A Flexible Adjustable Control Method and Device for a Distributed Photovoltaic System

Abstract

The present application provides a flexible adjustable control method and device for a distributed photovoltaic system, relating to the technical field of intelligent scheduling of power networks. The method includes: an edge node obtains multimodal sensing time-series data of a deployed photovoltaic array in a historical preset time period and weather forecast data corresponding to a future preset time period; based on a time-series prediction model, predicts the expected power generation amount and the expected power supply load demand of the corresponding photovoltaic array; based on a scheduling decision model, predicts the initial scheduling parameters of the corresponding photovoltaic array; sends the expected power generation amount, the expected power supply load demand, and the initial scheduling parameters to a core node, so that the core node updates the target scheduling parameters for each edge node; receives the corresponding target scheduling parameters from the core node, and regulates the corresponding photovoltaic array according to the target scheduling parameters. Thus, a global optimization dynamic cooperative control method is implemented in a large-scale distributed photovoltaic system, improving the reliability of global flexible scheduling.

Description

Technical Field

[0001] This application relates to the technical field of intelligent power network dispatching, and particularly to a flexible adjustable control method and device for a distributed photovoltaic system. Background Art

[0002] With the growing global demand for clean energy, distributed photovoltaic systems, as an important form of renewable energy, have been widely applied. Through distributed photovoltaic systems, solar energy resources can be utilized to directly generate electricity at the user side or near the user and supply power to the power grid. However, with the large-scale deployment of distributed photovoltaic systems, some technical challenges have gradually emerged.

[0003] Currently, most distributed photovoltaic systems adopt control strategies based on fixed parameters. Although various typical working conditions are considered in the initial design, in actual applications, due to the volatility of solar energy resources and the uncertainty of user load demands, the output of the photovoltaic system often needs to be dynamically adjusted. However, the existing fixed-parameter control strategies cannot effectively cope with these changes, resulting in frequent mismatches between the output of the photovoltaic system and the grid demand. This control rigidity not only reduces the power generation efficiency of the photovoltaic system but also increases the difficulty of grid dispatching.

[0004] In response to the above problems, the industry has not yet proposed a better technical solution. Summary of the Invention

[0005] This application provides a flexible adjustable control method, device, storage medium, computer program product, and electronic device for a distributed photovoltaic system, so as to at least solve the problem of the mismatch between the output of the photovoltaic system and the grid demand caused by the current control rigidity of the photovoltaic system.

[0006] In a first aspect, an embodiment of the present application provides a flexible adjustable control method for a distributed photovoltaic system. The distributed photovoltaic system includes multiple photovoltaic arrays, and a corresponding sensor network and edge node are respectively deployed for each photovoltaic array. The method includes: each edge node is respectively used to perform the following operations: obtaining multi-modal sensing time-series data of the deployed photovoltaic array in a historical preset time period and weather forecast data corresponding to a future preset time period; the multi-modal sensing time-series data includes multiple adjacent historical time steps and corresponding multi-modal sensing data, and the parameter types of the multi-modal sensing data include: solar radiation intensity, ambient temperature, power generation amount, grid operation parameters, and power supply load demand; inputting the multi-modal sensing time-series data and the weather forecast data into a time-series prediction model to predict the expected power generation amount and expected power supply load demand of the corresponding photovoltaic array in the future preset time period; the time-series prediction model adopts an LSTM (Long Short-Term Memory) model; processing the corresponding expected power generation amount and expected power supply load demand based on a scheduling decision model to predict the initial scheduling parameters of the corresponding photovoltaic array; the parameter types of the scheduling parameters include the target output power and the grid connection voltage regulation value; sending the expected power generation amount, the expected power supply load demand, and the initial scheduling parameters to a core node, so that the core node performs global analysis on the data sent by each edge node to generate target scheduling parameters for each edge node; receiving the corresponding target scheduling parameters from the core node, and regulating the corresponding photovoltaic array in the future preset time period according to the target scheduling parameters.

[0007] Second aspect, an embodiment of the present application provides a flexible adjustable control device for a distributed photovoltaic system. The distributed photovoltaic system includes multiple photovoltaic arrays, and a corresponding sensor network and edge node are respectively deployed for each photovoltaic array. The device includes an acquisition unit, a supply and demand prediction unit, an initial scheduling prediction unit, a data upload unit, and a target scheduling control unit arranged in the edge node; the acquisition unit is used to acquire the multi-modal sensing time-series data of the deployed photovoltaic array in a historical preset time period and the weather forecast data corresponding to a future preset time period; the multi-modal sensing time-series data includes multiple adjacent historical time steps and corresponding multi-modal sensing data, and the parameter types of the multi-modal sensing data include: solar radiation intensity, ambient temperature, power generation amount, grid operation parameters, and power supply load demand; the supply and demand prediction unit is used to input the multi-modal sensing time-series data and the weather forecast data into a time-series prediction model to predict the expected power generation amount and the expected power supply load demand of the corresponding photovoltaic array in the future preset time period; the time-series prediction model adopts an LSTM model; the initial scheduling prediction unit is used to process the corresponding expected power generation amount and the expected power supply load demand based on a scheduling decision model to predict the initial scheduling parameters of the corresponding photovoltaic array; the parameter types of the scheduling parameters include the target output power and the grid connection voltage regulation value; the data upload unit is used to send the expected power generation amount, the expected power supply load demand, and the initial scheduling parameters to the core node, so that the core node performs global analysis on the data sent by each edge node to generate the target scheduling parameters for each edge node; the target scheduling control unit is used to receive the corresponding target scheduling parameters from the core node and regulate the corresponding photovoltaic array in the future preset time period according to the target scheduling parameters.

[0008] Third aspect, an electronic device is provided, which includes: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and when the instructions are executed by the at least one processor, the at least one processor is enabled to execute the steps of the flexible adjustable control method for a distributed photovoltaic system according to any embodiment of the present application.

[0009] Fourth aspect, an embodiment of the present application provides a storage medium, on which a computer program is stored, and is characterized in that when the program is executed by a processor, the steps of the flexible adjustable control method for a distributed photovoltaic system according to any embodiment of the present application are implemented.

[0010] Fifth aspect, an embodiment of the present application provides a computer program product, including a computer program / instructions, and when the computer program / instructions are executed by a processor, the steps of the flexible adjustable control method for a distributed photovoltaic system according to any embodiment of the present application are implemented.

[0011] Through a flexible adjustable control method for a distributed photovoltaic system provided by this application, the following technical effects can be achieved at least:

[0012] (1) By introducing the joint input of multi-modal sensing time-series data in historical periods and weather forecast data in future periods, and using a time-series prediction model (i.e., the LSTM model) to accurately predict future photovoltaic power generation and power supply load demand, it can effectively cope with the volatility of solar energy resources and the uncertainty of user load demand, and can perform pre-scheduling based on this to optimize and adjust the output of the photovoltaic system in advance, avoiding frequent scheduling operations caused by the mismatch between load and power supply.

[0013] (2) Based on the global analysis of the data sent by each edge node by the core node, the target scheduling parameters of each photovoltaic array are updated. By comprehensively analyzing the predicted power generation and load demand of multiple photovoltaic arrays, the scheduling parameter configuration is optimized globally. Thus, it can not only balance the power supply pressure between different photovoltaic arrays, but also effectively avoid the decline of global performance caused by local optimality, ensuring the overall power generation performance of the photovoltaic system.

[0014] (3) By adopting the combination of multi-modal sensing data and time-series prediction model, even in the face of extreme weather changes or sudden load demands, the system can quickly adjust through the node prediction and global scheduling mechanism, ensuring the stable operation of the photovoltaic system, having stronger robustness and anti-interference ability, being able to effectively cope with the changing external environment, and ensuring the reliability of the system under complex working conditions.

[0015] Through this technical solution, edge network technology is used to achieve flexible adjustable control of the distributed photovoltaic system. The core node conducts global analysis on the initial scheduling parameters of each edge node to generate target scheduling parameters, comprehensively considering the synergy between different arrays, and achieving a globally optimal scheduling strategy. Thus, implementing a globally optimized collaborative control method in a large-scale distributed photovoltaic system can effectively reduce the mutual interference between arrays and enhance the overall coordination and reliability of the distributed photovoltaic system. Description of the Drawings

[0016] In order to more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of this application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0017] Figure 1 Shows a flowchart of an example of the flexible adjustable control method for a distributed photovoltaic system according to an embodiment of this application;

[0018] Figure 2 Shows an operation flowchart of an example of generating target scheduling parameters for each edge node based on a core node according to an embodiment of the present application;

[0019] Figure 3 Shows a schematic structural connection diagram of an example of a multi-scale attention LSTM model according to an embodiment of the present application;

[0020] Figure 4 Shows an operation flowchart of an example of a scheduling decision model predicting initial scheduling parameters of a corresponding photovoltaic array according to an embodiment of the present application;

[0021] Figure 5 Shows a schematic structural diagram of an example of a flexible adjustable control device for a distributed photovoltaic system according to an embodiment of the present application;

[0022] Figure 6 Is a schematic structural diagram of an embodiment of an electronic device of the present application. Specific embodiments

[0023] To make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present application. Apparently, the described embodiments are some, but not all, of the embodiments of the present application. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present application without creative efforts shall fall within the protection scope of the present application.

[0024] In the technical solutions of the present application, for the processing of collection, storage, use, processing, transmission, provision, and disclosure of user personal information involved, etc., they all comply with the provisions of relevant laws and regulations and do not violate public order and good customs.

[0025] Figure 1 Shows a flowchart of an example of a flexible adjustable control method for a distributed photovoltaic system according to an embodiment of the present application.

[0026] It should be understood that the edge computing network includes a core node and multiple edge nodes. Each edge node and the core node have a certain computing ability. The core node can be used to manage each edge node. The edge node can share a certain resource processing pressure for the core node, and can support the implementation of some local processing, and has the characteristics of being faster and more efficient.

[0027] In some examples of the embodiments of the present application, the distributed photovoltaic system includes multiple photovoltaic arrays, and a corresponding sensor network and edge node are respectively deployed for each photovoltaic array.

[0028] Regarding the execution entity of the method in the embodiments of the present application, it can be any controller or processor with computing or processing capabilities. By introducing a flexible adjustable control method based on a time series prediction model, comprehensively utilizing multi-modal data, dynamic time series prediction, globally optimized scheduling decisions, and the regulation of an edge computing network, the adaptability, efficiency, and stability of a distributed photovoltaic system are significantly improved, successfully solving the rigidity problem of fixed parameter control strategies in the prior art and achieving efficient, flexible, and intelligent management of the photovoltaic system.

[0029] In some examples, it can be an edge node, which is integrated and configured in an electronic device or terminal in a software, hardware, or software-hardware combination manner, and the types of terminals or electronic devices can be diverse, such as mobile phones, tablets, or desktop computers, etc.

[0030] As Figure 1 shown, in step S110, multi-modal sensing time series data of the deployed photovoltaic array in a historical preset time period and weather forecast data for a corresponding future preset time period are acquired.

[0031] Here, the time lengths of the historical preset time period and the future preset time period can be diverse and can be adjusted according to service requirements.

[0032] In some embodiments, a corresponding multi-modal sensor network is deployed for each photovoltaic array to continuously collect key parameters in the operating environment of the photovoltaic array, such as solar radiation intensity, ambient temperature, power generation, grid operating parameters, and power supply load demand, etc. Then, the edge node performs preliminary filtering and denoising processing on the data collected by the sensors to ensure the accuracy and reliability of the data. For each sensor, the data is stored in a time series format to form multi-modal sensing time series data, which contains multiple adjacent historical time steps and corresponding multi-modal sensing data. In addition, to ensure the accuracy of the data, weather forecast data can be obtained by interacting with a weather server based on a real-time update mechanism, which mainly contains prediction information on future solar radiation intensity and ambient temperature.

[0033] In step S120, the multi-modal sensing time series data and the weather forecast data are input into a time series prediction model to predict the expected power generation and expected power supply load demand of the corresponding photovoltaic array in a future preset time period. The time series prediction model uses an LSTM model.

[0034] Here, the LSTM model can have been offline trained with a large amount of historical data and continuously optimized. When applied online, the LSTM model will gradually predict the photovoltaic power generation and load demand within a preset future time period based on the input multi-modal time series data and weather forecast data. Since LSTM can effectively capture the long-term dependencies in the time series, it improves the prediction accuracy of photovoltaic power generation and power supply load demand and reduces the prediction error.

[0035] In step S130, the corresponding predicted photovoltaic power generation and predicted power supply load demand are processed based on the scheduling decision model to predict the initial scheduling parameters of the corresponding photovoltaic array. The parameter types of the scheduling parameters include the target output power and the grid connection voltage regulation value.

[0036] Here, the scheduling decision model can adopt various machine learning models, such as deep learning models or mathematical physics models, etc. Taking the predicted power generation and load demand as inputs, and combining the operating status of the photovoltaic array and the real-time parameters of the power grid, it calculates the initial scheduling parameters. Exemplarily, the scheduling decision model continuously optimizes the initial scheduling parameters according to the historical scheduling results and current prediction data to ensure that the optimal output power and voltage regulation can be provided under different array environmental conditions. Through the output power setting value, the best power generation utilization rate of the photovoltaic array is achieved. Through the grid connection voltage regulation value, the grid connection voltage between the photovoltaic array and the power grid is adjusted to ensure the stability of the power grid.

[0037] In step S140, the predicted photovoltaic power generation, predicted power supply load demand, and initial scheduling parameters are sent to the core node, enabling the core node to perform global analysis on the data sent by each edge node to generate the target scheduling parameters for each edge node.

[0038] In some embodiments, the edge nodes send their respective predicted photovoltaic power generation, power supply load demand, and initial scheduling parameters to the core node through a dedicated communication network, enabling the core node to collect the data of all edge nodes and use the global scheduling algorithm to uniformly analyze the initial scheduling parameters of each photovoltaic array, comprehensively considering the correlation between each photovoltaic array and the overall power grid demand, and generating globally optimized target scheduling parameters. Thus, the core node updates the target scheduling parameters of each photovoltaic array through global data analysis to make them more in line with the actual needs, effectively avoiding the problem of local optimality and improving the overall scheduling efficiency of the photovoltaic system.

[0039] In step S150, the corresponding target scheduling parameters are received from the core node, and the corresponding photovoltaic array is regulated according to the target scheduling parameters within the preset future time period.

[0040] In some embodiments, after receiving the target scheduling parameters sent by the core node, the edge node applies them to the corresponding photovoltaic array control system to adjust the output power and grid-connected voltage of each photovoltaic array in real time. In a preset future time period, the system continuously adjusts the operating state of the photovoltaic array according to the target scheduling parameters to ensure that the output meets the grid demand and maximally utilizes solar energy resources.

[0041] Through the embodiments of the present application, by applying the globally optimized target scheduling parameters, the photovoltaic array can achieve efficient operation under various working conditions, ensuring that the output power and grid-connected voltage meet the grid requirements. Through the real-time regulation and feedback mechanism of the edge node, the photovoltaic system can dynamically respond to environmental changes and grid demands, continuously optimize the operating state, and achieve the goal of flexible adjustable control.

[0042] Thus, by combining the advantages of edge computing and hierarchical decision-making, a multi-level scheduling from local rapid response to global optimization is achieved, which can maintain the efficient operation of the system in a complex and changeable environment and cope with the uncertainty of photovoltaic power generation.

[0043] Figure 2 The operation flowchart showing an example of generating target scheduling parameters for each edge node based on the core node according to the embodiments of the present application is shown.

[0044] As Figure 2 shown, in step S210, according to the predicted photovoltaic power generation amount, predicted power supply load demand, and initial scheduling parameters sent by each edge node, a photovoltaic scheduling graph structure is constructed.

[0045] Here, the photovoltaic scheduling graph structure includes multiple graph nodes and edge connections. Each graph node corresponds uniquely to an edge node, and the node characteristics of the graph node are defined by the corresponding predicted photovoltaic power generation amount, predicted power supply load demand, and initial scheduling parameters. In addition, the edge weight of each edge connection is determined by the supply-demand complementarity degree, power transmission loss rate, and historical cooperation efficiency between the corresponding photovoltaic arrays of the connected graph nodes.

[0046] Here, the magnitude of the edge weight can directly reflect the association strength between the corresponding photovoltaic arrays. The supply-demand complementarity degree is used to evaluate the complementarity between the power supply and demand of two photovoltaic arrays. The higher the complementarity degree, the greater the edge weight. The power transmission loss rate is used to calculate the loss rate during the power transmission process between two arrays. The smaller the power transmission loss rate, the greater the edge weight. The historical cooperation efficiency can be calculated through historical data, which can evaluate the cooperation efficiency between two arrays. The higher the cooperation efficiency, the greater the edge weight.

[0047] By constructing the photovoltaic scheduling graph structure, the complex relationships of the distributed photovoltaic system are abstracted into a graph model. The multi-dimensional calculation of edge weights ensures the comprehensiveness of the relationships between photovoltaic arrays, covering key factors such as supply-demand matching, power transmission loss, and historical collaboration, enabling better modeling of the relationships between photovoltaic arrays and exploring potential collaborative optimization opportunities.

[0048] In step S220, the photovoltaic scheduling graph structure is processed based on a graph neural network to update the node features of each graph node.

[0049] In some embodiments, advanced network architectures such as Graph Convolutional Network (GCN) or deep graph neural networks can be used. Through the graph neural network, the feature information of each node is connected, transmitted, and fused with the feature information of its adjacent nodes through the edges of the graph. In each layer of the graph neural network, the node feature vector is weighted and summed with the feature vectors of neighboring nodes. After being processed by multiple layers of the graph neural network, the initial node features are updated to more expressive high-order features. The graph neural network uses an information propagation mechanism, which not only considers the information of a single node but also integrates the information of neighboring nodes and the global graph structure, capturing the local and global information of each node and making the update of node features more comprehensive and accurate.

[0050] In step S230, the node features of each updated graph node are processed based on the global scheduling optimization model to determine the target scheduling parameters of each edge node.

[0051] Here, the global scheduling optimization model adopts a genetic algorithm model. The population individuals of the genetic algorithm model are defined by the set of scheduling parameters of all photovoltaic arrays in the distributed photovoltaic system, and each gene in the population individual is defined by the scheduling parameters of the corresponding photovoltaic array. Specifically, the set of scheduling parameters of all photovoltaic arrays in the distributed photovoltaic system is defined as a population individual, and each individual consists of a series of genes. In the genetic algorithm, a gene represents a specific parameter in the scheduling scheme, such as various scheduling parameters like the target output power and grid connection voltage regulation value. The initial population can be generated randomly or based on historical data.

[0052] In addition, the fitness function is defined based on the global supply-demand balance degree, global grid stability coefficient, and global transmission loss rate of the distributed photovoltaic system. The supply-demand balance degree reflects the matching degree between power generation and demand, the grid stability coefficient measures the operating stability of the entire power grid, and the transmission loss rate evaluates the power transmission loss. Furthermore, the fitness of each population individual (i.e., the set of scheduling parameters) is evaluated by the weighted sum of the above three. The higher the fitness, the more the set of scheduling parameters of this individual can meet the global optimization requirements of the system.

[0053] Furthermore, the next-generation population is generated through selection, crossover, and mutation operations. The selection operation preferentially selects individuals with higher fitness based on fitness. The crossover operation generates new individuals by exchanging parts of the genes of two parent individuals, and the mutation operation randomly changes certain genes to increase population diversity. Through multiple generations of iteration by the genetic algorithm, it gradually converges to the optimal solution, thereby determining the optimal individual and obtaining the target scheduling parameters for each photovoltaic array. Thus, by utilizing the global search and optimization of the genetic algorithm, which simulates the process of natural selection and can handle complex multi-objective optimization problems, it can search for scheduling parameters close to the global optimum in a vast solution space, enabling the system to generate an optimal combination of scheduling parameters and ensuring the efficient and coordinated operation of the photovoltaic system in a complex power grid environment.

[0054] Through the embodiments of the present application, the scheduling problem of the photovoltaic system is modeled as a graph structure and graph neural networks are used to extract features, combined with the genetic algorithm for global optimization. By comprehensively considering the supply-demand balance, grid stability, and global transmission loss rate, it can effectively achieve the global scheduling optimization of the distributed photovoltaic system, ensure the coordinated operation between each photovoltaic array, quickly determine the optimal scheduling parameters, and ensure the stability and efficient operation of the system.

[0055] In some examples of the embodiments of the present application, the time series prediction model can be enhanced and designed in combination with the technical scenario. More specifically, the time series prediction model can adopt a multi-scale attention LSTM model. Here, the multi-scale attention LSTM model combines multi-scale LSTM layers and a hybrid attention mechanism, and by processing data of different time scales and dynamically weighting input features, it improves the prediction accuracy of the power generation amount and power supply load demand in the distributed photovoltaic system.

[0056] Figure 3 The structural connection diagram of an example of the multi-scale attention LSTM model according to the embodiments of the present application is shown.

[0057] As Figure 3 shown, the multi-scale attention LSTM model 300 includes a multi-scale LSTM layer 310, a hybrid attention mechanism layer 320, and an output layer 330.

[0058] The multi-scale LSTM layer 310 includes multiple LSTM layers (3111, 3113…311n) and a fusion layer 312, and each LSTM layer has a corresponding time scale. Here, the LSTM layer of each time scale is respectively used to process the input data of different time windows to generate corresponding scale hidden states.

[0059] Here, through LSTM layers with different time scales, different patterns of photovoltaic power generation and load demand in the short, medium, and long terms are captured. These time scales correspond to different time windows. For example, the short time scale is hourly, the medium time scale is daily, and the long time scale is monthly or seasonal.

[0060] More specifically, the scale hidden states of each time window are calculated as follows:

[0061]

[0062] In the formula, represents the scale hidden state at the t-th time step under the time scale s, represents the multi-modal sensing time series data at the historical time step t′, represents the weather forecast data at the future time step t″; represents the hidden state at the (t - 1)-th time step.

[0063] By introducing multi-scale LSTM layers, the model can process input data at different time scales, thus finely capturing the different changing trends of photovoltaic power generation and power supply load demand in the short, medium, and long terms. For example, the LSTM layer with a short time scale can effectively respond to sudden weather changes, while the LSTM layer with a long time scale can identify seasonal changes and long-term trends, making the prediction more accurate and comprehensive.

[0064] The hybrid attention mechanism layer 320 includes a time attention module 321 and a feature attention module 323. The time attention module 321 dynamically adjusts the weight of each time step, and the feature attention module 323 weights according to the importance of different features.

[0065] More specifically, the hybrid attention mechanism is expressed as follows:

[0066]

[0067]

[0068] In the formula, represents the time attention weight at the t′-th historical time step under the time scale s, and respectively represent the weight matrix and bias term corresponding to the time attention mechanism for the time scale s; tanh represents the hyperbolic tangent function, represents the hidden state at the t′-th historical time step, T H represents the total number of historical time steps corresponding to the multi-modal sensing time series data; represents the feature attention weight of the j-th feature under the time scale s; represents the feature representation of the j-th feature in the historical multi-modal time series data over all historical time steps, is the feature representation of the j-th feature in the weather forecast data over all future time steps, is and the joint feature representation; and respectively represent the weight matrix and bias term corresponding to the time scale s of the feature attention mechanism, and n represents the total number of features; represents the weighted hidden state at time scale s, represents the hidden state corresponding to feature j at historical time step t′ and time scale s.

[0069] The time steps and input features are weighted through the hybrid attention mechanism to dynamically adjust the model's attention to important information. By applying the time attention mechanism, the weights of each time step are dynamically adjusted to ensure that the model can focus on the most critical time steps during prediction. By applying the feature attention mechanism and weighting according to the importance of different features, the model can dynamically adjust its attention to different features according to the actual situation, improving the prediction accuracy.

[0070] The output layer 330 is used to fuse the weighted hidden states of each scale to predict the expected photovoltaic power generation and the expected power supply load demand of the corresponding photovoltaic array in the preset future time period.

[0071] Here, by integrating information of different time scales together, a comprehensive representation is formed to predict future photovoltaic power generation and load demand.

[0072] More specifically, the structure of the output layer is expressed by the following formula:

[0073]

[0074] where, is the fused multi-scale weighted hidden state, γ (s) represents the scale fusion weight of time scale s, and S represents the total number of time scales corresponding to the multi-scale LSTM layer; represents the expected photovoltaic power generation and the expected power supply load demand of the final output of the model; W o and b o respectively represent the weight matrix and bias term of the output layer.

[0075] Here, by introducing the time scale weight γ (s) , the user can appropriately adjust the decision-making process of the model to control the influence of information of different time scales on the final prediction, making the model more flexible in practical applications and able to be adjusted according to specific requirements.

[0076] In some examples of the embodiments of the present application, the multi-scale attention LSTM model uses the mean squared error (MSE) as the loss function to minimize the error between the predicted value and the true value. In addition, in the training optimization process, an adaptive learning rate optimization algorithm (such as Adam) can be used for model training to accelerate convergence and avoid getting stuck in local optima.

[0077] Here, on the one hand, the scale fusion weight γ (s) can be set by prior knowledge. For example, if it is known that a certain time scale (such as seasonal changes) is particularly important for prediction, a higher weight can be artificially assigned to this time scale. On the other hand, it can also be obtained by calculating the time scale scores of each time scale. The calculation of this time scale score can be based on the hidden state of each time scale, and then the scores are converted into weights γ (s) by using the Softmax function, so as to ensure that the sum of the weights of all time scales is 1.

[0078] More preferably, in some examples of the embodiments of the present application, the scale fusion weight is obtained through the following calculation:

[0079]

[0080]

[0081] In the formula, g (s) represents the time scale score; A t represents the array state vector at time step t, and this array state vector is determined according to the power generation fluctuation amplitude and load change rate at time step t; W γ and b γ respectively represent the weight matrix and bias term related to the time scale, and V γ is the weight matrix related to the array state vector A t ; represents the sum of the exponential values of the scores of all time scales.

[0082] In the embodiments of the present application, by comprehensively considering the power generation fluctuation amplitude and load change rate through the array state vector A t it can more accurately reflect the actual operating state of the array at different time steps. The dynamic state representation enables the model to respond in a timely manner to the actual needs and changes of the system during prediction, thereby improving the real-time performance and accuracy of the prediction. For example, when the system load suddenly increases or the power generation fluctuates violently, the array state vector can guide the model to pay more attention to the changes on the short time scale, ensuring that the scheduling strategy can quickly respond to the changing environmental conditions and maintain the stability of the system. Thus, the array state vector dynamically adjusts γ in the multi-scale LSTM model (s), enabling the model to flexibly allocate weights between different time scales and improving the effect of multi-scale information fusion.

[0083] It should be noted that the relationship between photovoltaic power generation, power supply load demand, and scheduling parameters is usually non-linear. It may be difficult to accurately capture the system behavior directly using a linear model. By segmenting the non-linear relationship and performing linear approximation within each interval, the prediction accuracy can be improved while maintaining the computational efficiency.

[0084] In view of this, in some examples of the embodiments of the present application, for the scheduling decision model, it includes multiple local linear model modules, and each local linear model module has a corresponding supply-demand parameter interval. In a distributed photovoltaic system, the division of the supply-demand parameter interval is to capture the non-linear relationship between photovoltaic power generation and power supply load demand. By dividing these parameters into multiple intervals, a local linear model can be used for approximation processing within each interval, thereby improving the accuracy and adaptability of the scheduling decision.

[0085] Figure 4 The operation flowchart of an example showing the scheduling decision model according to the embodiments of the present application predicting the initial scheduling parameters of the corresponding photovoltaic array is shown.

[0086] As Figure 4 shown, in step S410, from multiple supply-demand parameter intervals, a target supply-demand parameter interval that matches the predicted photovoltaic power generation and the predicted power supply load demand is determined.

[0087] Specifically, multiple supply-demand matrix intervals are divided according to the photovoltaic power generation and the power supply load demand. Each supply-demand parameter matrix interval corresponds to a local linear model, which is used to describe the approximate linear relationship within this interval.

[0088] In some embodiments, by using historical data to perform statistical analysis on the photovoltaic power generation and the power supply load demand, according to the distribution characteristics of the data, the parameter range is divided into several intervals, such as equal-spacing division, equal-frequency division, or division based on data clustering.

[0089] In step S420, the predicted photovoltaic power generation and the predicted power supply load demand are input into the target local linear model module that matches the target supply-demand parameter interval to determine the first scheduling parameter.

[0090] Through numerical matching, the matching local linear model module can be quickly found. By segmenting the processing, the model can simplify the calculation in a linear manner within each interval, and at the same time, it can approximate the non-linear relationship well as a whole.

[0091] It should be noted that for the training of each local linear model module, it can be trained using historical data through the least squares method or other regression methods, so as to obtain the linear weights and biases of the model modules in each interval. In addition, during the operation of the model, the weights and biases of each interval can also be dynamically adjusted based on the actual feedback of the system to ensure that the model can continuously adapt to the changes in the actual situation.

[0092] In step S430, it is detected whether the first scheduling parameter meets the preset array safe operation condition.

[0093] In step S441, if the first scheduling parameter meets the array safe operation condition, the initial scheduling parameter is determined according to the first scheduling parameter.

[0094] In step S443, if the first scheduling parameter does not meet the array safe operation condition, the first scheduling parameter is updated according to the array safe operation condition to determine the initial scheduling parameter.

[0095] Here, the array safe operation condition includes the grid-connected voltage safety range and the power limit safety range. Exemplarily, if the grid-connected voltage or the output power exceeds the safety range, the voltage or the output power is adjusted accordingly to within the safety range. Through the grid-connected voltage safety range, the grid-connected voltage of the array must be kept within the safety range to avoid being too high or too low. Through the power limit safety range, the output power should be within the rated power range of the array equipment to avoid overloading. Thus, through the fusion rule-driven mechanism, the model can detect whether it meets the preset array safe operation condition after determining the initial scheduling parameter, ensuring the safety and stability of the system under various abnormal conditions.

[0096] Through the embodiments of the present application, by dividing the photovoltaic power generation amount and the power supply load demand into multiple supply-demand parameter intervals and adopting independent local linear models in each interval, the model can better capture the non-linear behavior of the photovoltaic system, and effectively improve the accuracy of the scheduling decision based on the segmented processing method. In addition, the piecewise linearization model can make corresponding scheduling decisions for different supply-demand conditions. Especially in the scenario where the photovoltaic power generation amount and the load demand fluctuate greatly, the model can adaptively select the appropriate local linear model for calculation, so as to ensure that the system can achieve optimal scheduling under various operating conditions. In addition, due to the combination of piecewise linearization and rule-driven, the design of the scheduling decision model has the characteristics of being lightweight, especially suitable for deployment on edge nodes. The model can operate efficiently with limited computing resources, reduce the consumption of system resources, and can quickly complete the pre-scheduling calculation under the condition of low power consumption.

[0097] It should be noted that in a distributed photovoltaic system, the correlations and differences between the states of each photovoltaic array and the operating environment are relatively large, which means that the impact of each photovoltaic array on the overall system scheduling is also different. To capture these differences and achieve better global scheduling, a Graph Attention Network (GAT) is introduced, enabling the network to dynamically adjust the node features of nodes according to the importance of each photovoltaic array.

[0098] In a distributed photovoltaic system, the contributions and impacts of each photovoltaic array on the global system scheduling are different. Through a graph neural network based on the attention mechanism, the system can automatically identify the key nodes that have a greater impact on the global scheduling and assign higher attention weights to these nodes. This mechanism ensures that the features of the key nodes can have a greater impact on the global scheduling during the update process, thereby optimizing the operating efficiency of the entire photovoltaic system.

[0099] In some examples of the embodiments of this application, the graph neural network based on the attention mechanism updates the node features of graph nodes in the following manner:

[0100]

[0101] In the formula, is the attention coefficient calculated by the m-th attention head, a T represents the transpose of the weight vector used to calculate the attention coefficient, LeakyReLU represents the leakyRelu activation function, W m represents the linear transformation matrix of the m-th attention head; l u and l v respectively represent the input feature vectors of graph nodes u and v, ‖ represents the concatenation operator, μ represents the adjustment coefficient; e uv represents the edge weight between graph node u and its neighbor graph node v, represents the set of neighbor nodes of graph node u; l u' represents the updated node feature vector of graph node u, M represents the total number of attention heads, σ represents the Sigmoid activation function; is the normalization factor, representing the result of accumulating the impacts of all neighbor nodes k of node u.

[0102] Regarding the description of the above formula, non-linearity is introduced through the LeakyReLU activation function to prevent overly smooth feature representations. To enhance the model's expressive power, a multi-head attention mechanism is introduced in GAT, that is, multiple independent attention coefficients are calculated in parallel in different attention heads, and then the outputs of these attention heads are averaged to improve the model's robustness and feature extraction ability while keeping the output dimension unchanged. In addition, during the node update process, the edge weights of different edge connections are comprehensively considered. By introducing edge weights and the multi-head attention mechanism, the relationship strength between nodes can be more accurately reflected, enabling the node update process to more precisely reflect the complex relationships between nodes.

[0103] By combining the attention mechanism with edge weights, the graph neural network can more meticulously express the complex relationships between each photovoltaic array during the node feature update process, achieving refined feature representation, which helps the system better coordinate the scheduling decisions of each photovoltaic array globally and ensure the supply-demand balance of the system and the stability of the power grid.

[0104] In some examples of the embodiments of the present application, the edge weight is calculated by the following formula:

[0105]

[0106] In the formula, C uv is the supply-demand complementarity degree, indicating the complementarity between the photovoltaic arrays corresponding to graph nodes u and v in terms of power generation and load demand; η uv is the power transmission loss rate, indicating the loss rate when power is transmitted from the photovoltaic array corresponding to graph node u to the photovoltaic array corresponding to graph node v; B uv is the historical cooperation efficiency, indicating the historical cooperation efficiency of the photovoltaic arrays corresponding to graph nodes u and v; λ1 represents the first adjustment parameter for balancing the relative importance of the supply-demand complementarity degree and the historical cooperation efficiency, and λ2 represents the second adjustment parameter for adjusting the influence of the transmission loss rate on the edge weight; J pred,u and J pred,v respectively represent the predicted power generation of the photovoltaic arrays corresponding to graph nodes u and v, L pred,u and L pred,v respectively represent the predicted power supply load demands corresponding to graph nodes u and v; δ uv represents the line transmission efficiency coefficient between the photovoltaic arrays corresponding to graph nodes u and v, and d uv represents the transmission line distance between the photovoltaic arrays corresponding to graph nodes u and v; ξ uv (t) represents the matching degree of the output power when the photovoltaic arrays corresponding to graph nodes u and v cooperate at historical time step t; T DRepresents the total number of historical time steps for the collaborative operation of the photovoltaic array.

[0107] Regarding the description of the above formula (11), the supply-demand complementarity degree, transmission efficiency, and historical collaboration efficiency adopt a multiplicative structure, which can better reflect the synergistic effect of these factors. If any one of the factors is zero, the edge weight will be significantly reduced. A linear weighted term of η uv is introduced in the denominator to make the negative impact of the transmission loss rate on the edge weight more prominent, especially in the case of a high transmission loss rate, so that the weight of the high-loss path can be more reasonably reduced.

[0108] By designing the edge weight calculation formula e uv , considering supply-demand complementarity, power transmission loss, and historical collaboration comprehensively, the calculation of the edge weight can provide more refined node feature update information, enabling the graph neural network to perform better in global scheduling optimization, more accurately reflecting the complex relationships between photovoltaic arrays in the photovoltaic scheduling graph structure, and helping to improve the overall operation efficiency and security of the system.

[0109] Specifically, by introducing the supply-demand complementarity degree, the system can give priority to those node pairs with good supply-demand matching, which helps to establish a more effective power scheduling channel between photovoltaic arrays, optimize the balance ability of the system, and reduce the supply-demand imbalance in power scheduling. Through the design of the power transmission loss rate part, it is ensured that the system can preferentially select paths with smaller transmission losses during scheduling, thereby reducing energy waste, improving the energy utilization efficiency of the entire distributed photovoltaic system, and further enhancing the economic and environmental benefits of the system. Through the historical collaboration efficiency part, the system can consider the successful collaborative operation modes in history in the scheduling decision. The node combinations that have been successfully collaborated in history will obtain higher weights in the edge weight calculation, thereby enhancing the reliability and stability of the scheduling strategy, enabling the system to better cope with the complex and changeable operating environment, and ensuring the efficient execution of the scheduling strategy.

[0110] In some examples of the embodiments of the present application, by designing the fitness function of the genetic algorithm model, and thus comprehensively considering supply-demand balance, grid stability, and transmission loss, the overall scheduling efficiency and operation stability of the system can be effectively improved.

[0111] More specifically, the fitness function F fitness of the global scheduling optimization model is expressed by the following formula:

[0112] F fitness =ψ1·G S +ψ2·G G +ψ3·G T , formula (15)

[0113]

[0114] In the formula, G S , G G and G T represent the global supply-demand balance degree, the global power grid stability coefficient, and the global transmission loss rate respectively. ψ1, ψ2, and ψ3 represent the corresponding importance adjustment weights respectively, and ∈ represents a preset constant; P out,u is the output power of the photovoltaic array corresponding to the u-th graph node after update, and L demand,u is the power supply load demand of the photovoltaic array corresponding to the u-th graph node. Z represents the total number of photovoltaic arrays in the distributed photovoltaic system; V grid,u is the grid-connected voltage of the photovoltaic array corresponding to the u-th graph node, and V nominal represents the nominal grid voltage.

[0115] Regarding the explanation of formula (16), represents the total output power of all photovoltaic arrays in the system, while represents the total power demand of the system. By normalizing the difference through the denominator , the supply-demand balance degree can be reflected as a standardized value, enabling G S to measure the matching degree between the total power generation and the total demand of the system. When the supply and demand are completely balanced (i.e., the total output power is equal to the total demand), G S takes 1, indicating that the system has achieved the best supply-demand balance; when the supply and demand are unbalanced, the value of G S will decrease.

[0116] Regarding the explanation of formula (17), measures the deviation of the actual grid-connected voltage of the u-th photovoltaic array from the nominal voltage. By summing up the deviations of all photovoltaic arrays, G G reflects the overall voltage stability of the system. G G is a reverse index. The larger the deviation in the denominator, the smaller the value of G G , indicating that the power grid stability is worse. The higher the value of G G , the more stable the power grid operation.

[0117] Regarding the explanation of formula (18), η uv ·e uv represents the weighted product of the transmission loss between nodes and the edge weight. By accumulating all possible node pairs, the total transmission loss of the entire system is calculated. Similar to G G , G T is a reverse index. The larger the weighted sum in the denominator, the smaller the value of G T , indicating a higher global transmission loss rate.

[0118] Through the fitness function in the embodiments of the present application, based on the global supply-demand balance degree G S , the model can continuously adjust the output power of each photovoltaic array to maximize the supply-demand balance of the entire system. Through the global power grid stability coefficient G G , the model can optimize the grid connection voltage of each photovoltaic array to ensure the stability of the system voltage in each region and prevent the impact of voltage fluctuations on the system operation. Through the global transmission loss rate G T , the model can preferentially select paths with low transmission losses during the global scheduling process, reduce energy waste, and improve the overall efficiency of the system. Thus, through evolutionary operations, the genetic algorithm continuously optimizes the population individuals. By comprehensively considering supply-demand balance, grid stability, and transmission losses, the scheduling parameters can adapt to changing system conditions, effectively improving the overall scheduling efficiency and operation stability of the system.

[0119] It should be noted that for the foregoing method embodiments, for the sake of simple description, they are all expressed as a combination of a series of actions. However, those skilled in the art should know that the present application is not limited by the described action sequence, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should also know that the embodiments described in the specification are all preferred embodiments, and the actions and modules involved are not necessarily essential to the present application. In the above embodiments, the descriptions of the various embodiments have their own emphases. For parts not detailed in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0120] Figure 5 FIG. shows a structural block diagram of an example of a flexible adjustable control device for a distributed photovoltaic system according to an embodiment of the present application.

[0121] The distributed photovoltaic system includes multiple photovoltaic arrays, and a corresponding sensor network and edge node are respectively deployed for each photovoltaic array. The flexible adjustable control device is arranged at the edge node.

[0122] As Figure 5 shown, the flexible adjustable control device 500 of the distributed photovoltaic system includes an acquisition unit 510, a supply-demand prediction unit 520, an initial scheduling prediction unit 530, a data upload unit 540, and a target scheduling control unit 550.

[0123] The acquisition unit 510 is used to acquire the multi-modal sensing time-series data of the deployed photovoltaic arrays in a historical preset time period and the weather forecast data corresponding to a future preset time period; the multi-modal sensing time-series data includes multiple adjacent historical time steps and corresponding multi-modal sensing data, and the parameter types of the multi-modal sensing data include: solar radiation intensity, ambient temperature, power generation, grid operation parameters, and power supply load demand.

[0124] The supply and demand prediction unit 520 is used to input the multi-modal sensing time-series data and the weather forecast data into a time-series prediction model to predict the expected photovoltaic power generation and the expected power supply load demand of the corresponding photovoltaic array in the preset future time period; the time-series prediction model adopts an LSTM model.

[0125] The initial scheduling prediction unit 530 is used to process the corresponding expected photovoltaic power generation and the expected power supply load demand based on a scheduling decision model to predict the initial scheduling parameters of the corresponding photovoltaic array; the parameter types of the scheduling parameters include the target output power and the grid connection voltage regulation value.

[0126] The data uploading unit 540 is used to send the expected photovoltaic power generation, the expected power supply load demand, and the initial scheduling parameters to the core node, so that the core node performs global analysis on the data sent by each edge node to generate the target scheduling parameters for each edge node.

[0127] The target scheduling control unit 550 is used to receive the corresponding target scheduling parameters from the core node and regulate the corresponding photovoltaic array in the preset future time period according to the target scheduling parameters.

[0128] In some embodiments, the present application provides a non-volatile computer-readable storage medium, in which one or more programs including execution instructions are stored, and the execution instructions can be read and executed by an electronic device (including but not limited to a computer, a server, or a network device, etc.) to be used for executing the steps of any one of the flexible adjustable control methods of the distributed photovoltaic system described above in the present application.

[0129] In some embodiments, the present application also provides a computer program product, the computer program product includes a computer program stored on a non-volatile computer-readable storage medium, the computer program includes program instructions, and when the program instructions are executed by a computer, the computer is enabled to execute the steps of any one of the flexible adjustable control methods of the distributed photovoltaic system described above.

[0130] In some embodiments, the present application also provides an electronic device, which includes: at least one processor, and a memory communicatively connected to the at least one processor, wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the steps of the flexible adjustable control method of the distributed photovoltaic system.

[0131] Figure 6 It is a schematic hardware structure diagram of an electronic device for executing the flexible adjustable control method of the distributed photovoltaic system provided by another embodiment of the present application, as Figure 6As shown, the device includes:

[0132] One or more processors 610 and a memory 620, Figure 6 Taking one processor 610 as an example.

[0133] The device for implementing the flexible adjustable control method of the distributed photovoltaic system may further include: an input device 630 and an output device 640.

[0134] The processor 610, the memory 620, the input device 630, and the output device 640 may be connected via a bus or other means, Figure 6 Taking connection via a bus as an example.

[0135] The memory 620, as a non-volatile computer-readable storage medium, can be used to store non-volatile software programs, non-volatile computer-executable programs, and modules, such as the program instructions / modules corresponding to the flexible adjustable control method of the distributed photovoltaic system in the embodiments of the present application. The processor 610 executes various functional applications and data processing of the server by running the non-volatile software programs, instructions, and modules stored in the memory 620, that is, implements the flexible adjustable control method of the distributed photovoltaic system in the above method embodiments.

[0136] The memory 620 may include a program storage area and a data storage area. Among them, the program storage area can store an operating system and application programs required for at least one function; the data storage area can store data created according to the use of the electronic device, etc. In addition, the memory 620 may include a high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, a flash memory device, or other non-volatile solid-state storage devices. In some embodiments, the memory 620 may optionally include a memory remotely set relative to the processor 610, and these remote memories can be connected to the electronic device through a network. Examples of the above networks include, but are not limited to, the Internet, an enterprise intranet, a local area network, a mobile communication network, and combinations thereof.

[0137] The input device 630 can receive input digital or character information, and generate signals related to the user settings and function control of the electronic device. The output device 640 may include a display device such as a display screen.

[0138] The one or more modules are stored in the memory 620, and when executed by the one or more processors 610, implement the flexible adjustable control method of the distributed photovoltaic system in any of the above method embodiments.

[0139] The above product can execute the method provided in the embodiments of the present application, and has the corresponding functional modules and beneficial effects for implementing the method. For technical details not described in detail in this embodiment, reference can be made to the method provided in the embodiments of the present application.

[0140] The electronic devices in the embodiments of the present application exist in various forms, including but not limited to:

[0141] (1) Mobile communication devices: These devices are characterized by having mobile communication functions and mainly aim to provide voice and data communication. Such terminals include: smart phones, multimedia phones, functional phones, and low-end phones, etc.

[0142] (2) Ultra-mobile personal computer devices: These devices belong to the category of personal computers, have computing and processing functions, and generally also have the characteristic of mobile Internet access. Such terminals include: PDA, MID, and UMPC devices, etc.

[0143] (3) Portable entertainment devices: These devices can display and play multimedia content. Such devices include: audio and video players, handheld game consoles, e-books, and intelligent toys and portable vehicle navigation devices.

[0144] (4) Other on-board electronic devices with data interaction functions, such as in-vehicle device installed on a vehicle.

[0145] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0146] Through the description of the above embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus a general hardware platform, and of course, it can also be implemented by hardware. Based on such an understanding, the essence of the above technical solution or the part that contributes to the related technology can be embodied in the form of a software product, and this computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disc, etc., including several instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0147] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application and are not intended to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of each embodiment of the present application.

Claims

1. A flexible adjustable control method for a distributed photovoltaic system. The distributed photovoltaic system includes multiple photovoltaic arrays, and a corresponding sensor network and edge node are respectively deployed for each photovoltaic array. The method includes: Each edge node is respectively used to perform the following operations: Obtain the multi-modal sensing time-series data of the deployed photovoltaic array in a historical preset time period and the weather forecast data corresponding to a future preset time period; the multi-modal sensing time-series data includes multiple adjacent historical time steps and corresponding multi-modal sensing data; Input the multi-modal sensing time-series data and the weather forecast data into a time-series prediction model to predict the expected power generation and the expected power supply load demand of the corresponding photovoltaic array in the future preset time period; Process the corresponding expected power generation and the expected power supply load demand based on a scheduling decision model to predict the initial scheduling parameters of the corresponding photovoltaic array; the parameter types of the scheduling parameters include the target output power and the grid connection voltage regulation value; Send the expected power generation, the expected power supply load demand, and the initial scheduling parameters to the core node, so that the core node performs global analysis on the data sent by each edge node to generate the target scheduling parameters for each edge node, and controls the corresponding photovoltaic array in the future preset time period according to the target scheduling parameters; Wherein, the core node is used to generate the target scheduling parameters for each edge node by performing the following operations: Construct a photovoltaic scheduling graph structure according to the expected power generation, the expected power supply load demand, and the initial scheduling parameters sent by each edge node; the photovoltaic scheduling graph structure includes multiple graph nodes and edge connections; each graph node corresponds uniquely to an edge node, and the node feature of the graph node is defined by the corresponding expected power generation, the expected power supply load demand, and the initial scheduling parameters; the edge weight of each edge connection is determined by the supply-demand complementarity, the power transmission loss rate, and the historical cooperation efficiency between the corresponding photovoltaic arrays of the connected graph nodes; Process the photovoltaic scheduling graph structure based on a graph neural network to update the node features of each graph node; Process the node features of each updated graph node based on a global scheduling optimization model to determine the target scheduling parameters of each edge node; the global scheduling optimization model adopts a genetic algorithm model.

2. The method according to claim 1, wherein The population individuals of the genetic algorithm model are defined by the set of scheduling parameters of all photovoltaic arrays in the distributed photovoltaic system, each gene in the population individual is defined by the scheduling parameters of the corresponding photovoltaic array, and the fitness function is defined according to the global supply-demand balance degree, the global power grid stability coefficient, and the global transmission loss rate of the distributed photovoltaic system.

3. The method according to claim 1, wherein The graph neural network adopts a graph neural network based on an attention mechanism, which is used to update the node features of the graph node in the following way: Wherein, is the attention coefficient calculated by the m-th attention head, a T represents the transpose of the weight vector used to calculate the attention coefficient, LeakyReLU represents the leakyRelu activation function, W m represents the linear transformation matrix of the m-th attention head; l u and l v represent the input feature vectors of graph nodes u and v respectively, ‖ represents the concatenation operator, μ represents the adjustment coefficient; e uv represents the edge weight between graph nodes u and neighbor graph node v, represents the set of neighbor nodes of graph node u; l u' represents the updated node feature vector of graph node u, M represents the total number of attention heads, σ represents the Sigmoid activation function; is the normalization factor, representing the result of accumulating the influences of all neighbor nodes k of node u.

4. The method according to claim 3, wherein The edge weight is calculated by the following formula: where C uv is the supply-demand complementarity degree, representing the complementarity of the photovoltaic arrays corresponding to graph nodes u and v in terms of power generation and load demand; η uv is the power transmission loss rate, representing the loss rate when power is transmitted from the photovoltaic array corresponding to graph node u to the photovoltaic array corresponding to graph node v; B uv is the historical cooperation efficiency, representing the historical cooperation efficiency of the photovoltaic arrays corresponding to graph nodes u and v; λ1 represents the first adjustment parameter for balancing the relative importance of the supply-demand complementarity degree and the historical cooperation efficiency, and λ2 represents the second adjustment parameter for regulating the influence of the transmission loss rate on the edge weight; J pred,u and J pred,v respectively represent the predicted power generation of the photovoltaic arrays corresponding to graph nodes u and v, L pred,u and L pred,v respectively represent the predicted power supply load demands corresponding to graph nodes u and v; δ uv represents the line transmission efficiency coefficient between the photovoltaic arrays corresponding to graph nodes u and v, d uv represents the transmission line distance between the photovoltaic arrays corresponding to graph nodes u and v; ξ uv (t) represents the matching degree of the output power when the photovoltaic arrays corresponding to graph nodes u and v cooperate at the historical time step t; T D represents the total number of historical time steps of the cooperation of the photovoltaic arrays.

5. The method according to claim 1, wherein The time-series prediction model adopts a multi-scale attention LSTM model, which includes a multi-scale LSTM layer, a hybrid attention mechanism layer, and an output layer; The multi-scale LSTM layer includes multiple LSTM layers and a fusion layer, and each of the LSTM layers has a corresponding time scale; the LSTM layer of each time scale is respectively used to process the input data of different time windows to generate corresponding scale hidden states: In the formula, represents the scale hidden state at the t-th time step under the time scale s, represents the multi-modal sensing time series data at the historical time step t', represents the weather forecast data at the future time step t''; represents the hidden state at the (t - 1)-th time step; The hybrid attention mechanism layer includes a temporal attention module and a feature attention module, dynamically adjusts the weights of each time step through the temporal attention module, and performs weighting according to the importance of different features through the feature attention module: In the formula, represents the temporal attention weight at the t'-th historical time step under the time scale s, and respectively represent the weight matrix and bias term of the temporal attention mechanism corresponding to the time scale s; tanh represents the hyperbolic tangent function, represents the hidden state at the t'-th historical time step, T H represents the total number of historical time steps corresponding to the multi-modal sensing time series data; represents the feature attention weight of the j-th feature under the time scale s; represents the feature representation of the j-th feature in all historical time steps in the historical multi-modal time series data, is the feature representation of the j-th feature in all future time steps in the weather forecast data, is and 's joint feature representation; and respectively represent the weight matrix and bias term of the feature attention mechanism corresponding to the time scale s, and n represents the total number of features; represents the weighted hidden state under the time scale s, represents the hidden state corresponding to the feature j at the historical time step t' and the time scale s; The output layer is used to fuse the weighted hidden states of each scale to predict the expected power generation amount and the expected power supply load demand of the corresponding photovoltaic array in the preset future time period: Wherein, is the fused multi-scale weighted hidden state, and γ (s) represents the scale fusion weight of time scale s, and S represents the total number of time scales corresponding to the multi-scale LSTM layer; represents the predicted photovoltaic power generation and the predicted power supply load demand of the final output of the model; W o and b o represent the weight matrix and the bias term of the output layer, respectively.

6. The method according to claim 5, wherein, The scale fusion weight is obtained through the calculation of the following formula: where, g (s) represents the time scale score; A t represents the array state vector at time step t, and the array state vector is determined according to the power generation fluctuation amplitude and the load change rate at time step t; W γ and b γ respectively represent the weight matrix and the bias term related to the time scale, V γ is the weight matrix related to the array state vector A t ; represents the sum of the exponential values of the scores of all time scales.

7. The method according to claim 1, wherein The scheduling decision model includes multiple local linear model modules, and each of the local linear model modules has a corresponding supply-demand parameter interval; the scheduling decision model is used to predict the initial scheduling parameters of the corresponding photovoltaic array by performing the following operations: Determine a target supply-demand parameter interval that matches the expected power generation amount and the expected power supply load demand from multiple supply-demand parameter intervals; Input the expected power generation amount and the expected power supply load demand into the target local linear model module that matches the target supply-demand parameter interval to determine the first scheduling parameter; Detect whether the first scheduling parameter meets the preset array safe operation conditions; the array safe operation conditions include the grid connection voltage safety range and the power limit safety range; If the first scheduling parameter meets the array safe operation conditions, determine the initial scheduling parameter according to the first scheduling parameter; And If the first scheduling parameter does not meet the array safe operation conditions, update the first scheduling parameter according to the array safe operation conditions to determine the initial scheduling parameter.

8. The method according to claim 1, wherein, The fitness function F of the global scheduling optimization model fitness is expressed by the following formula: F fitness = ψ1·G S + ψ2·G G + ψ3·G T , where G S , G G and G T respectively represent the global supply-demand balance degree, the global power grid stability coefficient, and the global transmission loss rate, ψ1, ψ2, and ψ3 respectively represent the corresponding importance adjustment weights, ∈ represents a preset constant; P out,u is the output power of the photovoltaic array corresponding to the u-th graph node after update, L demand,u is the power supply load demand of the photovoltaic array corresponding to the u-th graph node, Z represents the total number of photovoltaic arrays in the distributed photovoltaic system; V grid,u is the grid-connected voltage of the photovoltaic array corresponding to the u-th graph node, V nominal represents the grid nominal voltage.

9. A flexible adjustable control device for a distributed photovoltaic system, the distributed photovoltaic system includes multiple photovoltaic arrays, and a corresponding sensor network and edge node are respectively deployed for each photovoltaic array to implement the method according to any one of claims 1-8; the device includes an acquisition unit, a supply-demand prediction unit, an initial scheduling prediction unit, a data upload unit, and a target scheduling control unit provided in the edge node; The acquisition unit is used to acquire the multi-modal sensing time series data of the deployed photovoltaic array in the historical preset time period and the weather forecast data of the corresponding future preset time period; the multi-modal sensing time series data includes multiple adjacent historical time steps and corresponding multi-modal sensing data, and the parameter types of the multi-modal sensing data include: solar radiation intensity, ambient temperature, power generation amount, grid operation parameters, and power supply load demand; The supply-demand prediction unit is used to input the multi-modal sensing time series data and the weather forecast data into the time series prediction model to predict the expected power generation amount and the expected power supply load demand of the corresponding photovoltaic array in the preset future time period; the time series prediction model uses an LSTM model; An initial scheduling prediction unit, configured to process corresponding predicted photovoltaic power generation and predicted power supply load demand based on a scheduling decision model to predict initial scheduling parameters of a corresponding photovoltaic array; the parameter types of the scheduling parameters include a target output power and a grid connection voltage regulation value; A data uploading unit, configured to send the predicted photovoltaic power generation, the predicted power supply load demand, and the initial scheduling parameters to a core node, so that the core node performs global analysis on the data sent by each edge node to generate target scheduling parameters for each edge node; A target scheduling control unit, configured to receive corresponding target scheduling parameters from the core node and regulate the corresponding photovoltaic array according to the target scheduling parameters in a preset future time period.

Images