Adaptive optimization method of MPPT parameters for string inverters

Abstract

The present invention discloses a method for adaptively optimizing the MPPT parameters of a string inverter, comprising: parsing collected multi-channel real-time data of a string inverter to generate a string differential gradient feature matrix that characterizes the relative differences between the operating states of each photovoltaic string; based on the string differential gradient feature matrix, searching and matching in a preset scenario gene library to predictively generate a preliminary MPPT parameter adjustment strategy; and performing collaborative fusion and optimization of the preliminary MPPT parameter adjustment strategy in three dimensions of time, space, and frequency to obtain the optimal MPPT parameter adjustment instruction. By building differentiated perception capabilities, introducing a scenario prediction mechanism, and implementing multi-dimensional fusion optimization, the present invention achieves refined and forward-looking adaptive adjustment of MPPT parameters, thereby improving the power generation efficiency and stability of the system in various complex scenarios.

Description

Technical Field

[0001] The present invention relates to power electronics technology, in particular to a method for adaptively optimizing MPPT parameters of a string inverter. Background Art

[0002] As the "heart" of the photovoltaic system, one of the core tasks of the inverter is to perform maximum power point tracking (MPPT) to ensure that the photovoltaic modules can always output the maximum possible power under complex and changing environmental conditions. The performance of the MPPT algorithm directly determines the power generation and energy conversion efficiency of the photovoltaic system. Studies have shown that for every 1% increase in MPPT efficiency, for a large photovoltaic power station, its power generation revenue over its entire life cycle will differ by hundreds of thousands or even millions of yuan. Therefore, the development of more efficient, smarter, and more adaptable MPPT parameter optimization methods is not only a key technical bottleneck for improving the power generation efficiency of photovoltaic systems and reducing the levelized cost of electricity (LCOE), but also a core technological driving force for promoting high-quality development of the photovoltaic industry, improving project investment returns, and accelerating the realization of clean energy substitution. It has extremely important academic research value and broad engineering application prospects.

[0003] In related technologies, maximum power point tracking (MPPT) methods for string-type photovoltaic inverters mostly employ traditional feedback control strategies such as the perturbation-observation method or the incremental conductance method. These methods perform well under uniform and stable lighting conditions. However, in practical applications, such as large power plants or distributed rooftop scenarios, complex, non-uniform operating conditions often arise due to factors such as drifting clouds, obstruction by buildings or vegetation, dust or snow cover, and inconsistent aging and degradation of individual photovoltaic strings. Under these conditions, traditional MPPT methods have significant drawbacks: First, as a passive response mechanism, their regulation behavior has inherent lags, making it difficult to quickly track rapidly changing environments. Second, their optimization dimension is limited, typically treating the entire photovoltaic array or multiple strings as a whole for fuzzy regulation. This makes it impossible to identify and address differences between strings, and fails to balance global and local optimality, resulting in a significant reduction in the overall power generation efficiency of the system. Summary of the Invention

[0004] The purpose of the invention is to solve the technical problems of the existing technology under complex and non-uniform working conditions, such as the inability to perceive string differences and delayed response, resulting in a single MPPT optimization dimension and low overall power generation efficiency.

[0005] The technical solution is a method for adaptively optimizing string inverter MPPT parameters, including:

[0006] Analyze the collected multi-channel real-time data of the string inverter to generate a string difference gradient feature matrix that characterizes the relative differences between the operating states of each photovoltaic string;

[0007] Based on the cluster difference gradient feature matrix, search and match in the preset scene gene library to generate a preliminary MPPT parameter adjustment strategy;

[0008] The preliminary MPPT parameter adjustment strategy is subjected to coordinated fusion and optimization of time, space and frequency three-dimensional scales to obtain the optimal MPPT parameter adjustment instructions.

[0009] According to one aspect of the present application, the preliminary MPPT parameter adjustment strategy is subjected to coordinated integration and optimization of the three-dimensional scales of time, space and frequency, including:

[0010] The preliminary MPPT parameter adjustment strategy is deconstructed in three dimensions:

[0011] In the time dimension, it is decomposed into multiple time scale components corresponding to fast transient response, medium-speed dynamic regulation and low-speed trend tracking;

[0012] In the spatial dimension, according to the physical topology of the photovoltaic array, it is decomposed into spatial scale components corresponding to the local, regional and system levels respectively;

[0013] In the frequency dimension, it is decomposed into frequency scale components to characterize different dynamic characteristics;

[0014] The time, space and frequency scale components are combined to generate a three-dimensional scale decomposition parameter component matrix.

[0015] According to one aspect of the present application, the preliminary MPPT parameter adjustment strategy is deconstructed in three dimensions, including:

[0016] For the time dimension, empirical mode decomposition is used to separate multiple time scale components;

[0017] For spatial dimensions, multiple spatial scale components are analyzed through graph Fourier transform based on physical topology;

[0018] In the frequency dimension, variational mode decomposition is applied to finely separate multiple frequency scale components.

[0019] According to one aspect of the present application, after generating the three-dimensional scale decomposition parameter component matrix, the collaborative fusion and optimization further includes:

[0020] Construct an adaptive three-dimensional weight tensor based on the intensity of light changes, the difference in power output between strings, and the energy distribution characteristics of the system spectrum;

[0021] The three-dimensional weight tensor is used to weight each component in the three-dimensional scale decomposition parameter component matrix, and the weighted results are fused through a nonlinear optimization function to generate the optimal MPPT parameter adjustment instructions.

[0022] According to one aspect of the present application, based on the string difference gradient feature matrix, searching and matching are performed in the scene gene library, including:

[0023] Using a multi-dimensional similarity matching algorithm, the cluster difference gradient feature matrix is ​​compared with the scene templates in the scene gene library to calculate the numerical similarity between the two at the basic statistical feature level, the structural similarity at the dynamic pattern feature level, and the relationship similarity at the causal correlation feature level.

[0024] Numerical similarity, structural similarity and relational similarity are integrated to form a multi-dimensional scene similarity scoring matrix.

[0025] According to one aspect of the present application, after constructing the multi-dimensional scene similarity scoring matrix, the retrieval and matching step further includes:

[0026] Analyze the multi-dimensional scene similarity score matrix and select the scene with the highest score as the candidate;

[0027] Calculate the difference between the highest score of the candidate scene and the highest score of all other scene templates to evaluate the reliability of the match and generate a confidence assessment;

[0028] Based on the confidence assessment, when the reliability is higher than the preset threshold, a single scene match is determined; when the reliability is lower than the preset threshold, the mixed scene mode is activated to perform a weighted combination of multiple high-scoring scenes;

[0029] A scene recognition result is formed containing single scene or weighted combination scene information.

[0030] According to one aspect of the present application, after forming a scene recognition result, generating a preliminary MPPT parameter adjustment strategy includes:

[0031] Retrieve the basic regulation strategy corresponding to the scene recognition result from the scene gene library;

[0032] According to the deviation between the current string difference gradient feature matrix and the scene template standard feature, the basic adjustment strategy is overall modified;

[0033] The overall revised strategy is then personalized adapted to the specific gradient characteristics of each string to construct a preliminary MPPT parameter adjustment strategy.

[0034] According to one aspect of the present application, the step of generating a string difference gradient feature matrix further includes:

[0035] In the time dimension, the time change rate of each group of string data sequences is calculated through the difference algorithm to form the time dimension gradient;

[0036] In the spatial dimension, the string similarity weight matrix representing the physical and electrical characteristics between strings is calculated, as well as the state difference value between each string and its neighbors. The state difference value and the corresponding weight in the string similarity weight matrix are weighted and summed to form a spatial dimension gradient.

[0037] Integrate the time dimension gradient and the space dimension gradient, and characterize them to form a cluster difference gradient feature matrix;

[0038] According to one aspect of the present application, the temporal dimension gradient and the spatial dimension gradient are integrated and characterized, including:

[0039] The temporal gradient and spatial gradient of each string are combined to form their own gradient vector;

[0040] By quantifying the amplitude difference and direction difference between the gradient vectors of any two strings in parallel, the difference features between the strings are extracted;

[0041] A nonlinear weighted algorithm is used to fuse the amplitude difference and direction difference to construct the gradient feature matrix of group difference.

[0042] According to one aspect of the present application, before integrating and characterizing the temporal dimension gradient and the spatial dimension gradient, a noise reduction process is further included, specifically:

[0043] Analyze the gradient changes of all strings and obtain the benchmark coordinated change pattern;

[0044] Screen out abnormal clusters whose gradient dynamics are contrary to the baseline co-variation pattern;

[0045] With the help of the gradient information of the normal strings adjacent to the abnormal strings, the gradient is compensated and corrected to obtain the gradient data after noise reduction.

[0046] Beneficial effects: By constructing a gradient feature matrix of string differences, the relative operating status of each string is quantitatively characterized from the time and space dimensions, giving the system a refined ability to perceive differences, enabling differentiated and refined adjustments. Using a matching prediction method based on a scenario gene library, real-time operating condition characteristics are quickly matched with a preset template library. The system can predictively generate and call proven efficient strategies, transforming passive search into active prediction, thereby overcoming response lags. Through three-dimensional collaborative optimization of time, space, and frequency, the preliminary strategy is finally calibrated in real time to ensure that the adjustment instructions achieve a balance between rapid transients, local differences, and overall trends. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 It is a flow chart of the present invention.

[0048] Figure 2It is a flow chart of the present invention's implementation of the coordinated integration and optimization of the three-dimensional scales of time, space and frequency for the preliminary MPPT parameter adjustment strategy.

[0049] Figure 3 It is a flowchart of the present invention that deconstructs the preliminary MPPT parameter adjustment strategy in three dimensions.

[0050] Figure 4 It is a flow chart of collaborative fusion and optimization after generating a three-dimensional scale decomposition parameter component matrix of the present invention.

[0051] Figure 5 It is a flowchart of searching and matching in the scene gene library of the present invention. DETAILED DESCRIPTION

[0052] This embodiment aims to illustrate the overall process of the method for adaptively optimizing the MPPT parameters of a string inverter.

[0053] Example 1: According to one aspect of the present application, the step of generating a string difference gradient feature matrix includes:

[0054] Step S101 collects data and performs noise reduction processing. Specifically, the following steps are performed: analyze the gradient changes of all strings to obtain a benchmark coordinated change pattern; screen out abnormal strings whose gradient dynamics are contrary to the benchmark coordinated change pattern; and use the gradient information of normal strings adjacent to the abnormal strings to compensate and correct their gradients to obtain gradient data after noise reduction processing.

[0055] In one embodiment, real-time data from each channel of the string inverter is collected, primarily the voltage and current of each photovoltaic string, and a real-time power sequence Pi(t) is calculated, where i is the string number and t is time. In this embodiment, the data collection interval Δt is set to 1 second. Before calculating the gradient, the power data is subjected to noise reduction.

[0056] Specifically, the average rate of change of all string power at each time step is calculated, serving as a benchmark for coordinated change patterns. The power change rate of each string is then examined individually. If a string k exhibits a significant deviation (for example, exceeding three standard deviations from the average), it is marked as an abnormal string. For this abnormal string k, the gradient information of its neighboring normal strings is used to compensate for it.

[0057] For example, the gradient value can be corrected to a weighted average of the gradients of its adjacent normal strings, with the weights determined based on physical distance or the tightness of electrical connections. This step produces de-noised gradient data, improving the accuracy of subsequent calculations.

[0058] Step S102: Calculate the time dimension gradient and the space dimension gradient. Specifically:

[0059] In the time dimension, the time change rate of each group of string data sequences is calculated through the difference algorithm to form the time dimension gradient;

[0060] In the spatial dimension, the string similarity weight matrix representing the physical and electrical characteristics between strings is calculated, as well as the state difference value between each string and its neighbors. The state difference value and the corresponding weight in the string similarity weight matrix are weighted and summed to form a spatial dimension gradient.

[0061] Integrate the time dimension gradient and the space dimension gradient, and characterize them to form a cluster difference gradient feature matrix;

[0062] For the denoised power data, the gradients in both time and space dimensions are calculated in parallel.

[0063] Among them, the time dimension gradient G of string i T,i It reflects the speed of change of the power of a single string over time. Specifically, the first-order backward difference algorithm is used for calculation: G T,i =(Pi(t)-Pi(t-Δt)) / Δt;

[0064] The spatial dimension gradient G of string i S,i It reflects the difference in state between a single string and its surrounding strings. Specifically, first, based on the physical arrangement or electrical wiring diagram of the photovoltaic array, the neighboring string set N(i) of each string i is determined, that is, the set of strings N(i) that are physically or electrically adjacent to string i, and a string similarity weight matrix is ​​constructed to characterize the physical and electrical characteristics between strings, where the element W ij Indicates the influence weight of string j on string i, reflecting the degree of their mutual influence. The closer the distance or the stronger the electrical connection, the greater the weight value, and ∑ j∈N(i) W ij =1.

[0065] The spatial dimension gradient is calculated by the following formula: G S,i =∑ j∈N(i) W ij ·(P i (t)-P j (t)) This gradient value integrates the relative power differences of all related strings in the neighborhood.

[0066] Step S103 integrates and characterizes the spatiotemporal gradients to form a string difference gradient feature matrix. Specifically: the temporal gradient and spatial gradient of each string are combined to form their own gradient vectors; the difference characteristics between the strings are extracted by quantifying the amplitude difference and direction difference between the gradient vectors of any two strings in parallel; and a nonlinear weighted algorithm is used to fuse the amplitude difference and direction difference to construct a string difference gradient feature matrix.

[0067] Integrate the time gradient and spatial gradient of each string into a two-dimensional gradient vector V of string i i =[G T,i ,G S,i In order to characterize the relative difference between the operating states of any two strings (e.g. string i and string j), the amplitude difference D between their gradient vectors is quantified in parallel. mag (i,j) and direction difference D dir (i, j) to extract differential features.

[0068] The amplitude difference is calculated as: D mag (i,j)=|||V i ∥-∥V j ∥∣;

[0069] The direction difference is calculated as: D dir (i,j)=1-V i ·V j / (∥V i ∥·∥V j ∥);

[0070] D mag (i,j) represents the magnitude difference between the gradient vectors of strings i and j. dir (i, j) represents the directional difference between the gradient vectors of strings i and j.

[0071] Finally, a nonlinear weighted algorithm is used to fuse the amplitude difference and direction difference to construct the gradient feature matrix M of the string difference. The matrix element M ij The calculation is as follows: M ij =α·D mag (i,j)+β·D dir (i,j);

[0072] Among them, α and β are nonlinear weighting coefficients used to fuse amplitude differences and direction differences, which can be set according to historical data and expert experience. For example, when the light changes drastically, the weight of β can be increased to pay more attention to the synchronization of the string response.

[0073] For example, a photovoltaic array containing 16 strings will eventually generate a 16x16 string difference gradient feature matrix M through the above steps. Each element of this matrix M ij Both quantify the comprehensive difference between the operating states of strings i and j at the current moment. This matrix will serve as the input for subsequent scene recognition.

[0074] Through the above steps, this embodiment converts the original multi-channel time series data into a feature matrix that can comprehensively and finely characterize the dynamic differences between each group of strings, providing a high-quality data foundation for subsequent accurate scene recognition and strategy optimization.

[0075] Embodiment 2: This embodiment is based on the embodiment 1 and specifically describes the process of scene recognition and generation of a preliminary MPPT parameter adjustment strategy.

[0076] Step S201: perform multi-dimensional similarity matching, specifically:

[0077] Through a multidimensional similarity matching algorithm, the group difference gradient feature matrix is ​​compared with the scene templates in the scene gene library, and the numerical similarity at the basic statistical feature level, the structural similarity at the dynamic pattern feature level, and the relational similarity at the causal correlation feature level are calculated; the numerical similarity, structural similarity and relational similarity are integrated to form a multidimensional scene similarity scoring matrix.

[0078] The real-time cluster difference gradient feature matrix M generated in Example 1 is compared with multiple scene templates M in the preset scene gene library. templatek Compare, M templatek Represents the template feature matrix of the kth scene in the scene gene library. The scene gene library pre-stores feature matrix templates for various typical working conditions, such as uniform and gradual change, partial cloud shadow occlusion, array edge stain coverage, and string performance degradation. The comparison process is completed through a multi-dimensional similarity matching algorithm, which calculates similarity at three levels in parallel:

[0079] Numerical similarity Sim N : At the basic statistical feature level, calculate the real-time matrix M and the template matrix M templatek The element-wise difference between . For example, it can be calculated by taking the difference of the Frobenius norm of the two matrices or the cosine similarity.

[0080] Structural Similarity Sim S Compare the topological similarity of two matrices at the level of dynamic pattern features. For example, the matrices can be viewed as graph adjacency matrices, and their structural similarity can be assessed using graph edit distance or graph kernel functions to identify whether the difference patterns (e.g., localized or dispersed) are consistent.

[0081] Causal feature similarity Sim C : At the causal correlation feature level, the Granger causality between the original power sequences of the generated matrix is ​​analyzed and compared with the causal relationship diagram of the template scene to determine whether the dynamic influence transmission mode between the strings matches. Finally, the three similarities are integrated to form a multi-dimensional scene similarity score S for each template scene k k =w N Sim N +w S Sim S +w C SimC ; Among them, w N ,w S ,w C The preset weights.

[0082] Sim N ,Sim S ,Sim C Represents numerical similarity, structural similarity and causal similarity respectively. k Represents the total multi-dimensional scene similarity score between the real-time scene and the template scene k.

[0083] Step S202: Evaluate the matching reliability and determine the scenario, specifically:

[0084] Analyze the multi-dimensional scene similarity score matrix and select the scene with the highest score as the candidate;

[0085] Calculate the difference between the highest score of the candidate scene and the highest score of all other scene templates to evaluate the reliability of the match and generate a confidence assessment;

[0086] Based on the confidence assessment, when the reliability is higher than the preset threshold, a single scene match is determined; when the reliability is lower than the preset threshold, the mixed scene mode is activated to perform a weighted combination of multiple high-scoring scenes;

[0087] A scene recognition result is formed containing single scene or weighted combination scene information.

[0088] Analyze the multi-dimensional scene similarity score matrix obtained in the previous step and select the scene with the highest score as the candidate scene. In order to evaluate the reliability of the match, calculate the highest score S of the candidate scene. max The second highest score among all other scene templates second_max The gap between them is used to generate a confidence assessment C of the scene matching. conf =(S max -S second_max ) / S max ; Set a confidence threshold, such as 0.75. If C conf If C is higher than the threshold, the matching result is reliable and a single best matching scenario is determined. conf If the value is lower than the threshold, it indicates that the current working condition may be a superposition of multiple typical working conditions. The mixed scene mode is activated, and the highest-scoring scenes (for example, the top three) are weightedly combined according to their scores to form a mixed scene recognition result.

[0089] Step S203: Generate a preliminary MPPT parameter adjustment strategy, specifically:

[0090] Retrieve the basic regulation strategy corresponding to the scene recognition result from the scene gene library;

[0091] According to the deviation between the current string difference gradient feature matrix and the scene template standard feature, the basic adjustment strategy is overall modified;

[0092] The overall revised strategy is then personalized adapted to the specific gradient characteristics of each string to construct a preliminary MPPT parameter adjustment strategy.

[0093] Based on the scene recognition results of the previous step, a preliminary MPPT parameter adjustment strategy is generated. First, the basic adjustment strategy corresponding to the scene recognition result (whether it is a single scene or a mixed scene) is retrieved from the scene gene library. The basic strategy is a set of macro adjustment principles. For example, for a scene with local cloud shadow occlusion, the global MPPT scanning cycle should be reduced, and a refined scan should be started for the affected strings. Then, based on the current real-time string difference gradient feature matrix M and the matched scene template standard feature M templatek For example, if M shows a greater degree of difference than template M templatek If the disturbance is more intense, the adjustment strength will be increased accordingly (such as increasing the MPPT disturbance step size). Finally, for the overall revised strategy, the specific gradient characteristics of each string (i.e., the gradient vector V i ) for personalized adaptation. For example, more aggressive MPPT tracking parameters can be assigned to strings with particularly large gradient vector amplitudes. A specialized diagnostic subroutine can be triggered for strings whose gradient vector directions differ from those of the majority. For example, if string 3 is identified as experiencing performance degradation due to dust accumulation, the initial strategy would be to maintain normal MPPT parameters for the other strings but adjust the MPPT voltage search lower limit for string 3 and appropriately reduce its perturbation step size to avoid power oscillations.

[0094] Through the above steps, this embodiment can accurately identify the complex working conditions of the current photovoltaic system and generate a preliminary MPPT parameter adjustment strategy that is both macro-guidance and micro-targeted, laying the foundation for the final refined optimization.

[0095] Example 3: This example is based on Example 2 and specifically describes the process of performing coordinated integration and optimization of the three-dimensional scales of time, space and frequency on the preliminary MPPT parameter adjustment strategy.

[0096] The process includes the following steps:

[0097] Step S301: Deconstruct the preliminary adjustment strategy in three dimensions. Specifically:

[0098] The preliminary MPPT parameter adjustment strategy is deconstructed in three dimensions:

[0099] In the time dimension, it is decomposed into multiple time scale components corresponding to fast transient response, medium-speed dynamic regulation and low-speed trend tracking;

[0100] For the time dimension, empirical mode decomposition is used to separate multiple time scale components;

[0101] In the spatial dimension, according to the physical topology of the photovoltaic array, it is decomposed into spatial scale components corresponding to the local, regional and system levels respectively;

[0102] For spatial dimensions, multiple spatial scale components are analyzed through graph Fourier transform based on physical topology;

[0103] In the frequency dimension, it is decomposed into frequency scale components to characterize different dynamic characteristics;

[0104] In the frequency dimension, variational mode decomposition is applied to finely separate multiple frequency scale components.

[0105] The preliminary MPPT parameter adjustment strategy generated in Example 2 (which can be expressed as a set of parameter sequences for each string group that changes over time) is collaboratively deconstructed in three dimensions: time, space, and frequency.

[0106] In the temporal dimension, it is decomposed into components at multiple time scales. Specifically, empirical mode decomposition (EMD) is used to decompose time-varying parameters in the strategy (such as the perturbation voltage step size) into multiple intrinsic mode functions (IMFs). These IMF components correspond to different time scales. For example, high-frequency IMF components correspond to rapid responses to transient light changes, medium-frequency components correspond to the regulation of medium-speed dynamics such as cloud movement, and low-frequency residual components correspond to the slow trend tracking of sunlight intensity.

[0107] In the spatial dimension, the PV array's physical topology is decomposed into components at multiple spatial scales. Specifically, by constructing the array's graph Laplacian matrix and applying the physical topology-based Graph Fourier Transform (GFT), the regulation strategies applied to individual strings are decomposed into distinct spatial patterns. These patterns correspond to the local scale (independent regulation of a single string), the regional scale (coordinated regulation of a group of adjacent strings), and the system-level scale (global regulation of the entire array).

[0108] In the frequency dimension, in order to more finely separate different dynamic characteristics, variational mode decomposition (VMD) is applied to process the strategy signal and decompose it into multiple frequency scale components with specific center frequencies and compact bandwidths. Through the above decomposition, a three-dimensional scale decomposition parameter component matrix (or tensor) T is obtained. ijk , where the indices i, j, k correspond to the time, space and frequency scales respectively.

[0109] Step S302 : constructing an adaptive three-dimensional weight tensor. Specifically, the time, space, and frequency scale components are combined to generate a three-dimensional scale decomposition parameter component matrix.

[0110] In order to effectively fuse the decomposed components, it is necessary to dynamically determine their weights. This step constructs a ijk Adaptive three-dimensional weight tensor W of the same dimension ijk The element values ​​of this tensor are dynamically generated according to the real-time operating status of the system:

[0111] The severity of the light change is assessed by analyzing the high-frequency component energy of the input power. If the light change is severe, the weight of the time-scale component corresponding to the fast transient response (high-frequency IMF) is increased.

[0112] The degree of power output variability between strings is assessed by calculating the variance of all string powers. If the variability is large, the weights corresponding to the local and regional spatial scale components are increased to achieve differentiated regulation.

[0113] Energy distribution characteristics of the system spectrum: obtained by Fourier analysis of the total system power. If the energy is mainly concentrated in a specific frequency band, the weight of the corresponding frequency scale component is increased.

[0114] Step S303: weighted fusion and optimization to generate optimal instructions. Specifically, an adaptive three-dimensional weight tensor is constructed based on the intensity of illumination changes, the difference in power output between strings, and the energy distribution characteristics of the system spectrum.

[0115] The three-dimensional weight tensor is used to weight each component in the three-dimensional scale decomposition parameter component matrix, and the weighted results are fused through a nonlinear optimization function to generate the optimal MPPT parameter adjustment instructions.

[0116] Use the three-dimensional weight tensor W generated in the previous step ijk Decompose the parameter component matrix T of the three-dimensional scale ijk Then, the weighted results are fused through a preset nonlinear optimization function Fopt to generate the final optimal MPPT parameter adjustment instruction Cmd opt . Cmdopt =F opt (∑ i,j,k W ijk ·T ijk ) Nonlinear optimization function F opt It can be a multi-layer perceptron (MLP) or a radial basis function network (RBFN), which is trained offline to map the weighted multi-scale components to the optimal actual MPPT parameters (such as voltage step size, scanning period, etc.).

[0117] For example, in one scenario, monitoring of rapidly passing cloud fragments (dramatic changes in illumination) can cause significant power fluctuations in several strings (with high variability) in one corner of the array. In this case, the weight tensor Wijk automatically assigns high weights to components at the fast time scale, the local spatial scale, and the specific frequency scale corresponding to the fluctuation frequency. The resulting optimal command is a fast, refined regulation command for the string in that specific corner, while maintaining stable tracking for other unaffected strings, thereby maximizing global power generation efficiency.

[0118] Through the above steps, this embodiment refines the preliminary and relatively rough adjustment strategy into a final control instruction that achieves coordinated optimization in the three dimensions of time, space, and frequency, achieving ultimate adaptability to complex working conditions.

[0119] The method comprises the following steps:

[0120] Step S1: parsing real-time data to generate a string difference gradient feature matrix. Specifically, parsing the collected multi-channel real-time data of the string inverter to generate a string difference gradient feature matrix that characterizes the relative differences between the operating states of each photovoltaic string;

[0121] This step collects real-time voltage and current data from each PV string through the string inverter's multi-channel data interface and calculates power. The system first performs noise reduction on the data, concurrently calculating the temporal and spatial gradients for each string and integrating them into a gradient vector. By quantifying the amplitude and direction differences between any two string gradient vectors and performing a nonlinear weighted fusion, a string differential gradient feature matrix is ​​constructed that comprehensively characterizes the relative differences between the operating states of each PV string.

[0122] Step S2, searching and matching the scene gene library, generating a preliminary MPPT parameter adjustment strategy, specifically: based on the string difference gradient feature matrix, searching and matching in the preset scene gene library, generating a preliminary MPPT parameter adjustment strategy;

[0123] This step uses the real-time feature matrix generated in the previous step as input and performs high-speed search and matching within a pre-set scenario gene library. Using a multi-dimensional matching algorithm that includes numerical, structural, and causal similarity calculations, the system finds the single or mixed scenario that best matches the current operating conditions. The system evaluates the confidence level of the match and, based on the identified scenario, retrieves a basic adjustment strategy from the library. It then performs overall corrections and personalized adaptation based on the specific deviations from the current operating conditions, generating a preliminary MPPT parameter adjustment strategy.

[0124] Step S3: perform time-space-frequency three-dimensional collaborative optimization to obtain the optimal MPPT parameter adjustment instruction. Specifically: perform time-space-frequency three-dimensional collaborative fusion and optimization on the preliminary MPPT parameter adjustment strategy to obtain the optimal MPPT parameter adjustment instruction.

[0125] This step refines the initial control strategy. The system deconstructs the initial strategy across the three dimensions of time, space, and frequency, yielding a series of parameter components corresponding to different scales. Simultaneously, a three-dimensional weight tensor is dynamically constructed based on real-time illumination changes, string differences, and spectral characteristics. This weight tensor is used to adaptively weight the decomposed parameter components and fuse them using a nonlinear optimization function. This ensures that the final control command strikes a balance between rapid response, local adjustment, and global optimization.

[0126] Through the above steps, the method of the present invention achieves a complete closed loop from data perception to scene understanding, strategy generation, and refined optimization. Test data shows that under typical non-uniform lighting conditions (such as cloud cover and localized smudges), the MPPT efficiency of a string inverter system using this method can be improved by approximately 3% to 8% compared to the traditional perturbation-and-observe method, effectively increasing the actual power generation of the photovoltaic power station.

[0127] Example 5: According to one aspect of the present application, the string difference gradient feature matrix may also be generated as follows:

[0128] Step S110, calculating the time dimension gradient and the space dimension gradient.

[0129] In this embodiment, in order to more accurately capture the dynamic characteristics, the time dimension gradient is preferably calculated using a central difference algorithm.

[0130] Specifically, for the time gradient G of the jth parameter (such as power P) of the i-th string at time t, time [i,j,t], the calculation formula is: G time [i,j,t]=(D aligned [i,j,t+Δt]-D aligned [i,j,t-Δt]) / 2Δt where D aligned[i, j, t] is the standardized measured value of the jth parameter of the i-th string at time t; Δt is the time difference step, which can be set to one sampling period, for example.

[0131] The spatial dimension gradient is preferably calculated using the neighborhood weighted spatial difference method. Specifically, its calculation formula is: G space [i,j,t]=∑ N k=1 W[i,k]×(D aligned [k,j,t]-D aligned [i, j, t]); where N is the total number of strings in the PV array; k is the index used to traverse all strings; W[i, k] is the similarity weight between strings i and k, which can be determined based on factors such as the physical distance between strings, cable length, and orientation angle. The output of this step is the time dimension gradient matrix G time and the spatial dimension gradient matrix G space .

[0132] Step S120: differential characterization to construct a final feature matrix.

[0133] In this embodiment, in order to more robustly characterize the differences between strings, a nonlinear weighted algorithm is used to fuse the intensity and direction of the differences.

[0134] First, the temporal and spatial gradients of each string are combined into a two-dimensional gradient vector [G time [i,j,t],G space [i,j,t]].

[0135] Then, the magnitude difference and direction difference between the gradient vectors of any two strings (i and k) are calculated in parallel.

[0136] Amplitude difference calculation: First calculate the amplitude of each gradient vector Mag[i,j,t]=sqrt(G time [i,j,t] 2 +G space [i,j,t] 2 );

[0137] Then calculate the amplitude difference Diff mag [i,k,j,t]=∣Mag[i,j,t]-Mag[k,j,t]∣;

[0138] Direction difference calculation: First calculate the direction angle of each gradient vector Angle [i,j,t]=arctan2(G space [i,j,t],Gtime[i,j,t]);

[0139] Then calculate the direction difference Diffangle , and normalize it to the interval [0,π]: Diff angle [i,k,j,t]=min(| Angle [i,j,t]- Angle [k,j,t]|,2π-| Angle [i,j,t]- Angle [k,j,t]|);

[0140] Finally, the two differences are fused through a nonlinear weighting algorithm to construct the final cluster difference gradient feature matrix. Preferably, this fusion can include amplitude, direction, and the cross term between the two to capture nonlinear coupling effects.

[0141] Optionally, gradient denoising can be performed before the above steps. Specifically, by analyzing the correlation between gradient changes between strings, normal coordinated variation patterns can be identified. For abnormal strings that clearly do not conform to the variation patterns of other strings, a weighted average of their neighboring normal strings can be used to correct them, improving the quality of the gradient data.

[0142] Example 6: According to one aspect of the present application, the process of scene matching and preliminary strategy generation is further as follows:

[0143] Step S210 , performing multi-dimensional and multi-level similarity matching.

[0144] In this embodiment, to improve the accuracy of scene recognition, the similarity matching algorithm is performed from multiple dimensions and levels. The real-time generated string difference gradient feature matrix is ​​compared with the templates in the scene gene library, and the similarity at three levels is specifically calculated:

[0145] Numerical similarity: Calculates the similarity between the real-time gradient feature and the scene template in terms of basic statistical features such as mean and variance. Preferably, a weighted cosine similarity algorithm can be used.

[0146] Structural similarity: Calculates the similarity between the two in terms of dynamic pattern features such as change trends, periodicity, and frequency distribution. Preferably, dynamic time warping (DTW) or pattern correlation analysis algorithms can be used.

[0147] Relational similarity: Calculate the similarity between the causal relationship between the gradient and environmental factors (such as light and temperature). Preferably, an algorithm based on mutual information or transfer entropy can be used to quantify this relationship.

[0148] The three similarity scores are then adaptively weighted and fused to produce the final composite score. For example, the weights can be dynamically adjusted based on the quality of the current data, giving higher weights to numerical similarities when the data quality is high.

[0149] Step S220 , executing confidence-based scene recognition and strategy generation.

[0150] After obtaining the multi-dimensional scene similarity scoring matrix, the reliability of the matching results is first quantitatively evaluated. Specifically, the scene with the highest score is selected as a candidate, and the confidence is evaluated by calculating the normalized difference between its highest score and the second highest score. A confidence threshold is set, for example 0.15. If the score difference is greater than the threshold, it indicates that the matching result is highly reliable and is determined to be a single scene match. If the difference is less than the threshold, it indicates that the current working condition may be a mixed condition of multiple scenes, and the mixed scene mode is started at this time. In the mixed scene mode, multiple scenes with scores higher than a certain basic threshold (for example, 0.7) are weighted and combined according to their scores to form a weighted combination of scene recognition results.

[0151] After forming the scene recognition results, generate a preliminary MPPT parameter adjustment strategy. The specific steps are:

[0152] The basic MPPT parameter adjustment strategy corresponding to the identified scenario (single or mixed) is extracted from the scenario gene library.

[0153] According to the degree of difference between the current real-time gradient features and the standard scene template, the parameters of the basic strategy are fine-tuned.

[0154] Taking into account the differences between strings, further personalized parameter adjustments are performed on each string, and finally a preliminary MPPT parameter adjustment strategy vector is obtained.

[0155] Optionally, the scenario gene library can be dynamic. When the system encounters an unknown scenario that cannot be matched with any existing template with a high score, it can start a new scenario learning mode, form a new strategy through online optimization, and add it to the library as a new gene after verification, realizing the self-evolution of the gene library.

[0156] Example 7: According to one aspect of the present application, the process of collaborative fusion and optimization of the three-dimensional scales of time, space and frequency is specifically as follows:

[0157] Step S310 , performing refined time-space-frequency three-dimensional scaling decomposition.

[0158] In this embodiment, in order to achieve a refined deconstruction of the control problem, the preliminary MPPT parameter adjustment strategy is decomposed into three orthogonal dimensions.

[0159] Time dimension decomposition: Empirical mode decomposition (EMD) is preferably used to adaptively decompose the strategy's time series signal into multiple intrinsic mode functions (IMFs). These IMF components correspond to different time scales such as fast transient response, medium-speed dynamic adjustment, and low-speed trend tracking.

[0160] Spatial dimension decomposition: The strategy is preferably spatially parsed into multiple spatial scale components corresponding to local (single string), regional (adjacent strings) and system level (entire array) by constructing a spatial graph based on the physical topology of the PV array and applying the Graph Fourier Transform (GFT).

[0161] Frequency dimension decomposition: Variational mode decomposition (VMD) is preferably applied to perform fine frequency separation of the signal to obtain frequency scale components used to characterize different dynamic characteristics.

[0162] The output of this step is a three-dimensional scale decomposition parameter component matrix, which provides the basis for the next step of weighted fusion.

[0163] Step S320: performing adaptive weighting and nonlinear fusion based on real-time status.

[0164] To optimize based on the most pressing needs of the system, this step constructs an adaptive three-dimensional weight tensor and fuses the decomposed components. First, adaptive weights are constructed based on the real-time state of the system. For example:

[0165] If the rate of change of light intensity is detected to exceed a certain threshold (such as 500W / m² / s), the weight of the corresponding fast response component in the time dimension will be increased by 50%.

[0166] If it is detected that the maximum power difference between strings exceeds 20%, the weight of the corresponding local optimization component in the spatial dimension will be increased by 40%.

[0167] The weights are also adjusted based on the energy distribution characteristics of the system spectrum to emphasize the frequency range where energy is concentrated.

[0168] Then, the three-dimensional weight tensor is used to weight each component in the three-dimensional scale decomposition parameter component matrix, and the weighted results are fused through a nonlinear optimization function. The fusion process can be expressed as: G 3D (t)=∑ i ∑ j ∑ k w ijk (t)×G temporal [i]×G spatial [j]×G frequency [k]; where w ijk (t) is the adaptive weight at time t, G temporal [i],G spatial [j],G frequency [k] are the decomposed time, space and frequency components respectively.

[0169] Optionally, the nonlinear fusion process can also be implemented using a pre-trained small neural network to handle deeper nonlinear coupling relationships.

[0170] Finally, after generating the optimal MPPT parameter adjustment instructions, a parameter boundary constraint and safety check step is also included to ensure that the parameter change speed and value are within the hardware's safe tolerance range, avoiding system oscillation or instability, and ultimately obtaining MPPT parameter instructions that can be safely executed.

[0171] Example 8: This example takes a specific industrial rooftop distributed photovoltaic power station scenario as an example to fully illustrate the execution process of the method of the present invention.

[0172] Three PV strings (N=3) are connected to a string inverter. String 2 is shaded by a chimney, resulting in significantly lower power than the normal strings 1 and 3. String 3 (i=3) is operating normally, but is physically located farther from string 1 and adjacent to string 2.

[0173] Step S100: Generate a string difference gradient feature matrix. The system collects power data at three moments t=0, 1, and 2, where the power at t=1 is: P1=500W, P2=300W, and P3=490W.

[0174] Calculate the time gradient G time : At t=1, use the central difference method to calculate: G time [1]=12.5W / s, G time [2]=7.5W / s, G time [3]=12.5W / s.

[0175] Calculate the spatial gradient G space :Set the weight matrix W according to the physical proximity relationship. Calculate at time t=1: G space [1]=164, G space [2]=312, G space [3] = 148. The results clearly show that the spatial gradient of the shaded string 2 is a large positive value, indicating that its power is much lower than that of its neighbors.

[0176] Difference characterization: Comparison of cluster 1 (normal) and cluster 2 (occluded).

[0177] The gradient vector of string 1 is V_1 = [12.5, 164], and the gradient vector of string 2 is V_2 = [7.5, 312]. mag [1,2] is calculated as 147.8. Directional difference Diff angle [1,2] calculates to about 3.05 radians (close to 180 degrees).

[0178] The final feature matrix clearly shows that the gradient amplitude of string 2 is huge, and its gradient direction is almost completely opposite to that of the normal string, providing a strong criterion for subsequent scene recognition.

[0179] Step S200: scene gene library matching and strategy generation (see the above embodiment), assuming that two templates of uniform illumination change (scene A) and severe partial occlusion (scene B) are defined in the gene library.

[0180] The features currently calculated in real time (average spatial gradient amplitude of approximately 208, maximum directional difference of approximately 3.05) are highly matched with the feature vector of scene B [average spatial gradient > 100, maximum directional difference > 2.5].

[0181] The system calculated a similarity score of 0.95 with scene B, which is much higher than 0.1 of scene A, and the confidence level is extremely high.

[0182] The system determines that the current scene is severe local occlusion and retrieves the corresponding strategy from the gene library: for strings with large positive spatial gradients, deep voltage scanning is performed; for other normal strings, small step-size perturbations are maintained.

[0183] Generate a preliminary strategy: perform a deep scan on string 2 and maintain the original perturbation strategy for strings 1 and 3.

[0184] Step S300: 3D scale fusion optimization of time, space and frequency.

[0185] Decomposition: The preliminary strategy is decomposed into a local optimization component for string 2 and a system-level maintenance component for strings 1 and 3 in the spatial dimension.

[0186] Constructing adaptive weights: The system detects that the current core contradiction lies in the huge spatial differences, so the adaptive weight tensor will significantly increase the weight of the local optimization component in the spatial dimension. Assume that the weight w_local is set to 0.9.

[0187] Nonlinear fusion: Assuming the basic perturbation step size ΔV base =0.2V, initial step length of depth scan ΔV scan =5.0V. The final regulation command is generated by weighted fusion:

[0188] For strings 1 and 3, the command is close to ΔV_base, which is about 0.2 V. For string 2, the command is significantly biased towards deep scan: ΔV_final[2]=(10.9)×(0.2)+0.9×(5.0)=4.52 V.

[0189] The system outputs differentiated optimal MPPT parameter adjustment commands, sending a 0.2V voltage adjustment command to strings 1 and 3, and a 4.52V voltage adjustment command to string 2. This method successfully identifies partial shading scenarios and generates targeted adjustment commands, resolving the technical issue of power generation loss caused by traditional unified strategies.

[0190] In another embodiment of the present application, a method for adaptively optimizing the MPPT parameters of a string inverter is as follows:

[0191] Read multi-channel real-time data from string inverters, including parameters such as voltage, current, power, and temperature of each string. Through data cleaning and synchronization, a standardized multi-dimensional data matrix is ​​obtained.

[0192] Read the raw sensor data from each channel of the string inverter, including real-time measurement values ​​of each string's voltage, current, power, temperature, and ambient light intensity. Use the sensor calibration procedure to eliminate measurement deviations caused by equipment aging and environmental factors, and obtain a calibrated multi-channel raw data stream.

[0193] The calibrated multi-channel raw data stream is read, and statistical analysis methods are used to identify quality issues such as abnormal jumps, data missing, and random noise in the data. These abnormal data are repaired through intelligent interpolation and smoothing filtering technology to obtain a multi-channel data sequence of qualified quality.

[0194] Read the qualified multi-channel data sequence. Since the sampling frequencies of different sensors may be different, the timestamp alignment technology is used to unify all sensor data to the same time base. The precise time synchronization of data points is achieved through the intelligent interpolation method to obtain a time-synchronized multi-dimensional data matrix.

[0195] The time-synchronized multidimensional data matrix is ​​read. Since the numerical ranges of different physical quantities vary greatly, normalization processing is used to convert all data into a unified numerical range. At the same time, the data structure is reorganized to form a standard matrix format that is convenient for subsequent processing, thus obtaining a standardized multidimensional data matrix.

[0196] The standardized multi-dimensional data matrix is ​​read, and the improved string difference perception algorithm is used to calculate the three-dimensional gradient of light power and temperature of each string to obtain the string gradient feature matrix.

[0197] The standardized multidimensional data matrix is ​​read and the data subset corresponding to each string group is extracted from it. Through a precise time alignment algorithm, the data of each string group is ensured to be completely consistent in the time dimension, laying the foundation for subsequent difference analysis and obtaining the aligned string group data submatrix set.

[0198] Read the aligned string data sub-matrix set and use an innovative multi-dimensional gradient calculation method to analyze the change characteristics of each string:

[0199] The rate of change of the light intensity of each string is calculated to reflect the dynamic changes in the ambient light conditions. The rate of change of the power output of each string is calculated to reflect the real-time changes in the power generation efficiency. The rate of change of the temperature of each string is calculated to reflect the impact of thermal effects on the system. The difference weights are calculated based on the degree of difference in the power output of each string to highlight strings with abnormal performance. The change rate information is combined with the difference weights to form a weighted gradient matrix that can reflect the differences between strings. The string difference weighted gradient matrix is ​​obtained.

[0200] The weighted gradient matrix of string differences is read, and an intelligent noise reduction algorithm based on string correlation is used: the correlation of gradient changes between strings is analyzed to identify normal coordinated change patterns; abnormal strings that are obviously inconsistent with the change patterns of other strings are discovered, which may be caused by local occlusion or equipment failure; abnormal strings are corrected by the weighted average method of adjacent normal strings to maintain data continuity; key features are extracted from the corrected gradient data, including statistical features such as average change trend, change amplitude, change asymmetry and sharpness; and the denoised string gradient feature matrix is ​​obtained.

[0201] The denoised string gradient feature matrix is ​​read, and the standard normalization method is used to convert the gradient eigenvalues ​​into a standard distribution form to facilitate subsequent pattern matching and comparative analysis. At the same time, data compression technology is used to reduce storage space requirements to obtain a standardized string gradient feature matrix.

[0202] The string gradient feature matrix is ​​read, and scene recognition and parameter prediction are performed through the scene gene library similarity matching algorithm to obtain the scene matching results and preliminary parameter adjustment strategy.

[0203] The system reads a predefined multi-scenario simulation database, which contains the operating characteristics of photovoltaic systems in various typical application scenarios such as desert environments, industrial rooftops, and agricultural greenhouses. It uses cluster analysis technology to extract the most representative feature patterns from each scenario, establish a standard template library for scene recognition, and obtain a scene gene library feature template set.

[0204] Read the standardized string gradient feature matrix and scene gene library feature template set, and use an innovative multi-dimensional similarity calculation method to perform scene matching:

[0205] Calculate the similarity between the current gradient feature and each scene template in terms of numerical difference; calculate the similarity between the current gradient feature and each scene template in terms of change direction; calculate the similarity between the current gradient feature and each scene template in terms of time series dynamic characteristics; automatically adjust the importance weights of various similarity calculation methods according to the complexity of the current environment; intelligently fuse multiple similarity calculation results to obtain a comprehensive scene matching score; and obtain a multi-dimensional scene similarity score matrix.

[0206] The multi-dimensional scene similarity score matrix is ​​read and a confidence-based intelligent scene recognition algorithm is adopted: the scene with the highest similarity score is selected from all scene templates as the best matching scene; the reliability of the recognition result is evaluated by comparing the gap between the highest score and the second highest score; when the recognition confidence is high enough, the single scene matching result is confirmed; when the recognition confidence is low, it indicates that the current situation may be a mixture of multiple scenes, and the mixed scene processing mode is started; in the mixed scene mode, multiple high-scoring scenes are combined and processed according to the weights; the scene recognition result and confidence evaluation are obtained.

[0207] The scene recognition results and confidence assessment are read, and an intelligent parameter prediction method based on scene features is adopted: the MPPT parameter adjustment strategy corresponding to the identified scene is extracted from the scene gene library as the basic solution; the parameters are fine-tuned according to the degree of difference between the current actual gradient characteristics and the standard scene template; a predicted parameter adjustment solution suitable for the current specific situation is generated; considering the differences between each string in the string inverter, personalized parameter adjustments are made for each string; and a preliminary MPPT parameter adjustment strategy vector is obtained.

[0208] The scene matching results and preliminary parameter adjustment strategy are read, and the parameters are fine-tuned using the time-space-frequency three-dimensional scale decomposition and nonlinear fusion algorithm to obtain the optimal MPPT parameter adjustment instructions.

[0209] The preliminary MPPT parameter adjustment strategy vector is read, and the parameter adjustment strategy is refined using an innovative three-dimensional scale decomposition method: in the time dimension, the parameter adjustment is decomposed into a microsecond-level fast response component, a millisecond-level control loop response component, and a second-level environmental change response component; in the spatial dimension, the parameter adjustment is decomposed into a local optimization component at the single cell level, a coordinated optimization component at the component level, and a system optimization component at the string level; in the frequency dimension, the parameter adjustment is decomposed into a high-frequency transient response component, a medium-frequency dynamic adjustment component, and a low-frequency trend tracking component; accurate decomposition of each dimension is achieved through mathematical tools such as spectrum analysis and wavelet transform, and a three-dimensional scale decomposition parameter component matrix is ​​obtained.

[0210] The three-dimensional scale decomposition parameter component matrix is ​​read, and a weight adaptive algorithm based on the real-time status of the system is adopted: the importance weights of each component in the time dimension are dynamically adjusted according to the intensity of the illumination change. The faster the illumination changes, the greater the weight of the fast response component; the importance weights of each component in the spatial dimension are dynamically adjusted according to the degree of difference in the power output of each group of strings. The greater the difference, the greater the weight of the local optimization component; the importance weights of each component in the frequency dimension are dynamically adjusted according to the distribution characteristics of the system spectrum energy to highlight the frequency range where energy is concentrated; ensure the normalization of the weights of each dimension to maintain the stability of the system; construct a three-dimensional weight distribution to provide accurate weight guidance for subsequent fusion optimization; and obtain an adaptive three-dimensional weight tensor.

[0211] The three-dimensional scale decomposition parameter component matrix and the adaptive three-dimensional weight tensor are read, and a nonlinear fusion optimization algorithm is used: the parameter components obtained from the three-dimensional scale decomposition are weighted and fused according to the adaptive weights; the nonlinear optimization function is applied to the fusion results to avoid the performance limitations that may be caused by linear combinations; the parameters of the nonlinear function are automatically adjusted according to the dynamic response characteristics of the system; second-order compensation based on change trend prediction is added to improve the foresight of parameter adjustment; the final parameters are ensured to be within the safe operation range of the system; and the optimal MPPT parameter adjustment instructions are obtained.

[0212] Read the optimal MPPT parameter adjustment instructions and perform a comprehensive system safety check: check whether the parameter changes too quickly. If it exceeds the safe range, adopt a gradual adjustment method; verify through system stability theory that parameter adjustment will not cause system oscillation or instability; confirm that the parameter adjustment instructions will not exceed the inverter hardware's tolerance and design specifications; generate a complete execution instruction including parameter values, execution schedule and safety status identification; and obtain a safety-verified MPPT parameter execution instruction.

[0213] Example 9: According to one aspect of the present application, the detailed construction of the scenario gene library and the template generation process (usually completed offline before method deployment) are as follows:

[0214] Step S501: construct a simulation database of multiple working scenarios.

[0215] In this embodiment, to ensure the wide applicability of the gene library, it is first necessary to generate a large amount of photovoltaic system operation data covering various typical working conditions through simulation.

[0216] Specifically, using a MATLAB / Simulink-based simulation platform, the electrical characteristics of PV strings were accurately modeled using a dual-diode model. The environmental data used as input for the simulations was partially derived from the TMY3 (Typical Meteorological Year) dataset provided by the National Renewable Energy Laboratory (NREL), which contains hourly meteorological parameters for various geographic locations. To simulate non-uniform operating conditions, a dynamic shadowing model was incorporated into the simulation. For example, by modeling moving clouds (characterized by cloud masses of varying size, speed, and transparency) and static local shadowing (e.g., fixed shadows cast by chimneys, parapets, and vegetation), a variety of complex spatiotemporal irradiance distributions were generated. For each scenario (e.g., "desert midday, high-speed clouds passing" or "industrial rooftop, long chimney shadows in the afternoon"), dynamic simulations were conducted for at least eight hours, with voltage, current, power, temperature, and irradiance data recorded at a sub-second rate or higher for each string.

[0217] Step S502: batch generate cluster difference gradient feature matrices.

[0218] The original time series data of each scene generated by simulation in step S501 is used as input, and the method of embodiment 5 is applied to batch calculate and generate the corresponding time series of the group difference gradient feature matrix M.

[0219] Step S503: Feature extraction and dimensionality reduction. Directly performing cluster analysis on high-dimensional feature matrix time series is computationally expensive and ineffective. Therefore, it is necessary to extract a set of low-dimensional feature vectors from the feature matrix sequence of each scene that better characterizes its core characteristics. Specifically, for each scene's feature matrix sequence, the following statistics are calculated as its core features:

[0220] The mean, variance, and kurtosis of the spatial gradient amplitude. The mean and variance of the temporal gradient amplitude. The maximum and mean of the directional differences between strings. The sparsity and energy concentration of the feature matrix.

[0221] Through the above calculations, each simulation scenario lasting several hours is mapped into a feature vector containing 10-20 key features.

[0222] Step S504: performing cluster analysis to classify scenes.

[0223] The feature vector set of all simulation scenes is used as input, and cluster analysis is performed to classify scenes with similar features into one category. In this embodiment, DBSCAN (density-based spatial clustering algorithm) is preferably used. Compared with the K-Means algorithm that requires a preset number of categories, DBSCAN can automatically discover clusters of arbitrary shapes based on the compactness of the data distribution, and can effectively identify outliers (i.e., atypical scenes that cannot be classified). After processing by the DBSCAN algorithm, the feature vectors are divided into multiple clusters, each of which represents a typical operating scenario, such as "rapid and uniform change scenario", "continuous occlusion scenario of a single group of strings on the left", "intermittent occlusion scenario of multiple groups of strings", etc.

[0224] Step S505: Generate scene templates and associated basic adjustment strategies. Perform template processing on each divided scene cluster.

[0225] Generate scene template: Calculate the centroid of all feature vectors in the cluster (i.e., the average value of each feature dimension). This centroid vector is defined as the standard feature template for the scene category.

[0226] Associated Basic Regulation Strategy: Analyze all simulation data within the cluster and identify the optimal parameter combination by comparing the power generation efficiency under the operating conditions using different MPPT parameters (such as different perturbation step sizes, sweep cycles, and sweep range combinations). This optimal parameter combination, or its corresponding regulation rule, is defined as the basic regulation strategy associated with the scenario template.

[0227] Through the above steps, a scenario gene library containing multiple "scenario template-basic adjustment strategy" pairs is constructed. This gene library provides a solid foundation for subsequent real-time online matching and predictive adjustment.

[0228] Example 10: According to one aspect of the present application, after the scene matching is successful, the basic adjustment strategy retrieved from the gene library is overall modified according to the deviation between the real-time feature and the template feature, specifically:

[0229] After the system was online, it successfully matched the current working conditions to scenario B: severe partial occlusion, and retrieved the corresponding basic adjustment strategy from the gene library.

[0230] Using the scenario from Example 8, the current real-time calculated feature vector is F_realtime = [average spatial gradient amplitude: 208, maximum directional difference: 3.05]. The standard feature template for scene B in the scene gene library is F_template_B = [180, 2.8]. The basic adjustment strategy is: "For obscured strings, use a deep scan initial step size of -5.0V; for normal strings, use a regular perturbation step size of -0.2V."

[0231] Step S601: quantify the deviation between the real-time feature and the template feature.

[0232] Calculate the deviation vector ΔF between the real-time feature vector F_realtime and the template feature vector F_template_B. ΔF = F_realtime - F_template_B = [208 - 180, 3.05 - 2.8] = [28, 0.25]. This deviation vector ΔF intuitively indicates that the "degree of difference" (spatial gradient amplitude) of the current actual working condition is 28 units higher than the average situation defined in the standard template, and the "directional opposition" of the difference is also 0.25 units higher than the standard template. This indicates that the current occlusion situation may be more severe than the average situation represented by the scene template.

[0233] Step S602: defining a mapping function from deviation to correction amount.

[0234] One or more mapping functions are predefined to convert the quantized deviation ΔF into a correction for the key parameters in the basic adjustment strategy. In this embodiment, we correct the key parameter "initial step size of depth scan" ΔV_scan. The correction function f_correct(ΔF) is defined as: fcorrect(ΔF)=c1·ΔF[1]+c2·ΔF[2], where ΔF[1] is the deviation of the spatial gradient amplitude and ΔF[2] is the deviation of the directional difference. c_1 and c_2 are preset correction coefficients, for example, c_1=-0.01V / unit gradient and c_2=-0.5V / unit radian. The negative signs of these two coefficients indicate that the larger the deviation, the more severe the occlusion, and the larger the required scanning step size (the more negative it is).

[0235] Step S603: perform correction and generate a corrected policy.

[0236] Apply the mapping function from step S602 to calculate the specific correction for ΔV_scan. Correction = f_correct([28, 0.25]) = (-0.01)*28 + (-0.5)*0.25 = -0.28 - 0.125 = -0.405V. This correction is applied to the parameters of the basic regulation strategy: Corrected scan step ΔV'_scan = ΔV_scan + Correction = -5.0 + (-0.405) = -5.405V. The normal perturbation step for normal strings can be set to be unaffected by this deviation, or to have minimal weighting. Ultimately, the system generates a comprehensive, preliminary MPPT parameter regulation strategy: "For the shaded string (string 2), use a deep scan initial step of -5.405V; for the normal strings (strings 1 and 3), maintain a normal perturbation step of -0.2V."

[0237] Through this embodiment, the system not only calls the preset strategy, but also "quantitatively fine-tunes" the strategy according to the severity of the current working conditions, making the preliminary strategy more targeted and adaptable.

[0238] Example 11 describes the design and implementation of a nonlinear optimization function for optimal instructions.

[0239] According to one aspect of the present application, a preferred implementation of the "nonlinear optimization function" as the last step of the time-space-frequency three-dimensional optimization is to use a pre-trained neural network model.

[0240] The system has decomposed the revised preliminary adjustment strategy into multiple components in three dimensions: time, space, and frequency through the method of Example 7, and calculated the corresponding adaptive three-dimensional weight tensor based on the real-time status.

[0241] The implementation of this nonlinear optimization function includes two stages: offline training and online inference.

[0242] Phase 1: Offline training of the neural network.

[0243] Network structure design:

[0244] Input layer: The nodes of the input layer correspond to the weighted main components. For example, it can be designed as 9 input nodes, receiving:

[0245] Weighted time components (fast, medium, and slow responses);

[0246] weighted spatial components (local, regional, system-level responses);

[0247] Weighted frequency components (high frequency, mid frequency, and low frequency responses);

[0248] Hidden layers: Two fully connected hidden layers were designed. The first layer contained 16 neurons, and the second layer contained 8 neurons. All hidden layers used ReLU (rectified linear unit) as the activation function to introduce nonlinearity and accelerate training.

[0249] Output layer: The nodes in the output layer correspond to the MPPT parameters that need to be controlled. In this embodiment, there are two output nodes, representing:

[0250] Optimal voltage regulation step ΔV_opt;

[0251] The optimal perturbation period T_opt output layer adopts a linear activation function.

[0252] Training data set construction: The training data comes from the multi-scenario simulation database constructed in Example 9. Every second of data in the database is processed as follows:

[0253] Input label: Perform a full 3D decomposition and weighting to obtain a 9-dimensional input vector at that moment.

[0254] Output label (GroundTruth): Through backtracking analysis, we calculate which [ΔV, T] combination at that moment can achieve the power output closest to the theoretical maximum power point within the next short time window (such as 5 seconds). This [ΔV, T] combination that produces the best post-hoc result is defined as the "ground truth" output label for the training example.

[0255] Model training: The constructed "input-output" dataset pairs are used to train the designed neural network.

[0256] Loss function: Mean squared error (MSE) is used as the loss function to measure the gap between the network prediction value and the true value label.

[0257] Optimizer: Adam optimizer is used to adaptively adjust the learning rate.

[0258] Training process: The network weights and biases are continuously adjusted through the backpropagation algorithm until the loss function converges to a sufficiently small value. The trained model is then solidified and deployed to the inverter control system.

[0259] Phase 2: Online reasoning and instruction generation.

[0260] When the inverter is actually running, the trained neural network model is called as a "nonlinear optimization function".

[0261] The system calculates the weighted 9-dimensional input vector in real time.

[0262] This vector is fed into the neural network model for a forward propagation calculation, which typically completes within milliseconds through matrix multiplication and activation function calculations.

[0263] The network output layer directly gives the current optimal MPPT parameters [ΔV_opt, T_opt].

[0264] After the final safety check, the parameter forms the optimal MPPT parameter adjustment instruction and is issued for execution.

[0265] Through this embodiment, this pre-trained neural network can learn and fit the extremely complex nonlinear coupling relationship between components of different dimensions. Its output result is more accurate and optimized than simple linear weighted fusion, and is the key link to achieve ultimate performance.

[0266] The preferred embodiments of the present invention are described in detail above. However, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all fall within the scope of protection of the present invention.

Claims

1. A method for adaptively optimizing MPPT parameters of a string inverter, characterized in that: include: Analyze the collected multi-channel real-time data of the string inverter to generate a string difference gradient feature matrix that characterizes the relative differences between the operating states of each photovoltaic string; Based on the cluster difference gradient feature matrix, search and match in the preset scene gene library to generate a preliminary MPPT parameter adjustment strategy; The preliminary MPPT parameter adjustment strategy is subjected to coordinated integration and optimization of time, space and frequency in three dimensions to obtain the optimal MPPT parameter adjustment instructions; The preliminary MPPT parameter adjustment strategy is coordinated and optimized in three dimensions: time, space, and frequency, including: The preliminary MPPT parameter adjustment strategy is deconstructed in three dimensions: In the time dimension, it is decomposed into multiple time scale components corresponding to fast transient response, medium-speed dynamic regulation and low-speed trend tracking; In the spatial dimension, according to the physical topology of the photovoltaic array, it is decomposed into spatial scale components corresponding to the local, regional and system levels respectively; In the frequency dimension, it is decomposed into frequency scale components to characterize different dynamic characteristics; Combine the time, space and frequency scale components to generate a three-dimensional scale decomposition parameter component matrix; The step of generating a cluster difference gradient feature matrix further includes: In the time dimension, the time change rate of each group of string data sequences is calculated through the difference algorithm to form the time dimension gradient; In the spatial dimension, the string similarity weight matrix representing the physical and electrical characteristics between strings is calculated, as well as the state difference value between each string and its neighbors. The state difference value and the corresponding weight in the string similarity weight matrix are weighted and summed to form a spatial dimension gradient. The time dimension gradient and the space dimension gradient are integrated and characterized to form a cluster difference gradient feature matrix.

2. The method according to claim 1, characterized in that The preliminary MPPT parameter adjustment strategy is deconstructed in three dimensions, including: For the time dimension, empirical mode decomposition is used to separate multiple time scale components; For spatial dimensions, multiple spatial scale components are analyzed through graph Fourier transform based on physical topology; In the frequency dimension, variational mode decomposition is applied to finely separate multiple frequency scale components.

3. The method according to claim 1, characterized in that After generating the three-dimensional scale decomposition parameter component matrix, collaborative fusion and optimization also includes: Construct an adaptive three-dimensional weight tensor based on the intensity of light changes, the difference in power output between strings, and the energy distribution characteristics of the system spectrum; The three-dimensional weight tensor is used to weight each component in the three-dimensional scale decomposition parameter component matrix, and the weighted results are fused through a nonlinear optimization function to generate the optimal MPPT parameter adjustment instructions.

4. The method according to claim 1, wherein Based on the gradient feature matrix of group differences, search and match are performed in the scene gene library, including: Using a multi-dimensional similarity matching algorithm, the cluster difference gradient feature matrix is ​​compared with the scene templates in the scene gene library to calculate the numerical similarity between the two at the basic statistical feature level, the structural similarity at the dynamic pattern feature level, and the relationship similarity at the causal correlation feature level. Numerical similarity, structural similarity and relational similarity are integrated to form a multi-dimensional scene similarity scoring matrix.

5. The method according to claim 4, characterized in that After constructing the multi-dimensional scene similarity scoring matrix, the retrieval and matching steps also include: Analyze the multi-dimensional scene similarity score matrix and select the scene with the highest score as the candidate; Calculate the difference between the highest score of the candidate scene and the highest score of all other scene templates to evaluate the reliability of the match and generate a confidence assessment; Based on the confidence assessment, when the reliability is higher than the preset threshold, a single scene match is determined; when the reliability is lower than the preset threshold, the mixed scene mode is activated to perform a weighted combination of multiple high-scoring scenes; A scene recognition result is formed containing single scene or weighted combination scene information.

6. The method according to claim 5, characterized in that After the scene recognition results are formed, a preliminary MPPT parameter adjustment strategy is generated, including: Retrieve the basic regulation strategy corresponding to the scene recognition result from the scene gene library; According to the deviation between the current string difference gradient feature matrix and the scene template standard feature, the basic adjustment strategy is overall modified; The overall revised strategy is then personalized adapted to the gradient characteristics of each string to construct a preliminary MPPT parameter adjustment strategy.

7. The method according to claim 1, characterized in that Integrate the time dimension gradient and the space dimension gradient and characterize them, including: The temporal gradient and spatial gradient of each string are combined to form their own gradient vector; By quantifying the amplitude difference and direction difference between the gradient vectors of any two strings in parallel, the difference features between the strings are extracted; A nonlinear weighted algorithm is used to fuse the amplitude difference and direction difference to construct the gradient feature matrix of group difference.

8. The method according to claim 1, characterized in that Before integrating and characterizing the temporal dimension gradient and the spatial dimension gradient, noise reduction processing is also included, specifically: Analyze the gradient changes of all strings and obtain the benchmark coordinated change pattern; Screen out abnormal clusters whose gradient dynamics are contrary to the baseline co-variation pattern; With the help of the gradient information of the normal strings adjacent to the abnormal strings, the gradient is compensated and corrected to obtain the gradient data after noise reduction.

Images