Photovoltaic power generation intelligent evaluation system based on digital twinning

Abstract

The invention relates to the technical field of photovoltaic power generation, in particular to a photovoltaic power generation intelligent evaluation system based on digital twinning, and the system comprises a micro-macro coupling modeling module which is used for building a dynamic fouling probability field model of the surface of a photovoltaic module, the dynamic fouling probability field model simulates a non-uniform deposition process of pollutants on the surface of the photovoltaic module through a discrete element algorithm to generate a three-dimensional microscopic shielding topological graph; and carrying out space-time alignment on the three-dimensional microscopic occlusion topological graph and the macroscopic meteorological data by adopting a cross-scale feature fusion technology. According to the invention, through the microscopic-macroscopic coupling modeling module, a discrete element algorithm is utilized to simulate pollutant non-uniform deposition and generate a three-dimensional microscopic occlusion topological graph, and a cross-scale feature fusion technology is combined to align the three-dimensional microscopic occlusion topological graph with macroscopic meteorological data, so that the problem that a traditional model cannot quantify a microscopic pollutant occlusion effect is solved.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of photovoltaic power generation, in particular to a photovoltaic power generation intelligent evaluation system based on digital twinning. BACKGROUND

[0002] In the operation process of a photovoltaic power generation system, the power generation efficiency and equipment state of the photovoltaic power generation system need to be evaluated through real-time monitoring and environmental analysis. Traditional photovoltaic evaluation methods usually rely on data collected by physical sensors or historical statistical models, but these methods are difficult to dynamically reflect the real running state of the equipment under complex environments. Digital twinning technology, as a virtual mapping means, can provide new ideas for state monitoring and performance prediction of photovoltaic systems through real-time data driving and high-precision modeling. However, how to deeply combine digital twinning with real-time evaluation of photovoltaic power generation is still one of the key challenges currently faced by the technology.

[0003] When the prior art uses digital twinning to evaluate photovoltaic power generation, a specific problem is often ignored: the influence of the micro-pollutant distribution on the surface of a photovoltaic component on the power generation efficiency is difficult to be quantified by a traditional simulation model. The deposition of dust, pollen and other pollutants on the surface of the component presents non-uniformity, and the local shading effect thereof can cause slight but significant changes in light intensity distribution and hot spot effect. However, the existing digital twinning model usually adopts a homogenization assumption or simulates by using macro-environmental parameters, and cannot capture the dynamic shading difference at the micro scale, resulting in a large deviation between the evaluation result of the power generation efficiency and the actual situation. This limitation makes it difficult for the system to accurately identify the performance degradation caused by micro-pollution, thereby affecting the accuracy of operation and maintenance decisions. SUMMARY

[0004] In view of the deficiencies of the prior art, the application provides a photovoltaic power generation intelligent evaluation system based on digital twinning, which solves the problem that, in the prior art, the shading effect of the non-uniform distribution of micro-pollutants on the surface of a photovoltaic component cannot be quantified, resulting in a large deviation in the evaluation of power generation efficiency and difficulty in accurately identifying performance degradation.

[0005] To achieve the above purpose, the application is implemented by the following technical scheme: a photovoltaic power generation evaluation system based on multi-scale digital twinning and dynamic pollution feature migration, the system comprising:

[0006] A micro-macro coupling modeling module is configured to establish a dynamic pollution probability field model of the surface of a photovoltaic component. The dynamic pollution probability field model simulates the non-uniform deposition process of pollutants on the surface of the photovoltaic component by using a discrete element algorithm to generate a three-dimensional micro-shading topological graph. In addition, a cross-scale feature fusion technology is used to perform spatio-temporal alignment between the three-dimensional micro-shading topological graph and macro-weather data, and a multi-twin body correlation model is constructed. The multi-twin body correlation model has a local light intensity gradient analysis capability.

[0007] a bionic visual perception module for deploying a polarization-multiple spectral imaging module at the physical end to capture the polarization reflection characteristics and local thermal radiation distribution of the photovoltaic component surface; and developing a dirty feature extractor based on a generative adversarial network to convert the polarization reflection characteristics and the local thermal radiation distribution into a high-resolution dirt mask map;

[0008] an occlusion effect dynamic migration module for designing a virtual light particle tracking engine to inject a dynamic light cluster optimized by a Monte Carlo method into a digital twin to simulate the distortion of photon transmission path under microscopic occlusion; and establishing a neural differential equation for occlusion-power generation mapping to convert the dynamic occlusion effect into a time-varying disturbance term of current output;

[0009] a self-correcting evaluation closed-loop module for deploying a dual-channel difference perception device to continuously compare the microscopic fluctuation characteristics of the physical end power generation data and the digital twin prediction value; and configuring a small sample incremental learning module to update the dynamic dirt probability field model when a feature deviation caused by unknown contaminants is detected.

[0010] Further, the process of establishing a dynamic dirt probability field model includes:

[0011] acquiring initial geometric data and material parameters of the photovoltaic component surface;

[0012] setting the type, particle size distribution and physical properties of the contaminants;

[0013] based on the discrete element algorithm, simulating the non-uniform deposition trajectory and adhesion force distribution of the contaminants on the photovoltaic component surface over time;

[0014] According to the simulation results, a three-dimensional microscopic occlusion topology map with real-time dirt state on the photovoltaic component surface is generated.

[0015] Further, the process of constructing a multi-twin correlation model using cross-scale feature fusion technology includes:

[0016] Collecting real-time irradiance, ambient temperature, and ambient humidity and other macro-weather data at the location of the photovoltaic component;

[0017] High-precision spatiotemporal alignment of the local geometric features of the three-dimensional microscopic occlusion topology map with the macro-weather data;

[0018] By constructing the multi-twin correlation model, the local light intensity gradient distribution of the photovoltaic component surface formed by the combined action of microscopic occlusion and macro-weather is analyzed.

[0019] Further, the process of capturing the polarization reflection characteristics and local thermal radiation distribution of the photovoltaic component surface includes:

[0020] acquire the polarization information of the surface contamination by multi-angle polarization imaging of the photovoltaic module surface using the polarization-multiple spectral imaging module;

[0021] acquire the multi-spectral images of the photovoltaic module surface using the polarization-multiple spectral imaging module, and analyze the reflection characteristics at different wavelengths;

[0022] acquire the local thermal radiation distribution of the photovoltaic module surface by thermal infrared imaging using the polarization-multiple spectral imaging module, and identify potential hot spot regions.

[0023] Further, the process of developing a dirt feature extractor based on a generative adversarial network and converting the optical signals into a contamination mask map includes:

[0024] acquire the polarization reflection characteristics and the local thermal radiation distribution as input optical signals;

[0025] generate a preliminary contamination mask map from the input optical signals using the generator of the generative adversarial network;

[0026] distinguish the preliminary contamination mask map from a real contamination mask map using the discriminator of the generative adversarial network;

[0027] optimize the dirt feature extractor through adversarial training between the generator and the discriminator, and output a high-resolution contamination mask map.

[0028] Further, the process of designing a virtual light particle tracking engine to simulate the distortion of photon transmission paths under microscopic shading includes:

[0029] define the microscopic shading topological structure of the photovoltaic module surface in the digital twin;

[0030] generate dynamic light clusters with specific direction and energy distribution;

[0031] simulate the scattering, absorption and transmission behavior of the dynamic light clusters in the microscopic shading topological structure based on the Monte Carlo method;

[0032] track the path changes of each photon, and calculate the loss of light flux and distortion of light intensity distribution reaching the light-sensitive area of the photovoltaic module.

[0033] Further, the process of establishing a neural differential equation for shading-power generation mapping includes:

[0034] acquire the photon transmission path distortion data and the electrical parameters of the photovoltaic module;

[0035] define the initial state and dynamic evolution function of the neural differential equation;

[0036] fitting a non-linear mapping between the photonic transport path distortions and the photovoltaic assembly current output through end-to-end learning;

[0037] dynamically converting the shading effect into a time-varying perturbation term of the photovoltaic assembly current output at different time instants.

[0038] Further, the process of deploying the dual-channel difference perceiver to compare the microscopic fluctuation features of the physical end power generation data and the digital twin prediction values includes:

[0039] acquiring actual power generation current and voltage data of the photovoltaic assembly in real time from the physical end;

[0040] acquiring power generation current and voltage data predicted by a neural differential equation based on the shading-power generation mapping from the digital twin;

[0041] extracting features of time series of the actual power generation data and the predicted power generation data at a microscopic scale;

[0042] comparing the microscopic fluctuation features of the actual power generation data and the predicted power generation data to identify difference patterns.

[0043] Further, the process of configuring the small sample incremental learning module to update the dynamic soiling probability field model includes:

[0044] triggering the small sample incremental learning module when the dual-channel difference perceiver detects feature deviation caused by unknown contaminants;

[0045] incrementally training the dynamic soiling probability field model based on a small amount of newly acquired soiling sample data;

[0046] combining the original data and the new data to adaptively adjust the parameters of the soiling probability field, thereby improving the recognition and prediction of new soiling patterns by the model.

[0047] Further, the system further includes:

[0048] a performance evaluation module configured to comprehensively evaluate the real-time power generation efficiency, performance degradation degree, and soiling influence of the photovoltaic assembly by combining the multi-twin correlation model, the soiling mask map, the time-varying perturbation term of the current output, and the updated dynamic soiling probability field model.

[0049] The performance evaluation module is configured to generate a refined performance report of the photovoltaic assembly, indicating local performance loss caused by soiling and recommended maintenance measures.

[0050] Compared with the prior art, the present application has the following advantages:

[0051] The application solves the problem that the traditional model cannot quantify the micro-pollutant shielding effect by using the discrete element algorithm to simulate the non-uniform deposition of pollutants and generating a three-dimensional micro-shielding topology graph, and combining the cross-scale feature fusion technology to align it with the macro-weather data; the bionic visual perception module improves the dirt identification accuracy by extracting high-resolution dirt features through polarization-multiple spectrum imaging and generative adversarial network; the shielding effect dynamic migration module realizes the accurate mapping from micro-shielding to power generation change by means of virtual light particle tracking and neural differential equation; the self-correcting evaluation closed-loop module can dynamically update the model to adapt to unknown pollutants through double-channel comparison and small sample learning. These modules work together to significantly reduce the evaluation deviation of power generation efficiency, accurately identify the performance degradation caused by micro-pollution, provide reliable basis for operation and maintenance decision-making, and generate a refined report to guide maintenance. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The system structure diagram of the application. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the application.

[0054] Please refer to Figure 1 The application provides a photovoltaic power generation intelligent evaluation system based on digital twinning, which comprises:

[0055] The micro-macro coupling modeling module is used to establish a dynamic dirt probability field model of the surface of the photovoltaic module. The dynamic dirt probability field model simulates the non-uniform deposition process of pollutants on the surface of the photovoltaic module by the discrete element algorithm, and generates a three-dimensional micro-shielding topology graph. The three-dimensional micro-shielding topology graph is spatiotemporally aligned with the macro-weather data by using the cross-scale feature fusion technology, a multi-twin body correlation model is constructed, and the multi-twin body correlation model has local light intensity gradient analysis capability.

[0056] The bionic visual perception module is used to deploy a polarization-multiple spectrum imaging module on the physical end to capture the polarization reflection features and local thermal radiation distribution of the surface of the photovoltaic module, and develop a dirt feature extractor based on the generative adversarial network to convert the polarization reflection features and local thermal radiation distribution into a high-resolution dirt mask graph.

[0057] The occlusion effect dynamic migration module is used to design a virtual light particle tracking engine, inject dynamic light clusters optimized by the Monte Carlo method into a digital twin to simulate photon transmission path distortion under microscopic occlusion; and establish a neural differential equation for occlusion-power generation mapping to convert the dynamic occlusion effect into a time-varying perturbation term of current output.

[0058] The self-correcting evaluation closed-loop module is used to deploy a dual-channel difference sensor to continuously compare the micro-fluctuation characteristics of physical power generation data and digital twin predictions; and to configure a small-sample incremental learning module to update the dynamic contamination probability field model when feature shifts caused by unknown pollutants are detected.

[0059] Specifically, this embodiment solves the problem of uneven distribution of microscopic contaminants on the surface of photovoltaic modules leading to deviations in power generation efficiency assessment in the background art by having four core modules work together.

[0060] The implementation process of the micro-macro coupled modeling module is as follows. First, a dynamic contamination probability field model is established to obtain initial geometric data of the photovoltaic module surface, such as surface roughness parameters collected by a 3D laser scanner, and material parameters such as the friction coefficient of the glass cover. The types of contaminants are defined as dust and pollen, with dust particle size ranging from 0.5 to 10 micrometers and pollen particle size from 20 to 50 micrometers. Physical properties such as density are taken as 2650 kg / m³ for dust and 1200 kg / m³ for pollen. The contaminant deposition process is simulated based on the discrete element method (DEM), and the motion equation of the contaminant particles is:

[0061]

[0062] Where m i Let x be the mass of the i-th pollutant particle. i Let F be the spatial coordinates of the particle, t be time, and F be the time. g,i For gravity, F c,i F is the interparticle collision force. a,i The force represents the adhesion to the component surface. The impact force is calculated using the Hertzian contact model.

[0063]

[0064] In the formula, k is the stiffness coefficient, δ i Let be the particle overlap, and c be the damping coefficient. Adhesion is modeled using van der Waals forces.

[0065]

[0066] Where A is the Hamaker constant, r i h is the particle radius. iis the distance between the particle and the surface. The deposition trajectory and adhesion force distribution of the pollutants over time are simulated by the above equation, and a three-dimensional micro-shading topological map is generated.

[0067] Then, a cross-scale feature fusion technique is used to collect macro-weather data, including real-time irradiance ranging from 0 to 1000 watts per square meter, ambient temperature ranging from -20 to 40 degrees Celsius, and ambient humidity ranging from 10% to 90%. The local geometric features of the three-dimensional micro-shading topological map are spatiotemporally aligned with the macro-weather data, for example, by timestamp synchronization to ensure that the data collection time error is less than 0.1 seconds, and spatially by coordinate conversion to map the local coordinates of the micro-topological map to the monitoring area of the macro-weather data. A multi-twin body correlation model is constructed to analyze the local light intensity gradient distribution, for example, when the micro-shading area overlaps with a high irradiance area, the light intensity attenuation rate of that area is calculated.

[0068] The implementation process of the bionic visual perception module is as follows. A polarization-multiple spectral imaging module is deployed on the physical end, which includes a polarization camera, a multiple spectral camera, and a thermal infrared camera. Multi-angle polarization imaging is performed on the surface of the photovoltaic module, for example, from four angles of 0 degrees, 30 degrees, 60 degrees, and 90 degrees, to obtain polarization degree and polarization angle information. Multiple spectral images covering the 400 to 1000 nanometer waveband are collected, and the reflectivity difference at different wavelengths is analyzed, for example, dust has higher reflectivity than pollen at 800 nanometer waveband. The surface temperature distribution is obtained through thermal infrared imaging with a resolution of 640x512 pixels, and potential hot spot areas with a temperature higher than the surrounding temperature by more than 5 degrees Celsius are identified.

[0069] A dirt feature extractor based on a generative adversarial network is developed, the generator uses a U-Net structure, the input is a fusion matrix of polarization reflection features and local thermal radiation distribution, and the output is a preliminary dirt mask map. The discriminator is a convolutional neural network that classifies generated masks and real masks. The objective function of the adversarial training is:

[0070]

[0071] where x is the real dirt mask, z is the input optical signal, G is the generator, and D is the discriminator. After 100 rounds of training, a high-resolution dirt mask map with a resolution of 2048x2048 pixels is output.

[0072] The implementation process of the shading effect dynamic migration module is as follows. A virtual light particle tracking engine is designed, and the micro-shading topological structure of the photovoltaic module surface is defined in the digital twin, with a data format of a triangular facet model. A dynamic light cluster is generated containing 1 million light particles, each particle's direction follows the distribution determined by the solar elevation angle and azimuth angle, and the energy conforms to the AM1.5 standard spectrum. Based on the Monte Carlo method, the photon transmission is simulated, and the interaction probability between the photon and the shading object is:

[0073] P interact = 1 - exp(-τl),

[0074] where τ is the extinction coefficient of the obscuration, and l is the path length of the photon in the obscuration. The scattering angle follows the Henyey-Greenstein phase function:

[0075]

[0076] where g is the asymmetry factor, and θ is the scattering angle. The path of each photon is traced, and the loss of luminous flux reaching the photosensitive area is calculated.

[0077] A neural differential equation is established to map the obscuration-power generation, in the form of:

[0078]

[0079] where I(t) is the current output, Φ(t) is the luminous flux, θ is the neural network parameter, and ∈(t) is the time-varying disturbance term caused by dynamic obscuration. The nonlinear relationship is fitted through an LSTM network, converting the obscuration effect into a current disturbance.

[0080] The implementation process of the self-correcting evaluation closed-loop module is as follows. A dual-channel difference perceiver is deployed, with a physical end collecting current and voltage data at a sampling frequency of 1 kHz, and a digital twin outputting predicted data. Wavelet transform is used to extract the microscopic fluctuation characteristics of both, and decomposition is performed to the 1024 Hz frequency band. A difference threshold is set, for example, a prediction error exceeding 5% triggers the small sample incremental learning module. This module uses the MAML algorithm, trains based on 50 new contamination samples, and updates the parameters of the dynamic contamination probability field model, such as the adhesion coefficient, to adapt the model to the deposition characteristics of unknown contaminants, such as bird droppings.

[0081] This embodiment realizes precise modeling and dynamic evaluation of microscopic contamination distribution through multi-module cooperation, solves the problem that traditional methods cannot quantify microscopic obscuration effects, and reduces the power generation efficiency evaluation error to within 3%.

[0082] In this embodiment, the process of establishing a dynamic contamination probability field model includes:

[0083] Obtain the initial geometric data and material parameters of the photovoltaic module surface;

[0084] Set the type, particle size distribution, and physical properties of the contaminants;

[0085] Based on the discrete element algorithm, simulate the non-uniform deposition trajectory and adhesion distribution of contaminants on the surface of the photovoltaic module over time;

[0086] According to the simulation results, generate a three-dimensional microscopic obscuration topology map of the photovoltaic module surface with real-time contamination status.

[0087] Specifically, the initial geometric data of the photovoltaic module surface is obtained by a three-dimensional white light interferometer, and the root mean square roughness of the surface profile is 0.2-0.5 microns. Material parameters include the Young's modulus of glass 72 GPa and the Poisson's ratio 0.22. Set the type of pollutants, such as dust, to contain SiO2 particles accounting for 60%, Al2O3 particles accounting for 30%, and other components accounting for 10%. The particle size distribution conforms to the lognormal distribution d~logN(μ,σ 2 ) where μ=1.2, σ=0.3. Physical properties such as the density of SiO2 is 2650 kg per cubic meter.

[0088] Based on the discrete element algorithm to simulate the deposition process, first initialize the spatial position of 10000 pollutant particles, randomly distributed 1 meter above the surface of the module. The simulation time step is set to 1e-6 seconds, and the gravity, wind force and surface adhesion force acting on each particle are calculated. For example, the wind force adopts a Gaussian distribution model, when the wind speed is 2-8 meters per second, the horizontal force F w =0.5C d ρAv 2 where C d is the drag coefficient, 0.44, ρ is the air density, A is the particle wind area, and v is the wind speed. The simulation lasts for 24 hours, and the deposition trajectory and final position of the particles are recorded to generate a three-dimensional microscopic shading topology map, which is stored in point cloud format with a point spacing of 1 micron.

[0089] Through this embodiment, the deposition state of pollutants in different environments can be accurately simulated, providing high-fidelity microscopic shading data for subsequent evaluation.

[0090] In this embodiment, the process of constructing a multi-twin body correlation model using cross-scale feature fusion technology includes:

[0091] Collecting real-time irradiance, environmental temperature and environmental humidity and other macro-weather data at the location of the photovoltaic module;

[0092] High-precision spatiotemporal alignment of local geometric features of the three-dimensional microscopic shading topology map with macro-weather data;

[0093] By constructing a multi-twin body correlation model, the local light intensity gradient distribution formed by the combined action of microscopic shading and macro-weather on the surface of the photovoltaic module is analyzed.

[0094] Specifically, macro-weather data is collected through a weather station installed next to the photovoltaic array. Real-time irradiance measurement range 0-1500 W / m2, accuracy ±5 W / m2 Environmental temperature measurement range -40-80 degrees Celsius, accuracy ±0.5 degrees Celsius Environmental humidity measurement range 0-100% RH, accuracy ±3% RH Data sampling frequency is once every 1 minute.

[0095] The local geometric features of the three-dimensional micro-shading topological map are spatiotemporally aligned with the macro-weather data. The timestamps of the two are synchronized in time through the NTP protocol, ensuring an error of less than 1 second. In space, the local coordinates of the micro-topological map, such as millimeter-level coordinates with the lower left corner of the component as the origin, are converted into geographic coordinates of the macro-weather data, such as the WGS84 coordinate system. The conversion formula is:

[0096] X = X0 + x cos a - y sin a,

[0097] Y = Y0 + x sin a + y cos a,

[0098] where X0, Y0 are the geographic coordinates of the center of the component, x, y are the local coordinates, and a is the installation azimuth angle of the component.

[0099] The multi-twin body correlation model is constructed using a graph neural network. The micro-shading features and the macro-weather features are input as nodes, and the correlation weights of the two are learned through an attention mechanism. The model outputs the local light intensity gradient distribution, for example, the light intensity gradient can reach 50 watts per square meter per millimeter at the edge of the shading area.

[0100] This embodiment realizes the deep fusion of micro and macro data, enabling the model to capture both local shading details and overall environmental influences.

[0101] In this embodiment, the process of capturing the polarization reflection characteristics and local thermal radiation distribution of the photovoltaic module surface includes:

[0102] Use the polarization-multipolar imaging module to perform multi-angle polarization imaging on the surface of the photovoltaic module to obtain the polarization information of the surface contamination;

[0103] Use the polarization-multipolar imaging module to obtain the multi-spectral image of the surface of the photovoltaic module and analyze its reflection characteristics at different wavelengths;

[0104] Use the polarization-multipolar imaging module to perform thermal infrared imaging to obtain the local thermal radiation distribution of the surface of the photovoltaic module and identify potential hot spot areas.

[0105] Specifically, the polarization-multipolar imaging module is used for data acquisition. The module is installed on a rotatable gimbal and can be adjusted from 0 to 180 degrees. Multi-angle polarization imaging uses a linear polarization camera to capture images at 0, 45, 90, and 135 degrees, respectively, and calculates the degree of polarization where I max and I min are the light intensities in the maximum and minimum polarization directions.

[0106] Multi-spectral imaging selects four wavebands of 450 nm, 550 nm, 650 nm, and 850 nm, which are switched through a filter wheel, and analyzes the reflectivity of each waveband For example, the reflectivity of dust at 850 nanometer wavelength band is 20% to 30% higher than that of a clean surface.

[0107] The thermal infrared imaging adopts a non-cooled focal plane detector, with a resolution of 640x512 pixels, a temperature measurement range of-20 to 150 degrees Celsius, and an accuracy of ±2 degrees Celsius. By shooting, the surface temperature field is obtained, and the area with a temperature exceeding 50 degrees Celsius is identified as a potential hot spot, for example, the center temperature of the blocked area can be 10 to 20 degrees Celsius higher than the surrounding temperature.

[0108] The present embodiment provides rich raw data for the extraction of dirt features through multi-dimensional optical detection, thereby improving the accuracy of dirt identification.

[0109] In the present embodiment, the process of developing a dirt feature extractor based on a generative adversarial network and converting the optical signal into a dirt mask map includes:

[0110] The polarization reflection features and the local thermal radiation distribution are obtained as input optical signals;

[0111] A generator of the generative adversarial network is used to generate a preliminary dirt mask map from the input optical signals;

[0112] A discriminator of the generative adversarial network is used to distinguish the preliminary dirt mask map from the real dirt mask map;

[0113] Through the adversarial training between the generator and the discriminator, the dirt feature extractor is optimized, and a high-resolution dirt mask map is output.

[0114] Specifically, the input optical signals include a polarization reflection feature matrix with a size of 2048x2048x3, which contains a polarization degree, a polarization angle, and a reflectivity, and a local thermal radiation distribution matrix with a size of 2048x2048x1, which is a temperature value.

[0115] The generator adopts an improved U-Net structure, including an encoder and a decoder, each with 8 layers. The encoder extracts features through convolution operations, and the decoder restores the resolution through deconvolution. A skip connection is set between the encoder and the decoder. The output of the generator is a preliminary dirt mask map, with a pixel value range of 0 to 1, and 1 indicating the presence of dirt.

[0116] The discriminator is a 7-layer convolutional neural network, with an input of a mask map and an output of a probability value of 0 to 1, where 1 represents a real mask and 0 represents a generated mask. During adversarial training, the loss function of the generator is L G = E[log(1-D(G(z)))] and the loss function of the discriminator is L D =-E[log(D(x))+log(1-D(G(z)))] using the Adam optimizer with a learning rate of 0.0002 and 10,000 iterations.

[0117] The overlap degree between the dirt mask map output by the trained dirt feature extractor and the real dirt can reach more than 90%, realizing high-resolution dirt visualization.

[0118] In this embodiment, the process of designing a virtual light particle tracking engine to simulate the distortion of the photon transmission path under microscopic shading includes:

[0119] Defining the microscopic shading topology of the photovoltaic component surface in the digital twin;

[0120] Generating dynamic light clusters with specific direction and energy distribution;

[0121] Based on the Monte Carlo method, simulate the scattering, absorption and transmission behavior of dynamic light clusters in the microscopic shading topology;

[0122] Track the path changes of each photon, calculate the loss of light flux and the distortion of light intensity distribution reaching the light-sensitive area of the photovoltaic component.

[0123] Specifically, the microscopic shading topology is defined in the digital twin, and the three-dimensional microscopic shading topology graph is converted into a triangular mesh model, with each triangle having a side length of 1 micrometer, including vertex coordinate normal vector and reflectivity attribute.

[0124] When generating dynamic light clusters, the direction distribution of light particles is calculated based on the position of the sun, for example, when the solar elevation angle is 30 degrees and the azimuth angle is 180 degrees, the zenith angle of the light is 30 degrees and the azimuth angle is 180 degrees. The energy distribution of the light particles conforms to E(λ)=1.0W / m 2 / nmfor400nm≤λ≤1100nm.

[0125] Based on the Monte Carlo method, simulate photon transmission, the initial weight of each photon is 1, when the photon collides with the shading object, calculate the absorption and scattering ratio according to the material properties of the shading object, for example, the absorption rate of dust particles is 0.3, and the scattering rate is 0.7. The scattering angle is sampled by the Henyey-Greenstein function, and the g value is 0.7. When the photon weight is less than 0.01, stop tracking.

[0126] Calculate the light flux loss ΔΦ=Φ0-Φ received , where Φ0 is the incident light flux, Φ received is the light flux reaching the light-sensitive area. For example, under the shadow of a 50-micron-diameter shading object, the light flux loss can reach 80%.

[0127] This embodiment accurately simulates the impact of microscopic shading on light transmission, providing reliable light intensity distribution data for power generation assessment.

[0128] In this embodiment, the process of establishing a neural differential equation for shading-power generation mapping includes:

[0129] obtaining photonic transport path distortion data and electrical parameters of the photovoltaic module;

[0130] defining an initial state and a dynamic evolution function of the neural differential equation;

[0131] fitting a nonlinear mapping relationship between the photonic transport path distortion and the current output of the photovoltaic module through end-to-end learning;

[0132] dynamically converting the shading effect into a time-varying disturbance term of the current output of the photovoltaic module at different times.

[0133] Specifically, obtaining the photonic transport path distortion data includes light flux loss ΔΦ and light intensity distribution distortion δI(x, y) and electrical parameters of the photovoltaic module such as short-circuit current I sc = 8.2 A open-circuit voltage V oc = 42 V series resistance R s = 0.5 Ω parallel resistance R sh = 1000 Ω.

[0134] The initial state of the neural differential equation is defined as I(0) = I0, where I0 is the current without shading. The dynamic evolution function f(I(t), Φ(t), θ) adopts a 3-layer fully connected neural network, the input is I(t) and Φ(t), the output is the current change rate, and θ is the network parameter.

[0135] The time-varying disturbance term ∈(t) is calculated by:

[0136] ∈(t) = k·ΔΦ(t)·exp(-τt),

[0137] where k is the proportional coefficient and takes 0.02, and τ is the decay coefficient and takes 0.1. Through end-to-end learning, the neural network is trained by using the Adam optimizer, the loss function is L = ∑(I pred -I real ) 2 , and the training epochs is 50.

[0138] The present embodiment realizes accurate mapping of the shading effect to the change of power generation, and the current prediction error can be controlled within 2%.

[0139] In the present embodiment, the process of deploying a dual-channel difference perceiver to compare the microscopic fluctuation characteristics of the physical end power generation data and the predicted values of the digital twin includes:

[0140] real-time acquisition of actual power generation current and voltage data of the photovoltaic module from the physical end;

[0141] acquiring power generation current and voltage data predicted by the neural differential equation based on the shading-power generation mapping from the digital twin;

[0142] At the micro scale, the time series of actual power generation data and predicted power generation data are extracted for features;

[0143] The micro-fluctuation features of the actual power generation data and the predicted power generation data are compared to identify the difference patterns.

[0144] Specifically, the power generation data is obtained in real time from the physical end, which is realized through current sensors and voltage sensors connected in series and parallel on photovoltaic components, with a sampling frequency of 1 kHz and a data precision of 16 bits. The predicted data is obtained from the digital twin, with an output frequency consistent with that of the physical end.

[0145] At the micro scale, the time series features are extracted, and the current and voltage data are decomposed into four frequency bands of 128 Hz, 256 Hz, 512 Hz, and 1024 Hz using wavelet transform. The energy proportion and peak factor of each frequency band are calculated. For example, the energy proportion of the 128 Hz frequency band can reflect the low-frequency fluctuation caused by shading.

[0146] When comparing the micro-fluctuation features, the feature differences between the actual data and the predicted data in each frequency band are calculated, and a difference index D = ∑|E real,i -E pred,i |, where E real,i and E pred,i are the energy proportions of the i-th frequency band. When D exceeds 0.1, it is determined that there is a significant difference.

[0147] This embodiment can quickly identify and evaluate the deviation between the model and the actual state, providing a trigger signal for model correction.

[0148] In this embodiment, the process of configuring a small sample incremental learning module to update the dynamic soiling probability field model includes:

[0149] When the dual-channel difference perceiver detects feature deviation caused by unknown contaminants, the small sample incremental learning module is triggered;

[0150] Based on a small amount of newly acquired soiling sample data, the dynamic soiling probability field model is incrementally trained;

[0151] Combining the original data with the new data, the parameters of the soiling probability field are adaptively adjusted to improve the recognition and prediction of new soiling patterns by the model.

[0152] Specifically, when the dual-channel difference perceiver detects feature deviation and determines that it is caused by unknown contaminants, the small sample incremental learning module is triggered. The newly acquired soiling sample data includes 50 polar-multiple spectral images and corresponding actual power generation data, with a collection time interval of 1 hour.

[0153] Incremental training adopts a model parameter fine-tuning based method, freezing the first 5 layers of the dynamic soiling probability field model and only training the last 2 layers. The loss function is L = L old + aL new , where L old is the loss of original data, L new is the loss of new samples, and a is the weight taking 0.8.

[0154] The parameters of the adaptive soiling probability field such as the pollutant adhesion force coefficient and the deposition rate are adjusted, and the optimal parameter values are found by the Bayesian optimization algorithm to reduce the prediction error of the model to within 5% for new soiling patterns.

[0155] This embodiment enables the system to quickly adapt to unknown pollutant types and maintain long-term evaluation accuracy.

[0156] In this embodiment, the system further comprises:

[0157] The performance evaluation module is used to combine the multi-twin body correlation model, the soiling mask map, the current output time-varying disturbance term, and the updated dynamic soiling probability field model to comprehensively evaluate the real-time power generation efficiency, performance degradation degree, and soiling influence of the photovoltaic module.

[0158] The performance evaluation module is used to generate a refined performance report of the photovoltaic module, indicating the local performance loss caused by soiling and suggesting maintenance measures.

[0159] Specifically, the performance evaluation module combines the local light intensity gradient soiling mask map output by the multi-twin body correlation model, the current output time-varying disturbance term, and the updated dynamic soiling probability field model to perform comprehensive evaluation.

[0160] The real-time power generation efficiency is calculated as where I and V are the actual current and voltage, and P in is the incident light power. The performance degradation degree is obtained by comparing with the initial state efficiency Δη = η0- η. The soiling influence evaluation calculates the power loss caused by each soiling area ΔP = ∑(P clean -P soiled ).

[0161] The generated refined performance report contains a color-coded soiling influence heat map, for example, red represents areas with power loss exceeding 10%, and gives maintenance suggestions such as cleaning in areas with power loss exceeding 5%.

[0162] This embodiment provides comprehensive performance evaluation results, providing accurate guidance for operation and maintenance decisions.

[0163] In summary, the application solves the problem that the traditional model cannot quantify the micro-pollutant shielding effect by using the micro-macro coupling modeling module to simulate the non-uniform deposition of pollutants by the discrete element algorithm and generate a three-dimensional micro-shielding topology graph, and combining the cross-scale feature fusion technology to align it with the macro weather data; the bionic visual perception module improves the accuracy of the fouling identification by extracting high-resolution fouling features through polarization-multiple spectral imaging and generative adversarial networks; the shielding effect dynamic migration module realizes the accurate mapping of micro-shielding to power generation changes by means of virtual light particle tracking and neural differential equations; the self-correcting evaluation closed-loop module can dynamically update the model to adapt to unknown pollutants through double-channel comparison and small sample learning. These modules work together to significantly reduce the evaluation deviation of power generation efficiency, accurately identify the performance degradation caused by micro-pollution, provide reliable basis for operation and maintenance decision-making, and generate a detailed report to guide maintenance.

[0164] It should be noted that the relational terms herein, such as first and second, are used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any such actual relationship or order between such entities or actions. Moreover, the terms "comprises", "comprising", or any other variations thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but can include other elements not expressly listed or inherent to such process, method, article, or apparatus.

[0165] Although embodiments of the present application have been shown and described, it is to be understood that various modifications, substitutions, replacements and changes can be made to these embodiments without departing from the principles and spirit of the present application, and the scope of the present application is defined by the appended claims and their equivalents.

Claims

1. A photovoltaic power generation intelligent evaluation system based on digital twinning, characterized in that, The system comprises: a micro-macro coupling modeling module for establishing a dynamic contamination probability field model of a photovoltaic component surface, the dynamic contamination probability field model simulating a non-uniform deposition process of contaminants on the photovoltaic component surface by a discrete element algorithm to generate a three-dimensional micro-shading topological graph, and adopting a cross-scale feature fusion technology to perform spatio-temporal alignment of the three-dimensional micro-shading topological graph and macro-weather data, and to construct a multi-twin body correlation model, the multi-twin body correlation model having a local light intensity gradient analysis capability; a bionic visual perception module for deploying a polarization-multiple spectrum imaging module on a physical end to capture polarization reflection features and local thermal radiation distribution of the photovoltaic component surface, and developing a dirty feature extractor based on a generative adversarial network to convert the polarization reflection features and the local thermal radiation distribution into a high-resolution contamination mask graph; a shading effect dynamic migration module for designing a virtual light particle tracking engine to inject a dynamic light cluster optimized by a Monte Carlo method into a digital twin body to simulate a photon transmission path distortion under micro-shading, and establishing a neural differential equation of shading-power generation mapping to convert a dynamic shading effect into a time-varying disturbance term of current output; a self-correcting evaluation closed loop module for deploying a dual-channel difference perceiver to continuously compare micro-fluctuation features of physical end power generation data and digital twin body prediction values, and configuring a small sample incremental learning module to update the dynamic contamination probability field model when detecting feature deviation caused by unknown contaminants.

2. The photovoltaic power generation intelligent evaluation system based on digital twinning according to claim 1, characterized in that, The process of establishing the dynamic contamination probability field model comprises: acquiring initial geometric data and material parameters of the photovoltaic component surface; setting a contaminant type, a particle size distribution and physical properties; simulating a non-uniform deposition trajectory and adhesion force distribution of the contaminants on the photovoltaic component surface over time based on a discrete element algorithm; generating a three-dimensional micro-shading topological graph with real-time contamination states on the photovoltaic component surface according to the simulation results.

3. The photovoltaic power generation intelligent evaluation system based on digital twinning according to claim 1, characterized in that, The process of constructing the multi-twin body correlation model by the cross-scale feature fusion technology comprises: collecting real-time irradiance, environmental temperature and environmental humidity and other macro-weather data of a location where the photovoltaic component is located; performing high-precision spatio-temporal alignment of local geometric features of the three-dimensional micro-shading topological graph and the macro-weather data; analyzing a local light intensity gradient distribution of the photovoltaic component surface formed by the combined action of micro-shading and macro-weather by constructing the multi-twin body correlation model.

4. The photovoltaic power generation intelligent evaluation system based on digital twinning according to claim 1, characterized in that, The process of capturing the polarization reflection features and the local thermal radiation distribution of the photovoltaic component surface comprises: acquiring polarization information of surface contamination by multi-angle polarization imaging of the photovoltaic component surface using the polarization-multiple spectrum imaging module; acquiring a multi-spectrum image of the photovoltaic component surface using the polarization-multiple spectrum imaging module to analyze reflection characteristics thereof at different wavelengths; performing thermal infrared imaging using the polarization-multiple spectrum imaging module to acquire a local thermal radiation distribution of the photovoltaic component surface and identify potential hot spot areas.

5. The photovoltaic power generation intelligent evaluation system based on digital twinning according to claim 1, characterized in that, The process of developing the dirty feature extractor based on the generative adversarial network and converting the optical signals into the contamination mask graph comprises: acquiring the polarization reflection features and the local thermal radiation distribution as input optical signals; generating a preliminary contamination mask graph from the input optical signals using a generator of the generative adversarial network; The discriminator of the adversarial generative network distinguishes the preliminary dirt mask map from the real dirt mask map; Through the adversarial training between the generator and the discriminator, the dirt feature extractor is optimized to output a high-resolution dirt mask map.

6. The photovoltaic power generation intelligent evaluation system based on digital twinning according to claim 1, characterized in that, The process of simulating the distortion of the photon transmission path under micro-shading by the designed virtual light particle tracking engine includes: Defining the micro-shading topological structure of the photovoltaic component surface in the digital twin; Generating dynamic light clusters with specific direction and energy distribution; Based on the Monte Carlo method, simulating the scattering, absorption and transmission behavior of dynamic light clusters in the micro-shading topological structure; Tracking the path changes of each photon, calculating the loss of light flux and the distortion of light intensity distribution reaching the light-sensitive area of the photovoltaic component.

7. The photovoltaic power generation intelligent evaluation system based on digital twinning according to claim 1, characterized in that, The process of establishing a neural differential equation for shading-power generation mapping includes: Obtaining photon transmission path distortion data and electrical parameters of the photovoltaic component; Defining the initial state and dynamic evolution function of the neural differential equation; Through end-to-end learning, fitting the nonlinear mapping relationship between photon transmission path distortion and photovoltaic component current output; Dynamically convert shading effects into time-varying disturbance terms of photovoltaic component current output at different times.

8. The photovoltaic power generation intelligent evaluation system based on digital twinning according to claim 1, characterized in that, The process of deploying a dual-channel difference perceiver to compare the micro-fluctuation characteristics of physical end power generation data and digital twin prediction values includes: Real-time acquisition of actual power generation current and voltage data of the photovoltaic component from the physical end; Obtain the power generation current and voltage data predicted by the neural differential equation based on the shading-power generation mapping from the digital twin; On a micro scale, feature extraction is performed on the time series of actual power generation data and predicted power generation data; Compare the micro-fluctuation characteristics of actual power generation data and predicted power generation data to identify difference patterns.

9. The photovoltaic power generation intelligent evaluation system based on digital twinning according to claim 1, characterized in that, The process of configuring a small sample incremental learning module to update the dynamic dirt probability field model includes: When the dual-channel difference perceiver detects feature deviation caused by unknown contaminants, trigger the small sample incremental learning module; Based on a small amount of newly acquired dirt sample data, incrementally train the dynamic dirt probability field model; Combine the original data and new data to adaptively adjust the parameters of the dirt probability field and improve the model's recognition and prediction of new dirt patterns.

10. The photovoltaic power generation intelligent evaluation system based on digital twinning of claim 1, wherein The system also includes: A performance evaluation module for comprehensive evaluation of the real-time power generation efficiency, performance degradation and dirt impact of the photovoltaic component in combination with the multi-twin correlation model, the dirt mask map, the current output time-varying disturbance term, and the updated dynamic dirt probability field model. The performance evaluation module generates a refined performance report of the photovoltaic component, indicating the local performance loss caused by dirt and suggesting maintenance measures.

Images