A photovoltaic array dynamic reconstruction method and device, electronic equipment and storage medium

Abstract

The application discloses a photovoltaic array dynamic reconstruction method and device, electronic equipment and storage medium, relates to the photovoltaic power generation technical field, and includes: training an initial reconstruction model on a cloud server; the initial reconstruction model is compressed to generate a lightweight reconstruction model; the lightweight reconstruction model is deployed to an edge computing node; the edge computing node collects photovoltaic data in real time, calls the lightweight reconstruction model to generate a reconstruction strategy; the edge computing node controls the switch matrix to perform topology switching according to the reconstruction strategy. The technical effect of realizing low-delay response of photovoltaic array dynamic reconstruction, improving system operation reliability and data security, and reducing edge deployment computing power and storage threshold is achieved.

Description

Technical Field

[0001] This application relates to the field of photovoltaic power generation technology, and in particular to a method, apparatus, electronic device and storage medium for dynamic reconfiguration of a photovoltaic array. Background Technology

[0002] In recent years, artificial intelligence (AI) technology has been widely applied in fields such as photovoltaic (PV) power prediction and fault diagnosis. Introducing AI into the dynamic reconfiguration decision-making of PV arrays can achieve global optimization results superior to traditional rule-based methods. Currently, PV AI deployments mainly adopt cloud-based solutions and edge lightweight solutions. Cloud-based solutions have significant drawbacks: data transmission, cloud queuing, and result feedback lead to uncontrollable total latency, making it unable to meet the real-time reconfiguration requirements under scenarios of rapid cloud movement and sudden irradiance changes. Furthermore, system reliability is highly dependent on the network, and failure occurs directly when the network is down or congested. Simultaneously, the massive amount of raw data uploaded consumes significant bandwidth and poses security risks related to data privacy leaks. Existing edge lightweight technologies mostly focus on single tasks such as anomaly detection and power prediction, failing to fully consider the special needs of dynamic PV reconfiguration. Complex deep learning models have high requirements for computing power and storage, while edge computing nodes in PV power plants mostly use low-cost MCUs or lightweight FPGAs, with limited computing power, storage resources, and strict power consumption constraints, making it difficult to support real-time inference of complex models and achieve efficient dynamic reconfiguration of PV arrays. Summary of the Invention

[0003] This application provides a method, apparatus, electronic device, and storage medium for dynamic reconfiguration of a photovoltaic array. It can solve the problems in related technologies, such as high latency, strong network dependence, large bandwidth consumption, and high data privacy risks associated with cloud-based AI deployments, as well as the limited computing power of edge nodes, which cannot support real-time inference of complex AI models.

[0004] According to a first aspect of this application, a method for dynamic reconfiguration of a photovoltaic array is provided, comprising:

[0005] The initial reconstructed model is trained on a cloud server; The initial reconstruction model is compressed to generate a lightweight reconstruction model; Deploy the lightweight reconfiguration model to edge computing nodes; The edge computing node collects photovoltaic data in real time and calls the lightweight reconfiguration model to generate a reconfiguration strategy. The edge computing node controls the switch matrix to perform topology switching according to the reconstruction strategy.

[0006] According to a second aspect of this application, a photovoltaic array dynamic reconfiguration device is provided, comprising: The training module is configured to train the initial reconstructed model on a cloud server; A compression module is configured to compress the initial reconstructed model to generate a lightweight reconstructed model; The deployment module is configured to deploy the lightweight reconfiguration model to an edge computing node; The generation module is configured to collect photovoltaic data in real time from the edge computing node and call the lightweight reconfiguration model to generate a reconfiguration strategy. The execution module is configured so that the edge computing node controls the switch matrix to perform topology switching according to the reconstruction strategy.

[0007] According to a third aspect of this application, an electronic device is provided, comprising: At least one processor; and memory that is communicatively connected to at least one processor; The memory stores instructions that can be executed by at least one processor, which are executed by at least one processor to enable the at least one processor to perform the photovoltaic array dynamic reconfiguration method described in the first aspect above.

[0008] According to a fourth aspect of this application, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute the photovoltaic array dynamic reconfiguration method described in the first aspect above.

[0009] According to a fifth aspect of this application, a computer program product is provided, including a computer program that, when executed by a processor, implements the photovoltaic array dynamic reconfiguration method as described in the first aspect above.

[0010] This application addresses the challenges of cloud-based AI deployment, which involves training an initial reconstruction model, compressing it to generate a lightweight reconstruction model, and then deploying it to edge computing nodes. These edge computing nodes collect photovoltaic data in real time and use the lightweight model to generate reconstruction strategies, directly controlling the switching matrix to perform topology switching. This solves the problems associated with cloud-based AI deployment, such as high latency, strong network dependence, high bandwidth consumption, and high data privacy risks, as well as the limitations of edge node computing resources in supporting real-time inference of complex AI models. The goal is to achieve low-latency response for dynamic reconstruction of photovoltaic arrays, improve system reliability and data security, and simultaneously reduce the computational and storage barriers for edge deployment.

[0011] It should be understood that the description in this section is not intended to identify key or essential features of the embodiments of this application, nor is it intended to limit the scope of this application. Other features of this application will become readily apparent from the following description. Attached Figure Description

[0012] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating a photovoltaic array dynamic reconfiguration method provided in an embodiment of this application. Figure 2 This is a schematic diagram of a photovoltaic array dynamic reconfiguration device provided in an embodiment of this application. Detailed Implementation

[0014] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of this application, including various details to aid understanding. These should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this application. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0015] The following description, with reference to the accompanying drawings, describes a photovoltaic array dynamic reconfiguration method, apparatus, electronic device, and storage medium according to embodiments of this application.

[0016] Figure 1 This is a flowchart illustrating a photovoltaic array dynamic reconfiguration method provided in an embodiment of this application.

[0017] like Figure 1 As shown, the method includes the following steps: Step 101: Train the initial reconstruction model on the cloud server.

[0018] In some embodiments, the cloud server employs a high-performance server cluster with sufficient computing power and storage resources to support large-scale data processing and the entire training process of deep learning models. The initial model reconstruction for training begins with data acquisition and preprocessing. This involves collecting historical operational data from the entire photovoltaic power plant lifecycle, specifically including irradiance, ambient temperature, output voltage and current parameters of each photovoltaic string under different seasons and weather conditions, as well as array topology operation records and power generation data under corresponding operating conditions. The collected raw data is then cleaned to remove outliers and missing data caused by sensor malfunctions or communication interruptions. Parameters of different dimensions are then normalized to eliminate the interference of data scale differences on model training. Based on the processed standard data, a training sample set conforming to the model input format is constructed.

[0019] Based on the established training sample set, an initial reconstruction model adapted to the dynamic reconfiguration task of photovoltaic arrays is built. The model's input consists of the current electrical quantities and environmental quantities of the photovoltaic strings, along with their corresponding historical sequence data. The model's output is the optimal topology code of the photovoltaic array's switch matrix, which can be directly mapped to the on / off state of each switch in the switch matrix. Supervised learning is used to iteratively train the initial reconstruction model, enabling it to learn the intrinsic mapping relationship between the photovoltaic array's operating state and the optimal reconfiguration topology, thus possessing the basic ability to generate globally optimal reconfiguration decisions based on power plant operating data.

[0020] By leveraging the powerful computing capabilities of cloud servers to complete the initial model training, the operational patterns in massive amounts of historical data can be fully explored, ensuring that the initial reconstructed model has high decision-making accuracy and providing a reliable model foundation for subsequent real-time inference at the edge.

[0021] Step 102: Compress the initial reconstruction model to generate a lightweight reconstruction model.

[0022] In some embodiments, compressing the initial reconstructed model reduces its computational and storage requirements, making it adaptable to the hardware resource constraints of edge computing nodes. This step utilizes multiple model compression techniques to generate a lightweight reconstructed model suitable for edge deployment while preserving as much decision-making accuracy as possible from the initial model.

[0023] Knowledge distillation is one of the core compression techniques. It uses the initially reconstructed model, after training, as the teacher network to construct a lightweight student network with a simpler structure and fewer parameters. By designing a specialized loss function, the student network is guided to learn the output distribution and feature representation of the teacher network, enabling the student network to approximate the reconstruction decision performance of the teacher network.

[0024] Structured pruning is used to further simplify the model structure. By evaluating the contribution of each convolutional kernel and channel to the model output, redundant parameters with contributions below a preset threshold are removed. This method can directly reduce the size of the model's computation graph, thereby reducing the computational load and memory access overhead during inference.

[0025] INT8 quantization converts the weights and activation values, originally represented as 32-bit floating-point numbers, into 8-bit integers for storage and computation. The quantization process determines the optimal quantization parameters by calibrating the dataset, minimizing accuracy loss and enabling the model to directly adapt to the fixed-point computing architecture of edge hardware.

[0026] By combining the above compression techniques, the number of parameters in the final lightweight reconstructed model is reduced by more than 80% compared to the initial model, while maintaining the same global optimization decision-making capability as the initial model, which can meet the deployment requirements of low computing power and low storage at the edge.

[0027] By employing multi-dimensional collaborative model compression technology, the number of model parameters and computational complexity are significantly reduced while ensuring the accuracy of reconstruction decisions. This effectively solves the problem of deploying complex deep learning models on edge devices and provides a feasible model foundation for real-time reconstruction inference at the edge.

[0028] Step 103: Deploy the lightweight reconfiguration model to the edge computing node.

[0029] In some embodiments, edge computing nodes are deployed at the photovoltaic power plant site and can be integrated into string controllers, smart shut-off devices, or combiner boxes. The core hardware uses a field-programmable gate array (FPGA) or a neural processing unit (NPU). This type of hardware features low power consumption and high parallel computing capabilities, making it suitable for the harsh operating environment of photovoltaic power plants. It meets the power consumption constraints of edge devices, which are typically within a few watts, and is compatible with PoE or battery power supply methods.

[0030] The lightweight refactored model generated in the cloud is converted into an instruction set format supported by the target hardware, completing model compilation and hardware adaptation optimization, enabling the model to fully utilize the computing resources of the edge hardware. The compiled lightweight refactored model is then stored in the non-volatile memory of the edge computing node, ensuring that the model is fully preserved after the device is powered off and restarted. During the power-on initialization of the edge computing node, the lightweight refactored model is automatically loaded from the non-volatile memory into the runtime memory, completing the initialization of model parameters and the configuration of the runtime environment, putting the edge node into an inference-ready state.

[0031] Deploying the lightweight reconfiguration model directly to the field edge computing node achieves physical proximity between the model and the execution end, eliminating the latency risks caused by cloud-edge communication from an architectural perspective. At the same time, it reduces the system's dependence on network connectivity, effectively improving the operational reliability and environmental adaptability of the photovoltaic array dynamic reconfiguration system.

[0032] Step 104: The edge computing node collects photovoltaic data in real time and calls the lightweight reconfiguration model to generate a reconfiguration strategy.

[0033] In some embodiments, edge computing nodes collect real-time photovoltaic power plant operation data through a sensor network, with a sampling rate of no less than 10Hz, ensuring the ability to capture rapid changes in operating conditions under scenarios of sudden irradiance changes, such as rapid cloud movement. The collected basic data includes the output voltage and current of each photovoltaic string, as well as the irradiance and ambient temperature of the site environment. Considering the unique electrical characteristics of perovskite photovoltaic modules, the edge computing nodes extract key features from the collected basic data. Voltage drift rate reflects the slow voltage changes in perovskite modules due to ion migration; temperature coefficient changes indicate the degree of aging of the module material; dynamic response time reflects the speed of current response changes during sudden irradiance changes; and low-frequency noise spectrum is a unique operating characteristic of perovskite modules. These features enhance the model's adaptability to different types of photovoltaic modules.

[0034] Edge computing nodes preprocess the collected and extracted feature data, performing data normalization and format conversion to meet the input requirements of the lightweight reconstruction model. The preprocessed feature data is then input into the lightweight reconstruction model, utilizing the fixed-point arithmetic acceleration engine deployed within the edge computing nodes to achieve efficient parallel computation of convolutional and fully connected layers, with the latency of the entire inference process controlled within 10 milliseconds. Based on the input photovoltaic data at the current moment and its historical sequence, the model outputs the target state code of the switch matrix. This code represents the optimal reconstruction strategy for the photovoltaic array and can be directly mapped to the on / off commands of each switch in the switch matrix.

[0035] By completing data acquisition and model inference locally at edge nodes, real-time perception of the photovoltaic array's operating status and rapid generation of reconstruction decisions are achieved, meeting the core requirement of millisecond-level response under sudden irradiation scenarios. At the same time, it avoids the risks of bandwidth consumption and data privacy leakage caused by uploading massive amounts of raw data to the cloud.

[0036] Step 105: The edge computing node controls the switch matrix to perform topology switching according to the reconstruction strategy.

[0037] In some embodiments, the edge computing node receives the reconstruction strategy output by the lightweight reconstruction model, parses the target state code of the switch matrix, generates on / off control commands for each switch unit, and sends them to the high-voltage, high-current switch matrix through a dedicated communication interface. The high-voltage, high-current switch matrix, as the execution unit of the photovoltaic array, is directly connected to the output of each photovoltaic string, enabling flexible adjustment of the connection topology between strings.

[0038] The topology switching process employs a "connect first, disconnect later" control strategy. Edge computing nodes, following a preset switching sequence, first close the switch of the target path to establish a new current loop, and then disconnect the switch of the original redundant path. This strategy avoids power outages during switching, ensuring the continuous and stable operation of the photovoltaic power station and preventing inverter shutdowns or grid disturbances caused by momentary power outages.

[0039] Edge computing nodes monitor the execution status of the switching matrix in real time. After confirming that all switching units have completed their actions according to instructions, they update the current photovoltaic array topology record, providing basic data for subsequent operational status monitoring and reconfiguration decisions. Through topology switching, the series and parallel combination of photovoltaic strings can be dynamically adjusted, effectively mitigating power generation losses caused by module mismatch and improving the overall power generation efficiency of the photovoltaic power station.

[0040] By directly controlling the switching matrix through edge nodes to perform topology switching, the reconfiguration decision can be implemented quickly. The "connect first, disconnect later" strategy ensures power supply continuity, while dynamically adjusting the array topology can effectively reduce mismatch losses and improve the power generation efficiency of photovoltaic power plants.

[0041] Compared with related technologies, in this embodiment, an initial reconstruction model is trained on a cloud server; the initial reconstruction model is compressed to generate a lightweight reconstruction model; the lightweight reconstruction model is deployed to edge computing nodes; the edge computing nodes collect photovoltaic data in real time and call the lightweight reconstruction model to generate a reconstruction strategy; the edge computing nodes control the switching matrix to perform topology switching according to the reconstruction strategy. This can solve the problems in related technologies, such as high latency, strong network dependence, large bandwidth consumption, and high data privacy risks in cloud AI deployment, as well as the limited computing power resources of edge nodes that cannot support real-time inference of complex AI models. It achieves low-latency response for dynamic reconstruction of photovoltaic arrays, improves system reliability and data security, and reduces the technical barriers to computing power and storage for edge deployment.

[0042] As a specific implementation of this application, based on the basic scheme, the lightweight reconstruction model is further defined as being designed based on a temporal convolutional network or a Mamba architecture, wherein the Mamba architecture is a sequence modeling network based on a state-space model.

[0043] Specifically, the lightweight reconstruction model is designed using either a Temporal Convolutional Network (TCN) or Mamba architecture. Both architectures are optimized for the time-series data processing requirements of dynamic reconstruction of photovoltaic arrays, featuring low complexity and high inference efficiency, and can adapt to the hardware resource constraints of edge computing nodes.

[0044] Temporal convolutional networks model temporal data based on convolution operations. They employ causal convolutional structures to ensure temporal causality, guaranteeing that the model makes decisions using only historical and current data, thus avoiding logical errors caused by introducing future information. By expanding the model's receptive field through dilated convolution, they can effectively capture long-term temporal dependencies in photovoltaic operation data while maintaining high parallelism of convolutional operations, resulting in fast inference speed and low computational resource consumption.

[0045] The Mamba architecture is a sequence modeling network based on a state-space model, solving the problems of vanishing gradients in traditional recurrent neural networks and the high computational complexity of the Transformer. Its computational complexity is linear O(n), significantly lower than the Transformer's O(n²) complexity, with particularly significant advantages when handling long sequence inputs, reducing the number of parameters by more than 80%. Borrowing from the SegMamba approach, it employs a parallel multi-branch structure and a channel-sharing strategy, further simplifying the model structure and reducing computational and storage requirements while maintaining model expressiveness and reconstruction decision accuracy.

[0046] By employing temporal convolutional networks or Mamba architecture to design lightweight reconstruction models, the number of model parameters and computational complexity can be significantly reduced while ensuring the accuracy of reconstruction decisions. This fully adapts to the low computing power and low storage resource constraints of edge computing nodes, providing efficient architectural support for achieving millisecond-level real-time inference at the edge.

[0047] As a specific implementation of this application, based on the basic scheme, the compression of the initial reconstruction model is further defined to include one or more of the following methods: Knowledge distillation: Using the initial reconstructed model as the teacher network, a lightweight student network is trained; Structured pruning: Remove convolutional kernels or channels whose contribution to the output is below a threshold; INT8 quantization: Quantizes model weights and activations into 8-bit integers.

[0048] Specifically, the compression of the initial reconstruction model can be implemented using one or more of the following methods: knowledge distillation, structured pruning, and INT8 quantization. Different combinations of technologies can be flexibly adjusted according to the computing power and storage constraints of edge computing nodes to achieve the optimal balance between the accuracy of reconstruction decisions and the lightweight nature of the model.

[0049] Knowledge distillation uses the trained initial reconstruction model as the teacher network to construct a lightweight student network with significantly reduced parameter and computational costs. During training, the soft labels containing class probability distributions output by the teacher network are combined with the real topological labels of the photovoltaic array reconstruction task as a supervision signal. This guides the student network to learn the deep operational features and global decision-making logic extracted by the teacher network, rather than simply fitting discrete real labels. This allows the student network to approximate the reconstruction decision-making performance of the teacher network to the greatest extent possible while significantly simplifying the structure.

[0050] Structured pruning simplifies the structure of trained lightweight networks. By calculating the contribution of each convolutional kernel and channel to the model output on the validation set, a reasonable contribution threshold is set, and all redundant convolutional kernels and channels with contributions below the threshold are removed. This method directly regularizes and prunes the model's computation graph, avoiding sparse computational overhead and directly reducing the model's inference computation and memory usage, while maintaining the regularity of the model structure, facilitating subsequent hardware adaptation and acceleration.

[0051] INT8 quantization converts the model into a fixed-point computation format. The model is calibrated using a calibration dataset based on typical photovoltaic power plant operating conditions, and the distribution range of activation values ​​at each layer is statistically analyzed to determine the optimal quantization scaling factor and zero-point parameters. The weight parameters, originally stored as 32-bit floating-point numbers, and the runtime activation values ​​are uniformly converted to 8-bit integers for storage and computation. Error compensation techniques are employed during quantization to minimize accuracy loss and enable the model to directly adapt to the fixed-point computation architecture of edge hardware.

[0052] Through the synergistic application of the three compression techniques mentioned above, the number of parameters in the final lightweight reconstructed model is reduced by more than 80% compared to the initial model, while maintaining a global optimization decision-making capability comparable to the initial model.

[0053] Employing a variety of flexible model compression techniques, it can adapt to the resource constraints of different edge hardware. While ensuring the accuracy of reconstruction decisions, it achieves extreme model lightweighting, effectively solving the problem that complex deep learning models are difficult to deploy on resource-constrained edge devices, and providing a solid model foundation for real-time execution of reconstruction inference at the edge.

[0054] As a specific embodiment of this application, based on the basic solution, the embodiments of this application further include: The edge computing nodes cache real-time data and periodically upload it to the cloud server; The cloud server uses the uploaded incremental data to fine-tune the lightweight reconstruction model and generate an updated model. The cloud server will distribute the updated model to the edge computing node to replace the original model.

[0055] Specifically, while performing real-time reconfiguration tasks, edge computing nodes locally cache recently collected photovoltaic operation data and corresponding reconfiguration strategy implementation effect data. Operation data includes basic parameters such as string voltage, current, irradiance, and temperature, while reconfiguration effect data includes evaluation indicators such as changes in power generation before and after reconfiguration and the reduction in mismatch losses. Edge nodes periodically upload the cached incremental data to the cloud server according to a preset cycle, which can be set daily or weekly, flexibly adjusted according to the rate of change in the power plant's operating conditions. This combination of local caching and periodic uploading significantly reduces the amount of data in cloud-edge communication, lowers bandwidth consumption, and avoids the network load pressure caused by real-time uploads.

[0056] After receiving incremental data uploaded by edge nodes, the cloud server cleans and preprocesses the data, using it as new training samples to incrementally fine-tune the lightweight reconstructed model. Incremental training does not require re-initializing model parameters; it only updates parameters within a small range based on the original model. This allows for rapid adaptation to the latest operating conditions of photovoltaic power plants, including performance degradation due to component aging, changes in irradiance and temperature patterns caused by seasonal variations, and changes in local shading. After fine-tuning, the cloud server can compress the updated model again as needed to ensure that the model always maintains its lightweight characteristics suitable for edge deployment.

[0057] The cloud server distributes the generated updated lightweight reconfiguration model to the corresponding edge computing nodes. The edge nodes load the new model using a hot-loading method, without interrupting current tasks or restarting the equipment, enabling seamless model replacement during the continuous power generation of the photovoltaic power plant. After the model replacement is complete, the edge nodes will use the updated model to generate subsequent reconfiguration decisions, achieving continuous optimization of model performance.

[0058] By constructing a complete closed loop of edge-cached data, cloud-based incremental fine-tuning, and model hot updates, the lightweight reconfiguration model can continuously adapt to the changing operating conditions of photovoltaic power plants over time. This effectively solves the problems of insufficient generalization ability and performance degradation of existing lightweight models after long-term operation, ensuring that the dynamic reconfiguration system of photovoltaic arrays can maintain a stable power generation efficiency improvement effect in the long term.

[0059] As a specific implementation of this application, based on the basic scheme, the edge computing node is further defined as an FPGA or an edge NPU, which is internally deployed with a fixed-point arithmetic acceleration engine to accelerate the inference calculation of the lightweight reconstructed model. The fixed-point computation acceleration engine includes a convolution accelerator, a fully connected accelerator, and an activation function unit.

[0060] Specifically, edge computing nodes use field-programmable gate arrays (FPGAs) or edge neural network processors (NNZs) as their core computing units. These two types of hardware feature low power consumption and high parallel computing capabilities, meeting the power consumption constraints of typical photovoltaic power plant edge devices (typically within a few watts). They are compatible with PoE or battery power supply methods and can operate stably in outdoor environments with high temperature, high humidity, and strong electromagnetic interference. FPGAs support flexible customization of hardware logic, enabling deep optimization for specific model structures; edge NNZs, on the other hand, have a built-in dedicated neural network operation instruction set, enabling efficient execution of deep learning inference tasks.

[0061] The edge computing nodes are internally equipped with a dedicated fixed-point arithmetic acceleration engine, specifically designed for the computational characteristics of lightweight reconstruction models to accelerate the model's inference computation process. This engine directly supports INT8 fixed-point arithmetic, avoiding the overhead of dequantization and keeping model inference latency below 10 milliseconds. The convolution accelerator employs a systolic array or row buffer structure to achieve high-throughput parallel computation of convolutional layers, capable of simultaneously processing convolution operations on multiple input and output channels, significantly improving the execution efficiency of convolutional layers. The fully connected accelerator uses a matrix-vector multiplication unit, supporting batch inference mode and enabling rapid completion of matrix operations in fully connected layers. The activation function unit uses a piecewise linear approximation to implement commonly used activation functions such as ReLU and Sigmoid, ensuring computational accuracy while reducing hardware implementation complexity and resource consumption.

[0062] By employing dedicated hardware adapted to edge scenarios and deploying a customized fixed-point computing acceleration engine, the parallel computing capabilities of the hardware can be fully utilized to achieve millisecond-level inference of the lightweight reconstruction model, meeting the real-time requirements of dynamic reconstruction of photovoltaic arrays under irradiation mutation scenarios, while maintaining extremely low operating power consumption and adapting to the stringent resource constraints of edge devices.

[0063] As a specific implementation of this application, based on the basic scheme, the photovoltaic data is further defined to include string voltage, current, irradiance, temperature, and perovskite-related features extracted from these data. The perovskite properties include one or more of the following: voltage drift rate, temperature coefficient variation, dynamic response time, and low-frequency noise spectrum.

[0064] Specifically, edge computing nodes, through sensor networks deployed at photovoltaic power plants, collect real-time basic data reflecting the operating status of photovoltaic arrays. With a sampling rate of at least 10Hz, they can accurately capture fluctuations in operating conditions under scenarios of rapid cloud movement and sudden changes in irradiance. This basic data includes the output voltage and current of each photovoltaic string, as well as the irradiance and ambient temperature of the site. This data is the core basis for assessing the operating status of the photovoltaic array and generating reconfiguration decisions.

[0065] To address the unique electrical characteristics of perovskite photovoltaic modules compared to crystalline silicon modules, edge computing nodes further extract perovskite-related features from the collected basic data as supplementary model inputs. Voltage drift rate reflects the slow change trend of the output voltage caused by the internal ion migration effect of the perovskite module, characterizing the module's instantaneous operational stability. Temperature coefficient variation reflects the real-time law of perovskite module output performance changing with temperature, effectively indicating the aging degree of the module material. Dynamic response time refers to the time required for the module's output current to reach a stable value when irradiation changes abruptly; its variation reflects changes in the module's internal carrier transport characteristics. Low-frequency noise spectrum is a unique operating characteristic of perovskite modules, directly related to the defect density and ion migration activity within the module. These features can be selected and used in combination, depending on the type of module used in the power plant.

[0066] By incorporating the unique operating characteristics of perovskite modules, the lightweight reconfiguration model can more accurately perceive the actual operating status of the perovskite photovoltaic array, effectively improving the model's adaptability to perovskite power plants and ensuring the accuracy and effectiveness of reconfiguration decisions under different types of photovoltaic module scenarios.

[0067] As a specific implementation of this application, based on the basic scheme, the input of the lightweight reconstruction model is further defined as the photovoltaic data at the current moment and its historical sequence, and the output is the target state code of the switching matrix.

[0068] Specifically, the lightweight reconfiguration model takes current photovoltaic (PV) data and its historical sequence as input, designed to fully adapt to the temporal correlation of PV array operating states. Environmental parameters such as irradiance and temperature, as well as the electrical parameters of the strings, change continuously over time. Using only single-moment data cannot accurately characterize the trend of operating conditions or predict the array's performance after reconfiguration. Historical sequence data reflects the PV array's operating patterns and dynamic characteristics over a period of time, helping the model to more comprehensively assess the overall mismatch of the array and generate more forward-looking and stable reconfiguration decisions. The length of the historical sequence can be flexibly adjusted according to the dynamic response characteristics of the PV array and the storage resources of edge computing nodes. Based on a sampling rate of at least 10Hz, selecting several seconds to tens of seconds of historical data is usually sufficient to meet the decision accuracy requirements. The input data is arranged chronologically to form a standardized sequence tensor, which is directly input into the model for inference calculations.

[0069] The output of the lightweight reconfiguration model is the target state code of the switch matrix, which directly maps the on / off state of each switch unit in the switch matrix. Each independent switch unit in the switch matrix corresponds to a bit in the code, using binary encoding, with 0 representing the switch open and 1 representing the switch closed. The target state code output by the model is a digital representation of the globally optimal topology, comprehensively considering the operating state and electrical characteristics of all photovoltaic strings, and can maximize the overall power generation of the array. Edge computing nodes do not need to perform complex format conversions and logical processing on the model output; they can directly parse the target state code to generate independent control commands for each switch, effectively reducing intermediate processing steps and improving the execution efficiency of reconfiguration decisions.

[0070] By introducing time-series historical data as model input, the accuracy and foresight of reconstruction decisions can be improved; by using direct mapping switch state coding as output, the instruction generation process of edge nodes is simplified, intermediate processing delays are reduced, and the real-time performance and efficiency of dynamic reconstruction of photovoltaic arrays are further guaranteed.

[0071] Figure 2 This is a schematic diagram of the structure of a photovoltaic array dynamic reconfiguration device provided in an embodiment of this application, as shown below. Figure 2 As shown, it includes: training module 201, compression module 202, deployment module 203, generation module 204, and execution module 205.

[0072] Training module 201 is configured to train the initial reconstructed model on a cloud server; Compression module 202 is configured to compress the initial reconstruction model to generate a lightweight reconstruction model; Deployment module 203 is configured to deploy the lightweight reconfiguration model to an edge computing node; The generation module 204 is configured to collect photovoltaic data in real time from the edge computing node and call the lightweight reconstruction model to generate a reconstruction strategy; The execution module 205 is configured to allow the edge computing node to perform topology switching by controlling the switch matrix according to the reconstruction strategy.

[0073] It should be noted that other corresponding descriptions of the functional units involved in the photovoltaic array dynamic reconfiguration device provided in this embodiment can be found in [reference]. Figure 1 The corresponding descriptions in [the document] will not be repeated here.

[0074] Based on the above, Figure 1 The embodiment illustrates a dynamic reconfiguration method for a photovoltaic array. Correspondingly, this embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method. Figure 1This illustrates a method for dynamic reconfiguration of a photovoltaic array.

[0075] Based on the above, Figure 1 The embodiment illustrates a dynamic reconfiguration method for a photovoltaic array. Correspondingly, this embodiment also provides a computer program product storing a computer program that, when executed by a processor, implements the aforementioned method. Figure 1 This illustrates a method for dynamic reconfiguration of a photovoltaic array.

[0076] Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as CD-ROM, USB flash drive, mobile hard drive, etc.) and includes several instructions to cause a computer device (such as personal computer, server, or network device, etc.) to execute the methods of various implementation scenarios of this application.

[0077] Based on the above, Figure 1 A method for dynamic reconfiguration of a photovoltaic array is shown, and Figure 2 To achieve the above objectives, the present application also provides an electronic device, such as a personal computer or a server, in the illustrated virtual device embodiment. This device includes a storage medium and a processor; the storage medium stores a computer program; the processor executes the computer program to implement the above-described virtual device. Figure 1 This illustrates a method for dynamic reconfiguration of a photovoltaic array.

[0078] In some embodiments, the aforementioned physical device may further include a user interface, a network interface, a camera, radio frequency (RF) circuitry, sensors, audio circuitry, a Wi-Fi module, etc. The user interface may include a display screen, an input unit such as a keyboard, etc., and optionally, a USB interface, a card reader interface, etc. In some embodiments, the network interface may include a standard wired interface, a wireless interface (such as a Wi-Fi interface), etc.

[0079] The storage medium may also include an operating system and a network communication module. The operating system is a program that manages the hardware and software resources of the aforementioned physical device, supporting the operation of information processing programs and other software and / or programs. The network communication module is used to enable communication between the various components within the storage medium, as well as communication with other hardware and software in the information processing physical device.

[0080] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0081] The above are merely specific embodiments of this application, enabling those skilled in the art to understand or implement this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to these embodiments, but is to be accorded the widest scope consistent with the principles and novel features claimed herein.

Claims

1. A method for dynamic reconfiguration of a photovoltaic array, characterized in that, include: The initial reconstructed model is trained on a cloud server; The initial reconstruction model is compressed to generate a lightweight reconstruction model; Deploy the lightweight reconfiguration model to edge computing nodes; The edge computing node collects photovoltaic data in real time and calls the lightweight reconfiguration model to generate a reconfiguration strategy. The edge computing node controls the switch matrix to perform topology switching according to the reconstruction strategy.

2. The photovoltaic array dynamic reconfiguration method according to claim 1, characterized in that, The lightweight reconstruction model is designed based on a temporal convolutional network or the Mamba architecture, where the Mamba architecture is a sequence modeling network based on a state-space model.

3. The photovoltaic array dynamic reconfiguration method according to claim 1, characterized in that, Compression of the initial reconstructed model includes one or more of the following methods: Knowledge distillation: Using the initial reconstructed model as the teacher network, a lightweight student network is trained; Structured pruning: Remove convolutional kernels or channels whose contribution to the output is below a threshold; INT8 quantization: Quantizes model weights and activations into 8-bit integers.

4. The photovoltaic array dynamic reconfiguration method according to claim 1, characterized in that, Also includes: The edge computing nodes cache real-time data and periodically upload it to the cloud server; The cloud server uses the uploaded incremental data to fine-tune the lightweight reconstruction model and generate an updated model. The cloud server will distribute the updated model to the edge computing node to replace the original model.

5. The photovoltaic array dynamic reconfiguration method according to claim 1, characterized in that, The edge computing node is an FPGA or an edge NPU, which is equipped with a fixed-point arithmetic acceleration engine to accelerate the inference computation of the lightweight reconstructed model. The fixed-point computation acceleration engine includes a convolution accelerator, a fully connected accelerator, and an activation function unit.

6. The photovoltaic array dynamic reconfiguration method according to claim 1, characterized in that, The photovoltaic data includes string voltage, current, irradiance, temperature, and perovskite-related features extracted from these data. The perovskite properties include one or more of the following: voltage drift rate, temperature coefficient variation, dynamic response time, and low-frequency noise spectrum.

7. The photovoltaic array dynamic reconfiguration method according to claim 1, characterized in that, The lightweight reconstruction model takes the current photovoltaic data and its historical sequence as input and outputs the target state code of the switching matrix.

8. A photovoltaic array dynamic reconfiguration device, characterized in that, include: The training module is configured to train the initial reconstructed model on a cloud server; A compression module is configured to compress the initial reconstructed model to generate a lightweight reconstructed model; The deployment module is configured to deploy the lightweight reconfiguration model to an edge computing node; The generation module is configured to collect photovoltaic data in real time from the edge computing node and call the lightweight reconfiguration model to generate a reconfiguration strategy. The execution module is configured so that the edge computing node controls the switch matrix to perform topology switching according to the reconstruction strategy.

9. An electronic device, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the photovoltaic array dynamic reconfiguration method according to any one of claims 1-7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute the photovoltaic array dynamic reconfiguration method according to any one of claims 1-7.

Images