For the reasons mentioned above, the concept of edge computing has emerged as a new technology that can improve the processing capabilities of big data platforms. This is particularly important in the context of smart cities, because solar-based distributed energy generation is one of its basic characteristics. These distributed units need to be continuously and in real time monitored to ensure their performance and the stability of the entire power grid. Similar to any physical system, solar power plants may also experience anomalies, but manual intervention is not feasible, especially for large data sets and their distributed behavior. For this reason, machine learning has been widely developed and applied. When anomalies occur, they can be discovered in time, allowing operators to intervene quickly and make necessary adjustments or repairs, thereby greatly reducing performance losses.

[0005] However, deploying machine learning algorithms on resource-constrained edge devices is not a simple process. These devices are usually used for tasks such as data collection and preprocessing, and are not suitable for computationally intensive tasks required for machine learning. Therefore, various optimizations must be performed to ensure that the algorithm can run effectively on these devices. Many scholars have studied edge computing technologies within the scope of smart grids. These related works reveal the importance of combining anomaly detection with edge computing capabilities in smart grid applications and verify its feasibility in solar photovoltaic state monitoring applications. Some studies have discussed photovoltaic fault detection in smart grids using a wide range of machine learning techniques. However, previous work mainly focused on using detection, classification, or a hybrid combination of them to create a real-time monitoring framework. However, there is a lack of extensive discussion on the edge computing framework for anomaly detection in photovoltaic systems, and most of them are only applied to small-scale systems, and do not provide an architectural solution for large-scale deployment in smart cities or power grids.

[0006] Therefore, how to provide a sustainable edge computing framework construction method and system for abnormal state detection of distributed photovoltaic systems that can be deployed on a large scale in smart cities or power grids is an urgent problem that needs to be solved by technical personnel in this field. Summary of the invention

[0007] In view of this, the present invention proposes a method and system for constructing a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems.

[0008] In order to achieve the above object, the present invention adopts the following technical solution:

[0009] A method for constructing a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems, comprising:

[0010] Step 1: Create a preset number of synthetic data sets by changing the percentage of power amplitude as different abnormal scenarios to evaluate the abnormal state detection performance of the machine learning model;

[0011] Step 2: Use the training and prediction time and the F1 score under the macro-average technique as performance evaluation indicators to evaluate the abnormal state detection performance of the machine learning model on the central server under different abnormal scenarios; use the prediction time, RAM usage, and model size as performance evaluation indicators to evaluate the abnormal state detection performance of the machine learning model on the edge device under different abnormal scenarios;

[0012] Step 3: Based on the abnormal state detection performance of the machine learning model on the central server and edge devices, determine the optimal machine learning model and embed it into the edge node, integrate the object layer, edge layer, fog layer and cloud layer, and build a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems.

[0013] Optionally, in step 1, a systematic approach is used to collect data based on a comprehensive hierarchical framework consisting of three stages: perception, connection, and application, and a structured framework is used to clean the data.

[0014] Optional, comprehensive hierarchical framework, as follows:

[0015] Perception layer: It consists of measurement nodes deployed on site. The measurement nodes are responsible for capturing information related to the PV system, including solar irradiance, ambient temperature, string current, and string voltage.

[0016] Connection layer: responsible for facilitating data transmission, consisting of wired Ethernet and switching equipment;

[0017] Application layer: Contains the analysis and processing tasks that need to be performed on the collected data, including but not limited to prediction, anomaly detection, performance analysis, and condition monitoring.

[0018] Optionally, data cleaning, specifically: ensuring data synchronization, deleting duplicate values, night values, and current and voltage values ​​that are not within the preset current and voltage ranges, and using IIR filters and PR filters to remove R 2 and the number of days with PR less than 0.8.

[0019] Optionally, in step 1, the percentage of changing the power amplitude ranges from 20% to 90%.

[0020] Optionally, in step 1, the machine learning model is screened and determined by using the PyOd library with its default parameters based on the accuracy of the algorithm on the same abnormal data set, including: supervised learning models represented by multilayer perceptrons and unsupervised learning models represented by local outlier factors, K-nearest neighbor algorithms, and clustering-based local outlier factors.

[0021] Optionally, in step 2, the training time is used as a performance evaluation indicator to evaluate the abnormal state detection performance of the machine learning model on the central server under different abnormal scenarios, specifically:

[0022] For supervised learning models, training time is used as the performance evaluation indicator to evaluate the abnormal state detection performance of machine learning models on the central server under different abnormal scenarios;

[0023] For the unsupervised learning model, the fitting time is used as the performance evaluation indicator to evaluate the abnormal state detection performance of the machine learning model on the central server under different abnormal scenarios.

[0024] Optionally, in step 2, prediction time, RAM usage, and model size are used as performance evaluation indicators to evaluate the abnormal state detection performance of the machine learning model on the edge device under different abnormal scenarios, specifically:

[0025] Edge computing is performed by deploying a Raspberry Pi on the edge device, and the prediction time of the machine learning model on the edge device is calculated using the time library in Python. The RAM usage of the machine learning model on the edge device is calculated using the memory analysis library in Python. The model size of the machine learning model on the edge device is calculated using the pickle library in Python to evaluate the abnormal state detection performance of the machine learning model on the edge device under different abnormal scenarios.

[0026] Optionally, in step 3, the sustainable edge computing framework is as follows:

[0027] Physical layer: represents the equipment that generates raw data at the source end, including: photovoltaic data collectors that measure DC voltage and current in the photovoltaic system, weather sensors that measure solar irradiance and temperature, photovoltaic inverters that convert DC power to AC power, and drones that regularly perform thermal inspections on photovoltaic modules;

[0028] Edge layer: Located between the physical layer and the fog layer, it represents the edge nodes that process and analyze incoming data streams before sending the results to the cloud. It consists of a Raspberry Pi cluster deployed in each distributed PV system.

[0029] Fog layer: Located between the edge layer and the cloud layer, it consists of servers that centralize information from a set of edge nodes placed in the house and perform additional computation and data aggregation from the same location;

[0030] Cloud layer: The final layer for deploying analysis and decision support systems, responsible for collecting decisions, alerts, and status reports received from local nodes.

[0031] The present invention also provides a sustainable edge computing framework construction system for abnormal state detection of distributed photovoltaic systems using a sustainable edge computing framework construction method for abnormal state detection of distributed photovoltaic systems, characterized in that it includes:

[0032] Abnormal scenario building module: used to create a preset number of synthetic data sets by changing the percentage of power amplitude as different abnormal scenarios for evaluating the abnormal state detection performance of the machine learning model;

[0033] Abnormal state detection performance evaluation module: It is used to evaluate the abnormal state detection performance of machine learning models on central servers under different abnormal scenarios using training and prediction time and F1 score under macro-average technology as performance evaluation indicators; it is used to evaluate the abnormal state detection performance of machine learning models on edge devices under different abnormal scenarios using prediction time, RAM usage and model size as performance evaluation indicators;

[0034] Sustainable edge computing framework building module: It is used to determine the optimal machine learning model based on the abnormal state detection performance of the machine learning model on the central server and edge devices, and embed it into the edge node, integrate the object layer, edge layer, fog layer and cloud layer, and build a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems.

[0035] Through the above technical solutions, it can be known that compared with the prior art, the present invention proposes a sustainable edge computing framework construction method and system for abnormal state detection of distributed photovoltaic systems. By adopting a systematic method to collect data based on a comprehensive hierarchical framework consisting of three stages: perception, connection, and application, and using the training and prediction time of the central server and the F1 score under the macro-average technology and the prediction time, RAM usage, and model size of the edge device as performance evaluation indicators, the performance of the machine learning model on the central server and edge device is evaluated under different abnormal scenarios constructed by changing the power amplitude percentage, and the optimal model embedded in the edge node is obtained. The material layer, edge layer, fog layer, and cloud layer are integrated to construct a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems, which not only realizes the abnormal state detection edge computing architecture solution that can be deployed on a large scale in smart cities or power grids, but also effectively improves the edge computing accuracy of abnormal state detection of distributed photovoltaic systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0037] Figure 1 It is a schematic diagram of the method flow of the present invention.

[0038] Figure 2 It is a schematic diagram of different abnormal scenarios obtained by changing the percentage of power amplitude according to the present invention.

[0039] Figure 3 Schematic diagram of the f1 score value of each model on the central server under different abnormal scenarios of the present invention.

[0040] Figure 4 Schematic diagram of the training and prediction time required for each model on the central server under different abnormal scenarios of the present invention.

[0041] Figure 5 This is a schematic diagram of the prediction time, RAM usage, and model size of each model on the edge device under different abnormal scenarios of the present invention. DETAILED DESCRIPTION

[0042] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0043] Embodiment 1:

[0044] Embodiment 1 of the present invention discloses a method for constructing a sustainable edge computing framework for abnormal state detection of a distributed photovoltaic system, such as Figure 1 As shown, including:

[0045] Step 1: Create a preset number of synthetic datasets by changing the percentage of power amplitude as different abnormal scenarios to evaluate the abnormal state detection performance of the machine learning model.

[0046] The research object is a park equipped with many solar roofs. The photovoltaic array is interconnected with the grid-connected inverter in each house to convert direct current (DC) into alternating current (AC). The DC power is monitored by using a photovoltaic data logger with a data collection frequency of 15 minutes. The data logger uses pre-sampling technology to collect data, and the monitoring system retrieves the average value every 15 minutes.

[0047] A systematic approach was adopted to collect data based on a comprehensive hierarchical framework consisting of three stages: perception, connection, and application, and a structured framework was used for data cleaning.

[0048] Smart grids lack standardized communication and data architectures. The present invention is based on a comprehensive layered framework consisting of three stages: perception, connection, and application, which is designed to support the development of IoT smart grids. In this particular case, the perception layer consists of measurement nodes deployed in the field, which are responsible for capturing information related to the photovoltaic system, including solar irradiance, ambient temperature, string current, and string voltage. The connection layer is responsible for facilitating data transmission. Between the perception layer and the monitoring system, this layer consists of wired Ethernet and switching equipment as a means to establish and maintain connections. The application layer contains various analysis and processing tasks that need to be performed on the collected data. These tasks include but are not limited to prediction, anomaly detection, performance analysis, and condition monitoring. In addition, there is a security layer, which, although not involved in this invention, is also an important consideration in the design of smart grids.

[0049] The data cleaning process includes a set of filters to remove erroneous and invalid records, as well as to calculate performance indicators such as the daily performance ratio of the studied PV system. Data cleaning, specifically: ensuring data synchronization, removing duplicate values, night values, and current and voltage values ​​that are not within the preset current and voltage ranges (0<current<1.5*Isc, 0<voltage<1.2*Ns*Voc, Isc is the short-circuit current, Voc is the open-circuit voltage in the data table, and Ns is the number of modules in each series path), and using IIR filters and PR filters to remove R 2 and the number of days with PR less than 0.8 (because irradiance should be highly correlated with output power).

[0050] The range of the percentage of the power amplitude change is 20% to 90%. In the embodiment of the present invention, the power amplitude is changed to 90%, 70%, 50%, and 20%, respectively. Figure 2 As shown, four synthetic datasets are created, representing four different abnormal scenarios.

[0051] The machine learning model uses the PyOd library with its default parameters and is screened and determined based on the accuracy of the algorithm on the same abnormal data set, including: supervised learning models represented by multilayer perceptrons and unsupervised learning models represented by local outlier factors, K-nearest neighbor algorithms, and local outlier factors based on clustering.

[0052] Local Outlier Factor (LOF): Local Outlier Factor is an unsupervised technique that relies on neighborhood density to detect anomalies. It is calculated by dividing the density of the point by the density of its k nearest neighbors. Since neighborhood density is the basis of the method, LOF works well on imbalanced datasets. For each point, the algorithm calculates a parameter called k-distance that defines the distance between the point of interest and its farthest neighbor and then calculates the local outlier factor. Points with low density due to high outlier coefficient are considered outliers.

[0053] K-Nearest Neighbors (KNN): KNN is a supervised machine learning algorithm. However, since anomalies are considered rare events that are difficult to label, the model can also be turned into an unsupervised learning technique in anomaly detection use cases. KNN is a proximity-based clustering technique, which means that it relies on the distance of a data point to its kth neighbor to determine whether it is an anomaly. The farther the data point is from the nearest cluster, the higher the probability that it is an anomaly. In practice, it is necessary to find the optimal value for the hyperparameter k that represents the number of neighbors in the model. The KNN algorithm has two main parameters: the distance metric and the number of nearest neighbors k used in the calculation.

[0054] Cluster-based Local Outlier Factor (CBLOF): Cluster-based Local Outlier Factor is an unsupervised method that belongs to the family of clustering-based algorithms and aims to address some of the shortcomings of existing clustering methods. It has been effectively demonstrated to be competitive with other clustering techniques and single Local Outlier Factor (LOF) methods. To discover outliers, CBLOF uses the concepts of both size and distance. In fact, a data point can belong to a cluster but still be considered an outlier due to the size of this cluster.

[0055] Multilayer Perceptron (MLP): Multilayer Perceptron is a feedforward neural network inspired by the function of the human brain. It consists of an input layer, a hidden layer, and an output layer. The input layer consists of feature vectors, and the hidden layer consists of multiple neurons, located between the input layer and the output layer, which is responsible for delivering the classification results. Multilayer Perceptron operates in a similar way to a feedforward network. Data is transferred from the input layer to the output layer in a forward manner. Since MLPs are designed to approximate any continuous function, they can be used to solve nonlinear problems.

[0056] Step 2: Use the training and prediction time and the F1 score under the macro-average technique as performance evaluation indicators to evaluate the abnormal state detection performance of the machine learning model on the central server under different abnormal scenarios; use the prediction time, RAM usage, and model size as performance evaluation indicators to evaluate the abnormal state detection performance of the machine learning model on the edge device under different abnormal scenarios.

[0057] To evaluate performance on a central server:

[0058] Performance evaluation is the most important part of machine learning experiments. Depending on the problem you are solving, different metrics can be used to evaluate how good the model is. In classification problems, the most commonly used metrics are precision, recall, and f1 score (F1Score). Precision is a measure of how accurately the model classifies the positive class. Accuracy is highest when there are more correct positive classifications than incorrect positive classifications. Recall is more concerned with the sensitivity of the classification, and the best recall is achieved when all positive instances are correctly labeled as positive. The f1 score is an average metric between precision and recall, which gives a better conclusion about the performance of the model. An ideal classification task will get a score of 100% in all three metrics, so in a binary classification task, the score is judged by how close it is to the 100% ideal case, which means there are 2 classes: a normal class labeled 0 and an anomaly class labeled 1, so 4 different cases are encountered, the first two cases are called "True Positive (TP)" and "True Negative (TN)", which means that the predicted and ground truth labels are equal. This distinction is related to the type of classification. In the "TP" case, the target is the positive class, while the "TN" case represents the negative class. Typically, in anomaly classification tasks, the positive class (or class of interest) is the class with abnormal instances, the negative class represents the class with normal data points, and the other cases are called "false positive (FP)" and "false negative (FN)", which refer to the misclassification of normal instances as abnormal instances (FP) and abnormal instances as normal instances (FN). Therefore, it is important to specify the target class when calculating each metric. To calculate these metrics, the present invention can use classification reports to make a concise visual representation of the results, or use the precision-recall-f score support function in sklearn. In order to better understand the pros and cons of the model, the present invention uses a macro-average technique that considers two classes without weighing them. This is very important when the data set is unbalanced. Choosing a weighted average will always return a score close to 1 because the normal class has a greater weight than the abnormal class. In addition, all models of the experiment correctly classified the normal class. Therefore, relying on weighted average will lead to biased and incorrect interpretations. In order to obtain a correct interpretation, the present invention applies a macro-averaging technique to the f1-score, considering that this metric gives a more balanced indicator compared to precision and recall.

[0059] The four algorithms of LOF, KNN, CBLOF, and MLP were trained and tested on the central server, and their performance was evaluated. The evaluation indicators only included f1 score, training time (for supervised models), fitting time (for unsupervised models), and prediction time. Figure 3As shown, the f1 score values ​​of each model in different scenarios are plotted. The results show that ANN obtains the highest f1 score in all scenarios, followed by KNN. In the first and second scenarios, all models obtain high fl scores (100% in the first scenario and more than 95% in the second scenario, and it is noted that the difference in ANN scores is small). However, in the other two scenarios, the advantage of ANN begins to become more and more obvious, especially in the last scenario, where ANN reaches 79% in the f1 score indicator, while all other models do not exceed 59%.

[0060] In order to evaluate the computational footprint of the model, the training and prediction time required for each scenario was calculated. Figure 4 As shown in the figure, CBLOF takes the longest time in the training phase, followed by ANN, while the training time of LOF and KNN is almost not noticeable in the figure. In the prediction phase, KNN takes the longest time to predict new samples, followed by CBLOF, ANN and LOF. In summary, the model with the shortest training and prediction time is LOF, followed by KNN, ANN and CBLOF.

[0061] Edge device performance evaluation:

[0062] Since the goal of this invention is to implement the model on an embedded configuration, it is also important to consider other metrics, mainly prediction time, RAM usage, and model size. Prediction time refers to the time the model spends on predictions on the test set, which is calculated using the time library in Python. Model size refers to the size of the space used to store the pre-trained model, using the pickle library also in Python, and RAM usage is calculated using the memory profiler library in Python.

[0063] The present invention performs edge computing based on Raspberry Pi, such as Figure 5 As shown in Figure 1, the performance of machine learning models on edge devices is described, and prediction time, RAM usage, and model size are reported. It can be seen that on the Raspberry Pi, CBLOF takes the longest time to predict, followed by KNN, LOF, and ANN. CBLOF is a model with high memory requirements, while KNN, ANN, and LOF have a smaller memory footprint.

[0064] Step 3: Based on the abnormal state detection performance of the machine learning model on the central server and edge devices, determine the optimal machine learning model and embed it into the edge node, integrate the object layer, edge layer, fog layer and cloud layer, and build a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems.

[0065] From the results of the abnormal state detection performance of the machine learning models in step 2 on the central server as well as on the edge devices, it can be concluded that supervised and unsupervised techniques are different, and that unsupervised models, even without prior training, achieve better performance than supervised models; supervised models have access to labeled data compared to unsupervised models. According to the results, the unsupervised model is the best performer despite being considered the most difficult task. In fact, in all the above scenarios, the supervised model ANN outperformed only in the last scenario, and it is clear that more training is needed to detect this most difficult type of anomaly among the four scenarios. The second conclusion is related to deploying the right model in the edge architecture. As the results highlight, ANN is the best model in terms of accuracy, followed by KNN. Considering that KNN is an unsupervised model, it scores similarly to ANN in 3 out of 4 scenarios without prior training, while achieving the least RAM allocation, it can be concluded that KNN is the most suitable unsupervised model, so KNN is embedded in the edge nodes.

[0066] The sustainable edge computing framework is as follows:

[0067] Physical layer: represents the equipment that generates raw data at the source end, including: photovoltaic data collectors that measure DC voltage and current in the photovoltaic system, weather sensors that measure solar irradiance and temperature, photovoltaic inverters that convert DC power to AC power, and drones that regularly perform thermal inspections on photovoltaic modules;

[0068] Edge layer: Located between the physical layer and the fog layer, it represents the edge nodes that process and analyze incoming data streams before sending the results to the cloud. It consists of a Raspberry Pi cluster deployed in each distributed photovoltaic system. The edge layer can shorten processing time and enable devices to have artificial intelligence capabilities, thereby improving the efficiency of the monitoring system;

[0069] Fog layer: Located between the edge layer and the cloud layer, it consists of servers that centralize information from a group of edge nodes placed in the house and perform additional computation and data aggregation from the same location (e.g., community);

[0070] Cloud layer: The final layer for deploying analytics and decision support systems. Due to this architecture, many other tasks that were performed at this level in the past are moved to edge nodes. The cloud layer will only be responsible for collecting decisions, alerts, and status reports received from local nodes instead of acquiring the entire raw data set.

[0071] Distributed PV systems are interconnected in a traditional way, with data collectors sending raw data to a central monitoring server, which is responsible for processing and analyzing the data. However, in a smart decentralized configuration, the hazards of network congestion and delayed alarms make it infeasible to send raw data to a central cloud server for decision making. Therefore, this paper proposes a new edge framework that aims to solve this problem in a more sustainable and intelligent way.

[0072] The concept of edge computing emerged to address the challenges posed by network congestion caused by the transmission of large data sets to remote servers, especially in the context of big data and smart cities. Leveraging the capabilities of the edge can generate subsequent benefits: (1) Reduced latency: Since the analysis is performed close to the data source, the delay in data processing is minimized, and the distance that the data needs to traverse is reduced, resulting in increased response time. This is particularly important in applications that require real-time data processing, such as IoT sensors and distributed energy systems; (2) Reduced bandwidth: In the process of developing communication architectures, a comprehensive understanding of the bandwidth requirements associated with each communication protocol used is required. The transmission of data from one location to another requires the use of network connections, resulting in the consumption of part of the available capacity or bandwidth. Therefore, reducing bandwidth is an imperative research direction in this field, which can be effectively accomplished by implementing edge computing; 3) Reduced costs: The costs associated with communication in smart networks are directly related to the provision of data storage and analysis services on cloud platforms, which charge a large fee. These costs can be mitigated by transmitting only the relevant information to the central cloud instead of the entire raw data set.

[0073] Despite the many advantages and opportunities offered by edge computing, it is not without limitations that may pose challenges to its widespread implementation. These limitations are related to the inherent characteristics of edge computing and are described as follows: (1) Susceptibility to cyberattacks. Cyberattacks may be exacerbated by the distributed nature of edge nodes, which often involve deployment in remote and isolated areas. The vulnerability of these nodes is due to their limited resources, which force them to implement only limited defense measures and security protocols. (2) The implementation of edge computing requires a lot of adjustments to improve the reliability and consistency of its operation. In addition to hardware configuration, the deployment of trusted edge nodes also requires software improvements such as task loading and memory computing. (3) Limited computing: The limited processing power available in edge nodes creates considerable limitations and therefore creates problems when building complex machine learning or deep learning models. Therefore, it is necessary to conduct in-depth research on lightweight models and find a balance between computational efficiency and accuracy.

[0074] To deploy this edge computing framework, two steps are performed: (1) Offline mode. This mode consists of historical data collection from IoT devices, data processing, and benchmarking of existing supervised and unsupervised models. The best model is then trained and evaluated. (2) Online mode. The most effective model is selected based on the benchmark results, and real-time monitoring is performed to detect errors in a timely manner. After a fault is detected, the signal will be transmitted to the classification module, which aims to determine the precise category of the anomaly.

[0075] The present invention also provides a sustainable edge computing framework construction system for abnormal state detection of distributed photovoltaic systems using a sustainable edge computing framework construction method for abnormal state detection of distributed photovoltaic systems, characterized in that it includes:

[0076] Abnormal scenario building module: used to create a preset number of synthetic data sets by changing the percentage of power amplitude as different abnormal scenarios for evaluating the abnormal state detection performance of the machine learning model;

[0077] Abnormal state detection performance evaluation module: It is used to evaluate the abnormal state detection performance of machine learning models on central servers under different abnormal scenarios using training and prediction time and F1 score under macro-average technology as performance evaluation indicators; it is used to evaluate the abnormal state detection performance of machine learning models on edge devices under different abnormal scenarios using prediction time, RAM usage and model size as performance evaluation indicators;

[0078] Sustainable edge computing framework building module: It is used to determine the optimal machine learning model based on the abnormal state detection performance of the machine learning model on the central server and edge devices, and embed it into the edge node, integrate the object layer, edge layer, fog layer and cloud layer, and build a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems.

[0079] The embodiment of the present invention discloses a sustainable edge computing framework construction method and system for abnormal state detection of distributed photovoltaic systems. By adopting a systematic method to collect data based on a comprehensive hierarchical framework consisting of three stages: perception, connection, and application, and using the training and prediction time of the central server and the F1 score under the macro-average technology and the prediction time, RAM usage, and model size of the edge device as performance evaluation indicators, the performance of the machine learning model on the central server and edge device is evaluated under different abnormal scenarios constructed by changing the power amplitude percentage, and the optimal model embedded in the edge node is obtained. The material layer, edge layer, fog layer, and cloud layer are integrated to construct a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems, which not only realizes an abnormal state detection edge computing architecture solution that can be deployed on a large scale in smart cities or power grids, but also effectively improves the edge computing accuracy of abnormal state detection of distributed photovoltaic systems.

[0080] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the device disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0081] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A sustainable edge computing framework construction method for abnormal state detection of distributed photovoltaic systems, characterized in that: include: Step 1: Create a preset number of synthetic data sets by changing the percentage of power amplitude as different abnormal scenarios to evaluate the abnormal state detection performance of the machine learning model; Step 2: Using the training and prediction time and the F1 score under the macro-average technique as performance evaluation indicators, evaluate the abnormal state detection performance of the machine learning model on the central server under different abnormal scenarios; using the prediction time, RAM usage, and model size as performance evaluation indicators, evaluate the abnormal state detection performance of the machine learning model on the edge device under different abnormal scenarios; Step 3: Based on the abnormal state detection performance of the machine learning model on the central server and the edge device, determine the optimal machine learning model and embed it into the edge node, integrate the object layer, edge layer, fog layer and cloud layer, and build a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems.

2. A method for constructing a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems according to claim 1, characterized in that: In step 1, a systematic approach is used to collect data based on a comprehensive hierarchical framework consisting of three stages: perception, connection, and application, and a structured framework is used for data cleaning.

3. According to claim 1, a method for constructing a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems is characterized in that: The comprehensive hierarchical framework is as follows: Perception layer: It consists of measurement nodes deployed on site, which are responsible for capturing information related to the photovoltaic system, including solar irradiance, ambient temperature, string current, and string voltage; Connection layer: responsible for facilitating data transmission, consisting of wired Ethernet and switching equipment; Application layer: Contains the analysis and processing tasks that need to be performed on the collected data, including but not limited to prediction, anomaly detection, performance analysis, and condition monitoring.

4. The method for constructing a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems according to claim 1, characterized in that: The data cleaning is specifically to ensure data synchronization, delete duplicate values, night values, and current and voltage values ​​that are not within the preset current and voltage ranges, and use IIR filters and PR filters to remove R 2 and the number of days with PR less than 0.

8.

5. The method for constructing a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems according to claim 1, characterized in that: In step 1, the percentage of changing the power amplitude ranges from 20% to 90%.

6. The method for constructing a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems according to claim 1, characterized in that: In step 1, the machine learning model is screened and determined by using the PyOd library with its default parameters based on the accuracy of the algorithm on the same abnormal data set, including: a supervised learning model represented by a multilayer perceptron and an unsupervised learning model represented by a local outlier factor, a K-nearest neighbor algorithm, and a local outlier factor based on clustering.

7. A method for constructing a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems according to claim 6, characterized in that: In step 2, the training time is used as a performance evaluation indicator to evaluate the abnormal state detection performance of the machine learning model on the central server under different abnormal scenarios, specifically: For the supervised learning model, the training time is used as a performance evaluation indicator to evaluate the abnormal state detection performance of the machine learning model on the central server under different abnormal scenarios; For the unsupervised learning model, fitting time is used as a performance evaluation indicator to evaluate the abnormal state detection performance of the machine learning model on the central server under different abnormal scenarios.

8. The method for constructing a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems according to claim 1, characterized in that: In step 2, prediction time, RAM usage, and model size are used as performance evaluation indicators to evaluate the abnormal state detection performance of the machine learning model on the edge device under different abnormal scenarios, specifically: Edge computing is performed by deploying a Raspberry Pi on the edge device, and the prediction time of the machine learning model on the edge device is calculated using the time library in Python. The RAM usage of the machine learning model on the edge device is calculated using the memory analysis library in Python. The model size of the machine learning model on the edge device is calculated using the pickle library in Python to evaluate the abnormal state detection performance of the machine learning model on the edge device under different abnormal scenarios.

9. The method for constructing a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems according to claim 1, characterized in that: In step 3, the sustainable edge computing framework is as follows: Physical layer: represents the equipment that generates raw data at the source end, including: photovoltaic data collectors that measure DC voltage and current in the photovoltaic system, weather sensors that measure solar irradiance and temperature, photovoltaic inverters that convert DC power to AC power, and drones that regularly perform thermal inspections on photovoltaic modules; Edge layer: Located between the physical layer and the fog layer, it represents the edge nodes that process and analyze incoming data streams before sending the results to the cloud. It consists of a Raspberry Pi cluster deployed in each distributed PV system. Fog layer: Located between the edge layer and the cloud layer, it consists of servers that centralize information from a set of edge nodes placed in the premises and perform additional computation and data aggregation from the same location; Cloud layer: The final layer for deploying analysis and decision support systems, responsible for collecting decisions, alerts, and status reports received from local nodes.

10. A sustainable edge computing framework construction system for abnormal state detection of distributed photovoltaic systems using a sustainable edge computing framework construction method for abnormal state detection of distributed photovoltaic systems as described in any one of claims 1 to 9, characterized in that: include: Abnormal scenario building module: used to create a preset number of synthetic data sets by changing the percentage of power amplitude as different abnormal scenarios for evaluating the abnormal state detection performance of the machine learning model; Abnormal state detection performance evaluation module: used to evaluate the abnormal state detection performance of the machine learning model on the central server under different abnormal scenarios using the training and prediction time and the F1 score under the macro-average technology as performance evaluation indicators; and to evaluate the abnormal state detection performance of the machine learning model on the edge device under different abnormal scenarios using the prediction time, RAM usage and model size as performance evaluation indicators; Sustainable edge computing framework building module: used to determine the optimal machine learning model based on the abnormal state detection performance of the machine learning model on the central server and the edge device, and embed it into the edge node, integrate the object layer, edge layer, fog layer and cloud layer, and build a sustainable edge computing framework for abnormal state detection of distributed photovoltaic systems.

Citation Information

Cited By

  • Photovoltaic power station power generation monitoring system based on edge calculation

    CN119254143A

  • Distributed photovoltaic cluster cooperative control method based on edge calculation

    CN121076985A

  • A distributed photovoltaic cluster cooperative control method based on edge computing

    CN121076985B