Distributed photovoltaic power generation system optimization control method based on digital twinning

Abstract

The invention relates to the technical field of digital twinning, in particular to a distributed photovoltaic power generation system optimization control method based on digital twinning, and the method comprises the steps: constructing a photovoltaic module-level digital twinning body through multi-source data fusion, generating a digital mapping body through the training of historical fault data, and achieving the simulation of the operation state of a photovoltaic module; a multi-agent cluster is introduced, a multi-target optimization architecture framework coordinated with the multi-agent cluster is constructed, and balance and collaborative optimization among multiple targets are achieved through power generation efficiency optimization, equipment life management, power distribution coordination and other agents; and based on the candidate strategy after multi-target collaborative optimization, the improved grey wolf algorithm is used for further optimization processing, an optimal scheme is obtained, system optimization operation is realized, and power generation benefits and stability are improved.

Description

Technical Field

[0001] This invention relates to the field of digital twin technology, and in particular to an optimized control method for distributed photovoltaic power generation systems based on digital twins. Background Technology

[0002] Current photovoltaic (PV) power generation management still faces several challenges, including: traditional control strategies often struggle to balance multiple objectives (such as maximizing power generation efficiency, minimizing equipment losses, and ensuring grid security and stability), leading to poor optimization results. For example, pursuing maximum power generation efficiency may result in overuse of equipment, increasing equipment losses and failure risks; existing decision-making processes often rely on experience or simple rules, lacking data-driven decision support and failing to cope with complex and changing environmental conditions and system states. Therefore, this invention proposes an optimization control method for distributed PV power generation systems based on digital twins. Summary of the Invention

[0003] The purpose of this invention is to solve the problems in the background art by proposing an optimized control method for distributed photovoltaic power generation systems based on digital twins.

[0004] To achieve the above objectives, the present invention adopts the following technical solution: The optimization control method for distributed photovoltaic power generation systems based on digital twins includes: S1. Obtain a standardized digital twin input dataset, perform multi-source data fusion processing on the dataset to generate a component-level multi-dimensional time-series data representation, and use a deep learning model to construct a photovoltaic module-level digital twin. S2. Obtain historical fault data, inject the historical fault data into the constructed photovoltaic module-level digital twin, and use the feature embedding layer for training to finally generate a digital mapping body. S3. For digital mapping agents, introduce a multi-agent cluster and construct a multi-objective optimization architecture framework for multi-agent cluster collaboration to obtain the optimal solution.

[0005] Furthermore, the process of obtaining a standardized digital twin input dataset includes: By connecting to meteorological satellite data interfaces, installing sensors on inverters, and deploying an environmental sensor network, different types of data are acquired. Using the inverter timestamp as a benchmark, the time granularity of data with different resolutions is unified, and the spatial deviation of satellite remote sensing data is corrected, and the mapping relationship between sensors and photovoltaic modules is established. By constructing a spatiotemporal alignment algorithm, the spatiotemporal differences of the data are processed. Finally, the spatiotemporal aligned data is integrated and timestamps are added to form a standardized digital twin input dataset.

[0006] Furthermore, the process of generating component-level multidimensional time-series data representations by performing multi-source data fusion processing on the dataset, and constructing photovoltaic module-level digital twins using deep learning models includes: Data related to the photo-electric-thermal coupling field, including optical data, electrical data, and thermal data, are extracted from the standardized digital twin input dataset. Based on the established physical mapping relationship, the distributed sensor data are bound to the corresponding photovoltaic modules according to their corresponding numbers to form a module-level multidimensional time-series data representation, which is defined as a multidimensional time-series data cube. The deep learning model for processing coupled field data was determined, and a combination model of convolutional neural network (CNN) and recurrent neural network (RNN) was selected, namely the CNN-RNN hybrid network structure model, to capture the spatiotemporal features in the optical-electrical-thermal coupled field. The network structure of the deep learning model is constructed. The CNN-RNN hybrid network structure includes an input layer, hidden layers, and an output layer. The input layer receives all data from a multi-dimensional temporal data cube. The hidden layer consists of a CNN layer, an attention mechanism layer, and an RNN layer. The CNN layer is further subdivided into convolutional and pooling sub-layers. The convolutional sub-layers are responsible for extracting local spatiotemporal features, while the pooling sub-layers are used for dimensionality reduction and preservation of key features. An attention mechanism layer is inserted between the CNN and RNN layers to dynamically adjust the fusion weights of the light-electricity-thermal features. The RNN layer adopts an LSTM layer structure, which is an overall structure composed of multiple LSTM units used to process the entire sequence data. Each LSTM unit processes one time step in the sequence. The sequence data is a time-varying data sequence in the multi-dimensional temporal data cube, covering multiple dimensions of photovoltaic modules in the light-electricity-thermal coupling field. The core of the output layer is a fully connected layer used to output the implicit feature vector of the light-electricity-thermal coupling field. Based on the CNN-RNN hybrid network structure, the input layer, hidden layer, and output layer are configured; The extracted coupled field data is used to train the deep learning model, and the backpropagation algorithm and Adam optimizer are used to adjust the model parameters; finally, a photovoltaic module-level digital twin is constructed.

[0007] Furthermore, the process of acquiring historical fault data, injecting this data into the constructed photovoltaic module-level digital twin, and training it using a feature embedding layer to ultimately generate a digital mapping includes: Collect historical fault data of distributed photovoltaic power generation systems; organize and label the historical fault data: add corresponding labels to each fault sample, and set fault type labels through binary encoding; Add a feature embedding layer to the constructed photovoltaic module-level digital twin. This layer is designed to map high-dimensional fault features from historical fault data to a low-dimensional space. The parameters of the feature embedding layer are optimized by minimizing the loss function, enabling the feature embedding layer to accurately extract fault features; whereby minimizing the loss function... Using mean square error loss With feature embedding loss Combinations: In the formula, To balance the weighting coefficients of the two loss terms; mean squared error loss The formula used to measure the difference between the fault types predicted by the photovoltaic module-level digital twin and the actual fault types is as follows: In the formula, The number of fault samples. Number of fault types It is the first The first fault sample The true label of each fault type It is the first digital twin prediction at the photovoltaic module level. The first fault sample The probability of each fault type; Feature embedding loss The quality of feature embedding is measured using contrastive loss; for faulty samples... With fault samples Comparison of losses The calculation formula is: In the formula, For fault samples With fault samples Is it an indicator variable of the same type? For fault samples With fault samples Feature embedding and The distance between them The set boundary values; the final feature embedding loss. The average of the sum of the contrastive losses for all fault samples: ; A digital mapping volume is generated through deep learning fusion and feature embedding layer training; the digital mapping volume includes real-time optical-electrical-thermal state simulation results, aging decay trajectory and failure probability distribution.

[0008] Furthermore, for digital mapping agents, a multi-agent cluster is introduced, and a multi-objective optimization architecture framework for multi-agent cluster collaboration is constructed. The process of obtaining the optimal solution includes: Based on the generated digital mapping, different types of intelligent agents are defined and their functions are divided: the power generation efficiency optimization intelligent agent EEA adopts the deep reinforcement learning model DQN. Its input data comes from the real-time light-electric-thermal state of the digital mapping. By dynamically adjusting the maximum power point tracking (MPPT) parameters, it maximizes the instantaneous power output of the photovoltaic module. The intelligent agent EEA establishes a reward function based on the irradiance change rate and temperature gradient to balance power gain and energy consumption cost, and finally outputs MPPT step size adjustment suggestions. The LMA (Living Machine Management) agent is used for equipment life management, with the goal of minimizing the equipment aging rate. The LMA agent takes the aging degradation trajectory predicted by the digital mapping agent as input data and predicts the aging trend based on the LSTM-GRU hybrid model. At the same time, it combines the output results of the EEA (Energy Efficiency Optimization) agent to generate an equipment load reduction strategy that includes power limit thresholds and start-stop cycles through comprehensive analysis, and finally outputs dynamic load reduction instructions to guide equipment operation. The power allocation coordination agent PCA serves as the coordination hub, responsible for integrating the MPPT step size adjustment suggestions of the power generation efficiency optimization agent EEA, the equipment load reduction strategies of the equipment lifetime management agent LMA, and external grid dispatch instructions. Through comprehensive analysis and multi-objective optimization, it generates a set of power allocation strategies with equipment heterogeneity labels based on grid dispatch instructions, and finally outputs a set of candidate strategies after multi-objective collaborative optimization. A multi-agent cluster collaborative multi-objective optimization architecture framework is constructed, comprising three layers: agent layer, coordination layer, and decision layer. In the agent layer, agents with various functions are deployed to perform specific optimization tasks, namely power generation efficiency optimization, equipment lifetime management optimization, and power allocation optimization. The coordination layer is responsible for coordination and communication between agents. The decision layer is at the top of the architecture and is responsible for making the final decision based on the optimization results of the agents. The decision layer receives the candidate strategy set after multi-objective collaborative optimization output by the power allocation coordination agent PCA. Based on the candidate strategies after multi-objective collaborative optimization, the improved Grey Wolf algorithm is used for further optimization to obtain the optimal solution; the obtained optimal solution is then applied to the actual control of the distributed photovoltaic power generation system.

[0009] Furthermore, the process of further optimizing the candidate strategies based on multi-objective collaborative optimization using the improved Grey Wolf algorithm includes: The Grey Wolf algorithm updates the positions of candidate solutions by simulating wolf pack hunting behavior and integrates the set of candidate policies output by the power allocation coordinating agent PCA. The update formula is: In the formula, This represents the position vector of the new generation of gray wolves, i.e., the new strategy; This represents the current iteration number; These are the three optimal solutions at present; The exploration-development balance vector controls the gray wolf's movement direction: moving away during the exploration phase and moving closer during the development phase. Let be the prey distance vector. ,in For random perturbation vectors, The vector corresponding to the random candidate strategy; During the optimization process, multiple power allocation schemes in each candidate strategy are considered as candidate solutions in the Grey Wolf algorithm; a multi-attribute decision method is used to comprehensively evaluate the optimized candidate strategy set, i.e., a decision matrix is ​​constructed. For the candidate policy set One option, evaluate its performance. Performance under various optimization objectives: In the formula, For the first The plan is in Attribute values ​​under each optimization objective Indicates the number of options. This indicates the number of optimization objectives; based on the evaluation results, a preliminary selection is made from numerous options to meet the optimization objectives, and then the optimal option is selected from the preliminary selection.

[0010] Compared with existing technologies, the beneficial effects of this invention are as follows: By constructing a photovoltaic module-level digital twin, acquiring standardized datasets, fusing multi-source data, and using deep learning models to simulate the module's operating state, the system control can be based on a more accurate model, providing a solid foundation for subsequent optimized control, effectively avoiding control deviations, improving control accuracy and reliability, helping to identify potential problems in advance, and ensuring stable system operation; by introducing a multi-agent cluster and constructing a collaborative optimization architecture, it can simultaneously consider multiple objectives such as power generation efficiency, equipment lifespan, and power allocation, achieving balance and collaborative optimization; by using an improved gray wolf algorithm to further process candidate strategies after multi-objective collaborative optimization, and simulating wolf pack hunting behavior to update candidate solution positions, it comprehensively searches the optimal solution space, mines the optimal solution, and can efficiently find a solution that achieves optimal balance across multiple objectives, realizing optimized system operation, improving power generation efficiency and stability, and reducing operation and maintenance costs. Attached Figure Description

[0011] Figure 1 This is a flowchart of the optimized control method for a distributed photovoltaic power generation system based on digital twins proposed in this invention. Detailed Implementation

[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0013] Reference Figure 1 A digital twin-based optimization control method for distributed photovoltaic power generation systems includes: S1. Obtain a standardized digital twin input dataset, perform multi-source data fusion processing on the dataset to generate a component-level multi-dimensional time-series data representation, and use a deep learning model to construct a photovoltaic module-level digital twin. S2. Obtain historical fault data, inject the historical fault data into the constructed photovoltaic module-level digital twin, and use the feature embedding layer for training to finally generate a digital mapping body. S3. For digital mapping agents, introduce a multi-agent cluster and construct a multi-objective optimization architecture framework for multi-agent cluster collaboration to obtain the optimal solution.

[0014] It should be further explained that, in the specific implementation process, the process of obtaining a standardized digital twin input dataset, generating a component-level multidimensional time-series data representation by performing multi-source data fusion processing on the dataset, and constructing a photovoltaic module-level digital twin using a deep learning model involves: obtaining meteorological satellite remote sensing data through an interface with meteorological satellite data, including key parameters such as atmospheric optical thickness, cloud cover, and surface irradiance around the photovoltaic field; installing high-precision current, voltage, and power sensors on the inverters of the distributed photovoltaic power generation system to collect the inverter's operating parameters in real time, and transmitting the sensor data to the data acquisition terminal through an industrial communication protocol; and deploying a distributed sensor network (such as temperature sensors, humidity sensors, and wind speed sensors) around the distributed photovoltaic power generation system to monitor environmental parameters such as photovoltaic module surface temperature, wind speed, wind direction, and air temperature and humidity in real time, and acquiring environmental sensor data. Using the timestamps of inverter data as a benchmark, linear interpolation is performed on satellite remote sensing data (minute-level resolution), and downsampling and averaging are performed on environmental sensor data (second-level resolution) to unify the temporal granularity of all data. The regional irradiance of satellite remote sensing data is mapped to the geographic coordinate system of the photovoltaic field, and spatial deviations are corrected using a GIS (Geographic Information System). The locations of distributed sensors are calibrated, and their physical mapping relationship with photovoltaic modules is established using UWB positioning technology. The spatial correspondence between sensors and photovoltaic modules is clearly defined (e.g., module number C01 corresponds to latitude and longitude coordinates X, Y, and sensor number S01, etc.) to ensure the accuracy of subsequent data binding and processing. Addressing the temporal and spatial differences between meteorological satellite remote sensing data, inverter operating parameters, and environmental sensor data, a spatiotemporal alignment algorithm is constructed: temporally, time interpolation is used to unify data with different time resolutions to the same time step; spatially, GIS technology is used to perform spatial interpolation and registration of data with different spatial resolutions. The spatiotemporally aligned data is integrated into a single dataset, and a timestamp is added to each data point to form a standardized digital twin input dataset containing timestamps. Understandably, this dataset covers multi-dimensional information such as optical data (satellite remote sensing regional irradiance, component-level irradiance distribution after GIS correction), electrical data (inverter real-time power, DC / AC voltage and current), and thermal data (photovoltaic module surface temperature, ambient temperature and humidity). Data related to the photo-electric-thermal coupling field, including optical, electrical, and thermal data, is extracted from the standardized digital twin input dataset. Based on the established physical mapping relationship, the distributed sensor data is bound to the corresponding photovoltaic modules according to their corresponding numbers, forming a module-level multidimensional time-series data representation. This representation is defined as a multidimensional time-series data cube, which is set with [time step × number of sensor types × feature dimension] as the dimension. For example, [10 × 3 × 12] represents 10 time steps, 3 types of sensors, and 12 features for each type, comprehensively recording the feature information of the module under different times and different sensors. The deep learning model for processing coupled field data was determined, and a combination model of convolutional neural network (CNN) and recurrent neural network (RNN) was selected, namely the CNN-RNN hybrid network structure model, to capture the spatiotemporal features in the optical-electrical-thermal coupled field. The network structure of the deep learning model is constructed. The CNN-RNN hybrid network structure includes an input layer, hidden layers, and an output layer. The input layer receives all data from a multi-dimensional temporal data cube. The hidden layer consists of CNN layers, an attention mechanism layer, and RNN layers. The CNN layers are further subdivided into convolutional and pooling sub-layers. The convolutional sub-layers are responsible for extracting local spatiotemporal features (such as extracting local patterns of light intensity changes from optical data, for example, the rapid fluctuations in light intensity caused by cloud cover at different times of the day; extracting local features of voltage and current waveforms from electrical data, for example, the distorted waveforms of voltage and current when photovoltaic modules are partially shaded; and extracting local trends of temperature changes from thermal data, for example, the upward trend of temperature over time in locally overheated areas of the module). The pooling sub-layers are used for dimensionality reduction and to retain key features. An attention mechanism layer is inserted between the CNN and RNN layers to establish cross-domain connections. The fusion weights of light-electricity-thermal features are dynamically adjusted (for example, when the light intensity changes suddenly, the attention mechanism layer increases the weight of optical data features and correspondingly reduces the weight of electrical and thermal data features in the fusion process to highlight the impact of light changes on the overall coupled field; while when the component temperature continues to rise, the attention mechanism layer increases the weight of thermal data features so that the model pays more attention to the impact of temperature changes on the light-electricity conversion efficiency). The RNN layer adopts an LSTM layer structure. The LSTM layer is an overall structure composed of multiple LSTM units, used to process the entire sequence data. Each LSTM unit processes one time step in the sequence. The sequence data is a data sequence that changes over time in a multi-dimensional time-series data cube, covering multiple dimensions of photovoltaic module features in the light-electricity-thermal coupling field. The core of the output layer is a fully connected layer, used to output the implicit feature vector of the light-electricity-thermal coupling field, which comprehensively reflects the light-electricity-thermal coupling state of the module. Based on a CNN-RNN hybrid network structure, an input layer, hidden layers, and an output layer are set up. The number of nodes in the input layer is determined by the number of sensor types. The number of filters in the CNN layer gradually decreases from 64 to 32 (reducing the number of parameters while retaining core features). The number of LSTM units in the RNN layer is set to 128 to match the time series length of historical fault data and better capture long-term temporal dependencies. The input layer is directly connected to the CNN layer in the hidden layer, inputting the original multi-dimensional temporal data into the CNN layer for feature extraction. Inside the CNN layer, convolutional sub-layers and pooling sub-layers are connected in sequence, and the output of the CNN layer serves as the input to the attention mechanism layer. The attention mechanism layer receives the output of the CNN layer, calculates the weights of each feature through the attention mechanism, and outputs the weighted features to the RNN layer. The RNN layer receives the output of the attention mechanism layer, and its output serves as the input to the output layer. The fully connected layer is directly connected to the output of the RNN layer, mapping the output of the RNN layer to the desired output space through a linear transformation. The extracted coupled field data is used to train the deep learning model. The backpropagation algorithm and Adam optimizer are used to adjust the model parameters. The forward propagation process includes the coupled field data first passing through a CNN-RNN network to extract features. The high-dimensional features are then compressed into a low-dimensional space by a nonlinear encoder to generate a low-dimensional code representing the equipment fault. The features extracted by the CNN-RNN network include local spatiotemporal features and features after cross-field fusion. The cross-entropy error between the low-dimensional code and the fault label, as well as the MSE error between the coupled field prediction value and the true value, are compared to complete the loss calculation. The Adam optimizer is used to synchronously update the parameters of the CNN, RNN, and embedding layer to achieve gradient updates and continuously optimize the model. It is understood that the equipment mentioned in this embodiment are all photovoltaic modules and related power generation equipment. Ultimately, a photovoltaic module-level digital twin was constructed.

[0015] It should be further explained that, in the specific implementation process, the process of acquiring historical fault data, injecting the historical fault data into the constructed photovoltaic module-level digital twin, and using the feature embedding layer for training to finally generate the digital mapping is as follows: Historical fault data of distributed photovoltaic power generation systems is collected, including fault type, fault occurrence time, and related parameters. This historical fault data is then organized and labeled: a corresponding label is added to each fault sample, and fault type labels are set using binary encoding to represent fault types digitally, facilitating model processing. Assuming there are n fault types in the historical fault dataset, for the i-th fault sample, its fault type label... It is an n-dimensional binary vector, meaning that only the bit corresponding to the fault type is 1, and the rest are 0. For example, if there are 3 fault types (short circuit, open circuit, hot spot), and a sample has a short circuit fault type, then... ; Add a feature embedding layer to the constructed photovoltaic module-level digital twin. This layer is designed to map high-dimensional fault features from historical fault data to a low-dimensional space. Specifically, the high-dimensional fault features are compressed into a 16-dimensional latent space through a feature embedding layer to retain key fault criteria. During this process, the fault type label serves as a supervisory signal, guiding the feature embedding layer to learn and strengthen features sensitive to fault classification. The low-dimensional mapping method takes the high-dimensional fault features as input and sets them as... , For the set of real numbers, The number of fault parameters, Time step; Feature embedding layer It is a neural network layer, and its mapping function is expressed as: It is understandable that the specific computation process of completing the feature embedding layer through the fully connected layer includes: assuming the feature embedding layer has m neurons (m=16), high-dimensional fault features... After weight matrix With bias vector A linear transformation is performed, and then the mapping result is obtained through an activation function. ,Right now During this process, key fault criteria are retained, namely the abnormal temperature peak and the slope of the IV curve, where the abnormal temperature peak is the maximum temperature measured within a certain period before and after the fault occurs; the slope of the IV curve is obtained by fitting the IV curve using the least squares method: let the data points on the IV curve be... Then the slope of the IV curve : In the formula, For current, For voltage, For indexing data points, The number of data points, and ; By optimizing the parameters of the feature embedding layer by minimizing the loss function, the feature embedding layer can accurately extract fault features, thereby improving the fault identification capability of the photovoltaic module-level digital twin. It is understandable that minimizing the loss function... Using mean square error loss With feature embedding loss Combinations: In the formula, To balance the weighting coefficients of the two loss terms; Mean square error loss The formula used to measure the difference between the fault types predicted by the photovoltaic module-level digital twin and the actual fault types is as follows: In the formula, The number of fault samples. Number of fault types It is the first The first fault sample The true label of each fault type It is the first digital twin prediction at the photovoltaic module level. The first fault sample The probability of each fault type; Feature embedding loss The quality of feature embedding is measured using a contrastive loss method, specifically: feature embeddings of fault samples of the same class are closer together in the latent space, while feature embeddings of fault samples of different classes are farther apart; for fault samples... With fault samples Comparison of losses The calculation formula is: In the formula, For fault samples With fault samples Indicator variable for whether they belong to the same class (1 for the same class, 0 for different classes). For fault samples With fault samples Feature embedding and The distance between them The set boundary values; the total feature embedding loss. The average of the sum of the contrastive losses for all fault samples: ; A digital mapping volume is generated through deep learning fusion and feature embedding layer training; the digital mapping volume includes real-time optical-electrical-thermal state simulation results, aging decay trajectory and failure probability distribution; Specifically, for the real-time optical-electrical-thermal state simulation results, the power curve prediction for the next hour is used. For example, let's assume the current time is... Using time series forecasting methods to predict the next hour ( The power of ) is modeled as an autoregressive process, i.e.: In the formula, Let the order be the autoregressive order. For autoregressive order index, For time intervals, These are the autoregressive coefficients. These parameters are random error terms; a CNN-RNN model is used to learn these parameters to predict the power curve; the aging degradation trajectory is determined by the photovoltaic module efficiency. Over time The exponential decay curve was obtained as follows: In the formula, For initial efficiency, The attenuation coefficient is given by , where e is the base of the natural logarithm; for the failure probability distribution, the probability of hot spot effect occurring in the next 24 hours is used. For example, a logistic regression model is used for prediction. The form of the logistic regression model is as follows: In the formula, For the intercept term, The regression coefficients for each factor are denoted as . Factors influencing the occurrence of hot spot effect, , Index of factors influencing the occurrence of hot spot effect The number of factors that influence the occurrence of hot spot effect.

[0016] It should be further explained that, in the specific implementation process, for the digital mapping agent, a multi-agent cluster is introduced, and a multi-objective optimization architecture framework for multi-agent cluster collaboration is constructed to obtain the optimal solution. The process is as follows: Based on the generated digital mapping, different types of intelligent agents are defined and their functions are divided: the power generation efficiency optimization intelligent agent EEA adopts the deep reinforcement learning model DQN. Its input data comes from the real-time light-electric-thermal state of the digital mapping. By dynamically adjusting the maximum power point tracking (MPPT) parameters, it maximizes the instantaneous power output of the photovoltaic module. The intelligent agent EEA establishes a reward function based on the irradiance change rate and temperature gradient to balance power gain and energy consumption cost, and finally outputs MPPT step size adjustment suggestions. Specifically, the reward function The expression is: In the formula, Power gain is the difference between the instantaneous power output of the photovoltaic module after adjusting the MPPT parameters and the instantaneous power output before adjustment. Irradiance rate of change This represents the absolute value of the rate of change in irradiance. For temperature gradient, This represents the absolute value of the temperature gradient. Represents minute changes over time; This is the weighting coefficient for power gain, used to measure the importance of power gain in the reward function; These are the weighting coefficients for the rate of change of irradiance and the temperature gradient, used to balance the energy consumption cost. The LMA (Living Machine Management) agent is used for equipment life management, with the goal of minimizing the equipment aging rate. The LMA agent takes the aging degradation trajectory predicted by the digital mapping agent as input data and predicts the aging trend based on the LSTM-GRU hybrid model. At the same time, it combines the output results of the EEA (Energy Efficiency Optimization) agent to generate an equipment load reduction strategy that includes power limit thresholds and start-stop cycles through comprehensive analysis, and finally outputs dynamic load reduction instructions to guide equipment operation. Specifically, the power limiting threshold is calculated based on the current state of the equipment and the predicted aging and degradation trajectory. : In the formula, The rated power of the equipment; This is the power limiting factor, used to adjust the degree of power limiting; A mapping function from aging level to power reduction; start-stop cycle. The expression is: In the formula, The equipment aging cost under the start-stop cycle SC is related to the number of equipment start-stop cycles and the aging rate. The cost of power generation efficiency loss under start-stop cycle SC is related to power limitation and power fluctuations during start-stop. The power allocation coordination agent PCA serves as the coordination hub, responsible for integrating the MPPT step size adjustment suggestions of the power generation efficiency optimization agent EEA, the equipment load reduction strategies of the equipment lifetime management agent LMA, and external grid dispatch instructions. Through comprehensive analysis and multi-objective optimization, it generates a set of power allocation strategies with equipment heterogeneity labels based on grid dispatch instructions, and finally outputs a set of candidate strategies after multi-objective collaborative optimization. Specifically, a multi-objective optimization function is established: In the formula, For multi-objective optimization function vectors, The target is the power generation efficiency (obtained from the MPPT adjustment recommendations of the power generation efficiency optimization agent EEA). The equipment lifespan target is obtained from the load reduction strategy of the equipment lifespan management agent (LMA). The target of power grid dispatch (obtained through external power grid dispatch instructions, such as power fluctuation limits). The decision variables (such as power allocation ratio and equipment operating parameters) are taken from the power allocation strategy set; the candidate strategy set after multi-objective collaborative optimization is: Each strategy Includes device heterogeneity markers and optimization target values: , Indicates the candidate strategy index. Indicates the number of candidate strategies. Use it to mark the heterogeneity of equipment (such as equipment type, aging level). To optimize the objective value vector (power generation efficiency, lifetime loss, etc.); it is understandable that this solution set is an important basis for subsequent multi-objective optimization decisions, used to guide the screening and selection of subsequent power allocation strategies; Among them, the three types of intelligent agents achieve information interaction through message passing mechanism: the power generation efficiency optimization intelligent agent EEA and the equipment life management intelligent agent LMA send optimization suggestions to the power allocation coordination intelligent agent PCA in JSON format and Protobuf encoding respectively, while the power allocation coordination intelligent agent PCA issues global policies in XML format; A multi-agent cluster collaborative multi-objective optimization architecture framework is constructed, comprising three layers: agent layer, coordination layer, and decision layer. In the agent layer, agents with various functions are deployed to perform specific optimization tasks, namely power generation efficiency optimization, equipment lifetime management optimization, and power allocation optimization. The coordination layer is responsible for coordination and communication between agents. The decision layer is at the top of the architecture and is responsible for making the final decision based on the optimization results of the agents. The decision layer receives the candidate strategy set after multi-objective collaborative optimization output by the power allocation coordination agent PCA. Based on the candidate strategies optimized through multi-objective collaborative optimization, an improved Grey Wolf algorithm is used for further optimization. Specifically, the Grey Wolf algorithm updates the positions of candidate solutions by simulating wolf pack hunting behavior and integrates the candidate strategy set output by PCA. The update formula is as follows: In the formula, This represents the position vector of the new generation of gray wolves (candidate strategies), i.e., the new strategy; This is the current iteration number, used to mark the iteration round of the algorithm; These are the three optimal strategies, corresponding to the alpha wolf in the wolf pack. (Optimal Strategy), Suboptimal Wolf (Suboptimal strategy) and the third-best wolf (Third optimal strategy) The three solutions here are just examples, and the specific solution depends on the actual situation; To explore and develop the balance vector, which controls the gray wolf's movement direction: when When the time is right, it is the exploration phase (away from the prey), and the algorithm tends to explore underdeveloped regions in the candidate policy space; when At this time, it is the development stage (close to the prey), and the focus is on further exploring the current advantageous areas; Let be the prey distance vector. ,in For random perturbation vectors, The vector corresponding to the random candidate strategy; During the optimization process, multiple power allocation schemes within each candidate strategy are considered as candidate solutions (or strategy positions) in the Grey Wolf Algorithm. The Grey Wolf Algorithm iteratively updates the positions of these candidate solutions to find the optimal balance among multiple optimization objectives, including power generation efficiency, equipment lifespan, and power allocation. A multi-attribute decision-making method is then used to comprehensively evaluate the optimized candidate strategy set (i.e., a set containing multiple schemes optimized by the Grey Wolf Algorithm), thus constructing a decision matrix. For the candidate policy set One option, evaluate its performance. Performance under various optimization objectives: In the formula, For the first The plan is in Attribute values ​​(such as power generation efficiency value, lifetime loss value) under an optimization objective. Indicates the number of options. This indicates the number of optimization targets; based on the evaluation results, a preliminary selection of schemes that meet the optimization targets is made from numerous schemes, and the optimal scheme is selected from the preliminary selection; the optimal scheme is then applied to the actual control of the distributed photovoltaic power generation system to achieve optimized system operation.

[0017] It should be understood that in the various embodiments of this application, the order of the above-mentioned processes does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0018] It should be understood that determining B based on A does not mean determining B solely based on A; it also means determining B based on A and / or other information.

[0019] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0020] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An optimized control method for distributed photovoltaic power generation systems based on digital twins, characterized in that: S1. Obtain a standardized digital twin input dataset, perform multi-source data fusion processing on the dataset to generate a component-level multi-dimensional time-series data representation, and use a deep learning model to construct a photovoltaic module-level digital twin. S2. Obtain historical fault data, inject the historical fault data into the constructed photovoltaic module-level digital twin, and use the feature embedding layer for training to finally generate a digital mapping body. S3. For digital mapping agents, introduce a multi-agent cluster and construct a multi-objective optimization architecture framework for multi-agent cluster collaboration to obtain the optimal solution.

2. The optimized control method for a distributed photovoltaic power generation system based on digital twins according to claim 1, characterized in that: The process of obtaining a standardized digital twin input dataset includes: By connecting to meteorological satellite data interfaces, installing sensors on inverters, and deploying an environmental sensor network, different types of data are acquired. Using the inverter timestamp as a benchmark, the time granularity of data with different resolutions is unified, and the spatial deviation of satellite remote sensing data is corrected, and the mapping relationship between sensors and photovoltaic modules is established. By constructing a spatiotemporal alignment algorithm, the spatiotemporal differences of the data are processed. Finally, the spatiotemporal aligned data is integrated and timestamps are added to form a standardized digital twin input dataset.

3. The optimized control method for a distributed photovoltaic power generation system based on digital twins according to claim 1, characterized in that: The process of generating component-level multidimensional time-series data representations by performing multi-source data fusion processing on the dataset, and constructing photovoltaic module-level digital twins using deep learning models includes: Data related to the photo-electric-thermal coupling field, including optical data, electrical data, and thermal data, are extracted from the standardized digital twin input dataset. Based on the established physical mapping relationship, the distributed sensor data are bound to the corresponding photovoltaic modules according to their corresponding numbers to form a module-level multidimensional time-series data representation, which is defined as a multidimensional time-series data cube. The deep learning model for processing coupled field data was determined, and a combination model of convolutional neural network (CNN) and recurrent neural network (RNN) was selected, namely the CNN-RNN hybrid network structure model, to capture the spatiotemporal features in the optical-electrical-thermal coupled field. The network structure of the deep learning model is constructed. The CNN-RNN hybrid network structure includes an input layer, hidden layers, and an output layer. The input layer receives all data from a multi-dimensional temporal data cube. The hidden layer consists of a CNN layer, an attention mechanism layer, and an RNN layer. The CNN layer is further subdivided into convolutional and pooling sub-layers. The convolutional sub-layers are responsible for extracting local spatiotemporal features, while the pooling sub-layers are used for dimensionality reduction and to retain key features. An attention mechanism layer is inserted between the CNN and RNN layers to dynamically adjust the fusion weights of the light-electricity-thermal features. The RNN layer adopts an LSTM layer structure, which is an overall structure composed of multiple LSTM units used to process the entire sequence data. Each LSTM unit processes one time step in the sequence. The sequence data is a time-varying data sequence within the multi-dimensional temporal data cube, covering multiple dimensions of photovoltaic modules in the light-electricity-thermal coupling field. The core of the output layer is a fully connected layer used to output the implicit feature vector of the light-electricity-thermal coupling field. Based on the CNN-RNN hybrid network structure, the input layer, hidden layer, and output layer are configured; The extracted coupled field data is used to train the deep learning model, and the backpropagation algorithm and Adam optimizer are used to adjust the model parameters; finally, a photovoltaic module-level digital twin is constructed.

4. The optimized control method for a distributed photovoltaic power generation system based on digital twins according to claim 1, characterized in that: The process of acquiring historical fault data, injecting this data into a pre-constructed photovoltaic module-level digital twin, and training it using a feature embedding layer to ultimately generate a digital mapping includes: Collect historical fault data of distributed photovoltaic power generation systems; organize and label the historical fault data: add corresponding labels to each fault sample, and set fault type labels through binary encoding; Add a feature embedding layer to the constructed photovoltaic module-level digital twin. This layer is designed to map high-dimensional fault features from historical fault data to a low-dimensional space. The parameters of the feature embedding layer are optimized by minimizing the loss function, enabling the feature embedding layer to accurately extract fault features; whereby minimizing the loss function... Using mean square error loss With feature embedding loss Combinations: In the formula, To balance the weighting coefficients of the two loss terms; mean squared error loss The formula used to measure the difference between the fault types predicted by the photovoltaic module-level digital twin and the actual fault types is as follows: In the formula, The number of fault samples. Number of fault types It is the first The first fault sample The true label of each fault type It is the first digital twin prediction at the photovoltaic module level. The first fault sample The probability of each fault type; Feature embedding loss The quality of feature embedding is measured using contrastive loss; for faulty samples... With fault samples Comparison of losses The calculation formula is: In the formula, For fault samples With fault samples Is it an indicator variable of the same type? For fault samples With fault samples Feature embedding and The distance between them The set boundary values; the final feature embedding loss. The average of the sum of the contrastive losses for all faulty samples: ; A digital mapping volume is generated through deep learning fusion and feature embedding layer training; the digital mapping volume includes real-time optical-electrical-thermal state simulation results, aging decay trajectory and failure probability distribution.

5. The optimized control method for a distributed photovoltaic power generation system based on digital twins according to claim 1, characterized in that: For digital mapping agents, a multi-agent cluster is introduced, and a multi-objective optimization architecture framework for multi-agent cluster collaboration is constructed. The process of obtaining the optimal solution includes: Based on the generated digital mapping, different types of intelligent agents are defined and their functions are divided: the power generation efficiency optimization intelligent agent EEA adopts the deep reinforcement learning model DQN. Its input data comes from the real-time light-electric-thermal state of the digital mapping. By dynamically adjusting the maximum power point tracking (MPPT) parameters, it maximizes the instantaneous power output of the photovoltaic module. The intelligent agent EEA establishes a reward function based on the irradiance change rate and temperature gradient to balance power gain and energy consumption cost, and finally outputs MPPT step size adjustment suggestions. The LMA (Living Machine Management) agent is used for equipment life management, with the goal of minimizing the equipment aging rate. The LMA agent takes the aging degradation trajectory predicted by the digital mapping agent as input data and predicts the aging trend based on the LSTM-GRU hybrid model. At the same time, it combines the output results of the EEA (Energy Efficiency Optimization) agent to generate an equipment load reduction strategy that includes power limit thresholds and start-stop cycles through comprehensive analysis, and finally outputs dynamic load reduction instructions to guide equipment operation. The power allocation coordination agent PCA serves as the coordination hub, responsible for integrating the MPPT step size adjustment suggestions of the power generation efficiency optimization agent EEA, the equipment load reduction strategies of the equipment lifetime management agent LMA, and external grid dispatch instructions. Through comprehensive analysis and multi-objective optimization, it generates a set of power allocation strategies with equipment heterogeneity labels based on grid dispatch instructions, and finally outputs a set of candidate strategies after multi-objective collaborative optimization. A multi-agent cluster collaborative multi-objective optimization architecture framework is constructed, comprising three layers: agent layer, coordination layer, and decision layer. In the agent layer, agents with various functions are deployed to perform specific optimization tasks, namely power generation efficiency optimization, equipment lifetime management optimization, and power allocation optimization. The coordination layer is responsible for coordination and communication between agents. The decision layer is at the top of the architecture and is responsible for making the final decision based on the optimization results of the agents. The decision layer receives the candidate strategy set after multi-objective collaborative optimization output by the power allocation coordination agent PCA. Based on the candidate strategies optimized by multi-objective collaborative optimization, the improved Grey Wolf algorithm is used for further optimization to obtain the optimal solution; the obtained optimal solution is then applied to the actual control of the distributed photovoltaic power generation system.

6. The optimized control method for a distributed photovoltaic power generation system based on digital twins according to claim 5, characterized in that: The process of further optimizing the candidate strategies based on multi-objective collaborative optimization using the improved Grey Wolf algorithm includes: The Grey Wolf algorithm updates the positions of candidate solutions by simulating wolf pack hunting behavior and integrates the set of candidate policies output by the power allocation coordinating agent PCA. The update formula is: In the formula, This represents the position vector of the new generation of gray wolves, i.e., the new strategy; This represents the current iteration number; These are the three optimal solutions at present; The exploration-development balance vector controls the gray wolf's movement direction: moving away during the exploration phase and moving closer during the development phase. Let be the prey distance vector. ,in For random perturbation vectors, The vector corresponding to the random candidate strategy; During the optimization process, multiple power allocation schemes in each candidate strategy are considered as candidate solutions in the Grey Wolf algorithm; a multi-attribute decision method is used to comprehensively evaluate the optimized candidate strategy set, i.e., a decision matrix is ​​constructed. For the candidate policy set One option, evaluate its performance. Performance under various optimization objectives: In the formula, For the first The plan is in Attribute values ​​under each optimization objective Indicates the number of options. This indicates the number of optimization objectives; based on the evaluation results, a preliminary selection of the solutions that meet the optimization objectives is made from among many solutions, and the optimal solution is selected from the preliminary selection.

Images