MPPT parameter adaptive optimization method based on distributed control

Abstract

The application discloses an MPPT parameter adaptive optimization method based on distributed control, which comprises the following steps: collecting data to generate component-level state sensing results; inputting the component-level state sensing results into a photovoltaic shading sensing type Mask2Former to generate shading quantization results; judging the power-voltage curve state according to the shading quantization results, the component power change state and the historical maximum power point tracking results, and generating an adaptive MPPT parameter group; performing optimization through chaos initialization, an improved particle swarm optimization algorithm and an improved slime mold algorithm; adjusting a DC-DC Buck circuit and distributing MPPT control channels and power station-level power instructions; and generating hot spot effect early warning information, fault positioning information and an adaptive MPPT parameter group for the next control cycle based on operation feedback data. The application improves the stability, optimization accuracy and distributed regulation efficiency of maximum power point tracking under complex shading.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic power generation control technology, and in particular to an adaptive optimization method for MPPT parameters based on distributed control. Background Technology

[0002] With the rapid construction of distributed photovoltaic power stations and mountain photovoltaic projects, MPPT parameter adaptive optimization technology for addressing local shading, string mismatch, and multi-peak power curves has received widespread attention. Existing photovoltaic power optimization methods mainly rely on module-level operational data acquisition, traditional perturbation observation methods, incremental conductance methods, or single-group intelligent algorithms for maximum power point tracking. Some solutions combine image recognition or operation and maintenance monitoring to achieve shading detection and fault alarms. However, in practical applications, the following problems are commonly encountered: Local shading, tree occlusion, building occlusion, cloud occlusion, and module self-shading can cause photovoltaic (PV) module power-voltage curves to exhibit a multi-peak pattern. Traditional MPPT methods tend to get stuck at local maximum power points, and fixed search range, search step size, and algorithm parameters are difficult to adapt to dynamic changes in occlusion area and rate of change. Existing occlusion recognition methods mostly remain at the level of target detection or ordinary semantic segmentation, making it difficult to bind the occluded area with PV module number, string affiliation, module operation drop characteristics, and MPPT sensitivity. As a result, occlusion detection results cannot be directly converted into search range, inertia weight, learning factor, and Gaussian mutation parameters. Adaptive MPPT parameters are needed; existing particle swarm optimization algorithms or slime mold algorithms are usually used alone for global optimization, lacking parameter adjustment based on shading level, chaotic initialization, Gaussian mutation, arithmetic optimization operators, and candidate power point fusion evaluation mechanisms. Under complex multi-peak power curves, they suffer from slow convergence speed, insufficient ability to escape local optima, and low control stability. At the same time, there is a lack of closed-loop linkage between existing component-level power regulation, MPPT control channel allocation, hot spot effect early warning, and fault location. Operational feedback is difficult to correct the adaptive MPPT parameter set for the next control cycle in a timely manner, affecting the power generation efficiency and operational safety of distributed photovoltaic systems.

[0003] Therefore, how to provide an adaptive optimization method for MPPT parameters based on distributed control is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] One objective of this invention is to propose an adaptive optimization method for MPPT parameters based on distributed control. This invention fully utilizes the photovoltaic shading-sensing Mask2Former, the improved particle swarm optimization algorithm, and the improved slime mold algorithm, and describes in detail the process of component-level state sensing, adaptive optimization of MPPT parameters, power plant-level power allocation, and closed-loop optimization of operation feedback under complex shading. It has the advantages of high tracking accuracy, strong shading adaptability, and high distributed control efficiency.

[0005] An adaptive optimization method for MPPT parameters based on distributed control according to an embodiment of the present invention includes the following steps: Step 1: Collect module-level operating data, environmental status data, and shading image data according to the photovoltaic module number and string number, and generate module-level status perception results; Step 2: Input the component-level state perception results into the photovoltaic shading perception Mask2Former, and generate shading quantification results through component layout prior embedding, multi-source state condition guidance, shading mask segmentation and time-series shading evolution prediction. Step 3: Based on the shading quantification results, the power change status of the module, and the historical maximum power point tracking results, determine the current power-voltage curve status of the photovoltaic module, and generate an adaptive MPPT parameter set when entering the multi-peak power curve state; Step 4: Based on the adaptive MPPT parameter set, the candidate operating parameters are chaotically initialized. An improved particle swarm optimization algorithm is used to adjust the inertia weight and learning factor according to the occlusion level to update the candidate operating parameters. An improved slime mold algorithm is used to perturb the candidate operating parameters by combining Gaussian mutation and arithmetic optimization operator. The optimization results are fused to generate the target output voltage, target output current and target duty cycle corresponding to the global maximum power point. Step 5: Adjust the DC-DC Buck circuit based on the target output voltage, target output current, and target duty cycle to generate component-level power regulation results; Step 6: Based on the module-level power regulation results, shading quantification results, and photovoltaic unit adjustable capacity status, allocate the MPPT control channel and power plant-level power commands to generate power allocation results; Step 7: Generate operational feedback data based on component-level power regulation and power allocation results, and generate hot spot effect early warning information, fault location information, and adaptive MPPT parameter set for the next control cycle based on the operational feedback data.

[0006] Optionally, step one specifically includes: Based on the physical layout diagram of photovoltaic modules and the string wiring relationship, establish the binding relationship between photovoltaic module number, string number, photovoltaic optimizer number and distributed control node address; The centralized monitoring platform receives component-level operating data, environmental status data, and occlusion image data uploaded by each distributed control node according to a unified sampling time and sampling period. The component-level operating data includes component voltage, component current, output power, component temperature, DC-DC duty cycle status, wiring status, and arc fault status. Each distributed control node encapsulates component-level operational data, environmental status data, and occlusion image data into a component status data frame. The component status data frame includes the photovoltaic module number, string number, distributed control node address, sampling time, and data source marker. The centralized monitoring platform performs time alignment on the component status data frames based on the sampling time, and merges the time-aligned component status data frames according to the photovoltaic component number and string number to generate component-level status perception results.

[0007] Optionally, step two specifically includes: The photovoltaic shading-aware Mask2Former consists of a backbone network, pixel decoders, Transformer decoders, and an output head. It also includes a component layout prior embedding module, a multi-source state condition guidance module, a shading mask segmentation module, an MPPT sensitive weight generation branch, and a temporal shading evolution prediction module. The component layout prior embedding module encodes the physical layout topology of the photovoltaic power station into a layout prior map. The physical layout topology of the photovoltaic power station includes the component spatial boundary, installation posture and string affiliation relationship. On the input side of the pixel decoder, the layout prior map is fused with the multi-scale image features output by the backbone network through channel fusion, and pixel embedding features with component boundary constraints and string affiliation constraints are generated by point-by-point convolution. The multi-source state condition guidance module generates a state condition vector based on the operational drop characteristics, component temperature change, and environmental disturbance characteristics. The operational drop characteristics consist of the component current drop ratio, component voltage drop ratio, and output power drop ratio. The environmental disturbance characteristics consist of irradiance change, wind speed change, wind direction change, and humidity change. The state condition vector generates scaling and offset coefficients through the Transformer decoder, modulates the mask query features associated with the corresponding photovoltaic module channel by channel, and converts them into condition key features and condition value features to participate in cross-attention calculation, generating state constraint mask features. The occlusion mask segmentation module generates occlusion type and occlusion area mask based on state constraint mask features. The occlusion area mask is a binary pixel map composed of occluded pixels and unoccluded pixels. The MPPT sensitive weight generation branch extracts the feature vector of the shading area based on the shading area mask, and inputs the shading area feature vector, category embedding vector, location encoding, operational drop features and component temperature change into the weight generation network to generate MPPT sensitive weights. The category embedding vector is obtained by mapping the shading object type through an embedding table, and the location encoding is obtained by encoding the center coordinates, boundary coordinates and region number of the shading area mask in the effective power generation area of ​​the component. The weighted shading area ratio is generated based on the shading area mask, MPPT sensitivity weight, and effective power generation area of ​​the module. The time-series shading evolution prediction module generates a predicted shading mask based on the shading area mask, shading movement characteristics, and meteorological embedding vector. The meteorological embedding vector is obtained by encoding wind speed, wind direction, and irradiance changes, and outputs the shading area change rate, shading movement direction, shading level of the next control cycle, and risk coefficient of multi-peak power curve, generating shading quantification results.

[0008] Optionally, step three specifically includes: The current power-voltage curve status of the photovoltaic module is determined based on the weighted shading area ratio, shading area change rate, multi-peak power curve risk coefficient, shading level of the next control cycle, module power change status, and historical maximum power point tracking results. The benchmark MPPT parameters are formed based on historical stable operating cycles. The benchmark MPPT parameters include benchmark search range, benchmark search step size, benchmark particle swarm inertia weight, benchmark particle swarm learning factor, benchmark Gaussian mutation parameter, benchmark maximum number of iterations, and benchmark convergence judgment threshold. A parameter mapping table is formed based on historical calibration samples. The historical calibration samples include occlusion quantization results, component power change status, historical maximum power point tracking results, and actual output power change records. The parameter mapping table records the correspondence between occlusion area level and search range expansion coefficient, occlusion change level and search step size adjustment coefficient, multi-peak risk level and particle swarm inertia weight adjustment coefficient, particle swarm learning factor adjustment coefficient, and Gaussian mutation adjustment coefficient. The occlusion area level is determined based on the weighted occlusion area ratio, and the baseline search range is adjusted according to the search range expansion coefficient to generate the MPPT search range; The occlusion change level is determined based on the occlusion area change rate, and the baseline search step size is adjusted according to the search step size adjustment coefficient to generate the search step size; The risk level of the multi-peak power curve is determined based on the risk coefficient of the multi-peak power curve, and the baseline particle swarm inertia weight, baseline particle swarm learning factor and baseline Gaussian mutation parameter are adjusted according to the corresponding adjustment coefficient to generate the particle swarm inertia weight, particle swarm learning factor and Gaussian mutation parameter. The maximum number of iterations and the convergence threshold are adjusted according to the occlusion level of the next control cycle. The DC-DC duty cycle boundary is determined based on the historical maximum power point tracking results. The MPPT search range, search step size, particle swarm inertia weight, particle swarm learning factor, Gaussian mutation parameter, maximum number of iterations, convergence threshold and DC-DC duty cycle boundary are bound together to generate an adaptive MPPT parameter set.

[0009] Optionally, step four specifically includes: Candidate operating parameter boundaries are established based on the MPPT search range and DC-DC duty cycle boundary. Candidate output voltage, candidate output current, and candidate duty cycle are used as candidate operating parameters. Logistic chaotic mapping is used to generate an initial set of candidate operating parameters. The improved particle swarm optimization algorithm uses candidate operating parameters as particle positions and component output power as fitness values. It updates the candidate operating parameters according to the particle swarm inertia weight and particle swarm learning factor corresponding to the occlusion level in the next control cycle, and generates a set of candidate operating parameters for particle swarm. The improved slime mold algorithm generates slime mold weight coefficients based on the power ranking results, and perturbs the candidate running parameters by combining Gaussian mutation parameters and arithmetic optimization operator parameters to generate a set of slime mold candidate running parameters; The candidate operating parameter sets for particle swarm optimization and slime mold are merged, and candidate operating parameters that exceed the DC-DC duty cycle boundary are deleted to generate a fused candidate operating parameter set. An occlusion reduction coefficient is generated based on the historical maximum power point power, weighted occlusion area ratio, and MPPT sensitivity weight in the historical maximum power point tracking results. The occlusion reduction coefficient is obtained by multiplying the weighted occlusion area ratio and MPPT sensitivity weight and limiting it to between zero and one. The historical maximum power point power is then reduced according to the occlusion reduction coefficient to generate the occlusion correction target power. For candidate power points in the fusion candidate operating parameter set, a power evaluation term, a convergence evaluation term, a stability evaluation term, and a deviation evaluation term are generated respectively. The power evaluation term is obtained by normalizing the output power of the candidate component. The convergence evaluation term is obtained by inverse normalizing the number of convergence iterations. The stability evaluation term is obtained by normalizing the continuous stable time. The deviation evaluation term is obtained by inverse normalizing the target power deviation. The target power deviation is the absolute value of the difference between the output power of the candidate component and the occlusion correction target power. The power evaluation item, convergence evaluation item, stability evaluation item, and deviation evaluation item are multiplied by their respective weights and then summed to generate a fusion evaluation value for candidate operating parameters. The candidate operating parameter with the largest fusion evaluation value is selected as the target operating parameter corresponding to the global maximum power point, and the target output voltage, target output current, and target duty cycle are generated.

[0010] Optionally, step five specifically includes: Voltage deviation and current deviation are generated based on the target output voltage, target output current, actual output voltage, and actual output current; Multiply the voltage deviation by the voltage loop adjustment coefficient, multiply the current deviation by the current loop adjustment coefficient, and add the two together to generate the duty cycle adjustment amount; The target duty cycle is superimposed with the duty cycle adjustment amount, and the execution duty cycle is generated by limiting it according to the DC-DC duty cycle boundary. The PWM drive signal is generated based on the duty cycle to adjust the on-time of the switching devices in the DC-DC Buck circuit. The target output voltage is multiplied by the target output current to obtain the target power. When the power deviation between the actual output power and the target power is not greater than the convergence judgment threshold, the module-level power regulation result is generated according to the photovoltaic module number and string number.

[0011] Optionally, step six specifically includes: MPPT control channels and power plant-level power commands are allocated based on the component-level power regulation results, shading quantification results, and the adjustable capacity status of the photovoltaic unit. The adjustable capacity status of the photovoltaic unit includes the current output power, adjustable power upper limit, adjustable power lower limit, and power regulation direction. When the two photovoltaic strings have the same shading level in the next control cycle, the difference in the weighted shading area ratio does not exceed the preset shading difference threshold, and the difference in the risk coefficient of the multi-peak power curve does not exceed the preset risk difference threshold, the two photovoltaic strings will be assigned to the same MPPT control channel; otherwise, they will be assigned to different MPPT control channels. The adjustable capacity is adjusted upwards or downwards according to the direction of the power station-level power command, and the power station-level power command is allocated according to the proportion of the adjustable capacity of each photovoltaic unit to the total adjustable capacity, thus generating the power allocation result.

[0012] Optionally, step seven specifically includes: Based on the component-level power regulation and power allocation results, operational feedback data is generated, and the abnormal status of the component is determined based on the actual output power drop, power deviation duration, number of consecutive changes in duty cycle, component temperature change, wiring status, and arc fault status. When the component temperature change reaches the preset temperature rise threshold, or the actual output power decrease reaches the preset power decrease threshold and the power deviation duration reaches the preset duration threshold, a hot spot effect warning message is generated. When the wiring status is disconnected or the arc fault status is triggered, fault location information is generated. The adaptive MPPT parameter set for the next control cycle is updated based on the operation feedback data, hot spot effect early warning information and fault location information. The hot spot effect early warning information corresponds to tightening the DC-DC duty cycle boundary and reducing the target power adjustment amplitude. The continuous power deviation exceeding the limit corresponds to expanding the MPPT search range and increasing the search step size. The fault location information corresponds to reducing the power allocation ratio of the photovoltaic unit to which the faulty component belongs and triggering the allocation of the backup MPPT control channel.

[0013] The beneficial effects of this invention are: This invention addresses the problem that ordinary shading identification cannot directly serve MPPT control by embedding prior information about component layout in the photovoltaic shading-sensing Mask2Former, guiding multi-source state conditions, segmenting shading masks, and predicting time-series shading evolution. It binds the shading area with the photovoltaic component number, string affiliation, operational drop characteristics, and component temperature changes, enabling shading quantification results to generate a weighted shading area ratio, shading level for the next control cycle, and risk coefficient for the multi-peak power curve. A parameter mapping table maps the weighted shading area ratio, shading area change rate, and multi-peak power curve risk coefficient to the MPPT search range, search step size, particle swarm inertia weight, particle swarm learning factor, and Gaussian mutation parameter, respectively, solving the problem that fixed MPPT parameters cannot adapt to dynamic shading and improving tracking adaptability under multi-peak power curves. Through Logistic chaotic initialization, an improved particle swarm optimization algorithm, an improved slime mold algorithm, and a candidate power point fusion evaluation mechanism, it addresses the problem of single algorithms easily getting trapped in local optima and insufficient convergence stability, obtaining the target output voltage, target output current, and target duty cycle corresponding to the global maximum power point. Finally, it utilizes DC-DC... Buck circuit regulation, MPPT control channel allocation, hot spot effect early warning, and operation feedback update of the adaptive MPPT parameter group for the next control cycle form a closed-loop control from shading perception to power regulation and then to feedback correction. This is of great significance for improving the power generation efficiency of distributed photovoltaic power plants under complex shading conditions, reducing hot spot risk, and enhancing the safety of string-level operation. Attached Figure Description

[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an MPPT parameter adaptive optimization method based on distributed control proposed in this invention; Figure 2 This is a schematic diagram of an MPPT parameter adaptive optimization method based on distributed control proposed in this invention; Figure 3 This is a framework diagram of the photovoltaic shading sensing Mask2Former in the MPPT parameter adaptive optimization method based on distributed control proposed in this invention. Detailed Implementation

[0015] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0016] refer to Figures 1-3An adaptive optimization method for MPPT parameters based on distributed control includes the following steps: Step 1: Collect module-level operating data, environmental status data, and shading image data according to the photovoltaic module number and string number, and generate module-level status perception results; Step 2: Input the component-level state perception results into the photovoltaic shading perception Mask2Former, and generate shading quantification results through component layout prior embedding, multi-source state condition guidance, shading mask segmentation and time-series shading evolution prediction. Step 3: Based on the shading quantification results, the power change status of the module, and the historical maximum power point tracking results, determine the current power-voltage curve status of the photovoltaic module, and generate an adaptive MPPT parameter set when entering the multi-peak power curve state; Step 4: Based on the adaptive MPPT parameter set, the candidate operating parameters are chaotically initialized. An improved particle swarm optimization algorithm is used to adjust the inertia weight and learning factor according to the occlusion level to update the candidate operating parameters. An improved slime mold algorithm is used to perturb the candidate operating parameters by combining Gaussian mutation and arithmetic optimization operator. The optimization results are fused to generate the target output voltage, target output current and target duty cycle corresponding to the global maximum power point. Step 5: Adjust the DC-DC Buck circuit based on the target output voltage, target output current, and target duty cycle to generate component-level power regulation results; Step 6: Based on the module-level power regulation results, shading quantification results, and photovoltaic unit adjustable capacity status, allocate the MPPT control channel and power plant-level power commands to generate power allocation results; Step 7: Generate operational feedback data based on component-level power regulation and power allocation results, and generate hot spot effect early warning information, fault location information, and adaptive MPPT parameter set for the next control cycle based on the operational feedback data.

[0017] In this embodiment, step one specifically includes: Read the physical layout diagram of photovoltaic modules and the string wiring relationship. The physical layout diagram of photovoltaic modules is a record of module location established based on the power plant construction drawing, module installation coordinates and on-site number. Extract the photovoltaic module number, string number, module installation coordinates, photovoltaic optimizer number and distributed control node address, and establish the binding relationship between photovoltaic module number, string number, photovoltaic optimizer number and distributed control node address. According to the binding relationship, a unified sampling time and sampling period are issued to each distributed control node. Each distributed control node independently collects module-level operating data on the corresponding photovoltaic module side. The module-level operating data includes module voltage, module current, output power, module temperature, DC-DC duty cycle status, wiring status, and arc fault status. Environmental status data is collected according to the photovoltaic module installation area. The environmental status data includes light intensity, irradiance, wind speed, wind direction, humidity and weather forecast information. The environmental status data is matched to the corresponding photovoltaic module number and distributed control node address according to the module installation coordinates. Shading image data is collected according to the component installation coordinates. The shading image data consists of image frames covering the surface of the photovoltaic module and the surrounding shading area. The shooting time, shooting angle and image coverage area in the shading image data are matched with the photovoltaic module number. Each distributed control node encapsulates component-level operational data, environmental status data, and shading image data into a component status data frame. The component status data frame includes the photovoltaic module number, string number, distributed control node address, sampling time, and data source marker. The centralized monitoring platform receives status data frames from each component, performs time alignment based on the sampling time, merges data based on photovoltaic module number and string number, and generates component-level status perception results.

[0018] In this embodiment, step two specifically includes: The photovoltaic shading-aware Mask2Former consists of a backbone network, a pixel decoder, a Transformer decoder, and an output head. A component layout prior embedding module is set on the input side of the pixel decoder, a multi-source state condition guidance module is set in the Transformer decoder, and an shading mask segmentation module and an MPPT sensitive weight generation branch are set in the output head. The shading mask segmentation module includes a category prediction branch and a mask prediction branch. A temporal shading evolution prediction module is connected in series after the output head. The temporal shading evolution prediction module includes a ConvLSTM temporal module and a fully convolutional decoding layer. The occlusion image data is input into the backbone network, multi-scale image features are extracted, and the physical layout topology of the photovoltaic power station is read. The physical layout topology of the photovoltaic power station is a data structure that records the spatial boundaries, installation posture and string affiliation of the components. A layout prior map is generated based on the spatial dimensions of the multi-scale image features. On the input side of the pixel decoder, the multi-scale image features and the layout prior map are fused through the component layout prior embedding module, and the layout constraint image features are generated by compressing them to the original number of channels of the corresponding scale image features through pointwise convolution. The layout constraint image features are input into the pixel decoder and the pixel embedding features with component boundary constraints and string affiliation constraints are output. Read the operational drop characteristics, component temperature change, and environmental disturbance characteristics from the component-level state perception results. The operational drop characteristics are the component current drop ratio, component voltage drop ratio, and output power drop ratio obtained based on the historical unshaded operating conditions average. The environmental disturbance characteristics are external state quantities generated based on the changes in irradiance, wind speed, wind direction, and humidity at adjacent sampling times. After normalizing the operational drop characteristics, component temperature change, and environmental disturbance characteristics, input them into the multilayer sensor to generate a state condition vector. When the occlusion image data covers multiple photovoltaic modules, a corresponding state condition vector is generated according to the photovoltaic module number. The state condition vector is input into the multi-source state condition guidance module in the Transformer decoder. Scaling coefficient and offset coefficient are generated based on the state condition vector. Each state condition vector only acts on the mask query feature associated with the corresponding photovoltaic module. The mask query feature is modulated channel by channel to generate the modulated mask query feature. The state condition vector is converted into condition key features and condition value features, and then incorporated into pixel key features and pixel value features respectively. The modulated mask query features, the incorporated key features, and the incorporated value features are input into the cross-attention layer to generate state constraint mask features. The state constraint mask features are input to the output head. The occlusion type is output through the category prediction branch, and the occlusion region mask is output through the mask prediction branch. The occlusion types include tree occlusion, building occlusion, terrain occlusion, cloud shadow, foreign object coverage, dust accumulation, and component self-occlusion. The occlusion region mask is a binary pixel map with the same size as the component-level image region. Pixels with a value of one are occluded pixels, and pixels with a value of zero are non-occluded pixels. Based on the occlusion region mask, the pixel features corresponding to the occlusion region are extracted from the pixel embedding features, and the pixel features are subjected to region average pooling to generate the occlusion region feature vector. Read the category embedding vector corresponding to the type of shading object, the position code of the shading area in the effective power generation area of ​​the module, the operation drop feature, and the module temperature change. The category embedding vector is obtained by mapping the shading object type through the embedding table. The position code is obtained by encoding the center coordinates, boundary coordinates, and region number of the shading area mask in the effective power generation area of ​​the module. The shading area feature vector, category embedding vector, position code, operation drop feature, and module temperature change are concatenated to generate the MPPT sensitive feature vector. The MPPT sensitive feature vector is input into a weight generation network consisting of two fully connected layers, a normalization layer, and a Sigmoid activation layer. The output is an MPPT sensitive weight with a value ranging from zero to one. The MPPT sensitive weight is a quantized value of the degree of multi-peak distortion of the power-voltage curve caused by the masking of the occlusion region. During the training phase, power-voltage curve simulations were performed on photovoltaic module samples with different types of obstructions, obstruction locations, and obstruction areas. The characteristics of the number of power peaks, the difference between the maximum power peak and the second largest power peak, the voltage interval between adjacent power peaks, and the duration of obstruction were extracted. After normalizing the above characteristics, they were weighted according to the weights determined by the historical calibration samples to generate MPPT sensitive weight supervision labels. The weight generation network was trained based on the MPPT sensitive weight supervision labels. The historical calibration samples included obstruction quantification results, historical maximum power point tracking results, and actual output power change records. During the inference phase, the shading area mask is cropped based on the effective power generation area of ​​the module. The number of shading pixels in the cropped shading area mask is counted. The number of shading pixels is multiplied by the MPPT sensitivity weight to obtain the weighted number of shading pixels. The weighted shading area ratio is generated based on the ratio of the total number of weighted shading pixels in the same photovoltaic module to the number of pixels in the effective power generation area of ​​the module. The occlusion area mask within a continuous historical time window is read. Based on the centroid displacement and area change of the occlusion area mask at adjacent sampling times, occlusion motion features are generated. Wind speed, wind direction, and irradiance changes are encoded into meteorological embedding vectors. The occlusion area mask, occlusion motion features, and meteorological embedding vectors are input into the ConvLSTM temporal module to generate temporal hidden states. The predicted occlusion mask for the next control cycle is generated through a fully convolutional decoding layer. Based on the current shading area mask, predicted shading mask, weighted shading area ratio, MPPT sensitivity weight, and shading duration, the shading area change rate, shading movement direction, shading level for the next control cycle, and multi-peak power curve risk coefficient are generated, and shading quantification results are generated by binding the photovoltaic module number and string number. The photovoltaic shading sensing Mask2Former is similar to the original Mask2Former in that it retains the basic structure of backbone network, pixel decoder, Transformer decoder and output head. It achieves shading area recognition through multi-scale image feature extraction, mask query and mask segmentation, and can output the type of shading object and the shading area mask. The difference lies in the fact that the photovoltaic shading-aware Mask2Former adds a component layout prior embedding module to the input side of the pixel decoder, which encodes the physical layout topology of the photovoltaic power station into a layout prior map; it adds a multi-source state condition guidance module to the Transformer decoder, which allows operating drop features, component temperature changes and environmental disturbance features to participate in mask query modulation; and it adds an MPPT sensitive weight generation branch to the output head, and connects a time-series shading evolution prediction module after the output head. The above improvements enable occlusion recognition results to go beyond ordinary image segmentation, and further generate weighted occlusion area ratio, occlusion level for the next control cycle, and risk coefficient of multi-peak power curve. This solves the problem that ordinary occlusion detection is difficult to bind to component number, string affiliation, and MPPT control sensitivity, and provides a more accurate occlusion quantification basis for subsequent adaptive MPPT parameter group generation.

[0019] In this embodiment, step three specifically includes: Read the weighted shading area ratio, shading area change rate, multi-peak power curve risk coefficient and shading level for the next control cycle from the shading quantification results, and read the corresponding photovoltaic module power change status and historical maximum power point tracking results. The module power change status includes the output power change direction, output power decrease ratio and power fluctuation number within the same control cycle at adjacent sampling times. Read the benchmark MPPT parameters from the historical stable operating cycle. The historical stable operating cycle is the historical control cycle in which no hot spot effect warning information is generated and the number of power fluctuations does not exceed the preset fluctuation number threshold. The benchmark MPPT parameters include benchmark search range, benchmark search step size, benchmark particle swarm inertia weight, benchmark particle swarm learning factor, benchmark Gaussian mutation parameter, benchmark maximum number of iterations and benchmark convergence judgment threshold. Read the parameter mapping table generated from the historical calibration samples. The parameter mapping table records the correspondence between the occlusion area level and the search range expansion coefficient, the occlusion change level and the search step size adjustment coefficient, the multi-peak risk level and the particle swarm inertia weight adjustment coefficient, the particle swarm learning factor adjustment coefficient and the Gaussian mutation adjustment coefficient, and the occlusion level and the iterative convergence adjustment coefficient. The shading area level is determined based on the weighted shading area ratio, and the baseline search range is extended to the MPPT search range according to the parameter mapping table. The higher the shading area level, the wider the voltage range covered by the MPPT search range, and it does not exceed the voltage range corresponding to the DC-DC duty cycle boundary. The occlusion change level is determined based on the occlusion area change rate, and the baseline search step size is adjusted to the search step size according to the parameter mapping table. The higher the occlusion change level, the larger the search step size. The risk level of the multi-peak power curve is determined based on the risk coefficient of the multi-peak power curve, and the baseline particle swarm inertia weight, baseline particle swarm learning factor and baseline Gaussian mutation parameter are adjusted according to the parameter mapping table to generate the particle swarm inertia weight, particle swarm learning factor and Gaussian mutation parameter. The maximum number of iterations and the convergence threshold are adjusted according to the shading level of the next control cycle. The DC-DC duty cycle boundary is determined by combining the historical maximum power point tracking results. The adjusted parameters are then bound according to the photovoltaic module number and string number to generate an adaptive MPPT parameter group.

[0020] In this embodiment, step four specifically includes: Read the MPPT search range, search step size, particle swarm inertia weight, particle swarm learning factor, slime mold weight coefficient, Gaussian mutation parameter, arithmetic optimization operator parameter, maximum number of iterations, convergence judgment threshold and DC-DC duty cycle boundary from the adaptive MPPT parameter group, and define the candidate operating parameters as a parameter vector composed of candidate output voltage, candidate output current and candidate duty cycle; Candidate operating parameter boundaries are established based on the MPPT search range and DC-DC duty cycle boundary. A chaotic sequence is generated using Logistic chaotic mapping, and the chaotic sequence is mapped to the candidate operating parameter boundary to generate an initial set of candidate operating parameters. Logistic chaotic mapping means that the next chaotic value is equal to the preset chaotic control parameter multiplied by the previous chaotic value, then multiplied by one and subtracted from the previous chaotic value. The initial set of candidate running parameters is input into the improved particle swarm optimization algorithm. The occlusion level of the next control cycle is read from the occlusion quantization result. Based on the occlusion level of the next control cycle, the corresponding particle swarm inertia weight and particle swarm learning factor are selected from the adaptive MPPT parameter set. The candidate operating parameters are used as particle positions, the component output power corresponding to the candidate operating parameters is used as fitness values, the previous search direction of the particles is retained according to the particle swarm inertia weight, and the particles are guided to move towards the individual optimal candidate operating parameters and the swarm optimal candidate operating parameters according to the particle swarm learning factor, and the candidate operating parameters are updated. After each particle swarm update, candidate operating parameters that exceed the MPPT search range or DC-DC duty cycle boundary are removed, and the remaining candidate operating parameters are sorted from largest to smallest according to the component output power to generate a set of candidate operating parameters for particle swarm. The set of candidate running parameters for particle swarm optimization is input into the improved slime mold algorithm. Based on the power ranking of the candidate running parameters, slime mold weight coefficients are generated. The slime mold weight coefficients are used to strengthen the search direction close to high-power candidate running parameters and weaken the search direction close to low-power candidate running parameters, thereby generating slime mold-optimized candidate running parameters. Based on the Gaussian variation parameters, the candidate operating parameters for slime mold optimization are randomly perturbed. The random perturbation is generated by superimposing the perturbation amount generated by the Gaussian distribution on the three dimensions of candidate output voltage, candidate output current and candidate duty cycle to generate variant candidate operating parameters. Based on the parameters of the arithmetic optimization operator, the candidate running parameters of the mutation are subjected to expansion, contraction and fine-tuning processes. The expansion process increases the distance between the candidate running parameters and the current best candidate running parameters, the contraction process decreases the distance between the candidate running parameters and the current best candidate running parameters, and the fine-tuning process changes the candidate running parameters within the search step size limit to generate a set of slime mold candidate running parameters. The candidate operating parameter set for particle swarm optimization and the candidate operating parameter set for slime mold are merged, and duplicate candidate operating parameters and candidate operating parameters that do not meet the DC-DC duty cycle boundary are deleted to generate a fused candidate operating parameter set. For each candidate operating parameter in the fusion candidate operating parameter set, read its candidate output voltage, candidate output current and candidate duty cycle, and calculate the corresponding candidate power point. The candidate power point includes the candidate component output power, the number of convergence iterations, the continuous settling time and the target power deviation. The convergence iteration count is the number of iterations when the candidate power point first satisfies that the difference between the output power of two adjacent candidate components is not greater than the convergence judgment threshold. The continuous stabilization time is the sampling duration during which the candidate power point continuously satisfies that the power fluctuation amplitude is not greater than the preset stabilization fluctuation threshold within the current control cycle. Based on the historical maximum power point power in the historical maximum power point tracking results, the weighted occlusion area ratio in the occlusion quantization results, and the MPPT sensitive weight, an occlusion reduction coefficient is generated. The historical maximum power point power is then reduced according to the occlusion reduction coefficient to generate the occlusion correction target power. The ratio of the output power of the candidate component to the output power of the largest candidate component in the fusion candidate operating parameter set is used as the power evaluation term. The ratio of one minus the convergence iteration number to the maximum iteration number is used as the convergence evaluation term. The ratio of the continuous stable time to the duration of the current control cycle is used as the stability evaluation term. The ratio of one minus the target power deviation to the occlusion correction target power is used as the deviation evaluation term. The target power deviation is the absolute value of the difference between the output power of the candidate component and the occlusion correction target power. Read the power evaluation weight, convergence evaluation weight, stability evaluation weight, and deviation evaluation weight determined by historical calibration samples. The sum of the four weights is one. Multiply the power evaluation item, convergence evaluation item, stability evaluation item, and deviation evaluation item by their respective weights and then add them together to generate the candidate operating parameter fusion evaluation value. The candidate operating parameter with the largest fusion evaluation value is selected as the target operating parameter corresponding to the global maximum power point, and the target output voltage, target output current and target duty cycle are generated from the target operating parameter; This implementation generates an initial candidate operating parameter set through Logistic chaotic mapping, ensuring a more comprehensive distribution of candidate output voltage, candidate output current, and candidate duty cycle within the MPPT search range, thus reducing the risk of initial search points concentrating on local power peaks. An improved particle swarm optimization algorithm adjusts the particle swarm inertia weight and particle swarm learning factor according to the occlusion level of the next control cycle, addressing the issue of slow convergence or premature convergence of fixed parameters under dynamic occlusion. An improved slime mold algorithm, combined with Gaussian mutation parameters and arithmetic optimization operator parameters, perturbs the candidate operating parameters, enhancing the ability to escape local maximum power points. By fusing the candidate operating parameter set and generating a fusion evaluation value based on candidate component output power, convergence iteration count, continuous stability time, and target power deviation, the selection of the global maximum power point simultaneously considers power gain, convergence speed, and stability, thereby improving the reliability of the target output voltage, target output current, and target duty cycle.

[0021] In this embodiment, step five specifically includes: Read the target output voltage, target output current, target duty cycle, DC-DC duty cycle boundary and convergence threshold, and collect the actual output voltage, actual output current and actual output power at the output of the photovoltaic optimizer; The difference between the target output voltage and the actual output voltage is taken as the voltage deviation, and the difference between the target output current and the actual output current is taken as the current deviation. The voltage deviation is multiplied by the voltage loop adjustment coefficient, the current deviation is multiplied by the current loop adjustment coefficient, and the two are added together to generate the duty cycle adjustment amount. The target duty cycle is superimposed with the duty cycle adjustment amount, and the amplitude is limited according to the DC-DC duty cycle boundary to generate the execution duty cycle; The PWM drive signal is generated based on the execution duty cycle, and the conduction time of the switching devices in the DC-DC Buck circuit is adjusted so that the output voltage and output current of the photovoltaic module converge towards the target output voltage and target output current. When the power deviation between the adjusted actual output power and the target power corresponding to the target output voltage and target output current is not greater than the convergence judgment threshold, the actual output voltage, actual output current, actual output power, execution duty cycle and power deviation are recorded according to the photovoltaic module number and string number to generate the module-level power regulation result.

[0022] In this embodiment, step six specifically includes: Read the component-level power regulation results, shading quantization results, and photovoltaic unit adjustable capacity status. The photovoltaic unit adjustable capacity status includes the current output power, adjustable power upper limit, adjustable power lower limit, and power regulation direction. When the shading levels of two photovoltaic strings are the same in the next control cycle, and the difference in weighted shading area ratio does not exceed the preset shading difference threshold, and the difference in multi-peak power curve risk coefficient does not exceed the preset risk difference threshold, the shading status of the two photovoltaic strings is determined to be consistent. The photovoltaic strings with consistent shading status are assigned to the same MPPT control channel, and the photovoltaic strings with inconsistent shading status are assigned to different MPPT control channels. Based on the power increase direction of the power station-level power command, the difference between the adjustable power upper limit and the current output power is used as the upward adjustment of the adjustable capacity; based on the power decrease direction of the power station-level power command, the difference between the current output power and the adjustable power lower limit is used as the downward adjustment of the adjustable capacity. Power allocation commands are distributed according to the proportion of each photovoltaic unit's adjustable capacity to the total adjustable capacity. When the allocation result exceeds the corresponding adjustable capacity boundary, the excess portion is redistributed to photovoltaic units that still have adjustable capacity, thus generating a power allocation result.

[0023] In this embodiment, step seven specifically includes: Read the actual output power, power deviation, execution duty cycle and adjustment number from the component-level power regulation results, and read the MPPT control channel number, photovoltaic unit power command and adjustable capacity occupancy status from the power allocation results, and generate operation feedback data according to the photovoltaic module number and string number; Read the actual output power drop, power deviation duration, number of consecutive changes in execution duty cycle, component temperature change, wiring status and arc fault status from the operation feedback data to determine the abnormal status of the component. When the component temperature change reaches the preset temperature rise threshold, or the actual output power decrease reaches the preset power decrease threshold and the power deviation duration reaches the preset duration threshold, a hot spot effect warning message is generated. When the wiring status is disconnected or the arc fault status is triggered, fault location information is generated based on the photovoltaic module number, string number, distributed control node address, and module installation coordinates. Based on hot spot effect early warning information, fault location information, and operational feedback data, the occlusion level, MPPT search range, search step size, particle swarm parameters, slime mold parameters, and DC-DC duty cycle boundary are corrected for the next control cycle. When hot spot effect early warning information is generated, the DC-DC duty cycle boundary is tightened and the target power adjustment amplitude for the next control cycle is reduced. When the power deviation continues to exceed the preset duration threshold, the MPPT search range is expanded and the search step size is increased to generate an adaptive MPPT parameter set for the next control cycle.

[0024] Example 1: To verify the feasibility of this invention in practice, it was applied to a local shading management scenario at a distributed photovoltaic power station in a mountainous area. The power station selected four photovoltaic strings, totaling 96 photovoltaic modules, as the test objects. On-site conditions included tree shading, building shading, cloud shadows, and module self-shading. The centralized monitoring platform received module-level operational data, environmental status data, and shading image data uploaded by each distributed control node at a 5-second sampling cycle, and generated module-level status perception results according to the photovoltaic module number and string number. During implementation, the preset shading area level thresholds were 8% and 20%, the preset shading area change rate thresholds were 3% / min and 8% / min, the preset multi-peak power curve risk coefficient thresholds were 0.35 and 0.65, the preset shading difference threshold was 0.06, the preset risk difference threshold was 0.10, the preset temperature rise threshold was 8℃, the preset power decrease threshold was 15%, the preset duration threshold was 60 seconds, and the DC-DC duty cycle boundary was set to 0.18 to 0.92.

[0025] In actual operation, the component-level state perception results are input into the photovoltaic shading perception Mask2Former. The component layout prior embedding module encodes the physical layout topology of the photovoltaic power station into a layout prior map, and performs channel fusion with multi-scale image features at the input side of the pixel decoder to avoid shading areas from sticking across components; the multi-source state condition guidance module introduces the operating drop features, component temperature change, and environmental disturbance features into the Transformer decoder to make the shading mask segmentation results correspond to the actual operating state of the components; the MPPT sensitive weight generation branch generates MPPT sensitive weights based on the shading area feature vector, category embedding vector, position encoding, operating drop features, and component temperature change, and then combines them with the effective power generation area of ​​the component to generate a weighted shading area ratio; the time-series shading evolution prediction module outputs the shading level and multi-peak power curve risk coefficient for the next control cycle based on the shading motion features and meteorological embedding vector.

[0026] When the weighted shading area ratio reaches 20% or more and the multi-peak power curve risk coefficient reaches 0.65 or more, the system determines that the current photovoltaic module has entered the multi-peak power curve state and generates an adaptive MPPT parameter set according to the parameter mapping table. At this time, the MPPT search range is expanded from the interval near the historical maximum power point to the voltage interval corresponding to the DC-DC duty cycle boundary. The search step size increases with the rate of change of shading area, and the particle swarm inertia weight, particle swarm learning factor, and Gaussian mutation parameter increase with the multi-peak power curve risk coefficient. Subsequently, an initial candidate operating parameter set is generated using Logistic chaotic mapping, and then collaborative optimization is performed using an improved particle swarm optimization algorithm and an improved slime mold algorithm. A fusion evaluation value is generated according to the candidate module output power, convergence iteration number, continuous stability time, and target power deviation. The candidate operating parameter with the largest fusion evaluation value is selected as the target operating parameter corresponding to the global maximum power point, and the target output voltage, target output current, and target duty cycle are output.

[0027] To verify the effectiveness, this invention was compared with three other methods. Method A uses the traditional perturbation observation method with fixed MPPT parameters; Method B uses ordinary semantic segmentation to identify occlusion and then uses a fixed particle swarm optimization algorithm; Method C uses the original Mask2Former with a single particle swarm optimization algorithm, without the components layout prior embedding, multi-source state condition guidance, MPPT sensitive weight generation branch, and temporal occlusion evolution prediction module. The test period was 7 consecutive days, with the effective power generation period from 9:00 to 16:00 selected each day, and data were collected under three operating conditions: local shading, rapid changes in cloudy conditions, and fixed building occlusion.

[0028] Table 1. Comprehensive Test Data of the Invention and Other Methods

[0029] As shown in Table 1, the present invention outperforms methods A, B, and C in terms of occlusion recognition accuracy, multi-peak recognition accuracy, GMPPT tracking success rate, average convergence time, average daily power generation, and hot spot early warning accuracy. Compared with method A, the GMPPT tracking success rate of the present invention is increased from 82.1% to 97.2%, and the average convergence time is shortened from 4.8s to 2.6s, indicating that fixed MPPT parameters are difficult to escape the local maximum power point in time under local shading, while the present invention improves the tracking effect under multi-peak power curves through adaptive MPPT parameter sets and collaborative optimization mechanisms. Compared with method B, the occlusion recognition accuracy of the present invention is increased from 84.6% to 94.3%, indicating that although ordinary semantic segmentation can identify occlusion areas, it cannot fully bind the occlusion areas with component boundaries, string affiliation relationships, and operational drop characteristics. Compared with method C, the shading recognition accuracy of the present invention is improved by 6.2 percentage points and the multi-peak recognition accuracy is improved by 7.9 percentage points. This indicates that the component layout prior embedding, multi-source state condition guidance, MPPT sensitive weight generation branch, and time-series shading evolution prediction module in the photovoltaic shading sensing Mask2Former can more accurately reflect the impact of shading on the power-voltage curve.

[0030] In terms of power generation revenue, the average daily power generation of this invention reaches 394.6 kWh, which is 15.4 kWh higher than method C and 37.8 kWh higher than method A. This indicates that the invention not only improves the recognition accuracy but also transforms the occlusion quantization results into the basis for adjusting the MPPT search range, search step size, particle swarm inertia weight, particle swarm learning factor, and Gaussian mutation parameters. The improved particle swarm optimization algorithm is responsible for rapid searching based on the occlusion level of the next control cycle. The improved slime mold algorithm combines Gaussian mutation parameters and arithmetic optimization operator parameters to perturb candidate operating parameters. The fusion of these two methods reduces local optimum mistracking. The average convergence time is reduced to 2.6 s, indicating that the candidate operating parameter fusion evaluation mechanism takes into account the output power of candidate components, the number of convergence iterations, continuous stability time, and target power deviation, making the target output voltage, target output current, and target duty cycle more stable.

[0031] In terms of operational safety, the hot spot warning accuracy of this invention reaches 94.8%. When the component temperature change reaches 8°C, or the actual output power decreases by 15% and the power deviation lasts for 60 seconds, the system generates a hot spot effect warning and tightens the DC-DC duty cycle boundary and reduces the target power adjustment range in the next control cycle. When the wiring status is disconnected or the arc fault status is triggered, fault location information is generated, and the power allocation ratio of the photovoltaic unit to which the faulty component belongs is reduced. If necessary, the allocation of the backup MPPT control channel is triggered. Thus, this invention can form a closed-loop control from component-level state perception, shading quantification, adaptive MPPT parameter group generation, global maximum power point optimization, DC-DC Buck circuit adjustment to operational feedback updates. It has significant beneficial effects on improving power generation efficiency under complex shading conditions, reducing hot spot risk, and improving the operational stability of distributed photovoltaic power plants.

[0032] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An adaptive optimization method for MPPT parameters based on distributed control, characterized in that, The steps include the following: Step 1: Collect module-level operating data, environmental status data, and shading image data according to the photovoltaic module number and string number, and generate module-level status perception results; Step 2: Input the component-level state perception results into the photovoltaic shading perception Mask2Former, and generate shading quantification results through component layout prior embedding, multi-source state condition guidance, shading mask segmentation and time-series shading evolution prediction. Step 3: Based on the shading quantification results, the power change status of the module, and the historical maximum power point tracking results, determine the current power-voltage curve status of the photovoltaic module, and generate an adaptive MPPT parameter set when entering the multi-peak power curve state; Step 4: Based on the adaptive MPPT parameter set, the candidate operating parameters are chaotically initialized. An improved particle swarm optimization algorithm is used to adjust the inertia weight and learning factor according to the occlusion level to update the candidate operating parameters. An improved slime mold algorithm is used to perturb the candidate operating parameters by combining Gaussian mutation and arithmetic optimization operator. The optimization results are fused to generate the target output voltage, target output current and target duty cycle corresponding to the global maximum power point. Step 5: Adjust the DC-DC Buck circuit based on the target output voltage, target output current, and target duty cycle to generate component-level power regulation results; Step 6: Based on the module-level power regulation results, shading quantification results, and photovoltaic unit adjustable capacity status, allocate the MPPT control channel and power plant-level power commands to generate power allocation results; Step 7: Generate operational feedback data based on component-level power regulation and power allocation results, and generate hot spot effect early warning information, fault location information, and adaptive MPPT parameter set for the next control cycle based on the operational feedback data.

2. The MPPT parameter adaptive optimization method based on distributed control according to claim 1, characterized in that, Step one specifically includes: Based on the physical layout diagram of photovoltaic modules and the string wiring relationship, establish the binding relationship between photovoltaic module number, string number, photovoltaic optimizer number and distributed control node address; The centralized monitoring platform receives component-level operating data, environmental status data, and occlusion image data uploaded by each distributed control node according to a unified sampling time and sampling period. The component-level operating data includes component voltage, component current, output power, component temperature, DC-DC duty cycle status, wiring status, and arc fault status. Each distributed control node encapsulates component-level operational data, environmental status data, and occlusion image data into a component status data frame. The component status data frame includes the photovoltaic module number, string number, distributed control node address, sampling time, and data source marker. The centralized monitoring platform performs time alignment on the component status data frames based on the sampling time, and merges the time-aligned component status data frames according to the photovoltaic component number and string number to generate component-level status perception results.

3. The MPPT parameter adaptive optimization method based on distributed control according to claim 1, characterized in that, Step two specifically includes: The photovoltaic shading-aware Mask2Former consists of a backbone network, pixel decoders, Transformer decoders, and an output head. It also includes a component layout prior embedding module, a multi-source state condition guidance module, a shading mask segmentation module, an MPPT sensitive weight generation branch, and a temporal shading evolution prediction module. The component layout prior embedding module encodes the physical layout topology of the photovoltaic power station into a layout prior map. The physical layout topology of the photovoltaic power station includes the component spatial boundary, installation posture and string affiliation relationship. On the input side of the pixel decoder, the layout prior map is fused with the multi-scale image features output by the backbone network through channel fusion, and pixel embedding features with component boundary constraints and string affiliation constraints are generated by point-by-point convolution. The multi-source state condition guidance module generates a state condition vector based on the operational drop characteristics, component temperature change, and environmental disturbance characteristics. The operational drop characteristics consist of the component current drop ratio, component voltage drop ratio, and output power drop ratio. The environmental disturbance characteristics consist of irradiance change, wind speed change, wind direction change, and humidity change. The state condition vector generates scaling and offset coefficients through the Transformer decoder, modulates the mask query features associated with the corresponding photovoltaic module channel by channel, and converts them into condition key features and condition value features to participate in cross-attention calculation, generating state constraint mask features. The occlusion mask segmentation module generates occlusion type and occlusion area mask based on state constraint mask features. The occlusion area mask is a binary pixel map composed of occluded pixels and unoccluded pixels. The MPPT sensitive weight generation branch extracts the feature vector of the shading area based on the shading area mask, and inputs the shading area feature vector, category embedding vector, location encoding, operational drop features and component temperature change into the weight generation network to generate MPPT sensitive weights. The category embedding vector is obtained by mapping the shading object type through an embedding table, and the location encoding is obtained by encoding the center coordinates, boundary coordinates and region number of the shading area mask in the effective power generation area of ​​the component. The weighted shading area ratio is generated based on the shading area mask, MPPT sensitivity weight, and effective power generation area of ​​the module. The time-series shading evolution prediction module generates a predicted shading mask based on the shading area mask, shading movement characteristics, and meteorological embedding vector. The meteorological embedding vector is obtained by encoding wind speed, wind direction, and irradiance changes, and outputs the shading area change rate, shading movement direction, shading level of the next control cycle, and risk coefficient of multi-peak power curve, generating shading quantification results.

4. The MPPT parameter adaptive optimization method based on distributed control according to claim 1, characterized in that, Step three specifically includes: The current power-voltage curve status of the photovoltaic module is determined based on the weighted shading area ratio, shading area change rate, multi-peak power curve risk coefficient, shading level of the next control cycle, module power change status, and historical maximum power point tracking results. The benchmark MPPT parameters are formed based on historical stable operating cycles. The benchmark MPPT parameters include benchmark search range, benchmark search step size, benchmark particle swarm inertia weight, benchmark particle swarm learning factor, benchmark Gaussian mutation parameter, benchmark maximum number of iterations, and benchmark convergence judgment threshold. A parameter mapping table is formed based on historical calibration samples. The historical calibration samples include occlusion quantization results, component power change status, historical maximum power point tracking results, and actual output power change records. The parameter mapping table records the correspondence between occlusion area level and search range expansion coefficient, occlusion change level and search step size adjustment coefficient, multi-peak risk level and particle swarm inertia weight adjustment coefficient, particle swarm learning factor adjustment coefficient, and Gaussian mutation adjustment coefficient. The occlusion area level is determined based on the weighted occlusion area ratio, and the baseline search range is adjusted according to the search range expansion coefficient to generate the MPPT search range; The occlusion change level is determined based on the occlusion area change rate, and the baseline search step size is adjusted according to the search step size adjustment coefficient to generate the search step size; The risk level of the multi-peak power curve is determined based on the risk coefficient of the multi-peak power curve, and the baseline particle swarm inertia weight, baseline particle swarm learning factor and baseline Gaussian mutation parameter are adjusted according to the corresponding adjustment coefficient to generate the particle swarm inertia weight, particle swarm learning factor and Gaussian mutation parameter. The maximum number of iterations and the convergence threshold are adjusted according to the occlusion level of the next control cycle. The DC-DC duty cycle boundary is determined based on the historical maximum power point tracking results. The MPPT search range, search step size, particle swarm inertia weight, particle swarm learning factor, Gaussian mutation parameter, maximum number of iterations, convergence threshold and DC-DC duty cycle boundary are bound together to generate an adaptive MPPT parameter set.

5. The MPPT parameter adaptive optimization method based on distributed control according to claim 1, characterized in that, Step four specifically includes: Candidate operating parameter boundaries are established based on the MPPT search range and DC-DC duty cycle boundary. Candidate output voltage, candidate output current, and candidate duty cycle are used as candidate operating parameters. Logistic chaotic mapping is used to generate an initial set of candidate operating parameters. The improved particle swarm optimization algorithm uses candidate operating parameters as particle positions and component output power as fitness values. It updates the candidate operating parameters according to the particle swarm inertia weight and particle swarm learning factor corresponding to the occlusion level in the next control cycle, and generates a set of candidate operating parameters for particle swarm. The improved slime mold algorithm generates slime mold weight coefficients based on the power ranking results, and perturbs the candidate running parameters by combining Gaussian mutation parameters and arithmetic optimization operator parameters to generate a set of slime mold candidate running parameters; The candidate operating parameter sets for particle swarm optimization and slime mold are merged, and candidate operating parameters that exceed the DC-DC duty cycle boundary are deleted to generate a fused candidate operating parameter set. An occlusion reduction coefficient is generated based on the historical maximum power point power, weighted occlusion area ratio, and MPPT sensitivity weight in the historical maximum power point tracking results. The occlusion reduction coefficient is obtained by multiplying the weighted occlusion area ratio and MPPT sensitivity weight and limiting it to between zero and one. The historical maximum power point power is then reduced according to the occlusion reduction coefficient to generate the occlusion correction target power. For candidate power points in the fusion candidate operating parameter set, power evaluation terms, convergence evaluation terms, stability evaluation terms, and deviation evaluation terms are generated respectively. The power evaluation term is obtained by normalizing the output power of the candidate component. The convergence evaluation term is obtained by inverse normalizing the number of convergence iterations. The stability evaluation term is obtained by normalizing the continuous stable time. The deviation evaluation term is obtained by inverse normalizing the target power deviation. The target power deviation is the absolute value of the difference between the output power of the candidate component and the occlusion correction target power. The power evaluation item, convergence evaluation item, stability evaluation item, and deviation evaluation item are multiplied by their respective weights and then summed to generate a fusion evaluation value for candidate operating parameters. The candidate operating parameter with the largest fusion evaluation value is selected as the target operating parameter corresponding to the global maximum power point, and the target output voltage, target output current, and target duty cycle are generated.

6. The MPPT parameter adaptive optimization method based on distributed control according to claim 1, characterized in that, Step five specifically includes: Voltage deviation and current deviation are generated based on the target output voltage, target output current, actual output voltage, and actual output current; Multiply the voltage deviation by the voltage loop adjustment coefficient, multiply the current deviation by the current loop adjustment coefficient, and add the two together to generate the duty cycle adjustment amount; The target duty cycle is superimposed with the duty cycle adjustment amount, and the execution duty cycle is generated by limiting it according to the DC-DC duty cycle boundary. The PWM drive signal is generated based on the duty cycle to adjust the on-time of the switching devices in the DC-DC Buck circuit. The target output voltage is multiplied by the target output current to obtain the target power. When the power deviation between the actual output power and the target power is not greater than the convergence judgment threshold, the module-level power regulation result is generated according to the photovoltaic module number and string number.

7. The MPPT parameter adaptive optimization method based on distributed control according to claim 1, characterized in that, Step six specifically includes: MPPT control channels and power plant-level power commands are allocated based on the component-level power regulation results, shading quantification results, and the adjustable capacity status of the photovoltaic unit. The adjustable capacity status of the photovoltaic unit includes the current output power, adjustable power upper limit, adjustable power lower limit, and power regulation direction. When the two photovoltaic strings have the same shading level in the next control cycle, the difference in the weighted shading area ratio does not exceed the preset shading difference threshold, and the difference in the risk coefficient of the multi-peak power curve does not exceed the preset risk difference threshold, the two photovoltaic strings will be assigned to the same MPPT control channel; otherwise, they will be assigned to different MPPT control channels. The adjustable capacity is adjusted upwards or downwards according to the direction of the power station-level power command, and the power station-level power command is allocated according to the proportion of the adjustable capacity of each photovoltaic unit to the total adjustable capacity, thus generating the power allocation result.

8. The MPPT parameter adaptive optimization method based on distributed control according to claim 1, characterized in that, Step seven specifically includes: Based on the component-level power regulation and power allocation results, operational feedback data is generated, and the abnormal status of the component is determined based on the actual output power drop, power deviation duration, number of consecutive changes in duty cycle, component temperature change, wiring status, and arc fault status. When the component temperature change reaches the preset temperature rise threshold, or the actual output power decrease reaches the preset power decrease threshold and the power deviation duration reaches the preset duration threshold, a hot spot effect warning message is generated. When the wiring status is disconnected or the arc fault status is triggered, fault location information is generated. The adaptive MPPT parameter set for the next control cycle is updated based on the operation feedback data, hot spot effect early warning information and fault location information. The hot spot effect early warning information corresponds to tightening the DC-DC duty cycle boundary and reducing the target power adjustment amplitude. The continuous power deviation exceeding the limit corresponds to expanding the MPPT search range and increasing the search step size. The fault location information corresponds to reducing the power allocation ratio of the photovoltaic unit to which the faulty component belongs and triggering the allocation of the backup MPPT control channel.

Images