Micro inverter control method and system in photovoltaic system and storage medium

Abstract

The invention relates to the technical field of inverter control, and discloses a micro inverter control method and system in a photovoltaic system and a storage medium. The method comprises the following steps: carrying out parameter acquisition and processing on the photovoltaic micro inverter to obtain an initial parameter set; designing a radial basis function neural network compensation uncertainty parameter based on the initial parameter set; inputting the compensation parameter set into a particle swarm optimization algorithm to optimize control parameters; and training a back propagation neural network based on the optimized parameter set to realize maximum power point tracking control. According to the method, the uncertainty of the system is compensated by introducing the radial basis function neural network, the control parameters are optimized in combination with the particle swarm optimization algorithm, and the real-time prediction and adjustment of the parameters are realized by using the back propagation neural network; the technical problems that a traditional control method is low in control precision and poor in adaptability when facing nonlinear system characteristics, environmental condition changes and parameter uncertainty are effectively solved.

Description

Technical Field

[0001] The present application relates to the technical field of inverter control, and particularly to a control method, system and storage medium for a micro-inverter in a photovoltaic system. Background Art

[0002] Photovoltaic power generation, as a way of clean energy utilization, is playing an increasingly important role in the global energy structure transformation. In a photovoltaic system, micro-inverter technology has been widely used in recent years due to its advantages such as modular design, small impact of single-point faults, and reduced impact of shadow shading. Traditional control methods for photovoltaic micro-inverters mainly include PID control, digital control, open-loop control, closed-loop control, etc. based on linear control theory. By designing appropriate controller parameters, the control method adjusts the switching state of the micro-inverter to achieve the maximization of the output power of the photovoltaic module and stable grid connection. Currently, common maximum power point tracking (MPPT) control strategies include the perturbation and observation method (P&O), incremental conductance method (INC), constant voltage method (CV), etc., and these methods can achieve good tracking effects under standard test conditions.

[0003] However, the existing micro-inverter control methods face various challenges. The photovoltaic system has obvious non-linear characteristics, and its output characteristics are significantly affected by environmental factors such as temperature and irradiance intensity. Traditional linear control methods are difficult to cope with such complex non-linear characteristics. Secondly, the rapid change of environmental conditions will cause the maximum power point of the photovoltaic system to move frequently, and traditional MPPT algorithms often show problems such as slow tracking speed, large oscillation, and high steady-state error in a dynamically changing environment. Thirdly, the parameter uncertainties and external disturbances in the system, such as component aging, shadow shading, load change, etc., further increase the control difficulty. In addition, traditional control strategies usually adopt fixed parameter design and lack adaptability, making it difficult to maintain optimal performance under different working conditions. Finally, the power quality indexes of the micro-inverter output, such as harmonic distortion rate, power factor, etc., also need to be guaranteed while tracking the maximum power, which puts forward the requirement of multi-objective coordinated control for the control system. Summary of the Invention

[0004] The present application provides a control method, system and storage medium for a micro-inverter in a photovoltaic system, which effectively solves the technical problems of low control accuracy and poor adaptability of traditional control methods when facing non-linear system characteristics, environmental condition changes and parameter uncertainties by introducing a radial basis function neural network to compensate for system uncertainties, combining a particle swarm optimization algorithm to optimize control parameters, and using a backpropagation neural network to realize real-time prediction and adjustment of parameters.

[0005] In a first aspect, the present application provides a method for controlling a micro-inverter in a photovoltaic system. The method for controlling a micro-inverter in the photovoltaic system includes: collecting and processing parameters of a photovoltaic micro-inverter to obtain an initial parameter set, where the initial parameter set includes output parameters of photovoltaic modules and environmental condition parameters; designing a radial basis function neural network based on the initial parameter set to compensate for uncertain parameters and obtain a system compensation parameter set; inputting the system compensation parameter set into a particle swarm optimization algorithm, constructing a comprehensive performance evaluation function, optimizing system control parameters, and obtaining an optimized parameter set; constructing and training a backpropagation neural network based on the optimized parameter set, and using the trained backpropagation neural network to predict and adjust system parameters in real time.

[0006] In a second aspect, the present application provides a control system for a micro-inverter in a photovoltaic system. The control system for a micro-inverter in the photovoltaic system includes:

[0007] A modeling module for collecting and processing parameters of a photovoltaic micro-inverter to obtain an initial parameter set, where the initial parameter set includes output parameters of photovoltaic modules and environmental condition parameters;

[0008] A compensation module for designing a radial basis function neural network based on the initial parameter set to compensate for uncertain parameters and obtain a system compensation parameter set;

[0009] A construction module for inputting the system compensation parameter set into a particle swarm optimization algorithm, constructing a comprehensive performance evaluation function, optimizing system control parameters, and obtaining an optimized parameter set;

[0010] A training module for constructing and training a backpropagation neural network based on the optimized parameter set, and using the trained backpropagation neural network to predict and adjust system parameters in real time.

[0011] In a third aspect, there is provided a control device for a micro-inverter in a photovoltaic system, including: a memory and at least one processor, where instructions are stored in the memory; the at least one processor calls the instructions in the memory so that the control device for a micro-inverter in the photovoltaic system executes the above-mentioned method for controlling a micro-inverter in the photovoltaic system.

[0012] In a fourth aspect, there is provided a computer-readable storage medium, where instructions are stored in the computer-readable storage medium, and when the instructions are run on a computer, the computer is made to execute the above-mentioned method for controlling a micro-inverter in the photovoltaic system.

[0013] In the technical solution provided by this application, by collecting and processing the parameters of the photovoltaic micro-inverter, an initial parameter set is obtained, achieving a comprehensive understanding of the system state. The radial basis function neural network designed based on the initial parameter set can accurately compensate for uncertain parameters, effectively solving the problems of uncertain factors such as component aging and environmental changes that cannot be addressed by traditional control methods, and significantly improving the stability and robustness of control. Inputting the system compensation parameter set into the particle swarm optimization algorithm, constructing a comprehensive performance evaluation function, and optimizing the system control parameters, global optimization of the control parameters is achieved, avoiding the defect of traditional methods being prone to falling into local optima. At the same time, by considering three evaluation indicators: power tracking error, harmonic distortion rate, and switching loss, multi-objective balance optimization of the system performance is realized. The backpropagation neural network constructed and trained based on the optimized parameter set realizes fast and accurate tracking of the maximum power point of the micro-inverter by real-time predicting and adjusting the system parameters, greatly improving the dynamic response speed of the system, enabling the photovoltaic system to maintain the best working state under complex and variable environmental conditions. The present invention deeply integrates artificial intelligence algorithms with photovoltaic system control technology. The radial basis function neural network and the backpropagation neural network, as key algorithm features, effectively solve the limitations of traditional control methods in the face of system uncertainties and environmental changes. Through the online learning and adaptive adjustment mechanism, the control system has the capabilities of self-learning and self-adaptation, significantly improving the power generation efficiency of the micro-inverter, reducing system losses, and extending the service life of the equipment. BRIEF DESCRIPTION OF THE DRAWINGS

[0014] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for description in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on the drawings.

[0015] Figure 1 It is a schematic diagram of an embodiment of the control method for the micro-inverter in the photovoltaic system in the embodiment of this application;

[0016] Figure 2 It is a schematic diagram of an embodiment of the control system for the micro-inverter in the photovoltaic system in the embodiment of this application;

[0017] Figure 3 It is a structural schematic block diagram of the control device for the micro-inverter in the photovoltaic system in the embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] The embodiments of the present application provide a control method, system, and storage medium for a micro-inverter in a photovoltaic system. The terms "first", "second", "third", "fourth", etc. (if any) in the specification, claims, and the above-mentioned drawings of the present application are used to distinguish similar objects and do not necessarily describe a specific order or sequence. It should be understood that such data can be interchanged under appropriate circumstances so that the embodiments described herein can be implemented in an order other than that illustrated or described herein. In addition, the terms "include" or "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units does not necessarily have to be limited to those steps or units clearly listed, but may include other steps or units not clearly listed or inherent to the process, method, product, or device.

[0019] For ease of understanding, the specific process of the embodiments of the present application will be described below. Please refer to Figure 1 , an embodiment of the control method for the micro-inverter in the photovoltaic system in the embodiments of the present application includes:

[0020] Step S101: Collect and process the parameters of the photovoltaic micro-inverter to obtain an initial parameter set, which includes the output parameters of the photovoltaic module and the environmental condition parameters;

[0021] Step S102: Design a radial basis function neural network based on the initial parameter set to compensate for the uncertain parameters and obtain a system compensation parameter set;

[0022] Step S103: Input the system compensation parameter set into the particle swarm optimization algorithm, construct a comprehensive performance evaluation function, optimize the system control parameters, and obtain an optimized parameter set;

[0023] Step S104: Construct and train a backpropagation neural network based on the optimized parameter set, and use the trained backpropagation neural network to predict and adjust the system parameters in real time.

[0024] It can be understood that the execution entity of the present application can be the micro-inverter control system in the photovoltaic system, or it can also be a terminal or a server. Specifically, it is not limited here. The embodiments of the present application will be described by taking the server as the execution entity as an example.

[0025] Specifically, data is collected through a sensor network installed on the photovoltaic module. The sensors include voltage sensors, current sensors, temperature sensors, and irradiance sensors, which are respectively installed at key node positions between the photovoltaic module and the micro-inverter. The collected data includes the output voltage and output current of the photovoltaic module as output parameters of the photovoltaic module, as well as temperature and irradiance as environmental condition parameters. The collected raw data undergoes preprocessing steps such as signal filtering, data calibration, and outlier removal to ensure the accuracy and reliability of the data input into the system. When the output voltage of the photovoltaic module fluctuates, transient interference can be eliminated through filtering. For example, when the irradiance changes rapidly, causing the output voltage of the photovoltaic module to fluctuate from 28V to 30V and then back to 29V, a stable value of 29V is obtained after filtering. For abnormal data points collected, such as a suddenly appearing temperature value of 70°C (while surrounding data points are around 25°C), they will be removed through outlier detection algorithms to ensure data continuity and reliability.

[0026] Based on the preprocessed data, characteristic curves of the photovoltaic module are established, including I-V curves and P-V curves. The curves reflect the electrical characteristics of the photovoltaic module under different temperature and irradiance conditions and can identify the variation law of the maximum power point. At the same time, a model of the micro-inverter is established, including a power conversion circuit model, a filter circuit model, and a control system model, to fully describe the dynamic characteristics of the entire process from the DC output of the photovoltaic module to the injection into the AC grid. A mapping relationship matrix R between system parameters and environmental condition parameters is established. The matrix element rij represents the correlation degree between the i-th system parameter and the j-th environmental factor. For example, when the temperature rises from 25°C to 40°C, the open-circuit voltage of the photovoltaic module will decrease by approximately 5%. This corresponding relationship will be quantitatively expressed in the mapping matrix to facilitate subsequent adjustment of control strategies. Finally, an initial parameter set P0 is generated, which includes parameters such as conversion efficiency, maximum power point voltage, maximum power point current, etc., as well as the uncertainty range of each parameter.

[0027] Based on the initial parameter set P0, a radial basis function neural network compensation module is designed to accurately estimate and compensate for the uncertain parameters and external disturbances in the photovoltaic micro-inverter system. The input layer of this network receives system state information, including the output voltage, output current, temperature, irradiance intensity, etc. of the photovoltaic modules. The hidden layer consists of multiple neurons, and each neuron uses a Gaussian function as the activation function, where the center point is the center position of the radial basis function, and the width parameter determines the coverage range of the function. The output layer generates an estimated value of the system uncertainty, including the parameter deviation of the photovoltaic modules and external disturbances. When the efficiency of the photovoltaic modules decreases with the increase of the usage time, the radial basis function neural network can learn and compensate for this parameter drift. The weights of the neural network are updated in real time through an adaptive learning algorithm, and the weight adjustment amount is calculated through the learning rate, the system output error, and the neuron activation value. The number of hidden layer neurons is determined by the cross-validation method, and a regularization technique is implemented to prevent overfitting. After the neural network training is completed, a system compensation parameter set Pc is generated, which includes the parameter correction value ΔP and the uncertainty compensation function f(x).

[0028] The system compensation parameter set Pc is input into the particle swarm optimization algorithm, and a comprehensive performance evaluation function J is constructed. This function includes three evaluation indicators: power tracking error, harmonic distortion rate, and switching loss, and weight coefficients are set for each of the three indicators respectively. The particle encoding scheme of the particle swarm optimization algorithm includes the control parameter vector C to be optimized, that is, the controller parameters, the MPPT step size parameters, and the filter time constant. The initial population is generated by the Latin hypercube sampling method to ensure that the initial particles are evenly distributed in the parameter space. The Latin hypercube sampling determines the dimension and boundary of the parameter search space, and then divides the parameter range of each dimension into N equally spaced parts. N uniformly distributed random points are generated for each dimension, ensuring that the N sampling points on each dimension are not repeated. By randomly pairing the sampling points on each dimension, N particles with space-filling characteristics are generated. The iterative update of each particle follows the update rules of the particle velocity and position, and at the same time, an adaptive inertia weight strategy and a mutation operation are adopted. When the change of the global optimal value is less than the preset threshold ε for N consecutive iterations or the maximum iteration number tmax is reached, the optimized parameter set Copt is output, which includes the optimal values of all control parameters.

[0029] Based on the optimized parameter set Copt, a backpropagation neural network is constructed and trained for real-time prediction of the optimal control parameters of the system. This network adopts a multi-layer perceptron structure, including an input layer, two hidden layers, and an output layer. The input layer receives the current state information S of the system, including photovoltaic voltage, current, temperature, irradiance, grid voltage, and frequency; the output layer generates the optimal control parameter vector Cpredict, including controller parameters, maximum power point tracking step parameters, and filter time constants. The first hidden layer contains 16 neurons, the second hidden layer contains 8 neurons, both using the Leaky ReLU activation function, and the output layer uses a linear activation function. The network training adopts the batch gradient descent method, processing 64 samples per batch, and the loss function uses the mean square error. The collected training data is divided into a training set, a validation set, and a test set according to a ratio of 8:1:1, and Gaussian noise is added to the training data to enhance the robustness of the network. The trained BP neural network can quickly predict the optimal control parameters according to the current system state. For example, when the light intensity suddenly increases from 800 W / m 2 to 1000 W / m 2 , the neural network can output new optimal control parameters within milliseconds, achieving a rapid response to environmental changes without the need to run the computationally intensive PSO algorithm in real time.

[0030] In the embodiment of the present application, by collecting and processing the parameters of the photovoltaic micro-inverter, the initial parameter set is obtained, and a comprehensive understanding of the system state is achieved. The radial basis function neural network designed based on the initial parameter set can accurately compensate for uncertain parameters, effectively solving the problems of uncertain factors such as component aging and environmental changes that cannot be addressed by traditional control methods, and significantly improving the stability and robustness of control. The system compensation parameter set is input into the particle swarm optimization algorithm to construct a comprehensive performance evaluation function, optimize the system control parameters, achieve global optimization of the control parameters, avoid the defect of traditional methods being prone to falling into local optima, and at the same time, by considering three evaluation indicators of power tracking error, harmonic distortion rate, and switching loss, achieve multi-objective balanced optimization of the system performance. The backpropagation neural network constructed and trained based on the optimized parameter set realizes fast and accurate tracking of the maximum power point of the micro-inverter by real-time predicting and adjusting the system parameters, greatly improving the dynamic response speed of the system, enabling the photovoltaic system to maintain the best working state under complex and changeable environmental conditions. The present invention deeply integrates artificial intelligence algorithms with photovoltaic system control technologies. The radial basis function neural network and the backpropagation neural network, as key algorithm features, effectively solve the limitations of traditional control methods in the face of system uncertainties and environmental changes. Through the online learning and adaptive adjustment mechanism, the control system has the ability of self-learning and self-adaptation, significantly improving the power generation efficiency of the micro-inverter, reducing system losses, and extending the service life of the equipment.

[0031] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0032] Collect data through a sensor network, obtain the output voltage and output current of the photovoltaic module as the output parameters of the photovoltaic module, and obtain the temperature and irradiation intensity as the environmental condition parameters;

[0033] Preprocess the output parameters of the photovoltaic module and the environmental condition parameters, including signal filtering, data calibration, and outlier rejection, to obtain preprocessed data;

[0034] Based on the preprocessed data, establish the characteristics of the photovoltaic module, generate I-V curves and P-V curves, and determine the variation law of the maximum power point;

[0035] Establish a micro-inverter power conversion circuit, a filter circuit, and a control system to form the system dynamic characteristics;

[0036] Establish a mapping relationship matrix between the system parameters and the environmental condition parameters, where the elements of the mapping relationship matrix represent the correlation degree between the system parameters and the environmental condition parameters;

[0037] According to the system dynamic characteristics and the mapping relationship matrix, generate an initial parameter set, including the conversion efficiency, the maximum power point voltage, the maximum power point current, and the uncertainty range of each parameter.

[0038] Specifically, data is collected through a sensor network, which includes a voltage sensor, a current sensor, a temperature sensor, and an irradiation intensity sensor. The voltage sensor is installed at the output end of the photovoltaic module to collect the output voltage of the photovoltaic module; the current sensor is installed in the output loop to collect the output current of the photovoltaic module; the temperature sensor is attached to the back of the photovoltaic module to collect the surface temperature of the module; the irradiation intensity sensor is placed at the same angle as the photovoltaic module to collect the irradiation intensity of the incident light. These sensors collect data at regular time intervals (usually 100 milliseconds) to form a time series data stream. The raw data collected by the sensors directly reflects the real-time working state and environmental conditions of the photovoltaic module, but the raw data often contains noise and outliers and needs to be preprocessed.

[0039] Preprocess the output parameters of the photovoltaic module and the environmental condition parameters collected, including signal filtering, data calibration, and outlier rejection. Signal filtering uses a low-pass filter to weaken the high-frequency noise components and smooth the data fluctuations. Specifically, when implementing, for each sampling point, the weighted average of its five adjacent points before and after is taken as the filtering result of this point, and the weight coefficient is larger closer to the central point. Data calibration corrects the systematic errors existing in the sensors, and linearly or non-linearly converts the collected values according to the pre-calibrated calibration curve. Outlier rejection uses statistical methods to calculate the standard deviation of consecutive data points. If the deviation of a certain point from the adjacent points exceeds three times the standard deviation, it is determined as an outlier and replaced with the average value of the adjacent valid values. For example, when the irradiance sensor suddenly jumps from 800 W / m 2 to 1500 W / m 2 and then returns to 810 W / m 2 , the reading of 1500 W / m 2 will be recognized as an outlier and rejected. After preprocessing, a smooth, accurate, and continuous data sequence is obtained.

[0040] Based on the preprocessed data, the characteristics of the photovoltaic module are established to generate the I-V curve and the P-V curve. The I-V curve describes the relationship between the output current and the output voltage of the photovoltaic module under specific environmental conditions; the P-V curve represents the relationship between the output power and the voltage. By collecting the voltage and current values at different operating points, a complete I-V characteristic curve is depicted, and then the power values at each point are calculated by multiplying the voltage and the current to draw the P-V curve. There are three key points on the I-V curve: the open-circuit voltage point (the voltage value when the current is zero), the short-circuit current point (the current value when the voltage is zero), and the maximum power point (the point where the power reaches the maximum value). On the P-V curve, the maximum power point corresponds to the peak point of the curve, and the voltage at this point is called the maximum power point voltage, and the current is called the maximum power point current. By analyzing the I-V curve and the P-V curve under different temperature and irradiance intensity conditions, the variation law of the maximum power point is determined. For example, as the temperature increases, the maximum power point voltage decreases but the variation range is relatively stable, while as the irradiance intensity increases, the maximum power point current approximately linearly increases. A power conversion circuit, a filtering circuit, and a control system of the micro-inverter are established to form the system dynamic characteristics. The power conversion circuit includes a DC-DC boost circuit and a DC-AC inverter circuit. The former boosts the unstable DC voltage output by the photovoltaic module to a stable level, and the latter converts the DC power into AC power that meets the grid requirements. The filtering circuit includes an input-side filter capacitor and an output-side LC filter to reduce high-frequency noise and harmonic distortion. The control system includes a maximum power point tracking controller and a grid-connected controller, which are responsible for optimizing the output efficiency of the photovoltaic module and ensuring the grid-connected power quality respectively. The system dynamic characteristics refer to the response characteristics of the system under external disturbances and internal parameter changes, including indicators such as response time, overshoot, and settling time. By analyzing the transfer functions of each part of the circuit and combining the measured data, a state-space model or a difference equation model of the system is established to describe the dynamic behavior characteristics of the system under various working conditions.

[0041] A mapping relationship matrix between system parameters and environmental condition parameters is established. The elements of this matrix represent the correlation degree between system parameters and environmental condition parameters. The rows of the matrix correspond to system parameters (such as conversion efficiency, maximum power point voltage, etc.), and the columns correspond to environmental condition parameters (such as temperature, irradiance intensity, etc.). Each element value in the matrix is determined through statistical analysis and represents the sensitivity of a specific system parameter to a specific environmental factor. The specific calculation method is to change a single environmental factor while keeping other variables unchanged, record the change of the system parameter, and then calculate the ratio of the change amount to the change amount of the environmental factor. For example, through experiments, it is found that when the temperature increases by 10 °C, the open-circuit voltage of the photovoltaic module drops by 3%, then the mapping relationship value between temperature and open-circuit voltage is -0.3% / °C. The mapping relationship matrix helps to identify which environmental factors have the greatest impact on which system parameters and provides an important basis for the control strategy.

[0042] Generate an initial parameter set according to the system dynamic characteristics and the mapping relationship matrix, including the conversion efficiency, the maximum power point voltage, the maximum power point current, and the uncertainty ranges of each parameter. The initial parameter set is the starting point of system control and provides the reference parameter values of the system under standard conditions. The conversion efficiency refers to the efficiency of the photovoltaic module in converting light energy into electrical energy; the maximum power point voltage is the voltage value when the photovoltaic module outputs the maximum power; the maximum power point current is the corresponding current value. The uncertainty ranges of each parameter reflect the possible fluctuation ranges of the parameters in actual operation, taking into account the effects of measurement errors, environmental factor changes, and component aging. By analyzing historical operation data, combining component specification parameters and measured data, determine the expected value and standard deviation of each parameter to form a complete description of the parameter probability distribution.

[0043] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0044] Construct a radial basis function neural network, and set the input layer to receive system state information. The input layer receives the output parameters of the photovoltaic module and the environmental condition parameters from the initial parameter set;

[0045] Set multiple neurons in the hidden layer of the radial basis function neural network. Each neuron uses a Gaussian function as the activation function. The Gaussian function takes the square of the Euclidean distance between the neuron center point and the input vector, divides it by the square of twice the width parameter, and then takes the negative exponential operation;

[0046] Set the output layer of the radial basis function neural network to output the compensation value for system uncertainty, including the parameter deviation of the photovoltaic module and external interference;

[0047] Update the weight vector of the neural network through an adaptive learning algorithm. The update rule is the learning rate multiplied by the system output error and then multiplied by the neuron activation value;

[0048] Use the cross-validation method to determine the number of neurons in the hidden layer, regularize the weight vector, and automatically adjust the learning rate according to the system error change rate;

[0049] Take the output of the radial basis function neural network as the correction amount of the system parameters to generate a system compensation parameter set, including parameter correction values and uncertainty compensation functions.

[0050] Specifically, a radial basis function neural network is constructed based on the initial parameter set, and the input layer of the neural network is set. The input layer directly receives the state information collected from the photovoltaic micro-inverter system. The number of neurons in the input layer is equal to the dimension of the input features, including six key parameters: the output voltage of the photovoltaic module, the output current, the temperature, the irradiance intensity, and the grid voltage and current. Each input parameter needs to be normalized before entering the network to convert parameters with different dimensions into the same numerical range, usually the interval [0, 1] or [-1, 1]. The normalization process uses the maximum-minimum normalization method, that is, subtracting the minimum value from the original data and then dividing by the difference between the maximum value and the minimum value. For example, when the output voltage range of the photovoltaic module is 0 - 50V, the value obtained after normalizing the measured voltage of 30V is (30 - 0) / (50 - 0) = 0.6. The normalization process ensures that parameters with different dimensions have the same weight in the neural network, avoiding some parameters from dominating the learning process of the network due to their large numerical values.

[0051] Multiple neurons are set in the hidden layer of the radial basis function neural network, and each neuron uses the Gaussian function as the activation function. The calculation process of the Gaussian function calculates the Euclidean distance between the input vector and the center point of the neuron, then divides the square of this distance by the square of twice the neuron width parameter, and finally takes the negative exponent. The Euclidean distance is the straight-line distance between two points in a multi-dimensional space, and the calculation method is to add the squares of the differences in each dimension and then take the square root. For example, when the input vector is [0.6, 0.7, 0.5, 0.8, 0.4, 0.3] and the center point of the neuron is [0.5, 0.6, 0.4, 0.7, 0.5, 0.4], the Euclidean distance is calculated as √[(0.6 - 0.5)²+(0.7 - 0.6)²+(0.5 - 0.4)²+(0.8 - 0.7)²+(0.4 - 0.5)²+(0.3 - 0.4)²] = 0.3162. Assuming the width parameter of this neuron is 0.2, the output value of the Gaussian function is exp[-(0.3162²) / (2×0.2²)] = 0.4403. The number of neurons in the hidden layer directly affects the expression ability and generalization ability of the network. Too few neurons will lead to insufficient expression ability of the network, while too many neurons are prone to overfitting.

[0052] The output layer of the radial basis function neural network is responsible for generating the compensation value for system uncertainties, including photovoltaic module parameter deviations, inverter parameter deviations, and external disturbances. The calculation of the output layer neurons is the weighted sum of the outputs of each neuron in the hidden layer, and the weights are the weight vectors W connecting the hidden layer and the output layer. For example, assume that there are 5 neurons in the hidden layer, and their outputs are [0.4403, 0.6215, 0.3782, 0.5503, 0.4126] respectively, and the weight vector connecting the first output neuron is [0.25, -0.18, 0.32, 0.15, -0.22], then the output value of this output neuron is 0.4403×0.25 + 0.6215×(-0.18) + 0.3782×0.32 + 0.5503×0.15 + 0.4126×(-0.22) = 0.0920. The number of neurons in the output layer is equal to the number of uncertainty parameters to be compensated, which is usually 3 in this scheme, corresponding to photovoltaic module parameter deviations, inverter parameter deviations, and external disturbances respectively.

[0053] The training process of the neural network continuously updates the weight vector W through an adaptive learning algorithm. The update rule is that the weight increment is equal to the learning rate multiplied by the system output error and then multiplied by the neuron activation value. The system output error refers to the difference between the network output and the expected output, which is obtained by comparing the actually measured system parameters with the model prediction values. The learning rate is a key parameter that controls the step size of parameter updates. An overly large learning rate will cause the algorithm to be unstable, while an overly small learning rate will make the convergence process too slow. In actual operation, the learning rate will be automatically adjusted according to the system error change rate. When the error decreases rapidly, the learning rate is reduced to avoid missing the optimal solution. When the error changes slowly, the learning rate is increased to accelerate the convergence process. For example, when the error reduction rate exceeds 10% in three consecutive iterations, the learning rate is multiplied by 0.9; when the error reduction rate is lower than 1% in three consecutive iterations, the learning rate is multiplied by 1.1.

[0054] The number of hidden layer neurons is determined using the cross-validation method. This method divides the training dataset into a training set and a validation set, constructs networks with different numbers of hidden layer neurons, and evaluates the performance on the validation set, and selects the number of neurons that performs best on the validation set. For example, try network configurations with 5, 8, 12, 15, and 20 neurons. After each configuration is trained, calculate the mean square error on the validation set, and select the configuration with the smallest mean square error as the final solution. To prevent overfitting, regularization is performed on the weight vector, that is, an L2 norm term of the weight vector is added to the loss function to encourage the weights to take smaller values, thereby improving the generalization ability of the model.

[0055] After the radial basis function neural network is trained, its output is used as the correction amount of the system parameters to generate a system compensation parameter set, which includes parameter correction values and an uncertainty compensation function. The parameter correction values are directly used to adjust the parameters in the initial parameter set, and the uncertainty compensation function dynamically adjusts the control output during the subsequent control process. For example, when the neural network outputs a parameter deviation of the photovoltaic module of -0.05, it means that the actual output power of the photovoltaic module is 5% lower than the model prediction. At this time, this compensation value will be applied to the control algorithm to adjust the reference point of the maximum power point tracking to ensure that the maximum power point can still be accurately tracked under actual working conditions.

[0056] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0057] Construct a comprehensive performance evaluation function, which includes three evaluation indicators: power tracking error, harmonic distortion rate, and switching loss, and set weight coefficients for the three evaluation indicators respectively;

[0058] Design a particle coding scheme for the particle swarm optimization algorithm, where the particle coding scheme includes controller parameters, maximum power point tracking step parameters, and filter time constants;

[0059] Generate an initial population through the Latin hypercube sampling method. The initial population is evenly distributed in the parameter space, and each particle corresponds to a set of candidate control parameters;

[0060] Calculate the comprehensive performance evaluation function value for each particle in the initial population, and determine the fitness of each particle according to the system compensation parameter set;

[0061] Perform optimization calculations according to the iterative update rules of the particle swarm optimization algorithm. The iterative update rules include particle velocity update and particle position update, and at the same time, an adaptive inertia weight strategy and mutation operation are adopted;

[0062] When the change in the global optimal value is less than the preset threshold for multiple consecutive iterations or the maximum number of iterations is reached, output the optimized parameter set, which includes the optimal values of all control parameters.

[0063] Specifically, the comprehensive performance evaluation function includes three evaluation indicators: power tracking error, harmonic distortion rate, and switching loss, and weight coefficients are set for these three indicators respectively to form a weighted summation form. The power tracking error indicator reflects the deviation degree between the actual output power of the micro-inverter and the theoretical maximum power. The harmonic distortion rate indicator measures the quality of the output electric energy, and the switching loss indicator considers the energy conversion efficiency of the system. The mathematical expression of the comprehensive performance evaluation function is:

[0064] J = w1·E1 + w2·E2 + w3·E3

[0065] Among them, J represents the comprehensive performance evaluation value, E1 represents the power tracking error, and its calculation method is the absolute value of the difference between the actual output power and the theoretical maximum power divided by the theoretical maximum power; E2 represents the harmonic distortion rate, and its calculation method is the square root of the sum of the squares of the effective values of each harmonic current divided by the effective value of the fundamental wave current; E3 represents the switching loss, and its calculation method is the switching frequency multiplied by the energy loss of a single switching process; w1, w2, and w3 are weight coefficients, satisfying the constraint condition of w1 + w2 + w3 = 1.

[0066] The particle coding scheme design of the particle swarm optimization algorithm is the basis of the optimization process. This coding scheme includes controller parameters, the maximum power point tracking step size parameter, and the filter time constant. In the photovoltaic micro-inverter control system, the controller parameters include the proportional coefficient K1, the integral coefficient K2, and the differential coefficient K3. These parameters determine the response characteristics of the controller to errors; the maximum power point tracking step size parameters include the basic step size coefficient α and the minimum step size limit β. These parameters affect the speed and stability of the maximum power point tracking; the filter time constant τ determines the interference suppression ability of the system. The particle coding is in vector form C = [K1, K2, K3, α, β, fsw, τ], where fsw represents the switching frequency, and these parameters together constitute the parameter space to be optimized.

[0067] The initial population generation adopts the Latin hypercube sampling method, which can ensure that the initial particles are evenly distributed in the parameter space. The core idea of Latin hypercube sampling is to divide the value range of each parameter into N intervals (N is the population size), then randomly select a point in each interval, and finally form N particles through random combination. Determine the dimension and boundary of the parameter search space. The dimension is 7, corresponding to the 7 parameters in the particle coding, and the boundaries include the minimum and maximum values of each parameter. For example, the value range of the proportional coefficient K1 is [0.1, 10], the value range of the integral coefficient K2 is [0.01, 5], and so on. Then divide the parameter range of each dimension into N equally spaced parts, randomly select a point in each interval to form a set of N points. Finally, generate N particles with space-filling characteristics by randomly pairing the sampling points on each dimension to form the initial population. This method is more conducive to the rapid convergence of the algorithm than pure random initialization because it ensures the uniform distribution of the initial population in the parameter space and increases the coverage of the search.

[0068] Calculate the comprehensive performance evaluation function value for each particle in the initial population to determine the fitness of each particle. The system compensation parameter set is required in the fitness calculation process. This parameter set is generated by a radial basis function neural network and includes parameter correction values and uncertainty compensation functions. For each particle, its encoded vector is parsed into controller parameters, MPPT step parameters, and filter time constants. Then, these parameters are applied to construct a control system model. Finally, combined with the system compensation parameter set, the operation of the system under given conditions is simulated, and the values of three indicators, namely power tracking error, harmonic distortion rate, and switching loss, are calculated. The final fitness value is calculated through the comprehensive performance evaluation function. The smaller the fitness value, the better the parameter combination represented by the particle. The iterative process of the particle swarm optimization algorithm follows two basic rules: velocity update and position update. The velocity update rule takes into account three factors: the current velocity of the particle, the historical optimal position of the particle, and the global optimal position. The velocity update formula for each particle includes an inertia term, an individual cognitive term, and a social cognitive term. The new velocity of the particle is determined by the weighted sum of these three terms. The position update rule simply adds the current position of the particle to the velocity to obtain the new position. To improve the convergence performance of the algorithm, an adaptive inertia weight strategy is adopted, that is, as the number of iterations increases, the inertia weight gradually decreases from the maximum value to the minimum value. A large inertia weight in the initial stage is beneficial for global search, and a small inertia weight in the later stage is beneficial for local fine search. At the same time, to avoid the algorithm falling into a local optimum, a mutation operation is introduced. In each iteration, the positions of some particles are randomly perturbed with a certain probability to increase the search diversity of the algorithm.

[0069] During the iterative process, the individual optimal position and global optimal position of each particle are continuously updated until the termination condition is met. The termination condition is set as the change in the global optimal value being less than the preset threshold for consecutive multiple iterations or reaching the maximum number of iterations. After reaching the termination condition, the optimized parameter set is output, that is, the parameter vector corresponding to the global optimal position, which includes the optimal values of all control parameters.

[0070] In a specific embodiment, the process of performing the step of generating the initial population by the Latin hypercube sampling method may specifically include the following steps:

[0071] Determine the dimension and boundary of the parameter search space. The dimension corresponds to the number of parameters to be optimized, and the boundary includes the minimum and maximum values of each parameter;

[0072] Divide the parameter range of each dimension into N equally spaced parts, where N is the number of particles in the initial population;

[0073] Generate N uniformly distributed random points for each dimension, and each point is located within the corresponding interval;

[0074] Perform Latin hypercube sampling based on the random points to ensure that the N sampling points on each dimension are not repeated;

[0075] By randomly pairing the sampling points in each dimension, N particles with space-filling characteristics are generated, and the N particles form an initial population;

[0076] An initial velocity vector is assigned to each generated particle, and the magnitude of the initial velocity vector is one-tenth of the parameter range.

[0077] Specifically, the Latin hypercube sampling process determines the dimension and boundary of the parameter search space, where the dimension corresponds to the number of parameters to be optimized. In this solution, the parameters to be optimized include controller parameters (proportional coefficient K1, integral coefficient K2, derivative coefficient K3), maximum power point tracking step size parameters (base step size coefficient α, minimum step size limit β), switching frequency fsw, and filter time constant τ, a total of 7 parameters. Therefore, the dimension of the search space is 7. The boundaries refer to the minimum and maximum values of each parameter, and these boundaries are determined according to the physical characteristics and design requirements of the system. For example, the boundary of the proportional coefficient K1 may be set to [0.1, 10], the boundary of the integral coefficient K2 may be set to [0.01, 5], the boundary of the derivative coefficient K3 may be set to [0, 1], the boundary of the base step size coefficient α may be set to [0.01, 0.1], the boundary of the minimum step size limit β may be set to [0.0001, 0.01], the boundary of the switching frequency fsw may be set to [10000, 50000] Hz, and the boundary of the filter time constant τ may be set to [0.0001, 0.01] s.

[0078] After determining the dimension and boundary of the search space, the parameter range of each dimension is divided into N equally spaced parts, where N is the number of particles in the initial population. The purpose of the division is to ensure that the sampling points are evenly distributed throughout the parameter space and avoid the initial population concentrating in a certain area, resulting in too strong local search ability and insufficient global search ability. For example, if the size N of the initial population is set to 30, the value range of each parameter will be divided into 30 equally spaced parts. For the proportional coefficient K1, its value range [0.1, 10] will be divided into 30 equally spaced parts, and the width of each part is (10 - 0.1) / 30 = 0.33, forming intervals [0.1, 0.43], [0.43, 0.76],..., [9.67, 10].

[0079] After the division, N uniformly distributed random points are generated for each dimension, and each point is located within the corresponding interval. This step is the core of Latin hypercube sampling, ensuring that there is exactly one sampling point in each interval, thus guaranteeing the uniform distribution of sampling points in each dimension. The specific operation is to randomly generate a point within each interval, making it located at any position within that interval. Taking the proportionality coefficient K1 as an example, a point is randomly generated within the first interval [0.1, 0.43], such as 0.25; a point is randomly generated within the second interval [0.43, 0.76], such as 0.62; and so on, until a point is randomly generated within the 30th interval [9.67, 10], such as 9.85. In this way, there are 30 uniformly distributed random points on the parameter axis of K1. The same operation is performed on the other 6 parameter axes, and finally, there are 30 random points on each of the 7 parameter axes.

[0080] Based on the random points, Latin hypercube sampling is performed to ensure that the N sampling points in each dimension are non-repetitive. The key feature of Latin hypercube sampling is that there is only one sampling point in each interval in each dimension, and this constraint ensures the uniformity of sampling points in the dimension direction. The system will check whether the sampling points in each dimension are repetitive. If repetitions are found, random points are regenerated within the corresponding intervals until the non-repetitive condition is met.

[0081] By randomly pairing the sampling points in each dimension, N particles with space-filling characteristics are generated, and the N particles form an initial population. The operation method of this step is to randomly rearrange the N sampling points in each dimension, and then combine them one by one according to the positions to form N particles. For example, the parameter vector of the first particle is composed of the first sampling points on 7 parameter axes; the parameter vector of the second particle is composed of the second sampling points on 7 parameter axes; and so on. This random pairing method ensures the uniform distribution of particles in the multi-dimensional parameter space, has good space-filling characteristics, and helps the particle swarm optimization algorithm to perform global search. An initial velocity vector is assigned to each generated particle, and the magnitude of the initial velocity vector is one-tenth of the parameter range. The velocity vector represents the moving direction and rate of the particle in the parameter space and is an important part of the particle swarm optimization algorithm. The setting of the initial velocity vector affects the convergence speed and search ability of the algorithm. An overly large initial velocity may cause the particle to quickly cross the valuable parameter region, while an overly small initial velocity may cause the algorithm to converge too slowly. Setting it to one-tenth of the parameter range is an empirical choice. In this solution, for the proportionality coefficient K1, its parameter range is 10 - 0.1 = 9.9, so the magnitude of the initial velocity is 9.9 / 10 = 0.99; for the integral coefficient K2, its parameter range is 5 - 0.01 = 4.99, and the magnitude of the initial velocity is 4.99 / 10 = 0.499; and so on. The direction of the initial velocity vector is usually random, indicating that the initial moving trends of the particles are diverse.

[0082] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0083] Based on the optimization parameter set, determine the structure of the backpropagation neural network, and construct a neural network including an input layer, two hidden layers, and an output layer. The first hidden layer contains sixteen neurons, and the second hidden layer contains eight neurons;

[0084] Set the input layer of the backpropagation neural network to receive the output parameters of the photovoltaic module and the environmental condition parameters, set the output layer to output the control parameters, set the Leaky ReLU activation function for the neurons in the two hidden layers, and set the linear activation function for the output layer;

[0085] Use the optimization parameter set to create training samples, use the output parameters of the photovoltaic module and the environmental condition parameters as inputs, and use the control parameters in the optimization parameter set as the output target values;

[0086] Divide the training samples into a training set, a validation set, and a test set, and divide them according to the ratio of 8:1:1, and add noise to the training data to enhance the robustness of the network;

[0087] Train the backpropagation neural network using batch gradient descent, set the mean squared error as the loss function, and use the Adam optimizer for parameter updates;

[0088] Input the current output parameters and environmental condition parameters of the photovoltaic module into the trained backpropagation neural network, output the optimal control parameters, and apply the optimal control parameters to the maximum power point tracking control of the micro-inverter.

[0089] Specifically, construct a backpropagation neural network to achieve real-time parameter prediction. Determine the structure of the backpropagation neural network based on the optimized parameter set, and construct a neural network including an input layer, two hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the input features, including parameters such as the output voltage, output current, temperature, and irradiance intensity of the photovoltaic module; the first hidden layer contains sixteen neurons, and the second hidden layer contains eight neurons; the number of neurons in the output layer is equal to the number of control parameters to be predicted, including controller parameters, maximum power point tracking step parameters, and filter time constants. This hierarchical structure design fully considers the balance between the complexity of data features and the network expression ability. The first hidden layer has a relatively large number of neurons, which can capture rich features of the input data; the second hidden layer has a moderate number of neurons, which are responsible for integrating and abstracting these features, and finally the output layer directly generates the required control parameters.

[0090] Set the input layer of the backpropagation neural network to receive the output parameters and environmental condition parameters of the photovoltaic module. These parameters are normalized and converted to the same numerical range, usually in the interval [0, 1] or [-1, 1], which is convenient for network training and prediction. The output layer directly outputs the control parameters, including the proportional coefficient K1, integral coefficient K2, derivative coefficient K3, maximum power point tracking step coefficients α and β, and filter time constant τ. The neurons in both hidden layers use the Leaky ReLU activation function. For positive input values, this activation function directly passes the original value; for negative input values, it multiplies by a small coefficient (usually 0.01) and then passes. The advantage of the Leaky ReLU activation function is that it solves the problem of neuron death caused by the derivative of the traditional ReLU function being zero in the negative interval, while retaining the advantages of the ReLU function such as simple calculation and stable gradient. The output layer uses a linear activation function to directly output the weighted sum of the neurons. The reason for this design is that the control parameters usually do not have a fixed value range limit, and the linear activation function will not compress the output value and can generate prediction values in any range.

[0091] Create training samples using the optimized parameter set, with the output parameters of the photovoltaic module and the environmental condition parameters as inputs, and the control parameters in the optimized parameter set as the output target values. The creation of training samples requires running the particle swarm optimization algorithm under various environmental conditions and collecting a large number of optimization results. Specifically, during implementation, under different temperatures (such as from 0°C to 50°C, with an interval of 5°C) and irradiation intensities (such as from 200 W / m 2 to 1200 W / m 2 , with an interval of 100 W / m 2 ), run the particle swarm optimization algorithm on the photovoltaic micro-inverter system, and record the system state (input features) and the corresponding optimal control parameters (target output) under each condition. This process is carried out offline, collecting data through a large number of simulation calculations or experimental tests to form a complete training sample set. The quantity and quality of the training samples directly affect the prediction accuracy of the neural network, so it is necessary to ensure that the samples cover all possible operating conditions of the system. Divide the collected training samples into a training set, a validation set, and a test set, with a ratio of 8:1:1, that is, 80% of the samples are used to train the network parameters, 10% are used for network hyperparameter adjustment and early stopping judgment, and 10% are used for the final performance evaluation. This division method ensures that there is enough data for network training, while retaining independent validation and test sets for performance evaluation. Adding noise to the training data is an important means to enhance the robustness of the network. The specific implementation method is to add random noise to the input features of each sample in the training set. The noise usually follows a Gaussian distribution with a mean of zero and a standard deviation of 2% of the standard deviation of the original data. By adding noise, the sensor measurement errors and environmental disturbances in the actual system are simulated, and the trained neural network has a stronger resistance to small fluctuations in the input data.

[0092] The backpropagation neural network is trained using batch gradient descent, i.e., calculating the gradient and updating the network parameters using a batch of samples each time, rather than using all samples or a single sample. The batch size is usually set to 64 samples, which can provide stable gradient estimation without consuming too much computing resources. During the training process, the mean squared error is set as the loss function, i.e., the average of the squares of the differences between the network's predicted output and the target output. The Adam optimizer is used for parameter update. The Adam optimizer combines the advantages of the momentum method and the adaptive learning rate method, and can automatically adjust the learning rate of each parameter, accelerating convergence and improving the training effect. The initial learning rate is set to 0.001, and a cosine annealing learning rate scheduling strategy is adopted. As the number of training epochs increases, the learning rate gradually decreases to ensure that the network can converge to a better local optimal solution. At the same time, an early stopping strategy is implemented. When the loss on the validation set does not decrease for 10 consecutive epochs, the training is stopped to avoid overfitting. The trained backpropagation neural network is applied to an actual photovoltaic micro-inverter system. The current output parameters (voltage, current) of the photovoltaic modules and the environmental condition parameters (temperature, irradiance intensity) are collected. After normalizing these parameters, they are input into the trained neural network. Through forward propagation calculation, the neural network directly outputs the optimal control parameters under the current working conditions, including controller parameters, maximum power point tracking step parameters, and filter time constants. These parameters are then applied to the control system of the micro-inverter to execute the maximum power point tracking control algorithm, enabling the photovoltaic modules to always operate near the maximum power point and achieving the highest energy conversion efficiency. Compared with traditional methods, this neural network-based parameter prediction method can adjust the control parameters in real time according to changes in environmental conditions without the need to run computationally intensive optimization algorithms in real time, greatly reducing the computational burden on the system while maintaining high control performance.

[0093] In a specific embodiment, the process of performing the step of inputting the current output parameters of the photovoltaic modules and the environmental condition parameters into the trained backpropagation neural network may specifically include the following steps:

[0094] Collect the current operating state data of the micro-inverter, obtain the output voltage and output current of the photovoltaic modules as the current output parameters of the photovoltaic modules, and obtain the temperature and irradiance intensity as the environmental condition parameters;

[0095] Perform data preprocessing on the current output parameters of the photovoltaic modules and the environmental condition parameters, including normalization processing and outlier detection, to obtain standardized input data;

[0096] Input the standardized input data into the trained backpropagation neural network, and obtain the optimal control parameters through forward propagation calculation;

[0097] Post-process the optimal control parameters, including parameter range limitation and smooth transition processing, to ensure the stable change of control parameters;

[0098] Based on the post-processed optimal control parameters, set the controller parameters, maximum power point tracking step parameters, and filter time constant of the micro-inverter;

[0099] Apply the controller parameters, maximum power point tracking step parameters, and filter time constant to the micro-inverter control system to execute the maximum power point tracking control algorithm.

[0100] Specifically, collect the current operating state data of the micro-inverter and obtain relevant parameters in real time through the sensor network. The voltage sensor installed at the output end of the photovoltaic module collects the output voltage of the photovoltaic module, and the current sensor collects the output current. These two parameters together constitute the current output parameters of the photovoltaic module. The temperature sensor on the back of the photovoltaic module collects the surface temperature of the module, and the irradiance sensor placed parallel to the photovoltaic module collects the irradiance of the incident light. These two parameters constitute the environmental condition parameters. The collection process is carried out at a certain frequency, usually once every 100 milliseconds, to ensure that the control system can respond promptly to environmental and load changes. The collected raw data is transmitted to the processing unit through the data acquisition module and then enters the data preprocessing link.

[0101] Perform data preprocessing on the currently collected output parameters and environmental condition parameters of the photovoltaic module, including normalization processing and outlier detection, to obtain standardized input data. Normalization processing converts parameters with different dimensions and value ranges into a unified interval, usually [0,1] or [-1,1], so that each parameter has the same weight in the neural network. The specific normalization calculation uses the maximum-minimum normalization method. For example, when the preset range of the output voltage of the photovoltaic module is 0-50V and the measured value is 35V, the normalized value is (35-0) / (50-0)=0.7. Outlier detection identifies and processes abnormal data points through statistical methods. The moving window method is used to calculate the mean and standard deviation of the nearest n data points. If the deviation of the current data point from the mean exceeds k times the standard deviation, it is determined as an outlier. For the detected outliers, one processing method is to replace them with the previous valid value, and the other is to replace them with the mean of the nearest n valid values. After normalization processing and outlier detection, the obtained standardized input data has a unified value range and a smooth change trend, providing good input conditions for the accurate prediction of the neural network.

[0102] The standardized input data is input into the trained backpropagation neural network, and the optimal control parameters are obtained through forward propagation calculation. The backpropagation neural network consists of an input layer, two hidden layers, and an output layer. The first hidden layer contains 16 neurons, and the second hidden layer contains 8 neurons. The forward propagation calculation process is that the input data starts from the input layer, passes through each hidden layer in turn, and finally reaches the output layer. In the calculation of each layer, the output of the previous layer is multiplied by the weight matrix of the current layer, then the bias term is added, and finally the output of the current layer is obtained through the activation function. The LeakyReLU activation function is used in the hidden layer. This function keeps the original value for positive inputs and multiplies negative inputs by a small coefficient (usually 0.01). The linear activation function is used in the output layer, directly outputting the weighted sum of the neurons. After the forward propagation calculation is completed, the neural network outputs the optimal control parameter vector, including the controller parameters (K1, K2, K3), the maximum power point tracking step parameters (α, β), and the filter time constant (τ). These parameters are predicted by the neural network based on the current output state of the photovoltaic module and environmental conditions, and can enable the micro-inverter to reach the best working state under the current working conditions.

[0103] Post-process the optimal control parameters output by the neural network, including parameter range limitation and smooth transition processing, to ensure smooth changes in the control parameters. Parameter range limitation is to limit each control parameter within its reasonable value range to avoid parameter out-of-bounds caused by neural network prediction deviation. For example, the reasonable range of the proportional control parameter K1 is 0.1 - 10. If the neural network output value is 12, it is limited to the upper limit value of 10. Smooth transition processing ensures that the control parameters do not change drastically within adjacent control cycles, avoiding system instability. The specific implementation method is to calculate the change rate of the parameter. If the change rate exceeds the preset threshold, progressive adjustment is adopted, that is, the actual applied parameter value is the intermediate value between the old value and the target value. For example, if the K1 value output by the neural network at the current moment is 5.0, and the actual applied value in the previous control cycle is 2.0, with a change rate of 150%, exceeding the preset threshold of 30%, linear interpolation is used to obtain the actual applied value of 2.0+(5.0 - 2.0)×0.3 = 2.9, rather than directly using 5.0. This smooth transition mechanism ensures the stability of the control system and avoids system oscillations caused by parameter mutations. Based on the post-processed optimal control parameters, set the controller parameters, maximum power point tracking step size parameters, and filter time constant of the micro-inverter. The controller parameters include the proportional coefficient K1, integral coefficient K2, and derivative coefficient K3, corresponding to the three basic parameters in the PID controller respectively. The maximum power point tracking step size parameters include the base step size coefficient α and the dynamic adjustment coefficient β, which jointly determine the step size of the maximum power point tracking algorithm. The filter time constant τ determines the response speed of the system to changes in the input signal. A smaller time constant makes the system respond faster to input changes but has weaker anti-interference ability; a larger time constant makes the system respond slower but has stronger anti-interference ability. The setting of these parameters directly affects the control performance of the micro-inverter, including steady-state accuracy, dynamic response speed, and anti-interference ability. Through the real-time prediction and parameter optimization of the neural network, the micro-inverter can maintain good control performance under various working conditions.

[0104] Apply the controller parameters, maximum power point tracking step parameters, and filter time constant to the micro-inverter control system, and execute the maximum power point tracking control algorithm. The maximum power point tracking control algorithm adopts an improved perturbation observation method. This algorithm periodically changes the operating point of the micro-inverter, observes the change in output power, thereby determines the relative position of the current operating point and the maximum power point, and continuously adjusts the operating point to approach the maximum power point. The controller parameters K1, K2, and K3 determine the response characteristics of the control system to deviations, while the step parameters α and β determine the step size of the operating point adjustment. During the actual execution process, calculate the output power based on the current output voltage and current of the photovoltaic module, and then compare it with the power and voltage of the previous cycle to determine whether the perturbation direction is correct. If the power increases, continue to maintain the current perturbation direction; if the power decreases, reverse the perturbation direction. The perturbation step size consists of a basic step size α and a dynamic adjustment term β|dP / dV|, where |dP / dV| is the absolute value of the change rate of power with respect to voltage, reflecting the distance between the current operating point and the maximum power point. When |dP / dV| is large, it indicates that the current operating point is far from the maximum power point, and a larger step size is adopted to quickly approach; when |dP / dV| is small, it indicates that the maximum power point is approaching, and the step size is automatically reduced to improve the steady-state accuracy. The filter time constant τ is used to smooth the power measurement value and reduce the influence of environmental noise and measurement errors.

[0105] The above describes the micro-inverter control method in the photovoltaic system in the embodiments of the present application. Next, the micro-inverter control system in the photovoltaic system in the embodiments of the present application will be described. Please refer to Figure 2 , an embodiment of the micro-inverter control system in the photovoltaic system in the embodiments of the present application includes:

[0106] [[ID=***8]]The modeling module 201 is used to collect and process the parameters of the photovoltaic micro-inverter to obtain an initial parameter set, and the initial parameter set includes the output parameters of the photovoltaic module and the environmental condition parameters;

[0107] *** The compensation module 202 is used to design a radial basis function neural network based on the initial parameter set to compensate for the uncertain parameters and obtain a system compensation parameter set;

[0108] The construction module 203 is used to input the system compensation parameter set into the particle swarm optimization algorithm, construct a comprehensive performance evaluation function, optimize the system control parameters, and obtain an optimized parameter set;

[0109] The training module 204 is used to construct and train a backpropagation neural network based on the optimized parameter set, and use the trained backpropagation neural network to predict and adjust the system parameters in real time.

[0110] It should be noted that there seems to be an error in the content you provided. The text in line 8 has an asterisk added for the purpose of highlighting the error. The correct text should be: Modeling module 201, configured to collect and process parameters of the photovoltaic micro-inverter to obtain an initial parameter set, the initial parameter set including output parameters of the photovoltaic module and environmental condition parameters; Please check and correct the relevant content according to the actual situation. If you have any other questions, please feel free to let me know.Through the collaborative cooperation of the above-mentioned various components, by collecting and processing the parameters of the photovoltaic micro-inverter, an initial parameter set is obtained, achieving a comprehensive understanding of the system state. The radial basis function neural network designed based on the initial parameter set can accurately compensate for uncertain parameters, effectively solving the problems of uncertain factors such as component aging and environmental changes that cannot be addressed by traditional control methods, and significantly improving the stability and robustness of control. Inputting the system compensation parameter set into the particle swarm optimization algorithm, constructing a comprehensive performance evaluation function, and optimizing the system control parameters, the global optimization of the control parameters is achieved, avoiding the defect of traditional methods being prone to falling into local optima. At the same time, by considering three evaluation indicators of power tracking error, harmonic distortion rate, and switching loss, the multi-objective balanced optimization of the system performance is realized. The backpropagation neural network constructed and trained based on the optimized parameter set realizes the fast and accurate tracking of the maximum power point of the micro-inverter by real-time predicting and adjusting the system parameters, greatly improving the dynamic response speed of the system, enabling the photovoltaic system to maintain the best working state under complex and changeable environmental conditions. The present invention deeply integrates artificial intelligence algorithms with photovoltaic system control technology. The radial basis function neural network and the backpropagation neural network, as key algorithm features, effectively solve the limitations of traditional control methods in the face of system uncertainty and environmental changes. Through the online learning and adaptive adjustment mechanism, the control system has the capabilities of self-learning and self-adaptation, significantly improving the power generation efficiency of the micro-inverter, reducing system losses, and extending the service life of the equipment.

[0111] Above Figure 2 The micro-inverter control system in the photovoltaic system in the embodiment of the present invention is described in detail from the perspective of modular functional entities. Next, the micro-inverter control device in the photovoltaic system in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0112] Figure 3FIG. 0 is a schematic structural diagram of a micro-inverter control device in a photovoltaic system provided by an embodiment of the present invention. The micro-inverter control device 300 in the photovoltaic system may vary greatly due to different configurations or performances, and may include one or more processors (central processing units, CPUs) 310 (for example, one or more processors) and a memory 320, and one or more storage media 330 for storing application programs 333 or data 332 (for example, one or more mass storage device terminals). Among them, the memory 320 and the storage media 330 may be transient storage or persistent storage. The program stored in the storage media 330 may include one or more modules (not shown in the figure), and each module may include a series of instruction operations on the micro-inverter control device 300 in the photovoltaic system. Further, the processor 310 may be configured to communicate with the storage media 330 and execute a series of instruction operations in the storage media 330 on the micro-inverter control device 300 in the photovoltaic system to implement the steps of the micro-inverter control method in the above photovoltaic system.

[0113] The micro-inverter control device 300 in the photovoltaic system may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input / output interfaces 360, and / or one or more operating systems 331, such as Windows Serve, Mac OS X, Unix, Linux, FreeBSD, and so on. Those skilled in the art can understand that Figure 3 the shown structural diagram of the micro-inverter control device in the photovoltaic system does not limit the micro-inverter control device in the photovoltaic system provided by the present invention, and may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0114] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. Instructions are stored in the computer-readable storage medium, and when the instructions are run on a computer, the computer is caused to execute the steps of the micro-inverter control method in the photovoltaic system.

[0115] Those skilled in the art can clearly understand that for the convenience and brevity of description, the specific working processes of the above-described systems, systems, and units can refer to the corresponding processes in the foregoing method embodiments, and will not be described herein again.

[0116] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on such an understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions to enable a micro-inverter control device in a photovoltaic system (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical discs that can store program codes.

[0117] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention 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 for some of the technical features; and the modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A control method for a micro-inverter in a photovoltaic system, characterized in that, The method includes: Collect and process parameters of the photovoltaic micro-inverter to obtain an initial parameter set, where the initial parameter set includes photovoltaic module output parameters and environmental condition parameters; Design a radial basis function neural network based on the initial parameter set to compensate for uncertain parameters and obtain a system compensation parameter set; Input the system compensation parameter set into a particle swarm optimization algorithm, construct a comprehensive performance evaluation function, optimize the system control parameters, and obtain an optimized parameter set; Construct and train a backpropagation neural network based on the optimized parameter set, and use the trained backpropagation neural network to predict and adjust the system parameters in real time.

2. The control method of the micro-inverter in the photovoltaic system according to claim 1, wherein The step of collecting and processing parameters of the photovoltaic micro-inverter to obtain an initial parameter set, where the initial parameter set includes photovoltaic module output parameters and environmental condition parameters, includes: Collect data through a sensor network to obtain the output voltage and output current of the photovoltaic module as the photovoltaic module output parameters, and obtain the temperature and irradiation intensity as the environmental condition parameters; Preprocess the photovoltaic module output parameters and the environmental condition parameters, including signal filtering, data calibration, and outlier rejection, to obtain preprocessed data; Establish the characteristics of the photovoltaic module based on the preprocessed data, generate I-V curves and P-V curves, and determine the variation law of the maximum power point; Establish a power conversion circuit, a filtering circuit, and a control system of the micro-inverter to form the system dynamic characteristics; Establish a mapping relationship matrix between the system parameters and the environmental condition parameters, where the elements of the mapping relationship matrix represent the correlation degree between the system parameters and the environmental condition parameters; Generate the initial parameter set according to the system dynamic characteristics and the mapping relationship matrix, including the conversion efficiency, the maximum power point voltage, the maximum power point current, and the uncertainty range of each parameter.

3. The method for controlling a micro-inverter in a photovoltaic system according to claim 1, characterized in that, The step of designing a radial basis function neural network based on the initial parameter set to compensate for uncertain parameters and obtain a system compensation parameter set includes: Construct a radial basis function neural network, set the input layer to receive system state information, and the input layer receives the photovoltaic module output parameters and environmental condition parameters from the initial parameter set; Set multiple neurons in the hidden layer of the radial basis function neural network, and each neuron uses a Gaussian function as the activation function. The Gaussian function uses the square of the Euclidean distance between the neuron center point and the input vector divided by the square of twice the width parameter, and then takes the negative exponential operation; Set the output layer of the radial basis function neural network to output the compensation value for system uncertainty, including photovoltaic module parameter deviation and external interference; Update the weight vector of the neural network through an adaptive learning algorithm, and the update rule is the learning rate multiplied by the system output error and then multiplied by the neuron activation value; Use the cross-validation method to determine the number of hidden layer neurons, perform regularization processing on the weight vector, and automatically adjust the learning rate according to the system error change rate; Take the output of the radial basis function neural network as the correction amount of the system parameters, and generate the system compensation parameter set, including parameter correction values and uncertainty compensation functions.

4. The micro-inverter control method in the photovoltaic system according to claim 1, wherein Inputting the system compensation parameter set into the particle swarm optimization algorithm, constructing a comprehensive performance evaluation function, optimizing the system control parameters, and obtaining an optimized parameter set, including: Constructing a comprehensive performance evaluation function, where the comprehensive performance evaluation function includes three evaluation indicators: power tracking error, harmonic distortion rate, and switching loss, and weight coefficients are set for the three evaluation indicators respectively; Designing a particle coding scheme for the particle swarm optimization algorithm, where the particle coding scheme includes controller parameters, maximum power point tracking step parameters, and filter time constants; Generating an initial population through the Latin hypercube sampling method, where the initial population is evenly distributed in the parameter space, and each particle corresponds to a set of candidate control parameters; Calculating the comprehensive performance evaluation function value for each particle in the initial population, and determining the fitness of each particle according to the system compensation parameter set; Performing optimization calculations according to the iterative update rules of the particle swarm optimization algorithm, where the iterative update rules include particle velocity update and particle position update, and an adaptive inertia weight strategy and mutation operation are adopted at the same time; When the change in the global optimal value is less than a preset threshold for consecutive multiple iterations or the maximum number of iterations is reached, output the optimized parameter set, including the optimal values of all control parameters.

5. The method for controlling a micro-inverter in a photovoltaic system according to claim 4, characterized in that, The generating of the initial population through the Latin hypercube sampling method, where the initial population is evenly distributed in the parameter space, and each particle corresponds to a set of candidate control parameters, includes: Determining the dimension and boundary of the parameter search space, where the dimension corresponds to the number of parameters to be optimized, and the boundary includes the minimum and maximum values of each parameter; Dividing the parameter range of each dimension into N equally spaced parts, where N is the number of particles in the initial population; Generating N uniformly distributed random points for each dimension, and each point is located within the corresponding interval; Performing Latin hypercube sampling based on the random points to ensure that the N sampling points on each dimension are not repeated; Generating N particles with space filling characteristics by randomly pairing the sampling points on each dimension, and the N particles form the initial population; Assigning an initial velocity vector to each generated particle, and the magnitude of the initial velocity vector is one-tenth of the parameter range.

6. The method for controlling a micro-inverter in a photovoltaic system according to claim 1, wherein, Constructing and training a backpropagation neural network based on the optimized parameter set, and using the trained backpropagation neural network to predict and adjust system parameters in real time, including: Determining the structure of the backpropagation neural network based on the optimized parameter set, constructing a neural network including an input layer, two hidden layers, and an output layer. The first hidden layer includes sixteen neurons, and the second hidden layer includes eight neurons; Setting the input layer of the backpropagation neural network to receive photovoltaic module output parameters and environmental condition parameters, setting the output layer to output control parameters, setting the Leaky ReLU activation function for the neurons in the two hidden layers, and setting the linear activation function for the output layer; Creating training samples using the optimized parameter set, using the photovoltaic module output parameters and environmental condition parameters as inputs, and using the control parameters in the optimized parameter set as output target values; Divide the training samples into a training set, a validation set, and a test set, with a ratio of 8:1:1, and add noise to the training data to enhance the network's robustness. Train the backpropagation neural network using batch gradient descent, set the mean squared error as the loss function, and use the Adam optimizer to update the parameters. Input the current output parameters of the photovoltaic module and the environmental condition parameters into the trained backpropagation neural network, output the optimal control parameters, and apply the optimal control parameters to the maximum power point tracking control of the micro-inverter.

7. The method for controlling a micro-inverter in a photovoltaic system according to claim 6, wherein The step of inputting the current output parameters of the photovoltaic module and the environmental condition parameters into the trained backpropagation neural network, outputting the optimal control parameters, and applying the optimal control parameters to the maximum power point tracking control of the micro-inverter includes: Collect the current operating state data of the micro-inverter, obtain the output voltage and output current of the photovoltaic module as the current output parameters of the photovoltaic module, and obtain the temperature and irradiance intensity as the environmental condition parameters. Perform data preprocessing on the current output parameters of the photovoltaic module and the environmental condition parameters, including normalization processing and outlier detection, to obtain standardized input data. Input the standardized input data into the trained backpropagation neural network, and calculate the optimal control parameters through forward propagation. Perform post-processing on the optimal control parameters, including parameter range limitation and smooth transition processing, to ensure smooth changes in the control parameters. Based on the post-processed optimal control parameters, set the controller parameters, maximum power point tracking step parameters, and filter time constant of the micro-inverter. Apply the controller parameters, maximum power point tracking step parameters, and filter time constant to the micro-inverter control system to execute the maximum power point tracking control algorithm.

8. A micro-inverter control system in a photovoltaic system, characterized in that, A micro-inverter control method in a photovoltaic system for implementing any one of claims 1-7, wherein the micro-inverter control system in the photovoltaic system includes: A modeling module for collecting and processing parameters of the photovoltaic micro-inverter to obtain an initial parameter set, where the initial parameter set includes the output parameters of the photovoltaic module and the environmental condition parameters. A compensation module for designing a radial basis function neural network based on the initial parameter set to compensate for uncertain parameters and obtain a system compensation parameter set. A construction module for inputting the system compensation parameter set into a particle swarm optimization algorithm, constructing a comprehensive performance evaluation function, optimizing the system control parameters, and obtaining an optimized parameter set. A training module for constructing and training a backpropagation neural network based on the optimized parameter set, and using the trained backpropagation neural network to predict and adjust the system parameters in real time.

9. A micro-inverter control device in a photovoltaic system, characterized in that, It includes a memory and a processor. The memory stores a computer program that can run on the processor. When the processor executes the computer program, it implements the micro-inverter control method in the photovoltaic system according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, the processor is caused to execute the micro-inverter control method in the photovoltaic system according to any one of claims 1 to 7.

Images