Distributed photovoltaic safe operation method and device, computer equipment and storage medium

Abstract

The application relates to a distributed photovoltaic safety operation method and device, computer equipment and a storage medium. The method comprises the following steps: acquiring electrical quantity data and environmental parameters of a photovoltaic component in an edge system; performing time delay sensitivity evaluation on the electrical quantity data and the environmental parameters by using a trained support vector machine to obtain a target time delay sensitivity grade; determining a decision action space by using a trained decision network according to the target time delay sensitivity grade, communication bandwidth resources and a device load rate; the decision action space comprises a sampling frame rate, a data compression ratio and a model index; receiving a fault point position identification model matched with the model index issued by a cloud server, acquiring image data and electrical quantity sequences transmitted at the sampling frame rate and the data compression ratio, and performing fault point position identification by using the fault point position identification model to obtain a fault point position identification result. Efficient identification of distributed photovoltaic edge system faults is realized.

Description

Technical Field

[0001] This application relates to the field of new energy technology, and in particular to a method, apparatus, computer equipment, and storage medium for the safe operation and maintenance of distributed photovoltaic systems. Background Technology

[0002] With the accelerated global energy transition and the deepening implementation of "dual carbon" goals, renewable energy power generation systems, represented by distributed photovoltaic (PV) power, are experiencing explosive growth. Unlike centralized ground-mounted power plants, distributed PV is typically deployed in industrial park rooftops, residential homes, transportation hubs, and other locations, characterized by numerous locations, complex environments, and small individual unit capacities but a massive overall scale. However, with the exponential increase in installed capacity, the operation and maintenance of distributed PV faces unprecedented challenges. PV modules are exposed to the outdoors for extended periods, making them highly susceptible to hot spot effects, microcracks and breakage, dust accumulation and shading, and loose connectors. Among these, high-risk faults such as DC-side arcing, if not identified and interrupted within milliseconds, can easily lead to fires, seriously threatening people's lives and property.

[0003] Currently, photovoltaic (PV) operation and maintenance (O&M) technology has shifted from traditional "manual inspection" to "digital monitoring." However, significant technical bottlenecks remain in edge intelligence scenarios with large-scale and high real-time requirements. Traditional PV O&M relies heavily on an "edge-to-cloud" model, where sensors on the edge system collect data from the field and upload it to the cloud for processing and analysis. When a large number of distributed nodes need to process this data, severe communication bandwidth constraints arise. On one hand, the one-time upload of sensor data can easily cause channel congestion and incur expensive data charges. On the other hand, cloud-based data processing leads to significant transmission latency and network jitter, which can cause fault detection errors and even greater harm. Therefore, given the stringent resource constraints and the need for low-latency, high-precision, and secure response in distributed PV edge systems, a distributed PV security O&M method is urgently needed to achieve efficient fault identification in these systems. Summary of the Invention

[0004] Based on this, it is necessary to provide a method, device, computer equipment, computer-readable storage medium, and computer program product for the safe operation and maintenance of distributed photovoltaic systems, which can efficiently identify faults in distributed photovoltaic edge systems, in order to address the aforementioned technical problems.

[0005] Firstly, this application provides a method for the safe operation and maintenance of distributed photovoltaic systems, applied to edge systems, including:

[0006] Acquire electrical quantity data and environmental parameters of photovoltaic modules in the edge system;

[0007] Based on electrical quantity data and environmental parameters, a trained Support Vector Machine (SVM) is used to evaluate latency sensitivity and obtain the target latency sensitivity level.

[0008] Based on the target latency sensitivity level, communication bandwidth resources, and device load rate, a trained decision network is used to determine the decision action space; the decision action space includes the sampling frame rate, data compression ratio, and model index.

[0009] Send the model index to the cloud server to instruct the cloud server to send the fault location identification model that matches the model index.

[0010] The system receives the fault location identification model matched by the model index, acquires the image data and electrical quantity sequence transmitted at the sampling frame rate and data compression ratio, uses the fault location identification model to identify the fault location, and obtains the fault location identification result.

[0011] In one embodiment, based on electrical quantity data and environmental parameters, a trained support vector machine is used to evaluate latency sensitivity and obtain a target latency sensitivity level, including:

[0012] Determine the electrical characteristics of the electrical quantity data; the electrical characteristics include at least one of the following: standard deviation, absolute rate of change, and energy factor.

[0013] Construct a system state feature vector based on environmental parameters and electrical characteristics;

[0014] Based on the system state feature vector, a trained support vector machine is used to evaluate the latency sensitivity and obtain the target latency sensitivity level.

[0015] In one embodiment, the trained support vector machine includes multiple independent support vector machines, each used for evaluating multiple latency sensitivity levels. Based on the system state feature vector, the trained support vector machine is used to evaluate latency sensitivity to obtain the target latency sensitivity level, including:

[0016] Based on the system state feature vector, the decision function values ​​for multiple delay sensitivity levels are determined using the decision functions of multiple support vector machines. These multiple delay sensitivity levels include normal operation level, operation level requiring fault analysis and identification, and emergency operation level.

[0017] The target latency sensitivity level is determined among multiple latency sensitivity levels based on multiple decision function values.

[0018] In one embodiment, based on the target latency sensitivity level, communication bandwidth resources, and device load rate, a trained decision network is used to determine the decision action space, including:

[0019] The allowable latency corresponding to the target latency sensitivity level is determined by a preset mapping function;

[0020] Acquire communication bandwidth resources and device load rate, and use the target latency sensitivity level, allowable latency, communication bandwidth resources and device load rate as the state space;

[0021] The decision action space is determined based on the state space and the trained decision network.

[0022] In one embodiment, the step of determining the trained decision network includes:

[0023] Obtain state space samples, and based on the state space samples and the initial Actor network, obtain the predicted action space;

[0024] Obtain the total delay and the next state space after the predicted action space is executed, and determine the reward value based on the total delay and the preset reward function;

[0025] Determine the target Q value based on the next state space and the target network;

[0026] Based on the state space samples, the predicted action space, and the initial Critic network, the predicted Q value is obtained;

[0027] The network parameters of the initial Critic network are updated based on the difference between the target Q value and the predicted Q value to obtain the updated Critic network.

[0028] Based on the predicted action space and the updated Critic network, the updated Q value is obtained;

[0029] With the goal of maximizing the updated Q value, the network parameters of the initial Actor network are updated to obtain the updated Actor network, which is then used as the decision network after training.

[0030] In one embodiment, a fault location identification model is used to identify fault locations, and the fault location identification results are obtained, including:

[0031] Image features of image data transmitted at sampling frame rate and data compression ratio are extracted based on the visual branch network, and electrical quantity features of electrical quantity sequences transmitted at sampling frame rate and data compression ratio are extracted based on the temporal branch network.

[0032] Image features and electrical quantity features are input into the fault location identification model. Attention processing is performed on the image features and electrical quantity features to obtain attention weights. The image features and electrical quantity features are then fused according to the attention weights to obtain fused features. The fused features are then subjected to fully connected processing to obtain the fault location identification result output by the fault location identification model.

[0033] Secondly, this application also provides a distributed photovoltaic safety operation and maintenance device, comprising:

[0034] The acquisition module is used to acquire electrical quantity data and environmental parameters of photovoltaic modules in the edge system;

[0035] The evaluation module is used to evaluate the time delay sensitivity based on electrical quantity data and environmental parameters using a trained support vector machine, and obtain the target time delay sensitivity level.

[0036] The decision module is used to determine the decision action space based on the target latency sensitivity level, communication bandwidth resources, and device load rate, using the trained decision network. The decision action space includes the sampling frame rate, data compression ratio, and model index.

[0037] The matching module is used to send the model index to the cloud server, so as to instruct the cloud server to send the fault location identification model that matches the model index.

[0038] The identification module is used to receive the fault location identification model matched by the model index, acquire image data and electrical quantity sequences transmitted at the sampling frame rate and data compression ratio, identify the fault location using the fault location identification model, and obtain the fault location identification result.

[0039] Thirdly, this application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to perform the following steps:

[0040] Acquire electrical quantity data and environmental parameters of photovoltaic modules in the edge system;

[0041] Based on electrical quantity data and environmental parameters, a trained support vector machine is used to evaluate latency sensitivity and obtain the target latency sensitivity level.

[0042] Based on the target latency sensitivity level, communication bandwidth resources, and device load rate, a trained decision network is used to determine the decision action space; the decision action space includes the sampling frame rate, data compression ratio, and model index.

[0043] Send the model index to the cloud server to instruct the cloud server to send the fault location identification model that matches the model index.

[0044] The system receives the fault location identification model matched by the model index, acquires the image data and electrical quantity sequence transmitted at the sampling frame rate and data compression ratio, uses the fault location identification model to identify the fault location, and obtains the fault location identification result.

[0045] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the following steps:

[0046] Acquire electrical quantity data and environmental parameters of photovoltaic modules in the edge system;

[0047] Based on electrical quantity data and environmental parameters, a trained support vector machine is used to evaluate latency sensitivity and obtain the target latency sensitivity level.

[0048] Based on the target latency sensitivity level, communication bandwidth resources, and device load rate, a trained decision network is used to determine the decision action space; the decision action space includes the sampling frame rate, data compression ratio, and model index.

[0049] Send the model index to the cloud server to instruct the cloud server to send the fault location identification model that matches the model index.

[0050] The system receives the fault location identification model matched by the model index, acquires the image data and electrical quantity sequence transmitted at the sampling frame rate and data compression ratio, uses the fault location identification model to identify the fault location, and obtains the fault location identification result.

[0051] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, performs the following steps:

[0052] Acquire electrical quantity data and environmental parameters of photovoltaic modules in the edge system;

[0053] Based on electrical quantity data and environmental parameters, a trained support vector machine is used to evaluate latency sensitivity and obtain the target latency sensitivity level.

[0054] Based on the target latency sensitivity level, communication bandwidth resources, and device load rate, a trained decision network is used to determine the decision action space; the decision action space includes the sampling frame rate, data compression ratio, and model index.

[0055] Send the model index to the cloud server to instruct the cloud server to send the fault location identification model that matches the model index.

[0056] The system receives the fault location identification model matched by the model index, acquires the image data and electrical quantity sequence transmitted at the sampling frame rate and data compression ratio, uses the fault location identification model to identify the fault location, and obtains the fault location identification result.

[0057] The aforementioned distributed photovoltaic (PV) safety operation and maintenance method, device, computer equipment, computer-readable storage medium, and computer program product acquire electrical quantity data and environmental parameters of PV modules in the edge system. Based on this data, a trained support vector machine (SVM) is used to evaluate latency sensitivity, obtaining a target latency sensitivity level. The SVM effectively handles the coupling relationship between multimodal data such as electrical quantity data and environmental data of PV modules, thereby accurately evaluating the latency sensitivity of PV modules. Under the constraints of the target latency sensitivity level, communication bandwidth resources, and equipment load rate, a trained decision network is used to determine the optimal decision action space. The decision action space includes the sampled frames. By optimizing the sampling frame rate, data compression ratio, and model index, the system achieves optimal configuration of multi-dimensional resources such as communication, sensing, and computing. This helps ensure real-time fault identification and reduces data transmission pressure and computational load on the edge system. The edge system retrieves the fault location identification model matching the model index from the cloud server and transmits image data and electrical quantity sequences at the sampling frame rate and data compression ratio. This helps reduce the response latency of the edge system and effectively supports the accurate and efficient inference of the fault location identification model. By using the fault location identification model to identify fault locations and obtain fault location identification results, the system can efficiently identify faults in the distributed photovoltaic edge system, improving the efficiency and reliability of distributed photovoltaic safe operation and maintenance. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the drawings used in the description of the embodiments of this application or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 This is an application environment diagram of a distributed photovoltaic safety operation and maintenance method in one embodiment;

[0060] Figure 2 This is a flowchart illustrating a distributed photovoltaic safety operation and maintenance method in one embodiment;

[0061] Figure 3 This is a graph showing the convergence of reward values ​​during the training process of a decision network in one embodiment.

[0062] Figure 4 This is a latency compliance verification curve for one embodiment;

[0063] Figure 5 This is a structural block diagram of a distributed photovoltaic safety operation and maintenance device in one embodiment;

[0064] Figure 6This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0066] It should be noted that the terms "first," "second," etc., used in this application can be used to describe various elements, but these elements are not limited by these terms. These terms are only used to distinguish the first element from the second element. The terms "comprising" and "having," and any variations thereof, used in this application, are intended to cover non-exclusive inclusion. The term "multiple" used in this application refers to two or more. The term "and / or" used in this application refers to one of the embodiments, or any combination of multiple embodiments.

[0067] The distributed photovoltaic safety operation and maintenance method provided in this application embodiment can be applied to, for example, Figure 1 In the application environment shown, the edge system 102 communicates with the cloud server 104 via a network. A data storage system can store the data that the cloud server 104 needs to process. The data storage system can be integrated onto the cloud server 104 or placed in the cloud or on another network server. The edge system 102 acquires electrical quantity data and environmental parameters of the photovoltaic modules in the edge system; based on the electrical quantity data and environmental parameters, it uses a trained support vector machine to evaluate latency sensitivity and obtain a target latency sensitivity level; based on the target latency sensitivity level, communication bandwidth resources, and device load rate, it uses a trained decision network to determine the decision action space; the decision action space includes the sampling frame rate, data compression ratio, and model index; it sends the model index to the cloud server 104 to instruct the cloud server 104 to issue a fault location identification model matching the model index; it receives the fault location identification model matching the model index, acquires image data and electrical quantity sequences transmitted at the sampling frame rate and data compression ratio, and uses the fault location identification model to identify the fault location, obtaining the fault location identification result.

[0068] In one exemplary embodiment, such as Figure 2 As shown, a method for safe operation and maintenance of distributed photovoltaic systems is provided, which is applied to... Figure 1 The following steps are used as an example of the edge system 102 in the example, including steps 201 to 205.

[0069] Step 201: Obtain electrical quantity data and environmental parameters of photovoltaic modules in the edge system.

[0070] Photovoltaic modules are the core power generation unit of distributed photovoltaic (PV) systems, converting solar energy into electrical energy. They are composed of multiple photovoltaic cells connected in series or parallel. Edge systems refer to edge computing nodes deployed at the PV power plant site, used for data acquisition and processing of the PV modules. Edge systems utilize various sensors for data acquisition, such as visual sensors like cameras and infrared thermal imagers, electrical quantity acquisition devices like smart meters, and environmental sensors like irradiance meters. Visual sensors acquire visual images of the PV modules. Electrical quantity acquisition devices collect electrical quantity data from the PV modules. Environmental sensors collect environmental parameters of the PV modules.

[0071] For example, electrical quantity data is used to describe the electrical performance characteristics of photovoltaic modules, and may include voltage, current, power, and power generation. Environmental parameters are physical quantities that describe the environment in which the photovoltaic modules are located, and may include temperature, humidity, and wind speed.

[0072] Step 202: Based on electrical quantity data and environmental parameters, use the trained support vector machine to evaluate the time delay sensitivity and obtain the target time delay sensitivity level.

[0073] Among them, the trained support vector machine refers to the support vector machine model obtained by training with a large number of labeled samples. It is used to evaluate the time delay sensitivity based on the input electrical quantity data and environmental parameters, and obtain the target time delay sensitivity level.

[0074] In some embodiments, at least one electrical feature of the electrical quantity data is extracted, which may include time domain features and frequency domain features. The electrical features and environmental parameters are used together as multimodal system state features. A trained support vector machine is used to evaluate the time delay sensitivity to obtain the target time delay sensitivity level.

[0075] Latency sensitivity evaluation is used to assess the sensitivity of edge systems to data transmission and processing latency. The target latency sensitivity level is the result of evaluation using a trained support vector machine, based on the electrical quantity data of photovoltaic modules and environmental parameters.

[0076] For example, the target latency sensitivity level includes multiple levels, namely L0, L1, and L2. Among them, L0 level represents the normal operation level when the photovoltaic module is in steady-state operation or nighttime dormancy; L1 level represents the operation level when the photovoltaic module is operating under slight shadows such as cloud cover, and faults need to be analyzed and identified; L2 level is the emergency operation level that requires millisecond-level response when high-frequency noise or current surge is detected.

[0077] The higher the target latency sensitivity level, the higher the requirements of the edge system for real-time fault identification and data transmission and low latency.

[0078] Step 203: Based on the target latency sensitivity level, communication bandwidth resources, and device load rate, the trained decision network is used to determine the decision action space; the decision action space includes the sampling frame rate, data compression ratio, and model index.

[0079] Among them, communication bandwidth resources refer to the communication link capacity for data transmission between edge devices and cloud servers, which can be wireless communication bandwidth resources.

[0080] Device load rate is the utilization rate of CPU, memory, computing power, and other resources of the current device in the edge system. The higher the device load rate, the fewer remaining resources are available for fault identification.

[0081] The trained decision model refers to a neural network model trained through reinforcement learning. It takes the target latency sensitivity level, communication bandwidth resources, and device load rate as input and outputs the optimal combination of decision actions. For example, the initial model includes an initial Actor network and an initial Critic network. Reinforcement learning is performed on the initial model using state space samples to obtain the trained decision network. The state space samples include latency sensitivity level samples, communication bandwidth resource samples, and device load rate samples.

[0082] The decision action space refers to a set containing at least one decision action, each used to dynamically allocate resources for the communication, sensing, and computing behaviors of the edge system. Decision actions can include at least one of sampling frame rate, data compression ratio, and model index. Through the decision model, the collaborative allocation of communication, sensing, and computing resources of the edge system is achieved, optimizing service response latency.

[0083] The sampling frame rate refers to the number of data points collected by each sensor in an edge system per unit of time. A higher sampling frame rate results in higher real-time performance of the edge system. Examples include the number of electrical quantity data points collected per second by an electrical quantity acquisition device, and the number of visual images collected per second by a vision sensor.

[0084] Data compression ratio refers to the proportion of data (such as electrical quantity data, visual images, etc.) that is compressed. The higher the data compression ratio, the less data is transmitted, and the lower the data transmission latency of the edge system.

[0085] The model index is an identifier used to indicate different models. Each model has its own model index, which allows for quick identification of the corresponding model.

[0086] Step 204: Send the model index to the cloud server to instruct the cloud server to issue a fault location identification model that matches the model index.

[0087] The cloud server is a server cluster with large-capacity storage and high-performance computing power. The cloud server contains a model library that stores multiple models, each with its own model index.

[0088] The fault location identification model is a neural network model used to identify fault locations in photovoltaic modules. The edge system encapsulates the model index into request data and sends it to the cloud server. The cloud server parses the model index from the received request data, uses the model index to retrieve a matching fault location identification model from the model library, and then sends the model file of that model to the edge system.

[0089] Step 205: Receive the fault location identification model matched by the model index, acquire the image data and electrical quantity sequence transmitted at the sampling frame rate and data compression ratio, use the fault location identification model to identify the fault location, and obtain the fault location identification result.

[0090] The visual sensor uses the sampling frame rate of the decision action space to acquire image data of the photovoltaic module, while the electrical quantity acquisition device uses the sampling frame rate of the decision action space to acquire electrical quantity sequences of the photovoltaic module. Image data can be used to identify faults such as hot spots and damage in the photovoltaic module; electrical quantity sequences can be used to identify faults such as short circuits and open circuits in the photovoltaic module.

[0091] Image data is compressed using the data compression ratio of the decision action space to obtain a compressed image, which is then used as image data transmitted at the sampling frame rate and data compression ratio. Electrical quantity sequences are also compressed using the data compression ratio of the decision action space to obtain a compressed electrical quantity sequence, which is then used as an electrical quantity sequence transmitted at the sampling frame rate and data compression ratio.

[0092] The system receives a fault location identification model from a cloud server, matching the model index. This model is then used to identify fault locations. For example, image features from image data and electrical quantity features from electrical quantity sequences are extracted. The fault location identification model fuses these features to obtain fused features. Fault location identification is then performed based on these fused features to obtain the identification result. The fault location identification result can include multiple location points and the fault probability at each location point; for example, the result can be a fault probability distribution.

[0093] In the aforementioned distributed photovoltaic (PV) safety operation and maintenance method, electrical quantity data and environmental parameters of PV modules in the edge system are acquired. Based on this data, a trained support vector machine (SVM) is used to evaluate latency sensitivity and obtain the target latency sensitivity level. The SVM can effectively handle the coupling relationship between multimodal data such as electrical quantity data and environmental data of PV modules, thereby accurately evaluating the latency sensitivity of PV modules. Under the constraints of the target latency sensitivity level, communication bandwidth resources, and equipment load rate, a trained decision network is used to determine the optimal decision action space. The decision action space includes the sampling frame rate, data compression ratio, and model index. Achieving optimal configuration of multi-dimensional resources such as communication, sensing, and computing helps ensure real-time fault identification and reduces data transmission pressure and computational load on edge systems. Edge systems retrieve fault location identification models matching the model index from the cloud server, transmitting image data and electrical quantity sequences at a high sampling frame rate and data compression ratio. This reduces response latency and effectively supports accurate and efficient inference of the fault location identification model. Using the fault location identification model for fault location identification yields the identification results, enabling efficient fault identification in distributed photovoltaic edge systems and improving the efficiency and reliability of distributed photovoltaic safe operation and maintenance.

[0094] In an exemplary embodiment, a time delay sensitivity evaluation is performed using a trained support vector machine based on electrical quantity data and environmental parameters to obtain a target time delay sensitivity level. This includes: determining the electrical characteristics of the electrical quantity data; the electrical characteristics include at least one of standard deviation, absolute rate of change, and energy factor; constructing a system state feature vector based on environmental parameters and electrical characteristics; and performing a time delay sensitivity evaluation using a trained support vector machine based on the system state feature vector to obtain a target time delay sensitivity level.

[0095] The electrical characteristics of electrical quantity data include time-domain and frequency-domain characteristics. For example, time-domain characteristics include standard deviation and absolute rate of change, while frequency-domain characteristics include energy factor. Electrical quantity data can be current or voltage sequences acquired according to a preset sampling rate.

[0096] Since photovoltaic (PV) operational faults such as DC arcing and hot spots generally manifest as high-frequency noise or abrupt amplitude changes in electrical quantities, a sliding time window method is used to extract features from the collected electrical quantity data to capture pre-fault information of PV modules. For example, the window length is defined as... Then, the electrical quantity data can be extracted into N electrical quantity values ​​by following the sliding window. Composed of electrical quantity data sequences .

[0097] Furthermore, time-domain features of the electrical quantity data sequence X are extracted based on standard deviation and absolute rate of change to distinguish between slow changes and sudden, drastic fluctuations in electrical quantity data. Standard deviation It can be represented as:

[0098]

[0099]

[0100] in, The standard deviation is the average of multiple electrical quantity values. When there are microcracks or poor connections in the photovoltaic module, the electrical quantity data (such as current) will fluctuate continuously, causing the standard deviation to increase. It is significantly higher than the steady-state value.

[0101] absolute rate of change It can be represented as:

[0102]

[0103] in, It can effectively characterize the degree of instantaneous change in electrical quantity data. When electrical quantity data (such as current) experiences a step change in a very short time due to problems such as line faults, The value will increase significantly.

[0104] Furthermore, considering the characteristic that certain types of photovoltaic module faults can lead to unique high-frequency noise in electrical quantity data, a Fast Fourier Transform (FFT) is performed on the electrical quantity data sequence X to extract high-frequency energy features, yielding the energy factor. The spectral amplitude of the electrical quantity data sequence X is defined as F(K), and the high-frequency energy factor is... It can be represented as:

[0105]

[0106] in, The cutoff frequency is preset based on the fault type. The energy factor is the highest frequency at which the fault corresponding to this fault type occurs. It can effectively capture high-frequency energy that may be generated by photovoltaic module failures in the low-frequency current fluctuations caused by normal light changes.

[0107] Based on the feature extraction results above, environmental parameters are fused to construct a system state feature vector. :

[0108]

[0109] Among them, environmental parameters include irradiance G and ambient temperature. .

[0110] To establish a nonlinear mapping relationship between system state characteristics and business urgency, a latency sensitivity evaluation method based on support vector machines is proposed. Compared with traditional threshold judgment, SVM can effectively handle the coupling relationship between multi-dimensional features. The system state feature vector is used as the trained support vector machine, which outputs decision function values ​​for multiple latency sensitivity levels. Based on the decision function values, the target latency sensitivity level is determined among multiple latency sensitivity levels.

[0111] In this embodiment, by extracting multi-dimensional features of electrical quantity data in the time and frequency domains, the continuous jitter and instantaneous changes of photovoltaic modules are captured; combining the multi-dimensional features of electrical quantity data with environmental parameters to evaluate time delay sensitivity is beneficial for accurately identifying the real-time requirements of the current task and improving the accuracy of time delay sensitivity evaluation.

[0112] In an exemplary embodiment, the trained support vector machine includes multiple independent support vector machines, each used to evaluate multiple latency sensitivity levels. The process involves evaluating latency sensitivity using the trained support vector machines based on the system state feature vector to obtain a target latency sensitivity level. This includes: determining decision function values ​​for multiple latency sensitivity levels using the decision functions of each support vector machine based on the system state feature vector; the multiple latency sensitivity levels include a normal operation level, a fault-detection level, and an emergency operation level; and determining the target latency sensitivity level among the multiple latency sensitivity levels based on the multiple decision function values.

[0113] In this application, a single Support Vector Machine (SVM) is typically used for binary classification. The trained SVM in this embodiment comprises multiple independent SVMs, each used to evaluate different latency sensitivity levels. The decision function refers to the classification discriminant function of a single SVM. By inputting the system state feature vector to each SVM, the decision function value for the corresponding latency sensitivity level is obtained. The decision function value indicates the degree of matching between the system state feature vector and the corresponding latency sensitivity level; a larger decision function value indicates a higher degree of matching. For example, the decision function can be expressed as:

[0114]

[0115] in, is the Lagrange multiplier corresponding to the i-th support vector machine, reflecting the importance of the sample to the classification decision; This represents the latency sensitivity level corresponding to the i-th support vector machine; These are pre-collected and labeled training samples; is the radial basis kernel function to accommodate the nonlinear operating conditions of photovoltaic systems; j represents the index of the time delay sensitivity level; This represents the bias term indicating the delay sensitivity level of the j-th class. Specifically, it is represented as follows:

[0116]

[0117] The delay sensitivity level output by the decision function can include the normal operation level. The operational level of the fault needs to be analyzed and identified. and emergency operation level .

[0118] Among multiple delay sensitivity levels, the target delay sensitivity level with the highest decision function value is determined. The target delay sensitivity level is the delay sensitivity level that best matches the current system state.

[0119] In this embodiment, multiple independent support vector machines are used to divide the latency sensitivity into multiple levels, accurately matching the latency sensitivity of photovoltaic modules and improving the accuracy and efficiency of safe operation and maintenance of photovoltaic modules.

[0120] In an exemplary embodiment, the decision action space is determined using a trained decision network based on the target latency sensitivity level, communication bandwidth resources, and device load rate. This includes: determining the allowable latency corresponding to the target latency sensitivity level through a preset mapping function; obtaining the communication bandwidth resources and device load rate, and using the target latency sensitivity level, allowable latency, communication bandwidth resources, and device load rate as the state space; and determining the decision action space based on the state space and the trained decision network.

[0121] The edge system uses the target latency sensitivity level L obtained from SVM to perform dynamic resource allocation and output control commands. This enables adaptive mapping of multimodal acquisition strategies. A mapping function H(L) is introduced to transform the target latency sensitivity level L into a continuous-time variable, i.e., the allowable latency. :

[0122]

[0123] in, express The allowable latency corresponding to the level express The allowable latency corresponding to the level express Allowable latency corresponding to the level.

[0124] The edge system obtains the communication bandwidth resources and device load rate of the system at the current moment, and uses the target latency sensitivity level, allowable latency, communication bandwidth resources, and device load rate as the state space S. .in, For communication bandwidth resources, This refers to the equipment load rate.

[0125] Define the decision action space A. .in, The sampling frame rate is used to sense resource configuration, i.e., at what frequency data is collected; This is the data compression ratio, used for configuring computing and communication resources to reduce the consumption of computing and communication resources; It serves as a model index, enabling decisions on the complexity of the models used, and facilitating the allocation of computational resources to reduce computational resource consumption.

[0126] In some embodiments, the actual sampling frame rate can also be determined based on the sampling frame rate. , . These represent the preset maximum and minimum sampling rates, respectively.

[0127] The edge system uses the state space as input to the trained decision network to obtain the decision action space output by the decision network.

[0128] In this embodiment, by mapping the target latency sensitivity level to the corresponding allowable latency, a quantifiable constraint on resource scheduling decisions is formed; the target latency sensitivity level, allowable latency, communication bandwidth resources, and device load rate are integrated into a state space, which comprehensively covers the resource requirements of communication, sensing, and computing, and improves the accuracy and reliability of the decision-making action space.

[0129] In an exemplary embodiment, the step of determining the trained decision network includes: acquiring state space samples; obtaining a predicted action space based on the state space samples and an initial Actor network; acquiring the total delay and the next state space after the predicted action space is executed; determining a reward value based on the total delay and a preset reward function; determining a target Q value based on the next state space and a target network; obtaining a predicted Q value based on the state space samples, the predicted action space, and the initial Critic network; updating the network parameters of the initial Critic network based on the difference between the target Q value and the predicted Q value to obtain an updated Critic network; obtaining an updated Q value based on the predicted action space and the updated Critic network; and updating the network parameters of the initial Actor network with the goal of maximizing the updated Q value to obtain an updated Actor network, which is then used as the trained decision network.

[0130] The decision network is trained using state space samples to train the initial Actor and Critic networks. State space samples refer to historical state data samples collected from photovoltaic modules, obtained based on historical electrical quantity data and historical environmental parameters. For example, the historical electrical characteristics of the historical electrical quantity data are determined; these characteristics include at least one of standard deviation, absolute rate of change, and energy factor. A historical system state feature vector is constructed based on the historical environmental parameters and historical electrical characteristics. Based on this vector, a trained support vector machine is used to evaluate latency sensitivity, obtaining a target latency sensitivity level sample. A preset mapping function is used to determine the allowable latency sample corresponding to the target latency sensitivity level sample. Historical communication bandwidth resources and historical equipment load rates are acquired, and the target latency sensitivity level sample, allowable latency sample, historical communication bandwidth resources, and historical equipment load rates are used as state space samples.

[0131] The state space sample S is input into the initial Actor network to obtain the predicted action space a. The actions in the predicted action space a are then executed in the environment to obtain the total delay and the next state space. The reward value is calculated using a preset reward function and total delay, utilizing the next state space. The target Q-value is determined with the target network. The state space samples S and the predicted action space a are input into the initial Critic network to output the predicted Q-value. The network parameters of the initial Critic network are updated using the difference between the target Q-value and the predicted Q-value to obtain the updated Critic network. The predicted action space is input into the updated Critic network again to obtain the updated Q-value. The policy loss is calculated with the goal of maximizing the updated Q-value. The network parameters of the initial Actor network are updated using backpropagation in the gradient direction provided by the updated Critic network to obtain the updated Actor network, which is then used as the decision network after training.

[0132] The preset reward function can be expressed as follows:

[0133]

[0134] in, For action Corresponding expected fault identification accuracy; Total delay; This is the penalty function for exceeding the delay limit; Indicates the delay threshold; This represents the reward value.

[0135] For distributed photovoltaic (PV) security operation and maintenance scenarios, computing power simulations are conducted. For example... Figure 3The figure shows the reward value convergence curves during the training process of the decision network in some embodiments. As can be seen from the figure, the reward value gradually converges during the reinforcement learning training process, and the optimal decision network can be obtained at the highest reward value.

[0136] In some embodiments, latency compliance verification is performed on the decision network. If the actual total end-to-end response latency is... Then the time delay exceeding the limit deviation is .like This indicates that the decision-making action space meets the timeliness requirement; if This indicates that the decision-making action space does not meet the timeliness requirements, and this is fed back to the training steps of the decision network to continue the closed-loop iteration to update the decision network. For example... Figure 4 The figure shows latency compliance verification curves for some embodiments. The horizontal axis represents the number of emergency events, such as the number of times an emergency operation level occurs. The vertical axis represents the actual end-to-end response total latency, which can refer to the total time spent from data acquisition, transmission, model inference to fault identification. The figure shows that in the early stages of training the decision network, the output decision action space is not the optimal resource scheduling strategy, resulting in high total latency. As the model is updated, the total latency gradually decreases and stabilizes within a certain range, indicating that the trained decision network has high latency compliance and stability, meeting the requirements for real-time fault identification of photovoltaic modules.

[0137] In this embodiment, by constructing a decision network based on the Actor-Critic architecture, and based on the optimal decision action space that the trained decision network can output, the collaborative configuration of sensing, computing, and communication resources is realized, thereby optimizing service response latency.

[0138] In an exemplary embodiment, a fault location identification model is used to identify fault locations and obtain fault location identification results. This includes: extracting image features from image data transmitted at a sampling frame rate and data compression ratio using a visual branch network; extracting electrical quantity features from electrical quantity sequences transmitted at a sampling frame rate and data compression ratio using a temporal branch network; inputting the image features and electrical quantity features into the fault location identification model; performing attention processing on the image features and electrical quantity features to obtain attention weights; fusing the image features and electrical quantity features according to the attention weights to obtain fused features; and performing fully connected processing on the fused features to obtain the fault location identification results output by the fault location identification model.

[0139] Among them, the visual branch network can be an image feature extraction network based on CNN (convolutional neural network). The visual branch network performs operations such as convolution and pooling on image data and outputs image features of the image data, including information such as texture, shape and position, which are used to indicate appearance faults such as hot spots, cracks and occlusions.

[0140] The temporal branch network can be an LSTM (Long Short Term Memory) based network used to extract the temporal features of electrical quantity sequences, i.e., electrical quantity features.

[0141] The fault location identification model includes a fusion module and a fully connected processing module. The fusion module processes image features... and electrical quantity characteristics Attention processing is performed to obtain attention weights, where H and W are the spatial dimensions of the image features, and C is the number of channels. Image features and electrical quantity features are fused based on the attention weights to obtain fused features. The fused features are then processed using a fully connected processing module to obtain the fault location identification result.

[0142] For example, the fusion module could be a cross-attention module based on a cross-attention mechanism. The query matrix Q is defined from electrical quantity features, and the key matrix K and value matrix V are defined from image features, i.e.:

[0143]

[0144] in, These are the learnable weight matrices.

[0145] Attention weights can be represented as:

[0146]

[0147] Attention weights characterize the correlation between spatial locations in electrical quantity features and image features. The fused feature output after fusing image features and electrical quantity features can be expressed as:

[0148]

[0149] Fusion features Input a fully connected module and output the final fault probability distribution. That is, the result of fault location identification.

[0150] In this embodiment, by extracting image features from image data and electrical quantity features from electrical quantity sequences respectively, attention processing is performed on the image features and electrical quantity features to achieve the fusion of multi-dimensional data features. The fused features are then used for fault identification, improving the accuracy and reliability of fault identification. In addition, since image data and electrical quantity features are collected and transmitted through actions in the decision action space, the efficiency of fault identification is greatly improved, meeting the real-time requirements of edge systems in high latency-sensitive scenarios.

[0151] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages in other steps. It is understood that the steps in different embodiments can be freely combined as needed, and all non-contradictory solutions formed by such combinations are within the scope of protection of this application.

[0152] Based on the same inventive concept, this application also provides a distributed photovoltaic safety operation and maintenance device for implementing the distributed photovoltaic safety operation and maintenance method described above. The solution provided by this device is similar to the solution described in the above method; therefore, the specific limitations in one or more embodiments of the distributed photovoltaic safety operation and maintenance device provided below can be found in the limitations of the distributed photovoltaic safety operation and maintenance method described above, and will not be repeated here.

[0153] In one exemplary embodiment, such as Figure 5 As shown, a distributed photovoltaic safety operation and maintenance device 500 is provided, including: an acquisition module 510, an evaluation module 520, a decision-making module 530, a matching module 540, and an identification module 550.

[0154] The acquisition module 510 is used to acquire electrical quantity data and environmental parameters of photovoltaic modules in the edge system.

[0155] Evaluation module 520 is used to evaluate the time delay sensitivity based on electrical quantity data and environmental parameters using a trained support vector machine, and obtain the target time delay sensitivity level.

[0156] The decision module 530 is used to determine the decision action space based on the target latency sensitivity level, communication bandwidth resources and equipment load rate, using the trained decision network; the decision action space includes sampling frame rate, data compression ratio and model index.

[0157] The matching module 540 is used to send the model index to the cloud server to instruct the cloud server to issue the fault location identification model that matches the model index.

[0158] The identification module 550 is used to receive the fault location identification model matched by the model index, acquire the image data and electrical quantity sequence transmitted at the sampling frame rate and data compression ratio, use the fault location identification model to identify the fault location, and obtain the fault location identification result.

[0159] The aforementioned distributed photovoltaic (PV) safety operation and maintenance device acquires electrical quantity data and environmental parameters of PV modules in the edge system. Based on this data, it uses a trained support vector machine (SVM) to evaluate latency sensitivity and obtain a target latency sensitivity level. The SVM effectively handles the coupling relationship between multimodal data such as electrical quantity data and environmental data of PV modules, thus accurately evaluating the latency sensitivity of the PV modules. Under constraints of the target latency sensitivity level, communication bandwidth resources, and equipment load rate, a trained decision network determines the optimal decision action space. The decision action space includes the sampling frame rate, data compression ratio, and model index. Achieving optimal configuration of multi-dimensional resources such as communication, sensing, and computing helps ensure real-time fault identification and reduces data transmission pressure and computational load on edge systems. Edge systems retrieve fault location identification models matching the model index from the cloud server, transmitting image data and electrical quantity sequences at a high sampling frame rate and data compression ratio. This reduces response latency and effectively supports accurate and efficient inference of the fault location identification model. Using the fault location identification model for fault location identification yields the identification results, enabling efficient fault identification in distributed photovoltaic edge systems and improving the efficiency and reliability of distributed photovoltaic safe operation and maintenance.

[0160] In one embodiment, a time delay sensitivity evaluation is performed using a trained support vector machine based on electrical quantity data and environmental parameters to obtain a target time delay sensitivity level. The evaluation module 520 is further configured to: determine the electrical characteristics of the electrical quantity data; the electrical characteristics include at least one of standard deviation, absolute rate of change, and energy factor; construct a system state feature vector based on environmental parameters and electrical characteristics; and perform a time delay sensitivity evaluation using a trained support vector machine based on the system state feature vector to obtain a target time delay sensitivity level.

[0161] In one embodiment, the trained support vector machine includes multiple independent support vector machines, each used to evaluate multiple latency sensitivity levels. Based on the system state feature vector, the trained support vector machines are used to evaluate latency sensitivity and obtain a target latency sensitivity level. The evaluation module 520 is further configured to: determine the decision function values ​​for multiple latency sensitivity levels using the decision functions of each of the multiple support vector machines, based on the system state feature vector; the multiple latency sensitivity levels include normal operation level, operation level requiring fault analysis and identification, and emergency operation level; and determine the target latency sensitivity level among the multiple latency sensitivity levels based on the multiple decision function values.

[0162] In one embodiment, based on the target latency sensitivity level, communication bandwidth resources, and device load rate, a trained decision network is used to determine the decision action space. The decision module 530 is further configured to: determine the allowable latency corresponding to the target latency sensitivity level through a preset mapping function; obtain the communication bandwidth resources and device load rate, and use the target latency sensitivity level, allowable latency, communication bandwidth resources, and device load rate as the state space; and determine the decision action space based on the state space and the trained decision network.

[0163] In one embodiment, the decision module 530 further comprises the following steps in determining the trained decision network: acquiring state space samples; obtaining a predicted action space based on the state space samples and the initial Actor network; acquiring the total delay and the next state space after the predicted action space is executed; determining a reward value based on the total delay and a preset reward function; determining a target Q value based on the next state space and the target network; obtaining a predicted Q value based on the state space samples, the predicted action space, and the initial Critic network; updating the network parameters of the initial Critic network based on the difference between the target Q value and the predicted Q value; obtaining an updated Critic network based on the predicted action space and the updated Critic network; and updating the network parameters of the initial Actor network with the goal of maximizing the updated Q value, thereby obtaining an updated Actor network, which is then used as the trained decision network.

[0164] In one embodiment, a fault location identification model is used to identify fault locations and obtain fault location identification results. The identification module 550 is further configured to: extract image features of image data transmitted at sampling frame rate and data compression ratio according to the visual branch network, and extract electrical quantity features of electrical quantity sequences transmitted at sampling frame rate and data compression ratio according to the temporal branch network; input the image features and electrical quantity features into the fault location identification model, perform attention processing on the image features and electrical quantity features to obtain attention weights, fuse the image features and electrical quantity features according to the attention weights to obtain fused features, and perform fully connected processing on the fused features to obtain the fault location identification results output by the fault location identification model.

[0165] Each module in the aforementioned distributed photovoltaic safety operation and maintenance device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0166] In one exemplary embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 6 As shown, the computer device includes a processor, memory, input / output interfaces, a communication interface, a display unit, and an input device. The processor, memory, and input / output interfaces are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interfaces. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The input / output interfaces are used for exchanging information between the processor and external devices. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, Near Field Communication (NFC), or other technologies. When the computer program is executed by the processor, it implements a distributed photovoltaic safety operation and maintenance method. The display unit is used to form a visually visible image and can be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0167] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0168] In one exemplary embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0169] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above-described method embodiments.

[0170] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0171] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0172] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile memory and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, artificial intelligence (AI) processors, etc., and are not limited to these.

[0173] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0174] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for safe operation and maintenance of distributed photovoltaic systems, characterized in that, Applied to edge systems, the method includes: Acquire electrical quantity data and environmental parameters of photovoltaic modules in the edge system; Based on the electrical quantity data and the environmental parameters, a trained support vector machine is used to evaluate latency sensitivity and obtain a target latency sensitivity level. The latency sensitivity evaluation is used to assess the sensitivity of the edge system to data transmission latency and processing latency. The higher the target latency sensitivity level, the higher the real-time and low-latency requirements of the edge system for fault identification and data transmission. Based on the target latency sensitivity level, communication bandwidth resources, and device load rate, a trained decision network is used to determine the decision action space; the decision action space includes the sampling frame rate, data compression ratio, and model index. The model index is sent to the cloud server to instruct the cloud server to issue a fault location identification model that matches the model index. Receive the fault location identification model matched by the model index, acquire the image data and electrical quantity sequence transmitted at the sampling frame rate and the data compression ratio, use the fault location identification model to identify the fault location, and obtain the fault location identification result; The step of evaluating latency sensitivity using a trained support vector machine based on the electrical quantity data and the environmental parameters to obtain a target latency sensitivity level includes: Determine the electrical characteristics of the electrical quantity data; the electrical characteristics include at least one of standard deviation, absolute rate of change, and energy factor. Construct a system state feature vector based on the environmental parameters and electrical characteristics; The trained support vector machine includes multiple independent support vector machines, which are used to evaluate multiple latency sensitivity levels respectively. For each support vector machine, the system state feature vector is input to obtain the decision function value of the delay sensitivity level corresponding to the support vector machine; the multiple delay sensitivity levels include normal operation level, operation level requiring fault analysis and identification, and emergency operation level; the decision function refers to the classification discriminant function of a single support vector machine, and the decision function value indicates the degree of matching between the system state feature vector and the corresponding delay sensitivity level; The target latency sensitivity level is determined among multiple latency sensitivity levels based on multiple decision function values.

2. The method according to claim 1, characterized in that, The step of determining the decision action space using a trained decision network based on the target latency sensitivity level, communication bandwidth resources, and device load rate includes: The allowable latency corresponding to the target latency sensitivity level is determined by a preset mapping function; Obtain communication bandwidth resources and device load rate, and use the target latency sensitivity level, the allowable latency, the communication bandwidth resources, and the device load rate as the state space; Based on the state space and the trained decision network, the decision action space is determined.

3. The method according to claim 1, characterized in that, The steps for determining the trained decision network include: Obtain state space samples, and based on the state space samples and the initial Actor network, obtain the predicted action space; Obtain the total delay and the next state space after the predicted action space is executed, and determine the reward value based on the total delay and the preset reward function; Determine the target Q value based on the next state space and the target network; Based on the state space samples, the predicted action space, and the initial Critic network, the predicted Q value is obtained; The network parameters of the initial Critic network are updated based on the difference between the target Q value and the predicted Q value to obtain the updated Critic network. Based on the predicted action space and the updated Critic network, the updated Q value is obtained; With the goal of maximizing the updated Q value, the network parameters of the initial Actor network are updated to obtain the updated Actor network, which is then used as the trained decision network.

4. The method according to claim 1, characterized in that, The step of using the fault location identification model to identify fault locations and obtaining fault location identification results includes: Image features of image data transmitted at the sampling frame rate and the data compression ratio are extracted using the visual branch network, and electrical quantity features of electrical quantity sequences transmitted at the sampling frame rate and the data compression ratio are extracted using the temporal branch network. The image features and electrical quantity features are input into the fault location identification model. Attention processing is performed on the image features and electrical quantity features to obtain attention weights. The image features and electrical quantity features are fused according to the attention weights to obtain fused features. Fully connected processing is performed on the fused features to obtain the fault location identification result output by the fault location identification model.

5. A distributed photovoltaic safety operation and maintenance device, characterized in that, The device includes: The acquisition module is used to acquire electrical quantity data and environmental parameters of photovoltaic modules in the edge system; The evaluation module is used to evaluate latency sensitivity using a trained support vector machine based on the electrical quantity data and the environmental parameters, and obtain a target latency sensitivity level. The latency sensitivity evaluation is used to assess the sensitivity of the edge system to data transmission latency and processing latency. The higher the target latency sensitivity level, the higher the real-time and low-latency requirements of the edge system for fault identification and data transmission. The decision module is used to determine the decision action space based on the target latency sensitivity level, communication bandwidth resources, and device load rate, using a trained decision network; the decision action space includes the sampling frame rate, data compression ratio, and model index. The matching module is used to send the model index to the cloud server to instruct the cloud server to issue the fault location identification model that matches the model index; The identification module is used to receive the fault location identification model matched by the model index, acquire the image data and electrical quantity sequence transmitted at the sampling frame rate and the data compression ratio, use the fault location identification model to identify the fault location, and obtain the fault location identification result. The evaluation module is further configured to determine the electrical characteristics of the electrical quantity data; the electrical characteristics include at least one of standard deviation, absolute rate of change, and energy factor; construct a system state feature vector based on the environmental parameters and the electrical characteristics; the trained support vector machine includes multiple independent support vector machines, which are used to evaluate multiple latency sensitivity levels respectively; input the system state feature vector to each support vector machine to obtain the decision function value of the latency sensitivity level corresponding to the support vector machine; the multiple latency sensitivity levels include normal operation level, operation level requiring fault analysis and identification, and emergency operation level; the decision function refers to the classification discriminant function of a single support vector machine, and the decision function value indicates the degree of matching between the system state feature vector and the corresponding latency sensitivity level; and determine the target latency sensitivity level among the multiple latency sensitivity levels based on the multiple decision function values.

6. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

8. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 4.

Images