Control method of modularized low-voltage roof photovoltaic grid-connected phase-change switch device

Abstract

The invention relates to the technical field of low-voltage photovoltaic grid-connected load regulation, and discloses a control method of a modular low-voltage roof photovoltaic grid-connected phase-change switch device, which comprises the following steps of: acquiring operating parameters of a photovoltaic grid-connected system in real time through a system sampling module, and performing data preprocessing on the acquired operating parameters; based on the preprocessed data, constructing and training a probability hypergraph recurrent neural network model, and outputting an optimal commutation trigger time point of the next moment; according to the predicted optimal commutation trigger time point, adopting an adaptive hybrid grey wolf optimization and particle swarm optimization algorithm to optimize the conduction sequence and the trigger angle of a commutation switch on line to obtain an optimal commutation action; and executing the optimal commutation action at the predicted optimal commutation triggering time point to realize dynamic load balance adjustment of the photovoltaic grid-connected system. According to the invention, self-adaption, optimization and dynamic robust control of the commutation action can be realized, and the operation efficiency and the dynamic stability of the low-voltage photovoltaic grid-connected system are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of low-voltage photovoltaic grid-connected load regulation technology, and in particular to a control method for a modular low-voltage rooftop photovoltaic grid-connected phase-switching switch device. Background Technology

[0002] In low-voltage grid-connected photovoltaic (PV) systems, three-phase load imbalance is a long-standing key challenge. Due to factors such as random fluctuations in user loads, intermittent changes in PV output, and complex weather conditions, three-phase power often exhibits dynamic shifts, leading to grid voltage distortion, increased line losses, decreased power factor, and even affecting the stable operation of grid-connected inverters. Traditional commutation control strategies mainly rely on fixed threshold judgments or simple rule-based reasoning based on fuzzy logic. While these strategies can improve the imbalance problem to some extent, they have the following limitations:

[0003] 1. Lack of dynamic prediction capability: Traditional algorithms often rely only on parameters at the current moment and cannot effectively use historical data to identify the evolution trend of the system, resulting in a lag in the prediction of commutation timing.

[0004] 2. Insufficient control precision: When photovoltaic output fluctuates drastically or the load changes rapidly, the fixed threshold strategy is difficult to achieve the optimal match between the conduction sequence and the firing angle, and often requires multiple phase commutations to stabilize within the target range.

[0005] 3. The optimization algorithm is simple and the global convergence is poor: Existing control optimization methods based on particle swarm optimization (PSO) or genetic algorithm are prone to getting trapped in local optima and have low search efficiency under nonlinear and complex working conditions.

[0006] In recent years, with the development of artificial intelligence and swarm intelligence algorithms, intelligent commutation methods based on deep learning prediction and reinforcement optimization control have become a research hotspot. However, how to establish an efficient collaborative mechanism between commutation trigger moment prediction and adaptive control strategy optimization remains a key technical bottleneck in the field of intelligent control of photovoltaic grid-connected systems.

[0007] Therefore, there is an urgent need for a new commutation control method that combines time-series prediction, probabilistic modeling, and intelligent optimization decision-making to achieve efficient and stable dynamic load balance regulation. Summary of the Invention

[0008] This invention aims to overcome the problems of slow response, rigid action parameters, and insufficient global optimization capability in existing low-voltage photovoltaic grid-connected commutation control. It proposes a commutation control method based on a combination of intelligent prediction and reinforcement learning optimization to achieve adaptive, optimized, and dynamically robust control of commutation actions.

[0009] This invention provides a control method for a modular low-voltage rooftop photovoltaic grid-connected phase-switching switchgear, comprising the following steps:

[0010] Step S1: The system sampling module collects the operating parameters of the photovoltaic grid-connected system in real time and performs data preprocessing on the collected operating parameters. The operating parameters include the output current, output voltage, power factor, and module temperature of the photovoltaic array end; the three-phase input current, three-phase input voltage, and grid frequency of the grid-connected end; and the three-phase load power and three-phase load imbalance of the load end.

[0011] Step S2: Based on the preprocessed data, construct and train a probabilistic hypergraph recurrent neural network model to output the optimal commutation trigger time point for the next time step.

[0012] Step S3: Based on the predicted optimal commutation trigger time, the adaptive hybrid gray wolf optimization and particle swarm optimization algorithm is used to optimize the conduction sequence and trigger angle of the commutation switch online to obtain the optimal commutation action;

[0013] Step S4: Execute the optimal commutation action at the predicted optimal commutation trigger time to achieve dynamic load balance adjustment of the photovoltaic grid-connected system.

[0014] Furthermore, in step S1, the operating parameters include the output current, output voltage, power factor, and component temperature of the photovoltaic array, the three-phase input current, three-phase input voltage, and grid frequency of the grid-connected terminal, and the three-phase load power and three-phase load imbalance of the load terminal.

[0015] Furthermore, in step S1, the data preprocessing includes: performing filtering and noise reduction using a moving average filtering algorithm, and normalization processing to map the parameters to the interval between zero and one.

[0016] Furthermore, in step S2, the probabilistic hypergraph recurrent neural network model adopts an encoder-decoder architecture, including: a probabilistic hypergraph learning module and a hypergraph recurrent neural network unit;

[0017] The probabilistic hypergraph learning module includes a hypergraph structure consisting of a node set and a hyperedge set, and uses Gumbel reparameterization to model the association weights between nodes and hyperedges as a continuously differentiable probability distribution.

[0018] The hypergraph recurrent neural network unit achieves temporal dependency modeling through node-to-hyperedge aggregation, hyperedge-to-node aggregation, and hidden state updates.

[0019] Furthermore, the node set corresponds to the collected operating parameters, and the hyperedge set corresponds to the association combination between the parameters; the hidden state update adopts the residual connection method, which integrates real-time normalized features, historical hidden states, and hyperedge update features.

[0020] Furthermore, in step S2, the probabilistic hypergraph recurrent neural network model is trained using historical running data, and the training loss function is the mean square error between the predicted trigger time point and the actual optimal trigger time point.

[0021] Furthermore, in step S3, the optimization process of the adaptive hybrid gray wolf optimization and particle swarm optimization algorithm includes:

[0022] Define a reinforcement learning state space, action space, and reward function. The state space includes real-time operating state parameters of the system, the action space includes the conduction sequence and firing angle of the commutation switch, and the reward function integrates system imbalance, grid-connected energy efficiency ratio, and inrush current.

[0023] Initialize the particle swarm and the gray wolf population, wherein the gray wolf population selects the top three particles with the highest reward function values ​​from the particle swarm as elite wolves;

[0024] By fusing particle swarm optimization for velocity updates, gray wolf optimization for position updates, and virtual target guidance, the velocity and position of particles are iteratively updated to obtain the globally optimal action.

[0025] Furthermore, the particle swarm optimization velocity update is combined with adaptive inertia weights, which are dynamically adjusted based on individual evolution factors and population aggregation factors; the gray wolf optimization position update guides particle positions through elite wolves; and the virtual target guidance generates the optimal action direction based on K-means clustering or generates the obstacle avoidance action direction based on impact current constraints.

[0026] Furthermore, the reward function is a weighted function comprising an imbalance improvement term, a grid-connected energy efficiency ratio term, and an inrush current penalty term, wherein the imbalance improvement term has the largest weight coefficient.

[0027] Furthermore, in step S3, the adaptive hybrid gray wolf optimization and particle swarm optimization algorithm outputs the globally optimal action after iterating a preset number of times in each optimization process. The globally optimal action includes the optimal conduction sequence and the optimal trigger angle.

[0028] Compared with the prior art, the present invention achieves the following beneficial effects:

[0029] 1. This invention establishes a probabilistic hypergraph recurrent neural network (PHRNN) model: Based on historical operating data of photovoltaic grid-connected systems, a probabilistic hypergraph structure containing dynamic relationships between nodes and hyperedges is constructed. A recurrent neural network is used to model temporal dependencies, outputting the optimal commutation trigger time for the next moment. This model fully captures the high-order correlations and temporal characteristics between system parameters, achieving high accuracy and interpretability in commutation prediction.

[0030] 2. The present invention is based on the probabilistic association mechanism of Gumbel reparameterization technique: by introducing Gumbel distribution reparameterization into the hypergraph structure, the association weight of node-hyperedge is modeled as a continuously differentiable probabilistic variable, which effectively improves the model's adaptability to non-stationary factors such as photovoltaic power output fluctuations and load mutations, and avoids the problem of fixed association relationships and difficulty in optimization in traditional hypergraph structures.

[0031] 3. This invention designs an adaptive hybrid gray wolf optimization and particle swarm optimization algorithm (AR-GWOPSO): it integrates the elite guidance characteristics of the gray wolf optimization algorithm with the group cooperation mechanism of the particle swarm optimization algorithm, and combines real-time system feedback signals to optimize the conduction sequence and trigger angle of the commutation switch online, realizing dynamic adaptive adjustment of commutation action parameters, effectively improving the robustness and global optimization performance of the system.

[0032] 4. This invention designs an enhanced optimization mechanism based on elite wolf guidance and virtual target generation: by guiding particle position updates through elite wolves and referring to virtual targets generated by K-means clustering as the optimal action direction, a hybrid optimization framework with heuristic search and obstacle avoidance capabilities is constructed to achieve rapid convergence and safe optimization of the commutation strategy, ensuring optimal commutation safety and energy efficiency of the system under complex dynamic conditions.

[0033] In summary, this invention constructs an end-to-end intelligent control system for photovoltaic commutation by integrating three technologies: probabilistic graphical learning, time-series prediction, and swarm intelligent control optimization. This system enables self-learning, self-optimization, and adaptive decision-making regarding commutation timing, conduction sequence, and trigger angle, significantly improving the operating efficiency and dynamic stability of low-voltage photovoltaic grid-connected systems. Attached Figure Description

[0034] Other features, objects, and advantages of the invention will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings. The drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0035] Figure 1 This is a flowchart illustrating a control method for a modular low-voltage rooftop photovoltaic grid-connected phase-switching switch device provided in an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the overall system architecture and data flow of a control method for a modular low-voltage rooftop photovoltaic grid-connected phase-switching switch device provided in an embodiment of the present invention;

[0037] Figure 3 This is a schematic diagram of the internal structure and data flow of the Probabilistic Hypergraph Recurrent Neural Network (PHRNN) according to an embodiment of the present invention;

[0038] Figure 4 This is a schematic diagram of the optimization process of the Adaptive Hybrid Gray Wolf Optimization and Particle Swarm Optimization Algorithm (AR-GWOPSO) according to an embodiment of the present invention;

[0039] Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0040] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and not intended to limit it. Furthermore, it should be noted that, for ease of description, the accompanying drawings show only the parts relevant to the present invention, and not all of the structures.

[0041] Before discussing the exemplary embodiments in more detail, it should be mentioned that some exemplary embodiments are described as processes or methods depicted as flowcharts. Although the flowcharts describe the operations (or steps) as sequential processes, many of the operations (or steps) can be performed in parallel, concurrently, or simultaneously. Furthermore, the order of the operations can be rearranged. The process can be terminated when its operation is completed, but it may also have additional steps not included in the figures. The process may correspond to a method, function, procedure, subroutine, subroutine, etc.

[0042] The acquisition, storage, use, and processing of data in this application comply with relevant national laws and regulations. It should be noted that existing industry solutions such as software, components, or models may be mentioned in the embodiments of this application. These should be considered exemplary and intended only to illustrate the feasibility of implementing the technical solution of this application, but do not imply that the applicant has already used or necessarily used such a solution.

[0043] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in sequences other than those illustrated or described herein.

[0044] Figure 1 This is a flowchart illustrating a control method for a modular low-voltage rooftop photovoltaic grid-connected phase-switching switch device provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the overall system architecture and data flow of a control method for a modular low-voltage rooftop photovoltaic grid-connected phase-switching switchgear provided in an embodiment of the present invention. Figure 1 and Figure 2As shown, a control method for a modular low-voltage rooftop photovoltaic grid-connected phase-switching switchgear includes the following steps:

[0045] Step S1: Collect the operating parameters of the photovoltaic grid-connected system in real time through the system sampling module, and perform data preprocessing on the collected operating parameters; the operating parameters include the output current, output voltage, power factor, and module temperature of the photovoltaic array end; the three-phase input current, three-phase input voltage, and grid frequency of the grid-connected end; and the three-phase load power and three-phase load imbalance of the load end.

[0046] Step S1 is used to perform data acquisition and preprocessing.

[0047] Step S11: Data Acquisition

[0048] The system's built-in sampling module collects key parameters during the operation of the modular commutation switch in real time. These parameters include, but are not limited to:

[0049] Photovoltaic array end: Output current Output voltage Power factor Component temperature ;

[0050] Grid-connected end: Three-phase input current Three-phase input voltage Grid frequency ;

[0051] Load side: Three-phase load power Three-phase load imbalance (Initial calculations based on) ,in, , These are the three-phase currents within one power frequency cycle. The maximum, minimum, and average values ​​of the root mean square (RMS) of a given number.

[0052] Step S12: Data Preprocessing

[0053] Data preprocessing includes two parts: filtering and denoising, and normalization.

[0054] Step S121: Filtering and noise reduction processing

[0055] The moving average filtering algorithm is used to eliminate high-frequency noise in the sampled signal. The filtering formula is as follows:

[0056]

[0057] in, : The original sampled value at time t (such as the output current and voltage of the photovoltaic array, the three-phase current and voltage of the grid-connected terminal, and other parameters collected in real time); n: the length of the filter window (preferably ranging from 5 to 10, dynamically adjusted according to the sampling frequency, used to control the number of historical sampling points participating in the average calculation); k: the summation index (from 0 to n-1, representing the sequence number of the historical sampling point traced back from the current time). : Sum the original sampled values ​​at time t and the previous n-1 times (i.e. ); : The output value after moving average filtering at time t (the smoothed signal value after eliminating high-frequency noise).

[0058] Step S122: Normalization

[0059] Map the filtered parameters to For intervals, eliminating the influence of dimensions, the normalization formula is:

[0060]

[0061] in, These are the minimum and maximum values ​​of the parameter within a complete, representative historical operating cycle (e.g., the past 6 months); The normalized data serves as the input for subsequent commutation prediction models.

[0062] Step S2: Based on the preprocessed data, construct and train a probabilistic hypergraph recurrent neural network model to output the optimal commutation trigger time point for the next time step.

[0063] Step S2 is used to construct a commutation prediction model based on PHRNN to predict the commutation trigger time.

[0064] Figure 3 This is a schematic diagram illustrating the internal structure and data flow of the Probabilistic Hypergraph Recurrent Neural Network (PHRNN) according to an embodiment of the present invention. Figure 3 As shown, this embodiment of the invention proposes a probabilistic hypergraph recurrent neural network (PHRNN), which trains the model based on historical operating data and outputs the optimal commutation trigger time point for the next moment. The specific implementation process is as follows:

[0065] Step S21: PHRNN model structure design

[0066] Furthermore, in step S2, the probabilistic hypergraph recurrent neural network model adopts an encoder-decoder architecture, the core of which includes a probabilistic hypergraph learning module and a hypergraph recurrent neural network (RNN) unit. The probabilistic hypergraph learning module defines a hypergraph structure containing a set of nodes and a set of hyperedges, and uses the Gumbel reparameterization technique to model the association weights between nodes and hyperedges as a continuously differentiable probability distribution. The hypergraph recurrent neural network unit achieves temporal dependency modeling through node-to-hyperedge aggregation, hyperedge-to-node aggregation, and hidden state updates. Specifically, it includes the following steps:

[0067] Step S211: Set up probabilistic hypergraph learning

[0068] Define hypergraph Where V is the node set (corresponding to the 8 key parameters collected in step S1, i.e.) E represents the hyperedge set. Based on the physical correlation of parameters in the photovoltaic grid-connected system, six hyperedges are preset, each corresponding to a different set of parameter combinations, such as... (Used to characterize the direct relationship between three-phase current and unbalance). (Used to characterize the output characteristics of photovoltaic arrays) (Used to characterize the effect of temperature on photovoltaic output) (Used to characterize the state of the power grid side) (Used to characterize load distribution) (Used to characterize the relationship between photovoltaic power generation and grid-connected current); For probability correlation matrix ( The number of superedges is preset based on parameter correlation.

[0069] Using Gumbel's reparameterization technique, the node-hyperedge association weights are determined. The model is a probability distribution, and the formula is:

[0070]

[0071] : The probabilistic association weight between node v and hyperedge e (representing the probability that node v belongs to hyperedge e, with a value range of...) (used to quantify the dynamic relationship between the two in the parameter correlation of photovoltaic grid-connected systems). The sigmoid function maps input values ​​to the interval [0,1], ensuring... It conforms to the characteristics of a probability distribution; Parameters associated with node v and hyperedge e The natural logarithm of is used to transform the linear relationship between related parameters into a nonlinear relationship in the logarithmic domain, enhancing the model's ability to characterize the correlation between parameters; where It is a trainable parameter that is randomly initialized at the start of model training. Its value is learned and updated through the backpropagation algorithm during model training to ultimately dynamically determine the association strength between node v and hyperedge e. : Follows the standard Gumbel distribution (Gumbel The random variable (used to introduce randomness and simulate the dynamic fluctuations of the relationship between nodes and hyperedges in a photovoltaic system, such as the uncertainty of parameter association caused by sudden load changes and changes in illumination). : Probability that node v is not associated with hyperedge e ( The natural logarithm of ) (and Together they form the comparison term, and the logarithmic transformation enhances the discriminative power of the probability distribution. : Follows the standard Gumbel distribution (Gumbel Another random variable (with) In conjunction with this, the randomness of the probability distribution is further enhanced, enabling the model to adapt to the dynamic changes in the correlation of photovoltaic system parameters; numerator :node With super-edge The log-dominance term of the association (containing learned association parameters and random perturbations, reflecting the tendency of the association); denominator : The sum of the logarithmic advantages of node v and hyperedge e, both associated and unassociated (the numerator is transformed into a probability form by normalization).

[0072] Step S212: Set up the hypergraph RNN unit

[0073] The input to the hypergraph RNN is the normalized parameter values ​​after preprocessing in step S12. Hidden state from the previous moment The hidden state is updated through a two-level aggregation mechanism of "node-hyperedge-node". The hidden state update adopts a residual connection method, which integrates real-time normalized features, historical hidden states, and hyperedge update features.

[0074] Step S2121: Node-to-hyperedge aggregation

[0075] For each superedge The features of its associated nodes are aggregated, and the formula is:

[0076]

[0077] in, The aggregation characteristics of hyperedge e at time t (combining the characteristics of all nodes associated with hyperedge e to reflect the overall state of the hyperedge at the current time t, such as the comprehensive characteristics of a three-phase current-associated hyperedge). The number of nodes associated with the hyperedge e (used to normalize the aggregation results and avoid inconsistencies in feature scale caused by differences in the number of nodes). The sigmoid function maps aggregated feature values ​​to the [0,1] interval, enhancing nonlinear expressive power and suppressing the influence of outliers. : Sum the characteristics of all nodes v associated with the hyperedge e (summarize node features); Node-to-hyperedge weight matrix transpose (dimension is) ,in For the hidden layer dimension, (This refers to the hyperedge feature dimension, used to linearly transform node features into the hyperedge feature space). : Normalized parameter values ​​of node v at time t (such as normalized phase A current, photovoltaic output voltage, etc.); : The hidden state of node v at time t-1 (stores the historical feature information of node v); : The real-time normalized features of node v are concatenated with the hidden state of the previous time step (2D dimension, integrating real-time and historical information). : Concatenation and join function.

[0078] Step S2122: Hyperedge-to-Node Aggregation

[0079] For each node v, aggregate the features of its associated hyperedges, as shown in the formula:

[0080]

[0081] The update characteristics of node v at time t (combining the characteristics of all hyperedges associated with node v, reflecting the update state of node v at the current time). : The number of hyperedges associated with node v (used to average and normalize the aggregation results to avoid inconsistencies in feature scale caused by differences in the number of hyperedges); The sigmoid function (its function is the same as above, mapping hyperedge aggregation features to...) (interval); : For nodes Sum all associated hyperedges e (summarize hyperedge features); : Hyperedge to node weight matrix transpose (dimension is) ,in For the hyperedge feature dimension, Update the feature dimension for nodes to linearly transform the hyperedge features into the node feature space. : Aggregation features of hyperedge e at time t (output from step S2121); : The set of hyperedges associated with node v (defining the relationship between nodes and hyperedges, such as the photovoltaic voltage node being associated with "voltage-power" hyperedges, "voltage-temperature" hyperedges, etc.).

[0082] Step S2123: Update hidden state

[0083] Hidden state updates use residual joins, with the following formula:

[0084]

[0085] The hidden state of node v at time t (updated node features, used to store the comprehensive information at the current time and as input for the calculation at the next time step). Fully connected layer (performs linear transformation and non-linear mapping on input features, output dimension d=16, integrates real-time, historical and hyperedge updated features); : Real-time normalized features of node v, concatenation of the hidden state from the previous time step and the currently updated features (dimension: =16+16+16=48, fully integrating multi-dimensional information); d=16: hidden layer dimension (feature dimension of node hidden state, balancing model complexity and expressive power).

[0086] Step S22: PHRNN Model Training and Prediction

[0087] The probabilistic hypergraph recurrent neural network model is trained using historical running data. Specifically, it uses historical running data from the past 6 months, divided into training, validation, and test sets in a 7:2:1 ratio, with a time step of 5 minutes (consistent with the sampling frequency of the sampling module).

[0088] Mean squared error (MSE) is used, with the objective of minimizing the deviation between the predicted trigger time and the actual optimal trigger time. The training loss function is the mean squared error between the predicted and actual optimal trigger times, and the formula is:

[0089]

[0090] The loss function used for model training; T is the number of training samples; These are the model's predicted values; The actual optimal value is determined by offline calculation of the time point when the imbalance is minimal after commutation.

[0091] After the model training is complete, input real-time preprocessed data and output the optimal commutation trigger time point for the next moment (5 minutes later). .

[0092] Step S3: Based on the predicted optimal commutation trigger time, an adaptive hybrid gray wolf optimization and particle swarm optimization algorithm is used to optimize the conduction sequence and trigger angle of the commutation switch online to obtain the optimal commutation action.

[0093] Step S3 is used to construct an AR-GWOPSO reinforcement learning-based adaptive control strategy to optimize commutation actions.

[0094] Figure 4 This is a schematic diagram of the optimization process of the Adaptive Hybrid Gray Wolf Optimization and Particle Swarm Optimization (AR-GWOPSO) algorithm according to an embodiment of the present invention. Figure 4 As shown, step S3 uses the predictions made by the PHRNN model in step S2. Based on this, an adaptive hybrid gray wolf optimization and particle swarm optimization algorithm (AR-GWOPSO) is adopted to optimize the conduction sequence and firing angle of the commutation switch online, thereby achieving adaptive control of the commutation action.

[0095] Furthermore, the optimization process of the adaptive hybrid gray wolf optimization and particle swarm optimization algorithm includes: defining the reinforcement learning state space, action space, and reward function. The state space includes the real-time operating state parameters of the system, the action space includes the conduction sequence and trigger angle of the commutation switches, and the reward function integrates the system imbalance, grid-connected energy efficiency ratio, and inrush current; initializing the particle swarm and gray wolf populations, with the gray wolf population selecting the top three particles with the highest reward function values ​​from the particle swarm as elite wolves; and iteratively updating the particle velocity and position by fusing particle swarm optimization velocity updates, gray wolf optimization position updates, and virtual target guidance to obtain the globally optimal action. Specifically, as follows:

[0096] Step 31: Define the reinforcement learning space

[0097] The state space S is defined as the real-time operating state of the system, i.e., S = ,in, This represents the total load power.

[0098] The operating space A is defined as the adjustable parameter of the commutation switch, including:

[0099] Conduction sequence: 3 selectable modes ( );

[0100] Trigger angle: range of values Step length The action space dimension is .

[0101] like Indicates by Sequential conduction, trigger angle is .

[0102] reward function The system incorporates both stability and energy efficiency design, with a weighted sum reward function comprising an imbalance improvement term, a grid-connected energy efficiency ratio term, and an inrush current penalty term, where the imbalance improvement term has the largest weight coefficient. The formula is:

[0103]

[0104] : The reward value for performing action A in state S (to comprehensively evaluate the quality of the commutation operation; the larger the value, the better the action, and it is used to guide the reinforcement learning algorithm to optimize the commutation strategy). Unbalance weighting coefficient (value 0.6, the maximum weight, reflecting the objective of prioritizing the control of three-phase load imbalance). After executing action A in state S, the three-phase load imbalance in the low-voltage distribution area (calculated using the formula is...) ,in , These represent the maximum, minimum, and average values ​​of the three-phase currents after commutation, respectively, with the target control at... within); Imbalance reduction item (the smaller the imbalance, the larger this item value, the higher the reward, and the incentive algorithm reduces the imbalance). Energy efficiency weighting coefficient (value 0.3, secondary weight, used to balance system energy efficiency and imbalance control). : After performing action A in state S, the grid-connected power of the photovoltaic system (the effective power input to the grid). : Output power of the photovoltaic array (power generated by the photovoltaic panel) under state S. : Total load power of the low-voltage distribution area under state S (power consumed by the user side). Grid-connected energy efficiency ratio (reflects the matching efficiency between photovoltaic power and load power; the higher the ratio, the more efficient the energy utilization and the higher the reward). Impact current weighting coefficient (value 0.1, minimum weight, used to suppress current impact during commutation). : The absolute value of the inrush current generated at the moment of commutation when performing action A in state S (excessive inrush current will affect the life of the equipment; this item is a penalty item, and the larger the value, the lower the reward).

[0105] Step 32: AR-GWOPSO Algorithm Optimization Process

[0106] The algorithm integrates the "individual optimal - global optimal" guidance of Particle Swarm Optimization (PSO) with the "elite wolf" cooperation mechanism of Grey Wolf Optimization (GWO), while also introducing virtual target guidance and heuristic obstacle avoidance (corresponding to the constraints of phasing actions).

[0107] Furthermore, the particle swarm optimization velocity update incorporates adaptive inertia weights, which are dynamically adjusted based on individual evolution factors and population aggregation factors; the gray wolf optimization position update uses elite wolves to guide particle positions; and virtual target guidance generates optimal action directions based on K-means clustering or obstacle avoidance action directions based on impact current constraints. The specific steps are as follows:

[0108] Step S321: AR-GWOPSO algorithm initialization

[0109] Step S3211: Particle Swarm Initialization

[0110] Set the number of particles (Each particle corresponds to a set of actions) ), initial velocity ,Location (randomly from the action space) (Selected from the middle), individual optimal position Global optimal position .

[0111] : The number of particles in the particle swarm (set to 20, each particle represents a set of candidate actions for the commutation switch, balancing the search efficiency and computational complexity of the algorithm).

[0112] : No. The initial velocity of each particle (reflecting the particle's movement trend in the action space; the initial value is randomly generated and used for subsequent action updates).

[0113] : No. The initial position of each particle (corresponding to the action space) A specific set of actions, such as a combination of a certain conduction sequence and a trigger angle, is initially randomly selected from the action space.

[0114] : No. The initial optimal position of each particle (initially set to the particle's own initial position, and updated to the action with the highest historical reward in subsequent iterations).

[0115] The initial global optimal position of the particle swarm (via The calculation involves selecting the position of each particle from its initial position in the current state. Lower reward function (The position of the action with the highest value).

[0116] Step S3212: Initialize the gray wolf population

[0117] Select 3 "elite wolves" from the particle swarm. (Wolf), corresponding to the top 3 particle positions in the reward function

[0118] Wolves: Three elite wolves from the gray wolf pack (each corresponding to one of the top three particles in the reward function). The wolf is the best. Wolves are second best. The third function of the wolf is to guide other particles to search for high-quality actions.

[0119] : The particle position corresponding to the wolf (i.e. the action represented by the top 3 particles in the reward function, which serves as the core guiding information for the gray wolf optimization algorithm).

[0120] Step S322: Velocity and Position Update

[0121] Step S3221: PSO speed update

[0122] The formula is as follows: (Incorporating adaptive adjustment of inertia weights)

[0123]

[0124]

[0125]

[0126]

[0127] : The PSO velocity component of the i-th particle at time t (reflects the particle's movement trend guided by individual and global optima, used to update the particle's position); : Adaptive inertia weight at time t (controls the proportion of the particle's velocity inherited from the previous time step, and the ability to explore and utilize the balancing algorithm). : The velocity of the i-th particle at time t-1 (its motion state at the previous time step); : Learning factor (all values ​​are 2.0, controlling the influence of individual optimum and global optimum on particle velocity respectively). : follows a uniform distribution Random numbers (introducing randomness to enhance the search diversity of the algorithm); : The optimal position of the i-th particle at time t-1 (the particle's historical optimal action). : The position of the i-th particle at time t-1 (the current action of the particle); Individual optimal deviation term (guides the particle to move closer to its own historical best action); : The global optimal position at time t-1 (the historical optimal action of the entire particle swarm). : Global optimal deviation term (guides particles to move closer to the group's optimal action); G: Dominance coefficient (value 3.0, scales the overall size of the inertia weight); : Initial inertia weight (value 0.5, providing a baseline value for the inertia weight); k: Evolutionary factor coefficient (value 0.5, adjusting the influence of the evolutionary factor on the inertia weight); Evolutionary factor (reflects the degree of change in an individual's optimal state; a smaller value indicates a slower improvement in the individual's optimal state, requiring an increase in inertia weight to enhance exploration); ρ: Aggregation factor coefficient (value 4.0, adjusting the influence of the aggregation factor on the inertia weight). Aggregation factor (reflects the dispersion of the population; the larger the value, the more aggregated the population, and the inertia weight needs to be increased to promote dispersed exploration). The reward value for the optimal position of the i-th particle at time t; The reward value for the optimal position of the i-th particle at time t-1; The reward value for the globally optimal position at time t; : The average reward value of the particle swarm at any given time.

[0128] Step S3222: GWO Location Update

[0129] The formula for guiding particle positions using elite wolves is:

[0130]

[0131]

[0132]

[0133]

[0134]

[0135]

[0136] : The GWO position component of the i-th particle at time t (subject to (The wolf leads the new action position). :time The wolf's position (corresponding to the top 3 actions in the reward function); Convergence factor (range of values) Decreasing linearly from 2.0 to 0.0 controls the degree to which particles move closer to the elite wolf. Decreasing the value enhances convergence. : Represents the i-th particle and , , The distance of the wolf (quantifying the deviation between the particle's current position and the optimal action); Weighting factor (range of values) ,Depend on (Generate, randomly adjust the influence weight of elite wolf positions). : follows a uniform distribution Random numbers (used to calculate the convergence factor) ); : follows a uniform distribution Random numbers (used to calculate weighting factors) ); : Wolf-guided position component (the main guiding term for particle position); : The positional component guided by the wolf; : The positional component guided by the wolf.

[0137] Step S3223: Virtual Target Guidance

[0138] The virtual target (optimal action direction) is generated using K-means clustering, and the formula is as follows:

[0139]

[0140]

[0141] : The heuristic velocity component of the i-th particle at time t (generated based on virtual targets or obstacle avoidance requirements, guiding the particle to move in a better or safer direction). Virtual target guidance coefficient (value 2.0, controls the intensity of the virtual target's influence on particle velocity); : follows a uniform distribution Random numbers (introducing randomness to avoid excessive concentration of particles on a single virtual target); The virtual target at time t is generated as follows: From the globally optimal actions of the last 100 iterations, the corresponding trigger angle (a numerical feature) and conduction sequence (three features after one-hot encoding) are extracted to form a four-dimensional feature vector. Then, the K-means algorithm is used to cluster these feature vectors into three classes, and the cluster center of the cluster with the largest number of samples is selected. This cluster is then decoded and used as the virtual target. This is used to guide particles to search in the direction of historical high-frequency optimal solutions; : The position of the i-th particle at time t-1 (current action); Virtual target deviation term (guides particles to move closer to the virtual target, accelerating convergence to a high-quality action); Obstacle avoidance coefficient (dynamically adjusts the intensity of obstacle avoidance action; the closer the inrush current is to the danger threshold, the larger the obstacle avoidance coefficient). The heuristic obstacle avoidance position at time t, when the system detects the presence of an inrush current constraint. The following method is used to generate the action: A trigger angle of less than or equal to 30° is randomly selected from the action space A as a candidate action. If the predicted inrush current of this candidate action is evaluated to be lower than the safety threshold, then it is set as the candidate action. Otherwise, the selection process continues until a safe action is found. This mechanism aims to guide particles away from dangerous areas with high impact currents to ensure commutation safety. Obstacle avoidance deviation (guides particles away from high-impact actions to ensure commutation safety); Base obstacle avoidance coefficient (valued at 50, providing a baseline value for the obstacle avoidance coefficient). : Safety distance threshold (value 12, corresponding to the quantitative index of the safe range of the inrush current; if the value is greater than this, there is no need for obstacle avoidance). The danger distance at time t is calculated using the following formula: This value varies with the inrush current. It decreases linearly as it increases. When < (Right now When > 12 A), it is assumed that there is an inrush current constraint, triggering strong obstacle avoidance logic; Dangerous distance threshold (value 4, corresponding to the quantitative index of the dangerous range of the impact current; when it is less than this value, strong obstacle avoidance is triggered). The absolute value of the commutation inrush current; : Maximum safe threshold for inrush current (obstacle avoidance logic is triggered when this value is exceeded).

[0142] Step S3224: Total velocity and position update

[0143]

[0144]

[0145] The total velocity of the i-th particle at time t (integrating PSO velocity, GWO velocity, and heuristic velocity to comprehensively reflect the particle's movement trend). : The PSO velocity component of the i-th particle at time t (from step S3221); : The GWO velocity component of the i-th particle at time t (derived from the position update, reflecting the movement trend guided by the elite wolf). : The heuristic velocity component of the i-th particle at time t (from step S3223); : The new position of the i-th particle at time t (the updated action, obtained by superimposing the position and total velocity at the previous time step); : The position of the i-th particle at time t-1 (the action before the update).

[0146] After the position is updated, it needs to be mapped back to action space A to ensure that the conduction sequence and trigger angle match the actual values. Specifically, for the trigger angle, the closest integer value is taken; for the conduction sequence, the mode with the closest Euclidean distance is selected.

[0147] Step S323: Optimal Action Output

[0148] Furthermore, in step S323, the adaptive hybrid gray wolf optimization and particle swarm optimization algorithm outputs the globally optimal action after iterating a preset number of times in each optimization process. The globally optimal action includes the optimal conduction sequence and the optimal trigger angle. Specifically:

[0149] After 100 iterations, the algorithm selects the globally optimal position. corresponding actions = As a commutation switch The actions performed at any given moment.

[0150] The final global optimal position (the position of the action with the largest reward function value in the particle swarm after the iteration, corresponding to the optimal commutation strategy).

[0151] Optimal action ( For the optimal conduction sequence, The optimal firing angle is the final output of the algorithm.

[0152] Optimal commutation trigger time (predicted by the PHRNN model, at which the commutation switch performs its optimal action). ).

[0153] Step S4: Execute the optimal commutation action at the predicted optimal commutation trigger time to achieve dynamic load balance adjustment of the photovoltaic grid-connected system.

[0154] Step S4 is the final execution stage of the control method of the present invention. When the system real-time clock reaches the optimal commutation trigger time predicted in step S2... At this time, the system control module sends a command to the modular commutation switch device to execute the optimal commutation action obtained from step S3. .

[0155] For example: the central controller will include the optimal conduction sequence order. (e.g., a→b→c) and the optimal trigger angle. An execution command (e.g., 30°) is sent to the commutation switch's execution unit via a communication bus (e.g., CAN or RS485). The commutation switch's execution unit (typically composed of power semiconductor devices such as thyristors or IGBTs) receives the command. Based on the reference point of the grid voltage zero-crossing, the angle is delayed... After the corresponding electrical angle, strictly follow the order. The specified sequence triggers the switching elements of the corresponding phases to conduct sequentially. This commutation operation switches the load current from the current phase to the target phase. Its direct effect is a significant reduction in the imbalance of the three-phase load current. To improve the local absorption rate of photovoltaic energy (i.e., grid-connected energy efficiency ratio) and ensure the commutation inrush current. It is far below the safety threshold, thus achieving efficient, stable and safe operation of the system under dynamic conditions.

[0156] The control method for a modular low-voltage rooftop photovoltaic grid-connected commutation switchgear according to the above embodiments of the present invention constructs a probabilistic hypergraph recurrent neural network (PHRNN) and utilizes Gumbel reparameterization to dynamically capture the high-order nonlinear correlations and temporal evolution laws between system parameters, achieving high-precision and interpretable prediction of the optimal commutation timing and overcoming the response lag problem of traditional methods. Through adaptive hybrid gray wolf optimization and particle swarm optimization (AR-GWOPSO), the global search capability of swarm intelligence and the online decision-making advantages of reinforcement learning are integrated, and virtual target guidance and heuristic obstacle avoidance mechanisms are introduced to achieve fast, safe, and global optimization of commutation actions (conduction sequence and trigger angle), solving the defects of rigidity and easy local convergence of fixed parameter strategies. An integrated intelligent control architecture of prediction-optimization-execution is constructed, making the commutation strategy no longer an isolated response, but a forward-looking and adaptive decision based on system state prediction. Therefore, this invention significantly improves the dynamic response speed and control accuracy of the system, can stably control the three-phase imbalance within the target range, maximize grid-connected energy efficiency and suppress current surges, thereby comprehensively reducing system line losses and voltage distortion, enhancing the stability and power quality of photovoltaic grid connection, providing a solution for building intelligent, efficient and reliable distributed photovoltaic systems, and effectively solving the problem of controlling the dynamic imbalance of three-phase loads in low-voltage photovoltaic grid-connected systems.

[0157] Figure 5This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. The electronic device 10 is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processors, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0158] like Figure 5 As shown, the electronic device 10 includes at least one processor 11 and a memory, such as a read-only memory (ROM) 12 or a random access memory (RAM) 13, communicatively connected to the at least one processor 11. The memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes based on the computer program stored in the ROM 12 or loaded from storage unit 18 into the RAM 13. The RAM 13 can also store various programs and data required for the operation of the electronic device 10. The processor 11, ROM 12, and RAM 13 are interconnected via a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14.

[0159] Multiple components in electronic device 10 are connected to I / O interface 15, including: input unit 16, such as keyboard, mouse, etc.; output unit 17, such as various types of displays, speakers, etc.; storage unit 18, such as disk, optical disk, etc.; and communication unit 19, such as network card, modem, wireless transceiver, etc. Communication unit 19 allows electronic device 10 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0160] Processor 11 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. Processor 11 performs the various methods and processes described above, such as a control method for a modular low-voltage rooftop photovoltaic grid-connected commutation switchgear.

[0161] In some embodiments, a control method for a modular low-voltage rooftop photovoltaic grid-connected phase-connected switchgear can be implemented as a computer program tangibly contained in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the control method for a modular low-voltage rooftop photovoltaic grid-connected phase-connected switchgear described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to perform the control method for a modular low-voltage rooftop photovoltaic grid-connected phase-connected switchgear by any other suitable means (e.g., by means of firmware).

[0162] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0163] Computer programs used to implement the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, such that when executed by the processor, the computer programs cause the functions / operations specified in the flowcharts and / or block diagrams to be performed. The computer programs may be executed entirely on a machine, partially on a machine, or as a standalone software package, partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0164] In the context of this invention, a computer-readable storage medium can be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, apparatus, or device. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination thereof. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0165] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0166] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as data servers), or middleware components (e.g., application servers), or frontend components (e.g., user computers with graphical user interfaces or web browsers through which users can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., communication networks). Examples of communication networks include local area networks (LANs), wide area networks (WANs), blockchain networks, and the Internet.

[0167] A computing system can include clients and servers. Clients and servers are generally located far apart and typically interact through communication networks. The client-server relationship is created by computer programs running on the respective computers and having a client-server relationship with each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a hosting product within the cloud computing service system to address the shortcomings of traditional physical hosts and VPS services, such as high management difficulty and weak business scalability.

[0168] It should be understood that the various forms of processes shown above can be used, with steps reordered, added, or deleted. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0169] The above embodiments are merely illustrative examples and are not intended to limit the implementation. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is neither necessary nor possible to exhaustively list all possible implementations. However, obvious variations or modifications derived therefrom are still within the scope of protection of this invention.

Claims

1. A control method of a modular low voltage rooftop photovoltaic grid-tied commutation switch device, characterized in that, Includes the following steps: Step S1: The system sampling module collects the operating parameters of the photovoltaic grid-connected system in real time and performs data preprocessing on the collected operating parameters. The operating parameters include the output current, output voltage, power factor, and module temperature of the photovoltaic array end; the three-phase input current, three-phase input voltage, and grid frequency of the grid-connected end; and the three-phase load power and three-phase load imbalance of the load end. Step S2: Based on the preprocessed data, construct and train a probabilistic hypergraph recurrent neural network model to output the optimal commutation trigger time point for the next time step. Step S3: Based on the predicted optimal commutation trigger time, the adaptive hybrid gray wolf optimization and particle swarm optimization algorithm is used to optimize the conduction sequence and trigger angle of the commutation switch online to obtain the optimal commutation action; Step S4: Execute the optimal commutation action at the predicted optimal commutation trigger time to achieve dynamic load balance adjustment of the photovoltaic grid-connected system.

2. The method of claim 1, wherein, In step S1, the operating parameters include: output current, output voltage, power factor and module temperature at the photovoltaic array end; three-phase input current, three-phase input voltage and grid frequency at the grid connection end; and three-phase load power and three-phase load imbalance at the load end.

3. The method of claim 1, wherein, In step S1, the data preprocessing includes: filtering and denoising using a moving average filtering algorithm, and normalization processing that maps the parameters to a range of zero to one.

4. The method according to claim 1, characterized in that, In step S2, the probabilistic hypergraph recurrent neural network model adopts an encoder-decoder architecture, including: a probabilistic hypergraph learning module and a hypergraph recurrent neural network unit; The probabilistic hypergraph learning module includes a hypergraph structure consisting of a node set and a hyperedge set, and uses Gumbel reparameterization to model the association weights between nodes and hyperedges as a continuously differentiable probability distribution. The hypergraph recurrent neural network unit achieves temporal dependency modeling through node-to-hyperedge aggregation, hyperedge-to-node aggregation, and hidden state updates.

5. The method of claim 4, wherein, The node set corresponds to the collected operating parameters, and the hyperedge set corresponds to the association combination between the parameters; the hidden state update adopts the residual connection method, which integrates real-time normalized features, historical hidden states and hyperedge update features.

6. The method according to claim 4 or 5, characterized in that, In step S2, the probabilistic hypergraph recurrent neural network model is trained using historical running data, and the training loss function is the mean square error between the predicted trigger time point and the actual optimal trigger time point.

7. The method of claim 1, wherein, In step S3, the optimization process of the adaptive hybrid gray wolf optimization and particle swarm optimization algorithm includes: Define a reinforcement learning state space, action space, and reward function. The state space includes real-time operating state parameters of the system, the action space includes the conduction sequence and firing angle of the commutation switch, and the reward function integrates system imbalance, grid-connected energy efficiency ratio, and inrush current. Initialize the particle swarm and the gray wolf population, wherein the gray wolf population selects the top three particles with the highest reward function values ​​from the particle swarm as elite wolves; By fusing particle swarm optimization for velocity updates, gray wolf optimization for position updates, and virtual target guidance, the velocity and position of particles are iteratively updated to obtain the globally optimal action.

8. The method of claim 7, wherein, The particle swarm optimization velocity update is combined with adaptive inertia weights, which are dynamically adjusted based on individual evolution factors and population aggregation factors; the gray wolf optimization position update guides particle positions through elite wolves; the virtual target guidance generates the optimal action direction based on K-means clustering or generates the obstacle avoidance action direction based on impact current constraints.

9. The method of claim 7, wherein, The reward function is a weighted function that includes an imbalance improvement term, a grid-connected energy efficiency ratio term, and an inrush current penalty term, wherein the imbalance improvement term has the largest weight coefficient.

10. The method of claim 1, wherein, In step S3, the adaptive hybrid gray wolf optimization and particle swarm optimization algorithm outputs the globally optimal action after iterating a preset number of times in each optimization process. The globally optimal action includes the optimal conduction sequence and the optimal trigger angle.

Images