A distributed photovoltaic system grid-connected control method and system

Abstract

The present invention relates to the field of photovoltaic power generation technology, and in particular to a distributed photovoltaic system grid-connected control method and system. The method comprises the following steps: obtaining historical multi-source heterogeneous data, including historical power output data, historical meteorological data, battery aging data and array parameter data; extracting multimodal environmental features from the historical multi-source heterogeneous data, and capturing the long-term dependency of the time series based on a bidirectional LSTM network to obtain a LSTM modeling result sequence; using 5G edge computing technology and FPGA acceleration technology to determine the MPPT setting value, reactive compensation ratio, and anti-islanding control parameters of the inverter according to the multimodal power curve data to obtain distributed control strategy data. The present invention realizes dynamic optimization of the MPPT strategy through dual alternating reinforcement learning.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power generation, and in particular to a distributed photovoltaic system grid-connected control method and system. Background Art

[0002] With the rapid development of renewable energy, the proportion of distributed photovoltaic systems in the power system continues to increase. Grid-connected control refers to the process of managing and regulating power generation equipment and related equipment when a distributed power generation system (such as a photovoltaic system) is connected to the main grid. Its purpose is to ensure that the distributed power generation system can operate safely, stably and efficiently in coordination with the main grid. However, traditional centralized control methods have difficulty in coping with the challenges brought about by large-scale distributed photovoltaic grid connection, such as grid stability, power quality and system efficiency. Therefore, a more intelligent and flexible grid-connected control method is needed to adapt to this new energy landscape.

[0003] However, traditional distributed photovoltaic system grid-connected control methods often have the following difficulties: due to the dynamic changes in light intensity and temperature, it is difficult for the MPPT algorithm to accurately capture the true maximum power point, and there are usually certain oscillations and tracking errors; large-scale distributed systems have a large number of inverters connected to the grid. If the inverter's impact-free switching in and out is not properly controlled, it will cause voltage fluctuations and affect the stability of the power grid. Summary of the invention

[0004] Based on this, it is necessary for the present invention to provide a distributed photovoltaic system grid-connected control method and system to solve at least one of the above technical problems.

[0005] To achieve the above object, a distributed photovoltaic system grid-connected control method comprises the following steps:

[0006] Step S1: Acquire historical multi-source heterogeneous data, including historical power output data, historical meteorological data, battery aging data, and array parameter data; perform multimodal environmental feature extraction on the historical multi-source heterogeneous data, and capture the long-term dependency of the time series based on a bidirectional LSTM network to obtain a LSTM modeling result sequence;

[0007] Step S2: deploy a dual alternating reinforcement learning framework according to the LSTM modeling result sequence, and perform knowledge penetration to obtain MPPT strategy data; generate multimodal power curve data according to the MPPT strategy data;

[0008] Step S3: Using 5G edge computing technology and FPGA acceleration technology to determine the MPPT setting value, reactive power compensation ratio, and anti-islanding control parameters of the inverter according to the multi-modal power curve data, and obtain distributed control strategy data;

[0009] Step S4: Perform VGAE-based grid modeling according to preset real-time grid topology data to obtain a grid topology model; perform timing optimization on the grid connection path according to the grid topology model, and decompose the path into multi-phase grid connection vector data; generate grid connection control strategy data according to the multi-phase grid connection vector data;

[0010] Step S5: Acquire electromagnetic sensor measurement data, and inversely resolve the electromagnetic imaging therein into harmonic source distribution, extract harmonic current paths, and obtain path feature sequences; generate harmonic compensation strategy data according to the path feature sequences;

[0011] Step S6: Perform Transformer-XL fusion on the distributed control strategy data, the grid-connected control strategy data and the harmonic compensation strategy data to obtain an initial modulation mode sequence; encode the initial modulation mode sequence into a control instruction, and send it to the inverter through the industrial bus to execute SPWM.

[0012] The acquisition of multi-source heterogeneous data in the present invention can provide multiple data perspectives, comprehensively consider the impact of different factors on the system, and enhance the comprehensiveness and accuracy of system modeling; multimodal environmental feature extraction can extract representative features from different data sources, which helps to reveal the inherent laws and characteristics of the system; the bidirectional LSTM network can capture long-term dependencies in time series data, improving the model's understanding and prediction capabilities of historical data; the LSTM modeling result sequence can be used as input for subsequent steps to provide a basis for system optimization and control. The dual alternating reinforcement learning framework can achieve maximum power tracking of solar photovoltaic systems by learning the optimal MPPT strategy through interaction with the environment; knowledge penetration can apply previously learned knowledge and experience to new environments to improve the efficiency and performance of reinforcement learning; the generation of MPPT strategy data can guide photovoltaic systems to achieve maximum power output under different environmental conditions and improve the energy utilization efficiency of the system; multimodal power curve data contains photovoltaic power output conditions under different environments, providing a basis for control decisions in subsequent steps. Based on 5G edge computing technology and FPGA acceleration technology, efficient data processing and real-time control can be achieved, and the system response speed and control accuracy can be improved; the MPPT setting value, reactive compensation ratio and anti-islanding control parameters of the inverter can be determined according to the multi-modal power curve data, which can optimize the performance and stability of the photovoltaic system; the distributed control strategy data contains the control parameter configuration for different environmental conditions, which can realize the adaptive control of the photovoltaic system and improve the adaptability and robustness of the system. The grid modeling based on VGAE can learn and model the topological structure of the grid, extract the key characteristics of the grid, and provide a basis for the formulation of subsequent control strategies; by optimizing the grid-connected path in time sequence and decomposing the path into multi-phase grid-connected vector data, the time-varying and dynamic characteristics of the grid can be considered, and the accuracy and real-time performance of the grid-connected control can be improved; the multi-phase grid-connected vector data can represent the grid status and characteristics in different time periods, and provide diversified information for the formulation and adjustment of control strategies; the grid-connected control strategy data contains control instructions for different grid states and paths, which can achieve effective coordination and optimization of the photovoltaic system and the grid. By obtaining electromagnetic information in the power grid through electromagnetic sensor measurement data, the harmonic problems and abnormal conditions in the power grid can be revealed; by reversing the electromagnetic imaging into harmonic source distribution and extracting the harmonic current path, the propagation path and impact range of the harmonics can be determined, providing a basis for harmonic compensation; the path feature sequence contains the key features of the harmonic propagation path, providing a basis for the formulation and implementation of harmonic compensation strategies; the harmonic compensation strategy data can guide the inverter to actively suppress and compensate for harmonics, thereby improving the power quality and stability of the power grid.By integrating distributed control strategy data, grid-connected control strategy data, and harmonic compensation strategy data through Transformer-XL, the trade-offs and coordination of different strategies can be comprehensively considered to improve the overall performance of the system; the initial modulation mode sequence contains guidance for inverter control, which can achieve precise control and modulation of the inverter; encoding the initial modulation mode sequence as a control instruction and sending it to the inverter through the industrial bus can achieve real-time transmission and execution of the control strategy to ensure the stability and reliability of the system. In summary, the effects of the above steps include improving the comprehensiveness and accuracy of system modeling, improving the energy utilization efficiency of the photovoltaic system, improving the coordination and stability of the photovoltaic system and the power grid, and improving the power quality and reliability of the power grid; these effects help to achieve maximum power tracking of the photovoltaic system, optimize the operation and management of the power grid, and improve energy utilization efficiency and sustainable development.

[0013] The present invention also provides a distributed photovoltaic system grid-connected control system for executing the above-mentioned distributed photovoltaic system grid-connected control method, wherein the distributed photovoltaic system grid-connected control system comprises:

[0014] The LSTM modeling module is used to obtain historical multi-source heterogeneous data, including historical power output data, historical meteorological data, battery aging data, and array parameter data; extract multimodal environmental features from historical multi-source heterogeneous data, and capture the long-term dependency of time series based on a bidirectional LSTM network to obtain a sequence of LSTM modeling results;

[0015] The MPPT strategy optimization module is used to deploy a dual alternating reinforcement learning framework based on the LSTM modeling result sequence and perform knowledge penetration to obtain MPPT strategy data; and to generate multi-modal power curve data based on the MPPT strategy data;

[0016] The distributed control parameter module is used to determine the MPPT setting value, reactive power compensation ratio, and anti-islanding control parameters of the inverter according to the multi-modal power curve data using 5G edge computing technology and FPGA acceleration technology to obtain distributed control strategy data;

[0017] The grid-connected path planning module is used to perform VGAE-based grid modeling according to the preset real-time grid topology data to obtain a grid topology model; perform timing optimization on the grid-connected path according to the grid topology model, and decompose the path into multi-phase grid-connected vector data; generate grid-connected control strategy data according to the multi-phase grid-connected vector data;

[0018] The harmonic compensation strategy module is used to obtain electromagnetic sensor measurement data, and inversely interpret the electromagnetic imaging therein into harmonic source distribution, extract the harmonic current path, and obtain the path feature sequence; and generate harmonic compensation strategy data according to the path feature sequence;

[0019] The multi-strategy fusion module is used to perform Transformer-XL fusion on the distributed control strategy data, the grid-connected control strategy data and the harmonic compensation strategy data to obtain an initial modulation mode sequence; the initial modulation mode sequence is encoded into a control instruction and sent to the inverter through the industrial bus to execute SPWM.

[0020] The present invention can capture the time series dependency between data by using LSTM network to model historical multi-source heterogeneous data. LSTM network can handle long-term dependency, which is helpful to extract the characteristics of historical power output data, meteorological data, battery aging data and array parameter data. The LSTM modeling result sequence can provide an effective representation of historical data, providing a basis for subsequent MPPT strategy optimization and control decision. According to the LSTM modeling result sequence, the MPPT strategy is optimized using a dual alternating reinforcement learning framework. This method can learn the optimal MPPT strategy through interaction with the environment to achieve the maximum power output of the photovoltaic system; through knowledge penetration, the system can continuously improve and optimize the strategy to improve the energy conversion efficiency of the system, and the generated multimodal power curve data can provide a comprehensive analysis and evaluation of the power output under different working conditions. Using 5G edge computing technology and FPGA acceleration technology, the MPPT setting value, reactive compensation ratio and anti-islanding control parameters of the inverter are determined according to the multimodal power curve data. By analyzing the multimodal power curve data, the photovoltaic system can be optimized and controlled, and the performance and stability of the system can be improved. The distributed control strategy data can be used to dynamically adjust the inverter according to the real-time situation to adapt to different working conditions and grid requirements. Based on the preset real-time grid topology data, VGAE (Variational Graph Autoencoder) is used to model the grid and obtain the grid topology model; through timing optimization and path decomposition, the grid-connected path can be determined and multi-phase grid-connected vector data can be generated, which helps to optimize the grid-connected behavior of the photovoltaic system and ensure its stable connection with the grid. The grid-connected control strategy data provides guidance for achieving safe and reliable grid connection. By obtaining electromagnetic sensor measurement data and inverting electromagnetic imaging into harmonic source distribution, harmonic current path extraction is performed, so that the propagation path and characteristics of harmonic current can be identified. The generated path feature sequence can be used to formulate harmonic compensation strategies to reduce harmonic pollution in the system and improve power quality. Through Transformer-XL, the information of different strategies can be integrated and integrated to obtain the initial modulation mode sequence; this helps to comprehensively consider the impact and requirements of different strategies, improve the overall performance and efficiency of the system, and encode the initial modulation mode sequence into control instructions, which are sent to the inverter through the industrial bus to execute SPWM to achieve precise control and regulation of the photovoltaic system. In summary, the effects of the above steps include: extracting historical data features, optimizing MPPT strategies, determining distributed control parameters, optimizing grid-connected path planning, implementing harmonic compensation strategies, and integrating multiple strategies; these effects help to improve the energy conversion efficiency, stability, and power quality of photovoltaic systems, maximize the use of photovoltaic resources, and interconnect with the power grid safely and reliably. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] Other features, objects and advantages of the present invention will become more apparent from the detailed description of non-limiting embodiments thereof made with reference to the following drawings:

[0022] Figure 1 A schematic diagram of the steps of the distributed photovoltaic system grid-connected control method of the present invention;

[0023] Figure 2 for Figure 1 Detailed step flow diagram of step S1;

[0024] Figure 3 for Figure 1 Detailed step flow chart of step S2 in FIG. DETAILED DESCRIPTION

[0025] The technical method of the present invention is described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by technicians in this field without creative work are within the scope of protection of the present invention.

[0026] In addition, the accompanying drawings are only schematic illustrations of the present invention and are not necessarily drawn to scale. The same reference numerals in the figures represent the same or similar parts, and their repeated description will be omitted. Some of the block diagrams shown in the accompanying drawings are functional entities and do not necessarily correspond to physically or logically independent entities. The functional entities can be implemented in software form, or implemented in one or more hardware modules or integrated circuits, or implemented in different networks and / or processor methods and / or microcontroller methods.

[0027] It should be understood that, although the terms "first", "second", etc. may be used herein to describe various units, these units should not be limited by these terms. These terms are used only to distinguish one unit from another unit. For example, without departing from the scope of the exemplary embodiments, the first unit may be referred to as the second unit, and similarly the second unit may be referred to as the first unit. The term "and / or" used herein includes any and all combinations of one or more of the listed associated items.

[0028] To achieve this, please refer to Figures 1 to 3 The present invention provides a distributed photovoltaic system grid-connected control method, the method comprising the following steps:

[0029] Step S1: Acquire historical multi-source heterogeneous data, including historical power output data, historical meteorological data, battery aging data, and array parameter data; perform multimodal environmental feature extraction on the historical multi-source heterogeneous data, and capture the long-term dependency of the time series based on a bidirectional LSTM network to obtain a LSTM modeling result sequence;

[0030] Step S2: deploy a dual alternating reinforcement learning framework according to the LSTM modeling result sequence, and perform knowledge penetration to obtain MPPT strategy data; generate multimodal power curve data according to the MPPT strategy data;

[0031] Step S3: Using 5G edge computing technology and FPGA acceleration technology to determine the MPPT setting value, reactive power compensation ratio, and anti-islanding control parameters of the inverter according to the multi-modal power curve data, and obtain distributed control strategy data;

[0032] Step S4: Perform VGAE-based grid modeling according to preset real-time grid topology data to obtain a grid topology model; perform timing optimization on the grid connection path according to the grid topology model, and decompose the path into multi-phase grid connection vector data; generate grid connection control strategy data according to the multi-phase grid connection vector data;

[0033] Step S5: Acquire electromagnetic sensor measurement data, and inversely resolve the electromagnetic imaging therein into harmonic source distribution, extract harmonic current paths, and obtain path feature sequences; generate harmonic compensation strategy data according to the path feature sequences;

[0034] Step S6: Perform Transformer-XL fusion on the distributed control strategy data, the grid-connected control strategy data and the harmonic compensation strategy data to obtain an initial modulation mode sequence; encode the initial modulation mode sequence into a control instruction, and send it to the inverter through the industrial bus to execute SPWM.

[0035] In the embodiment of the present invention, reference Figure 1 The above is a schematic flow chart of the steps of a distributed photovoltaic system grid-connected control method of the present invention. In this example, the distributed photovoltaic system grid-connected control method includes the following steps:

[0036] Step S1: Acquire historical multi-source heterogeneous data, including historical power output data, historical meteorological data, battery aging data, and array parameter data; perform multimodal environmental feature extraction on the historical multi-source heterogeneous data, and capture the long-term dependency of the time series based on a bidirectional LSTM network to obtain a LSTM modeling result sequence;

[0037] In the embodiment of the present invention, historical power output data can be obtained through the monitoring equipment or data logger of the photovoltaic system; historical meteorological data can be obtained from data sources such as meteorological stations and weather forecasts; battery aging data can be obtained by periodic testing or monitoring battery parameters; array parameter data can be obtained from the specification sheet, design documents or monitoring equipment of the photovoltaic system. For multimodal environmental feature extraction, various data processing and feature extraction methods can be used, such as statistical feature extraction, spectrum analysis, wavelet transform, etc., selected and combined according to the nature of the data; then, the extracted features are modeled using a bidirectional LSTM network to capture long-term dependencies in the time series, and the LSTM modeling result sequence can be used as input data for subsequent steps.

[0038] Step S2: deploy a dual alternating reinforcement learning framework according to the LSTM modeling result sequence, and perform knowledge penetration to obtain MPPT strategy data; generate multimodal power curve data according to the MPPT strategy data;

[0039] The embodiment of the present invention uses a reinforcement learning algorithm (such as Q-learning, DQN, etc.) to establish a dual alternating reinforcement learning framework, in which one agent is responsible for interacting with the environment and the other agent is responsible for learning and optimization. The LSTM modeling result sequence is used as a representation of the environmental state, the action space and reward function are defined, and the agent is trained to optimize the MPPT strategy and perform knowledge penetration, that is, to integrate existing knowledge and experience into the reinforcement learning framework to accelerate learning and improve the effect of strategy optimization. The generation of multimodal power curve data can be obtained by applying the MPPT strategy to historical meteorological data and array parameter data, and using the mathematical model and power output model of the photovoltaic system for simulation calculation.

[0040] Step S3: Using 5G edge computing technology and FPGA acceleration technology to determine the MPPT setting value, reactive power compensation ratio, and anti-islanding control parameters of the inverter according to the multi-modal power curve data, and obtain distributed control strategy data;

[0041] The embodiment of the present invention uses 5G edge computing technology to upload multimodal power curve data to the cloud for processing and calculation. The cloud can provide high-performance computing resources and distributed algorithm support; using FPGA acceleration technology, some computing tasks are delegated to hardware accelerators to improve computing efficiency and real-time performance; on the cloud or edge server, optimization algorithms and control strategies are applied to determine the MPPT setting value, reactive compensation ratio and anti-islanding control parameters of the inverter based on the multimodal power curve data, and optimization algorithms (such as genetic algorithms, particle swarm algorithms, etc.) can be used for parameter optimization.

[0042] Step S4: Perform VGAE-based grid modeling according to preset real-time grid topology data to obtain a grid topology model; perform timing optimization on the grid connection path according to the grid topology model, and decompose the path into multi-phase grid connection vector data; generate grid connection control strategy data according to the multi-phase grid connection vector data;

[0043] The embodiment of the present invention uses preset real-time topological data of the power grid to construct a power grid topological model; VGAE (Variational Graph Autoencoder) and other graph neural network methods can be used to model the power grid. According to the power grid topological model, the timing optimization method is used to optimize the grid-connected path, and genetic algorithms, dynamic programming and other methods can be used to consider factors such as power grid stability and line capacity to obtain the optimal grid-connected path. The optimized optimal grid-connected path is segmented according to time, and the grid-connected path in each time period is represented as a vector. The vector can contain information such as path nodes, line capacity, and node voltage to describe the grid-connected situation in the time period. According to the grid-connected vector data in each time period, a grid-connected control strategy is designed. The strategy may include node voltage control, line capacity adjustment, and distributed generation scheduling of distribution networks. According to specific needs and conditions, a suitable control algorithm and parameter settings are selected to generate corresponding grid-connected control strategy data.

[0044] Step S5: Acquire electromagnetic sensor measurement data, and inversely resolve the electromagnetic imaging therein into harmonic source distribution, extract harmonic current paths, and obtain path feature sequences; generate harmonic compensation strategy data according to the path feature sequences;

[0045] The embodiment of the present invention uses electromagnetic sensors to obtain electromagnetic sensor measurement data in the power system. These data may include measurement results of current, voltage, power, etc.; electromagnetic imaging technology is applied to the measurement data, and the distribution of harmonic sources can be obtained by inversely solving electromagnetic imaging, which can help determine the harmonic problems existing in the power grid. According to the distribution of harmonic sources, the harmonic current path characteristic sequence is extracted. The harmonic current path characteristic sequence describes the propagation path and distribution of harmonic currents in the power system. Based on the harmonic current path characteristic sequence, a harmonic compensation strategy is designed. The compensation strategy may include measures such as the design of harmonic filters, injection and suppression of harmonic currents. According to specific harmonic problems and requirements, appropriate compensation methods and parameter settings are selected to generate harmonic compensation strategy data.

[0046] Step S6: Perform Transformer-XL fusion on the distributed control strategy data, the grid-connected control strategy data and the harmonic compensation strategy data to obtain an initial modulation mode sequence; encode the initial modulation mode sequence into a control instruction, and send it to the inverter through the industrial bus to execute SPWM.

[0047] The embodiment of the present invention uses the Transformer-XL model to fuse the distributed control strategy data, the grid-connected control strategy data, and the harmonic compensation strategy data. Transformer-XL is a sequence modeling model based on the self-attention mechanism, which can process long-term dependencies and sequence information. The fused data is input into the Transformer-XL model, and the associations and patterns between the data are learned by pre-training or fine-tuning; the fused data is encoded into control instructions. According to the specific application scenario and control requirements, appropriate encoding rules and formats are designed to convert the fused data into control instructions that the inverter can understand and execute. Use an industrial bus (such as Modbus, CAN, etc.) to send the control instructions to the inverter to perform SPWM (Sinusoidal Pulse Width Modulation) control; SPWM control is a commonly used inverter control method, which controls the inverter output waveform by adjusting the pulse width.

[0048] The acquisition of multi-source heterogeneous data in the present invention can provide multiple data perspectives, comprehensively consider the impact of different factors on the system, and enhance the comprehensiveness and accuracy of system modeling; multimodal environmental feature extraction can extract representative features from different data sources, which helps to reveal the inherent laws and characteristics of the system; the bidirectional LSTM network can capture long-term dependencies in time series data, improving the model's understanding and prediction capabilities of historical data; the LSTM modeling result sequence can be used as input for subsequent steps to provide a basis for system optimization and control. The dual alternating reinforcement learning framework can achieve maximum power tracking of solar photovoltaic systems by learning the optimal MPPT strategy through interaction with the environment; knowledge penetration can apply previously learned knowledge and experience to new environments to improve the efficiency and performance of reinforcement learning; the generation of MPPT strategy data can guide photovoltaic systems to achieve maximum power output under different environmental conditions and improve the energy utilization efficiency of the system; multimodal power curve data contains photovoltaic power output conditions under different environments, providing a basis for control decisions in subsequent steps. Based on 5G edge computing technology and FPGA acceleration technology, efficient data processing and real-time control can be achieved, and the system response speed and control accuracy can be improved; the MPPT setting value, reactive compensation ratio and anti-islanding control parameters of the inverter can be determined according to the multi-modal power curve data, which can optimize the performance and stability of the photovoltaic system; the distributed control strategy data contains the control parameter configuration for different environmental conditions, which can realize the adaptive control of the photovoltaic system and improve the adaptability and robustness of the system. The grid modeling based on VGAE can learn and model the topological structure of the grid, extract the key characteristics of the grid, and provide a basis for the formulation of subsequent control strategies; by optimizing the grid-connected path in time sequence and decomposing the path into multi-phase grid-connected vector data, the time-varying and dynamic characteristics of the grid can be considered, and the accuracy and real-time performance of the grid-connected control can be improved; the multi-phase grid-connected vector data can represent the grid status and characteristics in different time periods, and provide diversified information for the formulation and adjustment of control strategies; the grid-connected control strategy data contains control instructions for different grid states and paths, which can achieve effective coordination and optimization of the photovoltaic system and the grid. By obtaining electromagnetic information in the power grid through electromagnetic sensor measurement data, the harmonic problems and abnormal conditions in the power grid can be revealed; by reversing the electromagnetic imaging into harmonic source distribution and extracting the harmonic current path, the propagation path and impact range of the harmonics can be determined, providing a basis for harmonic compensation; the path feature sequence contains the key features of the harmonic propagation path, providing a basis for the formulation and implementation of harmonic compensation strategies; the harmonic compensation strategy data can guide the inverter to actively suppress and compensate for harmonics, thereby improving the power quality and stability of the power grid.By integrating distributed control strategy data, grid-connected control strategy data, and harmonic compensation strategy data through Transformer-XL, the trade-offs and coordination of different strategies can be comprehensively considered to improve the overall performance of the system; the initial modulation mode sequence contains guidance for inverter control, which can achieve precise control and modulation of the inverter; encoding the initial modulation mode sequence as a control instruction and sending it to the inverter through the industrial bus can achieve real-time transmission and execution of the control strategy to ensure the stability and reliability of the system. In summary, the effects of the above steps include improving the comprehensiveness and accuracy of system modeling, improving the energy utilization efficiency of the photovoltaic system, improving the coordination and stability of the photovoltaic system and the power grid, and improving the power quality and reliability of the power grid; these effects help to achieve maximum power tracking of the photovoltaic system, optimize the operation and management of the power grid, and improve energy utilization efficiency and sustainable development.

[0049] Preferably, step S1 comprises the following steps:

[0050] Step S11: extracting the output power record of the inverter through the data collection and monitoring system of the photovoltaic system to obtain historical power output data;

[0051] Step S12: collecting meteorological factors such as light intensity, temperature, and wind speed at the distributed photovoltaic system site to obtain historical meteorological data;

[0052] Step S13: estimating the battery aging degree based on the electrochemical model according to the battery cycle number, temperature history and voltage data monitored and recorded by the battery management system, thereby obtaining battery aging data;

[0053] Step S14: acquiring array parameter data including the structure, layout, orientation and tilt angle parameters of the photovoltaic array;

[0054] Step S15: integrating the historical power output data, historical meteorological data, battery aging data, and array parameter data with the labeled data sources, thereby obtaining historical multi-source heterogeneous data;

[0055] Step S16: extract multimodal environmental features from historical multi-source heterogeneous data, and capture long-term dependencies of time series based on a bidirectional LSTM network to obtain a sequence of LSTM modeling results.

[0056] As an embodiment of the present invention, refer to Figure 2 As shown, Figure 1 Detailed step flow diagram of step S1 in the embodiment of the present invention, step S1 includes the following steps:

[0057] Step S11: extracting the output power record of the inverter through the data collection and monitoring system of the photovoltaic system to obtain historical power output data;

[0058] In the embodiment of the present invention, an inverter monitoring device is installed in the photovoltaic system to record the output power data of the inverter. The inverter monitoring system usually provides data collection and recording functions. The inverter monitoring system will regularly collect and record the output power data of the inverter. These data may include power output values ​​per minute, hour, or day. The specific collection frequency depends on the demand and equipment settings. The historical power output data can be extracted from the inverter monitoring system by accessing the data interface of the monitoring system or using the corresponding data extraction tool.

[0059] Step S12: collecting meteorological factors such as light intensity, temperature, and wind speed at the distributed photovoltaic system site to obtain historical meteorological data;

[0060] In the embodiment of the present invention, meteorological monitoring equipment is deployed on site in the distributed photovoltaic system, such as light intensity sensors, temperature sensors, and wind speed sensors, etc. These devices are used to monitor the meteorological factors of the environment in real time. The meteorological monitoring equipment will regularly collect data such as light intensity, temperature, and wind speed, and record them. The collection frequency can be set as needed. To extract historical meteorological data from the meteorological monitoring equipment, data can be obtained by accessing the data interface of the equipment or using corresponding data extraction tools.

[0061] Step S13: estimating the battery aging degree based on the electrochemical model according to the battery cycle number, temperature history and voltage data monitored and recorded by the battery management system, thereby obtaining battery aging data;

[0062] In the embodiment of the present invention, a battery management system device is installed in the photovoltaic system to monitor and record the status and parameters of the battery, such as the number of battery cycles, temperature and voltage. The battery management system will regularly collect and record the relevant data of the battery, which may include the cumulative value of the number of battery cycles, the temperature change history and the voltage measurement results. To extract the historical battery aging data from the battery management system, the data can be obtained by accessing the system's data interface or using the corresponding data extraction tool.

[0063] Step S14: acquiring array parameter data including the structure, layout, orientation and tilt angle parameters of the photovoltaic array;

[0064] The embodiment of the present invention determines characteristic information of the array according to parameters such as the structure, layout, orientation and tilt angle of the photovoltaic array; and organizes and records parameter data of the array, including the quantity, model, installation method and location information of each component.

[0065] Step S15: integrating the historical power output data, historical meteorological data, battery aging data, and array parameter data with the labeled data sources, thereby obtaining historical multi-source heterogeneous data;

[0066] The embodiment of the present invention integrates historical power output data, historical meteorological data, battery aging data and array parameter data, and can use data processing tools or scripts to match and combine data from different data sources to form a comprehensive data set.

[0067] Step S16: extract multimodal environmental features from historical multi-source heterogeneous data, and capture long-term dependencies of time series based on a bidirectional LSTM network to obtain a sequence of LSTM modeling results.

[0068] The embodiment of the present invention can extract features from historical multi-source heterogeneous data using methods such as statistical features, frequency domain features, and time domain features; for example, statistical features such as the average, variance, maximum, and minimum values ​​can be calculated, or Fourier transform can be performed to extract frequency domain features. A bidirectional LSTM (Long Short-Term Memory) network is used to model time series data to capture long-term dependencies in the data. The LSTM network is a recurrent neural network suitable for processing sequence data. It can effectively process long-term dependencies in time series data. Through the bidirectional LSTM, past and future information can be considered simultaneously, thereby better modeling and predicting the behavior of the system. The bidirectional LSTM network is used to train historical multi-source heterogeneous data to obtain a sequence of LSTM modeling results, which can include prediction results for future power output, meteorological conditions, battery aging, etc.

[0069] The present invention can understand the power changes of the photovoltaic system in different time periods by extracting the historical power output data of the inverter; these data can be used to analyze the performance of the system, evaluate the stability and reliability of the system, and provide basic data for subsequent steps. By collecting meteorological factors data such as light intensity, temperature, wind speed, etc., the meteorological conditions of the environment in which the photovoltaic system is located can be understood; these data are of great significance for the performance analysis, power prediction and system optimization of the photovoltaic system. The battery aging degree can be estimated by the battery cycle number, temperature history and voltage data monitored and recorded by the battery management system; these data can help judge the health status of the battery and provide a reference for the maintenance and management of the battery. The array parameter data such as the structure, layout, orientation and tilt angle of the photovoltaic array can be obtained to provide the design and configuration information of the photovoltaic system; these data are important for the performance analysis, power prediction and system optimization of the photovoltaic system. The historical power output data, historical meteorological data, battery aging data and array parameter data are integrated into historical multi-source heterogeneous data, which can comprehensively consider the operation of the photovoltaic system under different conditions; the integration of these data can provide a comprehensive system perspective and provide a more accurate and comprehensive data basis for the analysis and modeling of the subsequent steps. By extracting multimodal environmental features from historical multi-source heterogeneous data and capturing the long-term dependencies of time series based on a bidirectional LSTM network, a model of system behavior and performance can be established; the LSTM modeling result sequence can provide the ability to understand and predict historical data, providing a basis for the optimization and control of subsequent steps. In summary, the effects of the above steps include: providing an evaluation of the performance and stability of the photovoltaic system, understanding the environmental conditions of the system, estimating the degree of battery aging, providing design and configuration information for the photovoltaic system, comprehensively considering the system operation of multiple data sources, and establishing a model of system behavior and performance. These effects help optimize the performance of the photovoltaic system and improve the reliability and sustainable development level of the system.

[0070] Preferably, step S16 comprises the following steps:

[0071] Perform multi-modal environmental feature extraction based on multi-branch CNN on historical multi-source heterogeneous data to obtain a multi-modal feature vector sequence, where the multi-modal feature vector sequence is a three-dimensional vector, the first dimension is the time step, the second dimension is the feature dimension, and the third dimension is the batch dimension;

[0072] The embodiment of the present invention uses a deep learning framework to create multiple parallel convolutional neural network branches, each branch corresponds to an input data source, such as power output data, meteorological data, and battery aging data; the multi-branch CNN model is trained using annotated historical data; the back propagation algorithm and optimizer (such as Adam) can be used to update the parameters of the model so that it gradually learns to extract features. Use the trained multi-branch CNN model to extract features from historical data, feed the input data into the model time by time step, obtain the corresponding multimodal feature vector, and finally obtain a three-dimensional feature vector sequence, in which the first dimension is the time step, the second dimension is the feature dimension, and the third dimension is the batch dimension.

[0073] The multimodal feature vector sequence is input into the preset bidirectional LSTM network model for back-propagation training of the model parameters to obtain a sequence of LSTM modeling results, in which the forward LSTM layer of the bidirectional LSTM network model sequentially captures the forward time dependency, and the reverse LSTM layer captures the reverse time dependency. The LSTM hidden layer states in two directions are concatenated at the same time step and then connected to a fully connected layer for nonlinear mapping of the hidden layer states.

[0074] The embodiment of the present invention uses a deep learning framework to create a bidirectional LSTM network model, including a forward LSTM layer and a reverse LSTM layer, and can set appropriate hyperparameters such as the number of LSTM units and the size of the hidden layer. In the training and inference stages, the hidden state of the forward LSTM layer and the reverse LSTM layer at the same time step is spliced ​​to obtain a comprehensive hidden state; the spliced ​​hidden state is sent to the fully connected layer, and an activation function (such as ReLU) can be added to introduce nonlinear mapping. Select a suitable loss function (such as mean square error) to measure the difference between the LSTM modeling result sequence and the labeled result sequence, and select a suitable optimizer (such as Adam) to update the model parameters so that it gradually adapts to the pattern of historical data. Use the labeled historical data to train the bidirectional LSTM network model, and obtain the LSTM modeling result sequence by inputting the multimodal feature vector sequence into the model; according to the error calculated by the loss function, use the back propagation algorithm to update the model parameters to improve the prediction ability of the model.

[0075] The present invention uses multi-modal environment feature extraction based on multi-branch CNN to extract multi-modal feature vector sequences from historical multi-source heterogeneous data; this method can simultaneously consider the features of different data sources, and perform feature extraction through a multi-branch CNN network; the multi-branch CNN can effectively capture the correlation and feature representation between different data sources, and improve the accuracy and expression ability of feature extraction; the multi-modal feature vector sequence is used as input, and time information, feature information and batch information can be retained. The multi-modal feature vector sequence is trained and modeled by a bidirectional LSTM network model, and the long-term dependency of time series data can be captured. The forward LSTM layer and the reverse LSTM layer process the forward and reverse time dependencies respectively, so that the model can fully understand the time series information of the data. The LSTM hidden layer states in the two directions are spliced ​​at the same time step, and the forward and reverse information can be comprehensively considered. The hidden layer states are nonlinearly mapped through the fully connected layer, thereby obtaining a LSTM modeling result sequence. Multimodal environment feature extraction can effectively extract feature vector sequences with rich expressive power from historical multi-source heterogeneous data, including information from different data sources, which helps to improve the accuracy and comprehensiveness of subsequent modeling; the bidirectional LSTM network model can capture the long-term dependencies of time series data, and obtain more comprehensive time information through forward and reverse processing, which helps to improve the model's understanding and prediction capabilities of time series data; the LSTM modeling result sequence can provide modeling and prediction capabilities for historical data, which helps to analyze the system's behavior patterns, predict future trends, and optimize the system's control strategies. In summary, the effects of the above steps include improving the accuracy and expressive power of feature extraction, capturing the long-term dependencies of time series data, and providing modeling and prediction capabilities for historical data; these effects help to deeply understand the operation of the system, predict system performance changes, and optimize system operation management.

[0076] Preferably, step S2 comprises the following steps:

[0077] Step S21: aligning the LSTM modeling result sequence with the real power output data, and constructing a supervision label to obtain a labeled modeling sequence;

[0078] Step S22: setting two reinforcement learning agents, namely the first agent and the second agent, and defining the state space, action space and reward function, so as to obtain an initialized reinforcement learning framework; performing MPPT strategy selection based on the state space on the first agent in the reinforcement learning framework to obtain the MPPT main strategy; performing auxiliary strategy selection based on the MPPT main strategy on the second agent in the reinforcement learning framework to obtain the MPPT auxiliary strategy;

[0079] Step S23: applying the MPPT main strategy and the MPPT auxiliary strategy to the annotation modeling sequence, and monitoring the environmental response to obtain the inductance state data;

[0080] Step S24: Calculate the reward based on the reward function for the inductor state data, optimize the MPPT main strategy and the MPPT auxiliary strategy, and perform periodic performance evaluation on the reinforcement learning framework, cyclically selecting strategy - applying strategy - responding to the environment - calculating reward - evaluating performance until the performance converges, and finally obtaining an optimized reinforcement learning framework;

[0081] Step S25: performing knowledge penetration on the optimization reinforcement learning framework according to the preset circuit principle and prior knowledge of photovoltaic principle, thereby obtaining MPPT strategy data;

[0082] Step S26: Apply the MPPT strategy data to the annotated modeling sequence, simulate the power output under different modes, and obtain multi-modal power curve data.

[0083] As an embodiment of the present invention, refer to Figure 3 As shown, Figure 1 Detailed step flow diagram of step S2 in the embodiment of the present invention, step S2 includes the following steps:

[0084] Step S21: aligning the LSTM modeling result sequence with the real power output data, and constructing a supervision label to obtain a labeled modeling sequence;

[0085] The embodiment of the present invention aligns the LSTM modeling result sequence with the data of the real power output according to the time step, and can use methods such as interpolation, alignment function or time alignment algorithm to implement the alignment operation. According to the aligned data, a rule or threshold is defined. For example, when the difference between the value of the LSTM modeling result sequence and the data of the real power output exceeds a certain threshold, the label is set to 1, indicating that adjustment is required; otherwise, the label is set to 0, indicating that no adjustment is required.

[0086] Step S22: setting two reinforcement learning agents, namely the first agent and the second agent, and defining the state space, action space and reward function, so as to obtain an initialized reinforcement learning framework; performing MPPT strategy selection based on the state space on the first agent in the reinforcement learning framework to obtain the MPPT main strategy; performing auxiliary strategy selection based on the MPPT main strategy on the second agent in the reinforcement learning framework to obtain the MPPT auxiliary strategy;

[0087] According to the specific situation of the problem and the task to be solved, the embodiment of the present invention defines the inductor state data as the state space, defines the range of adjusting the input power as the action space, and defines the reward function, for example, calculating the reward according to the difference between the expected power output and the actual power output. Use the reinforcement learning framework to create an environment including the state space, action space and reward function, construct the first agent and the second agent, and define their interaction mode and strategy selection mode in the environment.

[0088] Step S23: applying the MPPT main strategy and the MPPT auxiliary strategy to the annotation modeling sequence, and monitoring the environmental response to obtain the inductance state data;

[0089] The embodiment of the present invention loops through the labeled modeling sequence, uses the strategy selection method of the first agent and the second agent, and selects the corresponding action according to the current state. According to the selected action, the input power is adjusted; according to the adjusted input power, the response of the environment is observed, and the change of the inductance state data is recorded.

[0090] Step S24: Calculate the reward based on the reward function for the inductor state data, optimize the MPPT main strategy and the MPPT auxiliary strategy, and perform periodic performance evaluation on the reinforcement learning framework, cyclically selecting strategy - applying strategy - responding to the environment - calculating reward - evaluating performance until the performance converges, and finally obtaining an optimized reinforcement learning framework;

[0091] The embodiment of the present invention calculates the rewards of the first agent and the second agent in the current state and the selected action according to the reward function, and uses the optimization algorithm in the reinforcement learning framework, such as Q-learning or Policy Gradient method, to update the policy parameters of the first agent and the second agent according to the reward calculation results to improve their performance. The performance of the optimized agent is evaluated regularly, its performance in the environment is observed, and the results of different policy selections and optimization cycles are compared until the performance converges.

[0092] Step S25: performing knowledge penetration on the optimization reinforcement learning framework according to the preset circuit principle and prior knowledge of photovoltaic principle, thereby obtaining MPPT strategy data;

[0093] The embodiment of the present invention combines the preset circuit principle and the prior knowledge of photovoltaic principle to adjust the strategy selection and reward function in the reinforcement learning framework to make it more in line with the actual situation and better optimize the MPPT strategy. The reinforcement learning framework with knowledge penetration is applied to the annotation modeling sequence to simulate the power output under different modes and record the power curve data under different modes for further analysis and evaluation of the performance and adaptability of the MPPT strategy.

[0094] Step S26: Apply the MPPT strategy data to the annotated modeling sequence, simulate the power output under different modes, and obtain multi-modal power curve data.

[0095] The embodiment of the present invention ensures that the labeled modeling sequence and the data to which the MPPT strategy is applied have the same time step and aligned timestamps, and uses the modeling sequence of the supervised labels constructed in step S21, combined with the data to which the MPPT strategy is applied, to apply the model to each time step; by inputting different modal parameters or conditions into the model, the power output under different modes is simulated, and multimodal power curve data can be generated by looping through different modal parameters or conditions.

[0096] The present invention ensures the correspondence between the output sequence of the LSTM model and the real power output, so that the subsequent reinforcement learning process is more accurate and reliable; by constructing a supervision label, the target signal required by the reinforcement learning algorithm can be provided, which is helpful to guide the training process of the intelligent agent. By setting the intelligent agent and defining relevant parameters, an initial state is provided for the subsequent reinforcement learning process; by selecting a strategy based on the state space, an effective MPPT main strategy is obtained for maximum power point tracking; by selecting a strategy based on the MPPT main strategy, an auxiliary strategy is obtained, which can provide additional guidance and support to enhance the performance of the MPPT main strategy. The previously obtained MPPT main strategy and MPPT auxiliary strategy are applied to the labeled modeling sequence, and the response of the environment is monitored. The optimized MPPT main strategy and auxiliary strategy are applied to the actual labeled modeling sequence to verify its effect and feasibility in the actual environment; by monitoring the response of the environment, the performance of the MPPT main strategy and auxiliary strategy can be evaluated, and further optimized and adjusted. Through calculations based on reward functions, the MPPT main strategy and auxiliary strategy are evaluated to provide feedback signals for optimization; based on the results of reward calculations, the strategies are optimized and adjusted to improve the performance and effect of the MPPT main strategy and auxiliary strategy; through regular performance evaluations, the progress of the reinforcement learning framework can be monitored, and the strategies and parameters can be adjusted in time to achieve better maximum power point tracking effects. The prior knowledge of circuit principles and photovoltaic principles is combined with the optimized reinforcement learning framework to provide more accurate and reliable MPPT strategy data; by incorporating prior knowledge, the performance and learning effect of the reinforcement learning framework can be improved, making the MPPT strategy more in line with actual needs. By applying MPPT strategy data, different working modes can be simulated, including different environmental conditions and photovoltaic characteristics, to evaluate the adaptability and robustness of the MPPT strategy; through the simulation process, power curve data under multiple different modes can be obtained, providing a basis and reference for further analysis and optimization.

[0097] Preferably, step S261 includes the following steps:

[0098] Step S261: setting a plurality of different simulation scenarios according to the marked modeling sequence, simulating the operating status under different working conditions, and obtaining simulation scenario configuration data;

[0099] The embodiment of the present invention uses a programming language and environment (such as Python) to create a new project or script, and determines the parameters to be simulated according to the needs, such as light intensity, temperature, load change, etc.; sets multiple different simulation scenes according to the simulation parameters, and can use a loop structure to generate configuration data for multiple scenes, for example, setting different time points, different light intensity or temperature values, etc.; for each simulation scene, generates corresponding simulation scene configuration data according to the set parameters. These data may include information such as input power, current, voltage, etc.; save the generated simulation scene configuration data to an appropriate data structure (such as a CSV file) for subsequent use.

[0100] Step S262: applying the MPPT strategy to the simulation scenario configuration data, running in each simulation scenario to simulate power output, and obtaining power output data in each simulation scenario;

[0101] The embodiment of the present invention imports the simulation scenario configuration data generated previously, implements the corresponding algorithm or function according to the selected MPPT strategy, and uses a loop structure to traverse the configuration data of each simulation scenario; for each simulation scenario, the simulation scenario configuration data is input into the MPPT algorithm, the data of the maximum power point under the simulation scenario is obtained, and the power output data of each simulation scenario is recorded, which may include power, current, voltage and other information, and the obtained power output data is saved in a suitable data structure for subsequent use.

[0102] Step S263: Establishing a power-voltage curve based on the simulation scenario for the output power and voltage according to the power output data, thereby obtaining a power-voltage curve;

[0103] The embodiment of the present invention uses data processing and drawing tools (such as NumPy, Pandas and Matplotlib libraries in Python) to perform data analysis and drawing, and draws a power-voltage curve based on the power output data of each simulation scenario. It can be displayed in the form of scatter plots, line graphs, etc.; use appropriate statistical analysis and modeling methods, such as curve fitting, regression analysis, etc., to establish a power-voltage curve model based on the simulation scenario, and obtain the power-voltage curve in each simulation scenario based on the model parameters and curve fitting results.

[0104] Step S264: Summarize the power-voltage curves of all scenes and mark the corresponding working modes, so as to obtain multi-modal power curve data.

[0105] The embodiment of the present invention organizes the power-voltage curve data of each simulation scenario, including the numerical values ​​of power and voltage, and aggregates the power-voltage curve data of all simulation scenarios into one data set to ensure the consistency of data structure and variables; according to the characteristics and working mode of each simulation scenario, annotations are added to the aggregated power-voltage curve data to distinguish different working modes; data processing and drawing tools are used to visualize the multi-modal power curve, which can be displayed in the form of line graphs, scatter graphs, etc.

[0106] The present invention sets a plurality of different simulation scenes according to the marked modeling sequence to simulate the operating state of the photovoltaic system under different working conditions. By adjusting the parameters and environmental conditions of the simulation scene, different working conditions, such as different light intensity, temperature, shadow conditions, etc., can be simulated, so that the simulation scene configuration data can be obtained, which contains the parameter settings under various working conditions. The MPPT strategy obtained previously is applied to the simulation scene configuration data to simulate the power output of the photovoltaic system under various simulation scenes; by applying the MPPT strategy, the system can adjust the working point under different working conditions to achieve maximum power output, so that the power output data under various simulation scenes can be obtained, reflecting the performance of the system under different working conditions. According to the power output data, the power-voltage curve based on the simulation scene is established for the output power and voltage. By performing correlation analysis on the power output and voltage under different working conditions, the power-voltage curve of the simulation scene can be established; in this way, the relationship between the power and voltage of the photovoltaic system under different working conditions can be understood, revealing the characteristics and behavior of the system. The power-voltage curves of all simulation scenarios are summarized and the corresponding working modes are marked to obtain multi-modal power curve data. By integrating the power-voltage curves of different simulation scenarios, power curve data containing multiple working modes can be obtained. Each working mode corresponds to different working conditions and system states, which is helpful for analyzing the performance changes of the system, predicting the response of the system, and optimizing the design and control strategy of the system. In summary, the effects of the above steps include: simulating the operating state under different working conditions, obtaining simulation scenario configuration data, simulating the power output under each simulation scenario, establishing a power-voltage curve based on the simulation scenario, and obtaining multi-modal power curve data; these effects help to deeply understand the performance characteristics of the photovoltaic system under different working conditions, analyze the response and behavior of the system, and provide a reference for system design and optimization.

[0107] Preferably, step S3 comprises the following steps:

[0108] Step S31: using the 5G network to distribute the multimodal power curve data to the edge computing node, and preprocessing it to obtain standard power curve data;

[0109] The implementation of the present invention ensures that the edge computing node has a 5G network connection function and is able to receive and process data, and stores the multimodal power curve data in an appropriate data format, such as CSV, JSON, etc.; uses the 5G network to transmit the multimodal power curve data to the edge computing node, and can use a network communication library or API, such as HTTP, WebSocket, etc., to ensure reliable data transmission, and perform preprocessing operations on the edge computing node to obtain standard power curve data. The preprocessing may include data cleaning, data interpolation, outlier processing and other steps to ensure data quality and consistency, and the preprocessed standard power curve data is saved in the storage medium of the edge computing node for subsequent use.

[0110] Step S32: Perform a maximum power point quick search based on the standard power curve data through FPGA acceleration technology, thereby obtaining an MPPT set value sequence;

[0111] The present invention implements the steps of preparing an FPGA development board and ensuring that a development environment has been set up, loading standard power curve data into the FPGA development board, transmitting the data to the development board via a storage medium or a network connection, using an FPGA development tool to design and program the FPGA, implementing a selected MPPT algorithm in the FPGA, implementing a maximum power point search algorithm on the FPGA, obtaining an MPPT setting value sequence based on the standard power curve data, and saving the obtained MPPT setting value sequence in an appropriate data structure, such as an array or a file.

[0112] Step S33: obtaining current grid voltage data; performing optimal reactive power compensation ratio calculation based on a power factor optimization algorithm according to the MPPT set value sequence and the current grid voltage data, thereby obtaining a reactive power compensation ratio sequence;

[0113] The present invention implements the use of suitable sensors or monitoring equipment to collect voltage data of the current power grid in real time, inputs the MPPT set value sequence and voltage data into a calculation module, implements the selected power factor optimization algorithm in the calculation module, calculates the best reactive compensation ratio sequence according to the algorithm, and saves the calculated reactive compensation ratio sequence in an appropriate data structure.

[0114] Step S34: Obtain current grid frequency data, use the edge AI model to predict islanding risk based on the MPPT setting value sequence and current grid frequency data, and generate an anti-islanding control parameter sequence;

[0115] The present invention implements the use of suitable sensors or monitoring equipment to collect the frequency data of the current power grid in real time, inputs the MPPT setting value sequence, power grid frequency data and other relevant data into the edge AI model, implements the islanding risk prediction algorithm in the edge AI model, generates the corresponding anti-islanding control parameter sequence according to the risk prediction results, and saves the generated anti-islanding control parameter sequence in an appropriate data structure.

[0116] Step S35: performing parameter coordination based on the overall optimal performance on the MPPT setting value sequence, reactive power compensation ratio sequence and anti-islanding control parameter sequence, thereby obtaining parameter coordination optimization data;

[0117] The implementation of the present invention ensures that the photovoltaic inverter supports receiving real-time control instructions and has corresponding communication functions. The selected communication protocol is used to send the reactive compensation ratio sequence and the anti-islanding control parameter sequence to the photovoltaic inverter, and the corresponding control logic is implemented in the photovoltaic inverter. The output power of the inverter is adjusted according to the received real-time control instructions. The photovoltaic inverter performs reactive compensation and anti-islanding control according to the real-time control instructions to adapt to the needs of the power grid and prevent the occurrence of islanding events.

[0118] Step S36: Send the parameter coordination optimization data to the inverter in real time through the 5G network to obtain distributed control strategy data.

[0119] The present invention implements the installation of appropriate monitoring devices to monitor the output state of the photovoltaic inverter, such as current sensors, voltage sensors, etc., and obtains key parameters of the photovoltaic inverter such as output current and output voltage through these monitoring devices. The obtained output state parameters are transmitted to the control system for real-time monitoring and analysis. According to the results of real-time monitoring, they are compared with preset reactive power compensation ratios and anti-islanding control parameters. If the actual output state does not match the preset parameters, the control system will calculate corresponding control instructions, which are sent to the photovoltaic inverter through a communication protocol to adjust the output power to achieve the expected reactive power compensation and anti-islanding control effects. Repeated monitoring, analysis and adjustment are performed to achieve closed-loop control of the photovoltaic inverter.

[0120] The present invention uses 5G network to distribute multimodal power curve data to edge computing nodes, and performs preprocessing to process standard power curve data. By using 5G network, high-speed and low-latency data transmission can be achieved, and multimodal power curve data can be transmitted to edge computing nodes; preprocessing is performed at edge computing nodes, and data can be cleaned, filtered, normalized, and other operations can be performed to obtain accurate and reliable standard power curve data. By using FPGA acceleration technology, the maximum power point is quickly searched according to the standard power curve data. FPGA (field programmable gate array) is a programmable hardware device with high parallel computing capabilities and low latency characteristics; by implementing the maximum power point tracking algorithm on FPGA, the maximum power point of the photovoltaic system can be quickly and effectively found, and the corresponding MPPT (maximum power point tracking) set value sequence can be obtained. Obtain the current grid voltage data, and calculate the optimal reactive power compensation ratio based on the power factor optimization algorithm according to the MPPT set value sequence and the current grid voltage data. The power factor optimization algorithm can be calculated using the grid voltage data and the MPPT set value sequence to determine the optimal reactive power compensation ratio sequence. By optimizing the reactive power compensation ratio, the reactive power of the photovoltaic system can be controlled to improve the system's power factor and grid stability. Obtain the current grid frequency data, use the edge AI model to predict the islanding risk according to the MPPT set value sequence and the current grid frequency data, and generate an anti-islanding control parameter sequence; use the grid frequency data and the MPPT set value sequence, combined with the edge AI model to predict the islanding risk, and determine whether the photovoltaic system has the risk of islanding; generate the corresponding anti-islanding control parameter sequence based on the prediction results, which is used to control and protect the photovoltaic system and ensure its safe connection with the grid. The MPPT set value sequence, reactive compensation ratio sequence and anti-islanding control parameter sequence are coordinated based on the optimal overall performance; by comprehensively considering the influence of MPPT set value, reactive compensation ratio and anti-islanding control parameters, parameter coordination optimization is performed to achieve the optimization of overall performance. The optimized parameter coordination data can improve the efficiency, stability and safety of the photovoltaic system. The parameter coordination optimization data is sent to the inverter in real time through the 5G network to obtain the distributed control strategy data. The 5G network is used to transmit the parameter coordination optimization data to the inverter in real time to achieve real-time update of the distributed control strategy; the inverter adjusts the working state and control strategy of the photovoltaic system according to the received parameter coordination optimization data, thereby improving the performance and efficiency of the system.In summary, the effects of the above steps include: high-speed transmission and preprocessing of multi-modal power curve data, rapid search for maximum power points, optimization of reactive compensation ratios, prediction of islanding risks and generation of anti-islanding control parameters, parameter coordination optimization to achieve optimal overall performance, and real-time distribution of distributed control strategy data; these effects help improve the power output and efficiency of photovoltaic systems, optimize the power factor and grid stability of the system, predict and prevent the occurrence of islanding phenomena, and achieve distributed control and optimization of photovoltaic systems; at the same time, using 5G networks for data transmission and real-time distribution can achieve high-speed, low-latency communication and improve the response speed and real-time performance of the system. Overall, the effects of these steps can improve the performance, reliability and intelligence level of photovoltaic systems, and have a positive impact on the application and development of photovoltaic power generation.

[0121] Preferably, step S4 comprises the following steps:

[0122] Step S41: acquiring real-time topological data of a power grid, and cleaning and standardizing the real-time topological data of the power grid, thereby obtaining a power grid topological data set;

[0123] The embodiments of the present invention obtain real-time topological data of the power grid from power grid monitoring equipment, SCADA systems or other power grid data acquisition systems. These data generally include information on nodes, lines, transformers and other elements of the power grid. The acquired real-time topological data of the power grid is cleaned, including removing duplicate data, repairing erroneous data, processing missing data, etc., to ensure the accuracy and consistency of the data. The cleaned power grid topological data is standardized for subsequent modeling and analysis. Standardization may include conversion of node numbers, line attributes, transformer parameters, etc. into a unified format and unit.

[0124] Step S42: constructing a power grid topology modeling framework based on the variational graph autoencoder to obtain a VGAE power grid modeling framework; performing model training on the VGAE power grid modeling framework according to the power grid topology data set to obtain a VGAE power grid topology model;

[0125] The embodiment of the present invention uses a deep learning framework to build a VGAE power grid topology modeling framework, including components such as a graph convolution layer, an encoder, and a decoder; the VGAE model is trained using a power grid topology dataset, and the model parameters are optimized to enable it to accurately learn and represent the power grid topology structure.

[0126] Step S43: performing grid topology structure inference on the VGAE grid topology model, thereby obtaining a grid topology model;

[0127] The embodiment of the present invention uses the trained VGAE power grid topology model to infer the topology of the power grid, including detecting the connection relationship between potential power grid nodes and lines, etc. The potential topology representation is extracted from the VGAE model and decoded to obtain the topology information of the power grid.

[0128] Step S44: Optimizing the grid-connected path of the photovoltaic station according to the grid topology model, and obtaining the timing data of the optimized grid-connected path;

[0129] The embodiment of the present invention determines the grid-connected path between the photovoltaic station and the grid based on the grid topology model and the location information of the photovoltaic station; considers factors such as grid load and grid stability, and performs timing optimization on the grid-connected path of the photovoltaic station, which includes determining the time period for the photovoltaic station to be connected to the grid, the power regulation strategy, etc.

[0130] Step S45: decomposing the optimized grid-connected path time series data into multi-phase grid-connected vector data; and generating corresponding grid-connected control strategy data according to the multi-phase grid-connected vector data.

[0131] The embodiment of the present invention uses a time series processing tool to decompose the optimized grid-connected path time series data according to time to obtain data of multiple phases; the grid-connected path data of each phase is converted into grid-connected vector data to represent the grid-connected status of the photovoltaic station in that phase; and according to the grid-connected vector data, corresponding grid-connected control strategy data is generated, including power control, reactive power compensation, etc. of the photovoltaic station.

[0132] The present invention obtains real-time topological data of the power grid, cleans and standardizes it, so as to obtain an accurate and reliable power grid topological data set. The real-time topological data of the power grid reflects the connection relationship and topological structure between each node in the power system; by obtaining and processing the real-time topological data of the power grid, a complete power grid topological data set can be established, which provides a basis for subsequent modeling and optimization. A power grid topological modeling framework is built based on a variational graph autoencoder (VGAE), and a model is trained on the power grid topological data set to obtain a VGAE power grid topological model. VGAE is a machine learning model for graph data modeling, which can learn the potential representation of power grid topological data, capture the relationship and topological structure between nodes, and by training the VGAE model, a model that can effectively express the power grid topology can be obtained, which provides a basis for subsequent topological reasoning and optimization. The trained VGAE power grid topological model is subjected to the topological structure reasoning of the power grid, so as to obtain a more accurate power grid topological model; through the reasoning process, the power grid topological model can be further optimized and improved, and the implicit relationship and topological characteristics between nodes in the power system can be revealed. Such a topological model can be used for subsequent power system analysis, optimization and control. According to the grid topology model, the grid-connected path of the photovoltaic station is optimized to obtain the optimized grid-connected path timing data; by considering the grid topology and system operation status, the optimized grid-connected path can make the connection between the photovoltaic station and the grid more stable and reliable, improve the efficiency and power utilization of the system; the optimized grid-connected path timing data can guide the operation and scheduling of the photovoltaic station to achieve the best power interaction. The optimized grid-connected path timing data is decomposed into multi-phase grid-connected vector data, and the corresponding grid-connected control strategy data is generated based on these data; by decomposing and transforming the grid-connected path timing data, the grid-connected vector data that adapts to different phases can be obtained; these grid-connected vector data can be used to generate grid-connected control strategies, including power regulation of photovoltaic stations, control of grid voltage and frequency, etc.; through reasonable grid-connected control strategies, the coordinated operation and interaction between photovoltaic stations and the grid can be achieved, and the stability and responsiveness of the system can be improved. In summary, the effects of the above steps include: obtaining accurate and reliable grid topology data sets, establishing effective grid topology models, optimizing grid-connected path timing data, and generating grid-connected control strategy data; these effects help improve the operation and performance of the power system, improve the connection quality and stability between photovoltaic stations and the power grid, and achieve optimal scheduling and operation control of the system; at the same time, using machine learning and optimization algorithms, we can better analyze and understand the topological structure and operating characteristics of the power system, and provide support and guidance for research and development in the energy field.

[0133] Preferably, step S5 comprises the following steps:

[0134] Step S51: high-frequency sampling of on-site electromagnetic wave information of the photovoltaic system is performed by an electromagnetic sensor, thereby obtaining electromagnetic sensor measurement data;

[0135] The embodiment of the present invention uses an electromagnetic sensor to sample the electromagnetic wave information of the photovoltaic system. The electromagnetic sensor can be a device specially designed to measure the electromagnetic field, such as an electromagnetic induction coil or an antenna. By placing the sensor near the photovoltaic system, the electromagnetic radiation emitted by the system can be measured. During the sampling process, the sensor measures the electromagnetic waves at a high frequency and records the measurement data.

[0136] Step S52: filtering and denoising the electromagnetic sensor measurement data, and performing multi-sensor data alignment and amplitude calibration, so as to obtain a synchronously calibrated electromagnetic field data set;

[0137] The embodiment of the present invention processes the measurement data obtained from the electromagnetic sensor. First, the data is filtered and denoised to remove noise and unnecessary interference. Next, if multiple sensors are used for sampling, the data of the multiple sensors need to be aligned to ensure that they are synchronized in time. In addition, amplitude calibration may be required to eliminate the amplitude difference between different sensors. After these processing steps, a synchronously calibrated electromagnetic field data set is obtained.

[0138] Step S53: performing Fourier transform on the electromagnetic field data set and performing spectrum analysis to obtain frequency domain electromagnetic field distribution data;

[0139] The embodiment of the present invention performs Fourier transform on the electromagnetic field data set after synchronous calibration. The Fourier transform can convert time domain data into frequency domain data for analyzing the existence and energy distribution of different frequency components. By performing Fourier transform on the data, the representation of the electromagnetic field data in the frequency domain can be obtained. Then, spectrum analysis is performed on the frequency domain data to obtain the distribution of the electromagnetic field at different frequencies.

[0140] Step S54: using an inverse scattering algorithm to perform electromagnetic imaging inversion on the frequency domain electromagnetic field distribution data, thereby obtaining a preliminary harmonic source distribution map;

[0141] The embodiment of the present invention uses an inverse scattering algorithm to perform inversion processing on frequency domain electromagnetic field data to obtain a preliminary harmonic source distribution image. The inverse scattering algorithm is a method for inferring source distribution from electromagnetic field data. By applying the inverse scattering algorithm, the position and intensity distribution of harmonics generated in a photovoltaic system can be estimated.

[0142] Step S55: performing image enhancement on the preliminary harmonic source distribution map and extracting the harmonic source positions, thereby obtaining a harmonic source position coordinate sequence;

[0143] The embodiment of the present invention performs image enhancement processing on the preliminary harmonic source distribution map to improve the quality and clarity of the image. The image enhancement may include operations such as denoising, contrast enhancement, and edge detection. Subsequently, the position information of the harmonic source is extracted through image processing technology to obtain a sequence of harmonic source position coordinates.

[0144] Step S56: tracking and reconstructing the harmonic current path according to the harmonic source position coordinate sequence, thereby obtaining harmonic current path data; performing path feature extraction on the harmonic current path data, thereby obtaining a path feature sequence;

[0145] The embodiment of the present invention tracks and reconstructs the harmonic current path according to the harmonic source position coordinate sequence, which can be achieved by analyzing the positional relationship and connection mode between the harmonic sources. By tracking and reconstructing the harmonic current path, the propagation path of the harmonic current in the photovoltaic system can be determined. Next, the path feature extraction is performed on the harmonic current path data. The path feature may include information such as path length, angle, and branching. These features can reveal the propagation characteristics and distribution of the harmonic current in the system. By extracting the path features, a sequence of path features can be obtained.

[0146] Step S57: Generate harmonic compensation strategy data according to the path characteristic sequence.

[0147] The harmonic compensation strategy of the embodiment of the present invention is a measure taken to reduce or eliminate the adverse effects of harmonics on the photovoltaic system. According to the path characteristic sequence, the compensation strategy to be taken can be determined, such as adding a compensation device at a specific location or adjusting the system configuration. The generated harmonic compensation strategy data can be used to guide subsequent harmonic suppression measures.

[0148] The present invention uses electromagnetic sensors to perform high-frequency sampling of electromagnetic wave information at the photovoltaic system site, and can obtain measurement data of the electromagnetic sensors. These data reflect the electromagnetic field intensity and distribution around the photovoltaic system. Through high-frequency sampling, subtle electromagnetic wave changes can be captured, providing a basis for subsequent data processing and analysis. The electromagnetic sensor measurement data is filtered and denoised, and clean electromagnetic field data is extracted by filtering out noise and interference signals; at the same time, multi-sensor data alignment and amplitude calibration are performed to ensure data consistency and accuracy between different sensors. Reliable electromagnetic field intensity data can be obtained through the synchronously calibrated electromagnetic field data set, providing accurate input for subsequent analysis and processing. The electromagnetic field data set after synchronous calibration is Fourier transformed and spectrum analyzed. The Fourier transform converts the electromagnetic field data in the time domain into frequency domain data, and the electromagnetic field distribution at different frequencies can be obtained. The spectrum analysis can reveal the harmonic components and frequency domain characteristics existing in the photovoltaic system, and help identify and analyze the harmonic sources in the electromagnetic field. The inverse scattering algorithm is used to perform electromagnetic imaging inversion on the frequency domain electromagnetic field distribution data to obtain a preliminary harmonic source distribution map. The inverse scattering algorithm is based on the physical principle of electromagnetic wave propagation in the medium. By performing reverse analysis on the electromagnetic field data, the location and distribution of the harmonic source can be estimated. The preliminary harmonic source distribution map can provide spatial information about the generation of harmonics in the photovoltaic system, providing a basis for subsequent positioning and analysis. The preliminary harmonic source distribution map is image enhanced and the harmonic source position is extracted. Image enhancement technology can improve the clarity and contrast of the image and make the location of the harmonic source more obvious; by extracting the harmonic source position, the coordinate sequence of the harmonic source in the photovoltaic system can be obtained, that is, the spatial distribution of the harmonic source. The harmonic current path is tracked and reconstructed according to the harmonic source position coordinate sequence to obtain the harmonic current path data; by tracking the location of the harmonic source, the propagation path of the harmonic current in the photovoltaic system can be determined. The harmonic current path data can be used to analyze the harmonic propagation mechanism and path characteristics in the photovoltaic system. Harmonic compensation strategy data is generated based on the path characteristic sequence. The path characteristics include the length, direction, branching conditions and other information of the harmonic current path. By analyzing the path characteristic sequence, corresponding harmonic compensation strategies can be formulated, such as selecting appropriate compensation devices, adjusting system configuration or control strategies, to reduce the impact of harmonics on the photovoltaic system and improve the stability and reliability of the system.In summary, the effects of the above steps include: obtaining accurate electromagnetic field data, performing harmonic imaging and harmonic source positioning, tracing harmonic current paths, extracting path characteristics, and generating harmonic compensation strategy data; these effects help to identify and analyze harmonic problems in photovoltaic systems, provide support for the positioning and analysis of harmonic sources, and by formulating appropriate harmonic compensation strategies, reduce the impact of harmonics on photovoltaic systems and improve system operating efficiency and stability; these steps combine electromagnetic sensors, signal processing, imaging algorithms, path analysis and other technologies to provide a comprehensive method and tool for diagnosing and optimizing harmonic problems in photovoltaic systems.

[0149] Preferably, step S6 comprises the following steps:

[0150] Step S61: unifying the data formats of the distributed control strategy data, the grid-connected control strategy data, and the harmonic compensation strategy data, thereby obtaining three types of strategy data;

[0151] The embodiment of the present invention cleans and preprocesses the original policy data, including removing noise, processing missing values, processing outliers, etc., to ensure the accuracy and consistency of the data. The policy data in different formats are converted into a unified data format, which can include converting the data into a specific file format (such as CSV, JSON, etc.) or using a unified data structure (such as a table, dictionary, etc.). The data is standardized so that different types of policy data can be compared and integrated. Standardization can include normalization, standardization or other suitable data transformation methods.

[0152] Step S62: designing a Transformer-XL architecture suitable for multi-strategy fusion according to the three types of strategy data, thereby obtaining Transformer-XL model structure data;

[0153] The embodiment of the present invention determines the input representation of each policy data, such as word embedding, time series features, etc., which will be used as the input of the Transformer-XL model. The structural parameters of the Transformer-XL model, such as the number of layers, hidden layer dimensions, and attention mechanism, are designed. These parameters will affect the representation ability and computational efficiency of the model. A multi-policy fusion mechanism is designed to effectively combine the three types of policy data. The use of attention mechanism, fusion layer or other effective methods can be considered to achieve fusion.

[0154] Step S63: performing data alignment and vector representation conversion based on timestamps on the three types of strategy data, and performing strategy fusion according to the Transformer-XL model structure data, thereby obtaining fusion strategy representation data;

[0155] The embodiment of the present invention aligns the distributed control strategy data, the grid-connected control strategy data, and the harmonic compensation strategy data according to the timestamp information of the strategy data, ensuring that they are synchronized in time, and facilitating the continuation of the subsequent step S63: converting the aligned strategy data into a vector representation, which can use different methods, such as word embedding, time series feature extraction, etc., to represent each strategy data as a vector of a fixed length. Using the designed Transformer-XL model structure data, the converted vector representation is input into the model for strategy fusion. The Transformer-XL model can effectively learn the relationship and weight between strategies and generate fused strategy representation data.

[0156] Step S64: generating an initial modulation mode sequence from the fusion strategy representation data;

[0157] An embodiment of the present invention defines a set of modulation modes, including parameters such as modulation waveform and modulation frequency. The selection of these modulation modes should be based on specific control needs and system requirements. According to the fusion strategy representation data, an initial modulation mode sequence is generated using appropriate algorithms or rules, which may involve decoding and mapping the fusion strategy representation data to determine the modulation mode at each time point.

[0158] Step S65: Encode the initial modulation mode sequence into a control instruction and send it to the inverter via the industrial bus to execute SPWM.

[0159] The embodiment of the present invention uses an inverter control algorithm to generate corresponding control instructions based on the initial modulation mode sequence, and these instructions generally include the switch state and pulse width of each phase. The generated control instructions are encoded to adapt to the communication protocol and requirements of the industrial bus, which may involve converting the instructions into a specific data format or encapsulating them using a communication protocol. The encoded control instructions are sent to the inverter via the selected industrial bus, which can be achieved through appropriate hardware and communication equipment to ensure reliable transmission of the control instructions.

[0160] The present invention unifies the data format of distributed control strategy data, grid-connected control strategy data and harmonic compensation strategy data, and can convert different types of strategy data into the same data format; the advantage of doing so is to simplify the data processing and fusion operations in the subsequent processing steps. Through the unification of data formats, effective comparison, fusion and analysis can be performed between strategy data. According to the three types of strategy data, a Transformer-XL architecture suitable for multi-strategy fusion is designed to obtain Transformer-XL model structure data. Transformer-XL is a powerful sequence modeling method suitable for processing long sequence data and multi-strategy fusion. The Transformer-XL architecture can effectively capture the association and dependency between strategy data and improve the modeling ability of strategy data. The three types of strategy data are aligned and transformed based on timestamps, and strategy fusion is performed according to the Transformer-XL model structure data to obtain fusion strategy representation data. Data alignment can ensure the consistency of different strategy data in time, so that they can be fused at the same time point; vector representation conversion converts strategy data into numerical vector form, which is convenient for model processing and calculation. Through the Transformer-XL model structure data, strategy fusion can combine the information of different strategy data to obtain more comprehensive and comprehensive strategy representation data. The fusion strategy representation data is used to generate an initial modulation pattern sequence, which is a pattern sequence that describes the inverter's working state and control instructions. By generating an initial modulation pattern sequence from the fusion strategy representation data, complex strategy information can be converted into a pattern sequence that the inverter can understand and execute, preparing for the subsequent control instruction generation. The initial modulation pattern sequence is encoded into a control instruction and sent to the inverter through the industrial bus to execute SPWM (Sinusoidal Pulse Width Modulation). SPWM is a commonly used modulation technology that controls the output current and voltage of the photovoltaic system connected to the grid by adjusting the pulse width of the inverter output voltage. By converting the initial modulation pattern sequence into a control instruction and sending it to the inverter through the industrial bus, precise control and regulation of the photovoltaic system can be achieved to meet the grid requirements and harmonic compensation strategies. In summary, the effects of the above steps include: achieving unified processing of different types of strategy data, improving the modeling capability and fusion effect of strategy data, obtaining comprehensive fusion strategy representation data, and generating an initial modulation pattern sequence and encoding it into control instructions. It is possible to achieve precise control and regulation of the photovoltaic system to meet the requirements of the power grid and harmonic compensation strategies. These steps combine data processing, model design, and control instruction generation technologies to provide a comprehensive method and tool for intelligent control and harmonic compensation of photovoltaic systems.

[0161] The present invention also provides a distributed photovoltaic system grid-connected control system for executing the above-mentioned distributed photovoltaic system grid-connected control method, wherein the distributed photovoltaic system grid-connected control system comprises:

[0162] The LSTM modeling module is used to obtain historical multi-source heterogeneous data, including historical power output data, historical meteorological data, battery aging data, and array parameter data; extract multimodal environmental features from historical multi-source heterogeneous data, and capture the long-term dependency of time series based on a bidirectional LSTM network to obtain a sequence of LSTM modeling results;

[0163] The MPPT strategy optimization module is used to deploy a dual alternating reinforcement learning framework based on the LSTM modeling result sequence and perform knowledge penetration to obtain MPPT strategy data; and to generate multi-modal power curve data based on the MPPT strategy data;

[0164] The distributed control parameter module is used to determine the MPPT setting value, reactive power compensation ratio, and anti-islanding control parameters of the inverter according to the multi-modal power curve data using 5G edge computing technology and FPGA acceleration technology to obtain distributed control strategy data;

[0165] The grid-connected path planning module is used to perform VGAE-based grid modeling according to the preset real-time grid topology data to obtain a grid topology model; perform timing optimization on the grid-connected path according to the grid topology model, and decompose the path into multi-phase grid-connected vector data; generate grid-connected control strategy data according to the multi-phase grid-connected vector data;

[0166] The harmonic compensation strategy module is used to obtain electromagnetic sensor measurement data, and inversely interpret the electromagnetic imaging therein into harmonic source distribution, extract the harmonic current path, and obtain the path feature sequence; and generate harmonic compensation strategy data according to the path feature sequence;

[0167] The multi-strategy fusion module is used to perform Transformer-XL fusion on the distributed control strategy data, the grid-connected control strategy data and the harmonic compensation strategy data to obtain an initial modulation mode sequence; the initial modulation mode sequence is encoded into a control instruction and sent to the inverter through the industrial bus to execute SPWM.

[0168] The present invention can capture the time series dependency between data by using LSTM network to model historical multi-source heterogeneous data. LSTM network can handle long-term dependency, which is helpful to extract the characteristics of historical power output data, meteorological data, battery aging data and array parameter data. The LSTM modeling result sequence can provide an effective representation of historical data, providing a basis for subsequent MPPT strategy optimization and control decision. According to the LSTM modeling result sequence, the MPPT strategy is optimized using a dual alternating reinforcement learning framework. This method can learn the optimal MPPT strategy through interaction with the environment to achieve the maximum power output of the photovoltaic system; through knowledge penetration, the system can continuously improve and optimize the strategy to improve the energy conversion efficiency of the system, and the generated multimodal power curve data can provide a comprehensive analysis and evaluation of the power output under different working conditions. Using 5G edge computing technology and FPGA acceleration technology, the MPPT setting value, reactive compensation ratio and anti-islanding control parameters of the inverter are determined according to the multimodal power curve data. By analyzing the multimodal power curve data, the photovoltaic system can be optimized and controlled, and the performance and stability of the system can be improved. The distributed control strategy data can be used to dynamically adjust the inverter according to the real-time situation to adapt to different working conditions and grid requirements. Based on the preset real-time grid topology data, VGAE (Variational Graph Autoencoder) is used to model the grid and obtain the grid topology model; through timing optimization and path decomposition, the grid-connected path can be determined and multi-phase grid-connected vector data can be generated, which helps to optimize the grid-connected behavior of the photovoltaic system and ensure its stable connection with the grid. The grid-connected control strategy data provides guidance for achieving safe and reliable grid connection. By obtaining electromagnetic sensor measurement data and inverting electromagnetic imaging into harmonic source distribution, harmonic current path extraction is performed, so that the propagation path and characteristics of harmonic current can be identified. The generated path feature sequence can be used to formulate harmonic compensation strategies to reduce harmonic pollution in the system and improve power quality. Through Transformer-XL, the information of different strategies can be integrated and integrated to obtain the initial modulation mode sequence; this helps to comprehensively consider the impact and requirements of different strategies, improve the overall performance and efficiency of the system, and encode the initial modulation mode sequence into control instructions, which are sent to the inverter through the industrial bus to execute SPWM to achieve precise control and regulation of the photovoltaic system. In summary, the effects of the above steps include: extracting historical data features, optimizing MPPT strategies, determining distributed control parameters, optimizing grid-connected path planning, implementing harmonic compensation strategies, and integrating multiple strategies; these effects help to improve the energy conversion efficiency, stability, and power quality of photovoltaic systems, maximize the use of photovoltaic resources, and interconnect with the power grid safely and reliably.

[0169] Therefore, the embodiments should be regarded as illustrative and non-restrictive from all points, and the scope of the present invention is limited by the appended claims rather than the above description, and it is therefore intended that all changes falling within the meaning and range of equivalent elements of the application documents are included in the present invention.

[0170] The above description is only a specific embodiment of the present invention, so that those skilled in the art can understand or implement the present invention. Various modifications to these embodiments will be 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 the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but should conform to the widest scope consistent with the principles and novel features invented herein.

Claims

1. A distributed photovoltaic system grid-connected control method, characterized in that: The following steps are involved: Step S1: Acquire historical multi-source heterogeneous data, including historical power output data, historical meteorological data, battery aging data, and array parameter data; perform multimodal environmental feature extraction on the historical multi-source heterogeneous data, and capture the long-term dependency of the time series based on a bidirectional LSTM network to obtain a LSTM modeling result sequence; Step S2: deploy a dual alternating reinforcement learning framework according to the LSTM modeling result sequence, and perform knowledge penetration to obtain MPPT strategy data; generate multimodal power curve data according to the MPPT strategy data; Step S3: Using 5G edge computing technology and FPGA acceleration technology to determine the MPPT setting value, reactive power compensation ratio, and anti-islanding control parameters of the inverter according to the multi-modal power curve data, and obtain distributed control strategy data; Step S4: Perform VGAE-based grid modeling according to preset real-time grid topology data to obtain a grid topology model; perform timing optimization on the grid connection path according to the grid topology model, and decompose the path into multi-phase grid connection vector data; generate grid connection control strategy data according to the multi-phase grid connection vector data; Step S5: Acquire electromagnetic sensor measurement data, and inversely resolve the electromagnetic imaging therein into harmonic source distribution, extract harmonic current paths, and obtain path feature sequences; generate harmonic compensation strategy data according to the path feature sequences; Step S6: Perform Transformer-XL fusion on the distributed control strategy data, the grid-connected control strategy data and the harmonic compensation strategy data to obtain an initial modulation mode sequence; encode the initial modulation mode sequence into a control instruction, and send it to the inverter through the industrial bus to execute SPWM.

2. The distributed photovoltaic system grid-connected control method according to claim 1, characterized in that: Step S1 includes the following steps: Step S11: extracting the output power record of the inverter through the data collection and monitoring system of the photovoltaic system to obtain historical power output data; Step S12: collecting meteorological factors such as light intensity, temperature, and wind speed at the distributed photovoltaic system site to obtain historical meteorological data; Step S13: estimating the battery aging degree based on the electrochemical model according to the battery cycle number, temperature history and voltage data monitored and recorded by the battery management system, thereby obtaining battery aging data; Step S14: acquiring array parameter data including the structure, layout, orientation and tilt angle parameters of the photovoltaic array; Step S15: integrating the historical power output data, historical meteorological data, battery aging data, and array parameter data with the labeled data sources, thereby obtaining historical multi-source heterogeneous data; Step S16: extract multimodal environmental features from historical multi-source heterogeneous data, and capture long-term dependencies of time series based on a bidirectional LSTM network to obtain a sequence of LSTM modeling results.

3. The distributed photovoltaic system grid-connected control method according to claim 2, characterized in that: Step S16 includes the following steps: Perform multi-modal environmental feature extraction based on multi-branch CNN on historical multi-source heterogeneous data to obtain a multi-modal feature vector sequence, where the multi-modal feature vector sequence is a three-dimensional vector, the first dimension is the time step, the second dimension is the feature dimension, and the third dimension is the batch dimension; The multimodal feature vector sequence is input into the preset bidirectional LSTM network model for back-propagation training of the model parameters to obtain a sequence of LSTM modeling results, in which the forward LSTM layer of the bidirectional LSTM network model sequentially captures the forward time dependency, and the reverse LSTM layer captures the reverse time dependency. The LSTM hidden layer states in two directions are concatenated at the same time step and then connected to a fully connected layer for nonlinear mapping of the hidden layer states.

4. The distributed photovoltaic system grid-connected control method according to claim 3, characterized in that: Step S2 includes the following steps: Step S21: aligning the LSTM modeling result sequence with the real power output data, and constructing a supervision label to obtain a labeled modeling sequence; Step S22: setting two reinforcement learning agents, namely the first agent and the second agent, and defining the state space, action space and reward function, so as to obtain an initialized reinforcement learning framework; performing MPPT strategy selection based on the state space on the first agent in the reinforcement learning framework to obtain the MPPT main strategy; performing auxiliary strategy selection based on the MPPT main strategy on the second agent in the reinforcement learning framework to obtain the MPPT auxiliary strategy; Step S23: applying the MPPT main strategy and the MPPT auxiliary strategy to the annotation modeling sequence, and monitoring the environmental response to obtain the inductance state data; Step S24: Calculate the reward based on the reward function for the inductor state data, optimize the MPPT main strategy and the MPPT auxiliary strategy, and perform periodic performance evaluation on the reinforcement learning framework, cyclically selecting strategy - applying strategy - responding to the environment - calculating reward - evaluating performance until the performance converges, and finally obtaining an optimized reinforcement learning framework; Step S25: performing knowledge penetration on the optimization reinforcement learning framework according to the preset circuit principle and prior knowledge of photovoltaic principle, thereby obtaining MPPT strategy data; Step S26: Apply the MPPT strategy data to the annotated modeling sequence, simulate the power output under different modes, and obtain multi-modal power curve data.

5. The distributed photovoltaic system grid-connected control method according to claim 4, characterized in that: Step S261 includes the following steps: Step S261: setting a plurality of different simulation scenarios according to the marked modeling sequence, simulating the operating status under different working conditions, and obtaining simulation scenario configuration data; Step S262: applying the MPPT strategy to the simulation scenario configuration data, running in each simulation scenario to simulate power output, and obtaining power output data in each simulation scenario; Step S263: Establishing a power-voltage curve based on the simulation scenario for the output power and voltage according to the power output data, thereby obtaining a power-voltage curve; Step S264: Summarize the power-voltage curves of all scenes and mark the corresponding working modes, so as to obtain multi-modal power curve data.

6. The distributed photovoltaic system grid-connected control method according to claim 5, characterized in that: Step S3 includes the following steps: Step S31: using the 5G network to distribute the multimodal power curve data to the edge computing node, and preprocessing to obtain standard power curve data; Step S32: Perform a fast maximum power point search based on standard power curve data using FPGA acceleration technology to obtain an MPPT set value sequence; Step S33: obtaining current grid voltage data; performing optimal reactive power compensation ratio calculation based on a power factor optimization algorithm according to the MPPT set value sequence and the current grid voltage data, thereby obtaining a reactive power compensation ratio sequence; Step S34: Obtain current grid frequency data, use the edge AI model to predict islanding risk based on the MPPT setting value sequence and current grid frequency data, and generate an anti-islanding control parameter sequence; Step S35: performing parameter coordination based on the overall optimal performance on the MPPT setting value sequence, reactive power compensation ratio sequence and anti-islanding control parameter sequence, thereby obtaining parameter coordination optimization data; Step S36: Send the parameter coordination optimization data to the inverter in real time through the 5G network to obtain distributed control strategy data.

7. The distributed photovoltaic system grid-connected control method according to claim 6, characterized in that: Step S4 includes the following steps: Step S41: acquiring real-time topological data of a power grid, and cleaning and standardizing the real-time topological data of the power grid, thereby obtaining a power grid topological data set; Step S42: constructing a power grid topology modeling framework based on the variational graph autoencoder to obtain a VGAE power grid modeling framework; performing model training on the VGAE power grid modeling framework according to the power grid topology data set to obtain a VGAE power grid topology model; Step S43: performing grid topology structure inference on the VGAE grid topology model, thereby obtaining a grid topology model; Step S44: Optimizing the grid-connected path of the photovoltaic station according to the grid topology model, and obtaining the timing data of the optimized grid-connected path; Step S45: decomposing the optimized grid-connected path time series data into multi-phase grid-connected vector data; and generating corresponding grid-connected control strategy data according to the multi-phase grid-connected vector data.

8. The distributed photovoltaic system grid-connected control method according to claim 7, characterized in that: Step S5 includes the following steps: Step S51: high-frequency sampling of on-site electromagnetic wave information of the photovoltaic system is performed by an electromagnetic sensor, thereby obtaining electromagnetic sensor measurement data; Step S52: filtering and denoising the electromagnetic sensor measurement data, and performing multi-sensor data alignment and amplitude calibration, so as to obtain a synchronously calibrated electromagnetic field data set; Step S53: performing Fourier transform on the electromagnetic field data set and performing spectrum analysis to obtain frequency domain electromagnetic field distribution data; Step S54: using an inverse scattering algorithm to perform electromagnetic imaging inversion on the frequency domain electromagnetic field distribution data, thereby obtaining a preliminary harmonic source distribution map; Step S55: performing image enhancement on the preliminary harmonic source distribution map and extracting the harmonic source positions, thereby obtaining a harmonic source position coordinate sequence; Step S56: tracking and reconstructing the harmonic current path according to the harmonic source position coordinate sequence, thereby obtaining harmonic current path data; performing path feature extraction on the harmonic current path data, thereby obtaining a path feature sequence; Step S57: Generate harmonic compensation strategy data according to the path characteristic sequence.

9. The distributed photovoltaic system grid-connected control method according to claim 8, characterized in that: Step S6 includes the following steps: Step S61: unifying the data formats of the distributed control strategy data, the grid-connected control strategy data, and the harmonic compensation strategy data, thereby obtaining three types of strategy data; Step S62: designing a Transformer-XL architecture suitable for multi-strategy fusion according to the three types of strategy data, thereby obtaining Transformer-XL model structure data; Step S63: performing data alignment and vector representation conversion based on timestamps on the three types of strategy data, and performing strategy fusion according to the Transformer-XL model structure data, thereby obtaining fusion strategy representation data; Step S64: generating an initial modulation mode sequence from the fusion strategy representation data; Step S65: Encode the initial modulation mode sequence into a control instruction and send it to the inverter via the industrial bus to execute SPWM.

10. A distributed photovoltaic system grid-connected control system, characterized in that: Used to execute the distributed photovoltaic system grid-connected control method according to claim 1, the distributed photovoltaic system grid-connected control system comprises: The LSTM modeling module is used to obtain historical multi-source heterogeneous data, including historical power output data, historical meteorological data, battery aging data, and array parameter data; extract multimodal environmental features from historical multi-source heterogeneous data, and capture the long-term dependency of time series based on a bidirectional LSTM network to obtain a sequence of LSTM modeling results; The MPPT strategy optimization module is used to deploy a dual alternating reinforcement learning framework based on the LSTM modeling result sequence and perform knowledge penetration to obtain MPPT strategy data; and to generate multi-modal power curve data based on the MPPT strategy data; The distributed control parameter module is used to determine the MPPT setting value, reactive power compensation ratio, and anti-islanding control parameters of the inverter according to the multi-modal power curve data using 5G edge computing technology and FPGA acceleration technology to obtain distributed control strategy data; The grid-connected path planning module is used to perform VGAE-based grid modeling according to the preset real-time grid topology data to obtain a grid topology model; perform timing optimization on the grid-connected path according to the grid topology model, and decompose the path into multi-phase grid-connected vector data; generate grid-connected control strategy data according to the multi-phase grid-connected vector data; The harmonic compensation strategy module is used to obtain electromagnetic sensor measurement data, and inversely interpret the electromagnetic imaging therein into harmonic source distribution, extract the harmonic current path, and obtain the path feature sequence; and generate harmonic compensation strategy data according to the path feature sequence; The multi-strategy fusion module is used to perform Transformer-XL fusion on the distributed control strategy data, the grid-connected control strategy data and the harmonic compensation strategy data to obtain an initial modulation mode sequence; the initial modulation mode sequence is encoded into a control instruction and sent to the inverter through the industrial bus to execute SPWM.

Images