A photovoltaic dust thickness real-time inversion method and system

Abstract

The application discloses a photovoltaic dust thickness real-time inversion method and system, relates to the intelligent operation and maintenance technical field of photovoltaic power stations, and solves the problems of poor environmental adaptability and insufficient generalization ability caused by the lack of physical model constraints in the existing photovoltaic dust thickness monitoring technology; the application comprises the following steps: collecting multi-modal data such as thermal infrared and SAR; extracting thermal characteristics and dielectric characteristics based on a thermal-dust physical coupling model; dynamically and adaptively fusing the multi-modal characteristics; inputting the fused characteristics into a physically constrained neural network to realize real-time inversion of dust thickness; and outputting a cleaning instruction according to the inversion result; the application deeply fuses a physical model and a data-driven method, adopts a multi-modal dynamic weight mechanism, significantly improves the precision, robustness and physical credibility of the inversion result, realizes low-power consumption and real-time monitoring at an edge end, and has important engineering application value.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance technology for photovoltaic power plants, specifically to a method and system for real-time inversion of photovoltaic dust thickness. Background Technology

[0002] With the growing global demand for renewable energy, photovoltaic (PV) power generation, especially large-scale PV power plants built in desert regions rich in sunshine, has become an important part of energy structure transformation. However, frequent sandstorms in desert areas lead to the accumulation of dust on the surface of PV modules. This not only blocks sunlight and significantly reduces photoelectric conversion efficiency but also creates hot spots, accelerating module aging and severely impacting the power generation efficiency and equipment lifespan of PV power plants. Therefore, real-time and accurate monitoring of the dust thickness on the surface of PV modules is a key technological prerequisite for achieving intelligent and refined clean operation and maintenance, improving power generation, and ensuring the long-term stable operation of power plants. To this end, the industry has developed various monitoring technologies, but many challenges remain.

[0003] Specifically, to address the aforementioned dust monitoring issues, the mainstream monitoring technologies currently include the following categories:

[0004] The first category is optical remote sensing monitoring technology. For example, Chinese patent CN108801198A discloses a method that uses a visible light or near-infrared camera to capture images of a component surface and then uses a convolutional neural network (CNN) to analyze image grayscale changes to estimate dust thickness. A significant drawback of this method is its high dependence on stable lighting conditions. In environments with frequent dust storms or drastic changes in lighting, atmospheric particulate scattering can severely reduce the image signal-to-noise ratio (SNR) (as Li et al., Solar Energy Materials 2021, indicated, the SNR can decrease by more than 60%), thus affecting the accuracy of the inversion. Furthermore, its calibration model typically does not consider the thermal conductivity of dust on the component temperature, resulting in thickness estimation errors exceeding 1 mm in scenarios with large diurnal temperature variations. Simultaneously, deploying dedicated optical sensing networks is costly, limiting its large-scale application.

[0005] The second category is contact sensing technology. For example, the technology disclosed in US patent application US20200348115A1 uses a micron-scale piezoelectric thin-film array to directly measure pressure changes caused by sand and dust deposition. While this method can achieve high accuracy at sensor locations, the sensor placement is typically sparse (e.g., 5cm intervals), leading to a high rate of missed detections (exceeding 30%) for irregularly shaped dust deposits. More seriously, in harsh environments such as deserts, the sensors themselves are susceptible to wear and tear, resulting in a high annual failure rate (statistically up to 17.3%). Furthermore, environmental factors such as strong winds can introduce mechanical vibration and noise, interfering with measurement accuracy (as described in Zhang et al., IEEE Sensors Journal 2022). The high cost of single-station maintenance also limits its engineering practicality.

[0006] The third category is based on thermal infrared monitoring technology. For example, the research by Liang Jianfeng et al. (2024) detects dust accumulation areas by capturing anomalies in the temperature field of photovoltaic panels. The main problem is that existing models are mostly static temperature threshold models, failing to establish a deep understanding of the dynamic physical coupling relationship between the thermal resistance effect of dust and factors such as photovoltaic panel tilt angle, ambient temperature, and solar irradiance. This leads to a sharp increase in the false alarm rate during the midday high-temperature period, as the model cannot accurately decouple dust coverage from normal temperature rise.

[0007] The fourth category is data-driven deep learning methods. For example, the research by Niu Xiaoyu et al. (2024) utilizes deep learning architectures such as U-Net to process multimodal data. These methods perform well in laboratory environments, but they have two core problems: First, the model's generalization ability is insufficient. Since the training data is mostly artificially simulated, it has poor adaptability to changes in sand particle size, humidity, and component tilt angle in real desert environments, leading to a significant increase in field test errors. Second, it is highly dependent on computing resources. Complex models (such as those with more than 10M parameters) usually require cloud GPU support, which brings problems of high power consumption (exceeding 200W) and high latency (inference time can reach the second level), which is seriously inconsistent with the operation and maintenance requirements of photovoltaic power plants for edge-end, low-power, and real-time response (millisecond level).

[0008] However, after in-depth research and practice on existing technologies, it was found that the current photovoltaic dust thickness monitoring technology system has the following systemic technical defects:

[0009] (1) Poor accuracy and environmental adaptability: All kinds of single sensing technologies have obvious shortcomings: optical methods are limited by light and weather, contact methods have problems of incomplete spatial coverage and poor reliability, while traditional thermal infrared methods are prone to misjudgment in complex working conditions due to oversimplification of the model.

[0010] (2) Lack of completeness of model mechanism: Existing technologies generally lack in-depth modeling of the thermodynamic coupling mechanism of dust and photovoltaic panels. Whether it is empirical calibration curves or pure data-driven deep learning models, they are not constrained by basic physical laws such as heat conduction, resulting in insufficient generalization ability in the ever-changing real environment and difficulty in ensuring the physical credibility of the inversion results.

[0011] (3) The contradiction between real-time performance and power consumption: The complex algorithm model relies on cloud computing and cannot meet the low power consumption and real-time decision-making requirements of the edge.

[0012] (4) Poor engineering economics: The cloud computing-dependent solution has the disadvantages of high latency and high power consumption, which cannot meet the needs of real-time decision-making at the edge; while the high-cost dedicated sensor network solution limits its large-scale deployment and application in large photovoltaic power plants. Summary of the Invention

[0013] To address the problems existing in the prior art, this invention provides a real-time photovoltaic dust thickness inversion method and system, which solves the problems of poor environmental adaptability and insufficient generalization ability due to the lack of physical model constraints in existing photovoltaic dust thickness monitoring technologies.

[0014] A method for real-time inversion of photovoltaic dust thickness includes:

[0015] Step 1: Acquire multimodal data, including at least thermal infrared image data and synthetic aperture radar data;

[0016] Step 2: Based on the pre-built thermal-dust physics coupling model, extract thermal and dielectric features from the multimodal data;

[0017] Step 3: Perform dynamic weighted fusion of the thermal and dielectric characteristics to obtain the fused characteristics;

[0018] Step 4: Input the fused features into a pre-trained physical constraint neural network to perform real-time inversion of sand and dust thickness;

[0019] Step 5: Output cleaning instructions and perform closed-loop feedback analysis based on the real-time inversion results.

[0020] Furthermore, the multimodal data also includes environmental calibration parameter data, involving ambient temperature, relative humidity, wind speed, and real-time total solar irradiance.

[0021] The synthetic aperture radar data is obtained by acquiring the backscattering coefficient matrix of the photovoltaic module array surface through a low-power, miniaturized SAR sensor. This data is sensitive to the surface dielectric constant and micro-roughness.

[0022] The thermal infrared image data is obtained by using a thermal infrared camera to capture a two-dimensional temperature field distribution map of the photovoltaic modules in the same area. This data reflects the thermal state of the module surface.

[0023] The environmental calibration parameter data are obtained from meteorological stations or local sensors, and are environmental parameters closely related to the heat exchange process.

[0024] Furthermore, the spatiotemporal registration preprocessing includes: using the pixel coordinate system of the thermal infrared image as a reference, and utilizing pre-calibrated camera intrinsic and extrinsic parameters, registering the SAR data matrix onto the thermal infrared image to achieve sub-pixel-level spatial alignment; simultaneously, adding a synchronized timestamp to all data; subsequently, using the Lee filtering algorithm to suppress speckle noise on the registered SAR data, using the median filtering algorithm to remove random thermal noise on the thermal infrared image, and finally normalizing all data to form a standardized multimodal input dataset.

[0025] Further, step 2 includes:

[0026] Step 2.1: Establish a three-dimensional unsteady-state heat conduction equation: The photovoltaic module covered with sand and dust is abstracted into a multi-layer composite structure, including an encapsulation layer, solar cells, backsheet, and sand and dust layer. Based on Fourier's law of heat conduction, a partial differential equation is established to describe the temperature field T(x, y, z, t) inside and on the surface of the composite structure as a function of time. The thermal conductivity, specific heat capacity, and density of the sand and dust layer are incorporated into the model as key parameters.

[0027] Step 2.2: Quantifying the relationship between dust thickness and thermal resistance: Treating the dust deposition layer as an additional thermal resistance layer, its equivalent thermal resistance ( The relationship between the thickness of the sandstorm (δ) and the thickness of the sandstorm is directly proportional, i.e. ,in Let δ be the equivalent thermal conductivity of the sand. Therefore, the change in sand thickness δ will directly affect the heat conduction process of the entire system, and will ultimately be reflected in the distribution and dynamic rate of change of the temperature field T(x, y, t) on the component surface;

[0028] Step 2.3: Dynamic Boundary Condition Calibration: To ensure the equations have a solution and a unique solution, boundary conditions need to be set. The boundary conditions for the upper surface (in contact with the atmosphere) are complex hybrid boundary conditions, comprehensively considering solar radiation absorption, long-wave radiation heat dissipation to the sky and environment, and convective heat transfer. These processes are all directly related to environmental parameters such as solar irradiance, ambient temperature, and wind speed collected in Step 1. In particular, this invention introduces a photovoltaic panel tilt angle (θ) compensation factor to correct the solar irradiance component and convective heat transfer coefficient received at different tilt angles, thereby enabling the model to adapt to photovoltaic modules with different installation angles.

[0029] Furthermore, the extraction of multimodal features in step 3 includes thermal features and dielectric features;

[0030] The thermal features include two key features extracted from the thermal infrared image sequence: the spatial temperature gradient (∇T) and the rate of temperature change over time (∂T / ∂t). These two features directly correspond to key terms in the heat conduction equation and have clear physical meaning.

[0031] The dielectric characteristics include obtaining the dielectric constant (ε) distribution map of the photovoltaic panel surface by inverting preprocessed SAR data using an improved dielectric model (such as the Oh model). Dust cover can significantly alter the dielectric constant of the surface.

[0032] Furthermore, the dynamic weight adaptive fusion described in step 3 includes:

[0033] Within a specific local region, the data variance of thermal and dielectric characteristics is calculated. A larger variance generally indicates a stronger signal response, richer information content, and a higher signal-to-noise ratio for that mode. Therefore, features of modes with larger variances are assigned higher weights. The fused feature vector... It can comprehensively reflect the dual physical effects of sand and dust in terms of both thermal and electromagnetic aspects.

[0034] Furthermore, the physical constraint neural network described in step 4 employs an encoder-decoder structure. The encoder is responsible for extracting deep abstract features from the multimodal feature map fused in step 3, while the decoder is responsible for mapping the features extracted by the encoder to a pixel-level dust thickness distribution map. The entire network is deeply optimized through techniques such as convolution kernel sharing, channel pruning, and model quantization, compressing the number of model parameters to below 1M to ensure low-power and high-efficiency operation on edge computing devices.

[0035] Furthermore, the training loss function of the physical constraint neural network described in step 4 is:

[0036]

[0037] λ is a hyperparameter used to balance the relationship between data fitting and adherence to physical laws.

[0038] Data-driven loss ( ): The standard mean squared error (MSE) or mean absolute error (MAE) is used to measure the difference between the thickness map predicted by the network and a small amount of real labeled data.

[0039] Physical constraint loss ( The residual terms of the partial differential equation for heat conduction established in step 2 are added as penalty terms to the loss function. Specifically, the dust thickness map output by the network is used as the penalty term. Substituting into the heat conduction equation, calculate the imbalance on both sides of the equation. This imbalance is the physical residual. If the network output does not conform to physical laws, the physical residual will be large, thus imposing a penalty gradient on the network during backpropagation, forcing it to learn a mapping relationship that conforms to the law of heat conduction.

[0040] Furthermore, step 5 includes:

[0041] Based on a preset dust thickness threshold (set to an average thickness exceeding 0.5 mm or a local thickness exceeding 1 mm), a cleaning instruction is automatically generated. The instruction may include the photovoltaic module number that needs cleaning, the coordinates of the dust accumulation area, and the suggested cleaning priority, and this information is sent to the photovoltaic power plant's operation and maintenance control system or intelligent cleaning robot.

[0042] The retrieved dust thickness results are correlated with the actual power generation data of the photovoltaic array. If a systematic deviation is found between the retrieved dust accumulation degree and the theoretical model of power loss, a self-calibration procedure is initiated to fine-tune the key parameters of the heat-dust coupling model in step two (such as the equivalent thermal conductivity of dust, k_dust), thereby forming a data-driven online model optimization closed loop to continuously improve the retrieval accuracy.

[0043] A real-time photovoltaic dust thickness inversion system, including

[0044] The data acquisition unit is used to acquire multimodal data, including at least thermal infrared image data and synthetic aperture radar (SAR) data;

[0045] The feature extraction unit is configured to extract thermal and dielectric features from the multimodal data based on a pre-built thermal-dust physics coupling model.

[0046] A feature fusion unit is configured to perform dynamic weighted fusion of the thermal features and dielectric features to obtain fused features;

[0047] The inversion unit includes a pre-trained physical constraint neural network for receiving the fused features and performing real-time inversion of the dust thickness.

[0048] The beneficial effects of this invention include:

[0049] (1) Significantly improved inversion accuracy and physical reliability: This invention provides a solid theoretical foundation for dust thickness inversion by constructing a physical model based on the heat conduction equation. Furthermore, by embedding this physical equation as a constraint term into the loss function of the neural network, it ensures that the inversion results follow physical laws under any operating conditions, fundamentally avoiding the physical deviation problem caused by pure data-driven models in scenarios with extreme temperature differences or insufficient training data coverage. Experimental results show that under a wide temperature range of -25℃ to 70℃, the mean absolute error (MAE) can be stably maintained within 0.07mm, which is more than 5 times higher than the 0.35mm of the traditional method, ensuring the physical reliability of the monitoring results.

[0050] (2) Enhanced robustness and adaptability in complex environments: This invention innovatively designs a multimodal dynamic weight fusion mechanism. By adaptively adjusting the fusion weights of SAR dielectric features and thermal infrared temperature features, it can automatically prioritize more reliable signal sources in severe weather conditions such as sandstorms and high humidity, where the signal-to-noise ratio of a single sensor drops sharply, significantly improving the stability and anti-interference capability of the features. In addition, the dynamic tilt angle compensation algorithm introduced in the model enables this method to be accurately applied to photovoltaic modules with any tilt angle within the range of 15° to 45°, solving the problem that existing technologies are usually only effective under fixed or small-range tilt angles, and greatly expanding the application scenarios.

[0051] (3) Real-time edge monitoring and low-power operation are achieved: This invention designs a lightweight physical constraint neural network, PhysCNN, by employing technologies such as model compression and operator fusion, and successfully deploys it on a low-power edge computing device. This allows the entire process of data processing and thickness inversion to be completed on-site, achieving a real-time response of 95 milliseconds and a data update frequency of minutes. Compared with cloud computing solutions that often have delays of several seconds or even minutes, the real-time performance of this invention is improved by several orders of magnitude, providing instant data support for the path planning of cleaning robots. At the same time, the operating power consumption of the entire system is as low as 5W, solving the problem of high power consumption of hundreds of watts in cloud solutions, enabling it to be independently powered by a small photovoltaic / energy storage system, suitable for remote desert power stations not covered by the power grid.

[0052] (4) Reduced overall system deployment and maintenance costs: This invention adopts a non-contact remote sensing monitoring method, eliminating the need to install any physical sensors on the photovoltaic panel surface. This fundamentally avoids the damage, malfunctions, and high maintenance costs associated with contact sensors due to harsh environments. Compared to dedicated optical sensor networks with high deployment costs, this invention utilizes relatively low-cost small SAR and thermal infrared modules, combined with edge computing nodes, to reduce the hardware cost per station to 1 / 5 of the traditional optical solution, and the average annual maintenance cost is less than 1 / 3 of the traditional contact solution, demonstrating excellent engineering economics and laying the foundation for large-scale application. Attached Figure Description

[0053] Figure 1 This is a schematic diagram of the overall method flow involved in the embodiments of this application.

[0054] Figure 2 This is a schematic diagram of the thermal-dust physical coupling model involved in the embodiments of this application.

[0055] Figure 3 This is a block diagram of the mechanism for dynamic fusion of multimodal features involved in the embodiments of this application.

[0056] Figure 4 This is a schematic diagram of the computation process of the Physically Constrained Neural Network (PhysCNN) involved in the embodiments of this application.

[0057] Figure 5 This is a schematic diagram of the closed-loop feedback and control process involved in the embodiments of this application. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely represents selected embodiments of this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0059] Example 1

[0060] The following is in conjunction with the appendix Figure 1 Specific embodiments of the present invention will be described in detail;

[0061] This embodiment provides a specific method for high-precision real-time inversion of photovoltaic dust thickness in a photovoltaic power station environment in the Gobi Desert. The hardware environment of this embodiment includes: an NVIDIA Jetson AGXXavier edge computing device deployed next to the photovoltaic array, which is connected to a Ku-band miniaturized synthetic aperture radar (SAR) sensor and a FLIR A65 thermal infrared camera, and simultaneously accesses meteorological station data from the power station via the Modbus protocol. The software environment is an Ubuntu 20.04 operating system, and the algorithm is implemented based on the PyTorch 2.5 deep learning framework and the OpenCV 4.12.0 library.

[0062] like Figure 1 As shown, this method is executed on an edge computing device and includes five core steps:

[0063] Step 1: Perform multimodal data acquisition and preprocessing, receiving externally input SAR data, thermal infrared data, and environmental parameter data;

[0064] Step 2: Construct a thermal-dust physical coupling model;

[0065] Step 3: Perform dynamic fusion of multimodal features;

[0066] Step 4: Perform real-time inversion using a Physically Constrained Neural Network (PhysCNN);

[0067] Step 5: Output the final dust thickness map and generate cleaning instructions or feedback calibration signals.

[0068] The specific implementation process is as follows:

[0069] Step 1: Synchronous Acquisition and Preprocessing of Multimodal Data

[0070] Data Acquisition: The edge computing device synchronously triggers the SAR sensor and thermal infrared camera at a frequency of 1Hz to acquire thermal infrared images with a resolution of 640x512 pixels and the corresponding SAR backscattering coefficient matrix of the region. Simultaneously, it reads ambient temperature, relative humidity, wind speed, and total solar irradiance data from the weather station every minute.

[0071] Spatiotemporal registration: Using the spatial transformation matrix of the thermal infrared camera and SAR sensor obtained in advance by the checkerboard calibration method, an affine transformation is performed on each frame of SAR data to achieve pixel-level alignment with the thermal infrared image in space.

[0072] Noise suppression and normalization: The registered SAR data were processed using a 7x7 window Lee filter algorithm to remove speckle noise; the thermal infrared images were processed using a 5x5 kernel median filter algorithm to remove random thermal noise. Subsequently, all data (including environmental parameters) were linearly mapped to the [0,1] interval using the min-max normalization method.

[0073] Step 2: Construct a thermal-dust physics coupling model, such as Figure 2 As shown, the model uses solar irradiance, ambient temperature, wind speed, and photovoltaic panel tilt angle θ as dynamically changing boundary conditions as inputs. The core of the model is a three-dimensional unsteady-state heat conduction equation incorporating a dust thermal resistance model, where the dust thickness δ directly affects the thermal resistance. The final output of the model is the theoretically calculated temperature field distribution T(x,y,t) on the photovoltaic panel surface, which establishes a physical relationship between the observable quantity (temperature) and the unknown quantity (dust thickness).

[0074] A three-dimensional unsteady-state finite element model of the photovoltaic module's thermal conduction was pre-built in COMSOL Multiphysics software. In the model, the equivalent thermal conductivity of the dust layer... According to the reference, its thermal resistance is set at 0.25 W / (m·K). It is directly proportional to the thickness δ. That is... ,in Let δ be the equivalent thermal conductivity of the sand. Therefore, the change in sand thickness δ will directly affect the heat conduction process of the entire system, and will ultimately be reflected in the distribution and dynamic rate of change of the temperature field T(x, y, t) on the component surface;

[0075] The environmental parameters collected in step one, namely ambient temperature, irradiance, and wind speed, are used as real-time boundary conditions inputs for the model. In particular, the photovoltaic panel tilt angle θ (30° in this embodiment) is used to correct the received solar irradiance component (multiplied by cosθ) and the convective heat transfer coefficient.

[0076] Specifically, a nonlinear relationship is established between the dust thickness δ and the photovoltaic panel temperature field T:

[0077]

[0078] Due to the thickness of the sandstorm, , , For calibration coefficients, For temperature gradient, It is the dielectric constant (SAR inversion value). Effective irradiance (solar radiation value after dust attenuation). Let be the rate of change of the temperature field over time; where the parameter values ​​are set as follows:

[0079] (Effective irradiance attenuation model)

[0080] In the formula, Total irradiance;

[0081] (Temperature difference between surface and environment)

[0082] In the formula, The surface temperature of the photovoltaic panel. The ambient temperature;

[0083] Calibration coefficient range:

[0084] Introducing a photovoltaic panel tilt angle compensation factor (θ∈[15°,45°]), the corrected formula is as follows:

[0085]

[0086] The above formula introduces a compensation factor related to the installation tilt angle (θ) of the photovoltaic panel to correct the received solar radiation and convective heat transfer coefficient, which is the key to achieving full tilt angle adaptability.

[0087] Step 3: Multimodal feature fusion guided by a physical model, such as... Figure 3 As shown, the raw information from SAR data and thermal infrared data is extracted using dielectric and thermal features, respectively. Then, based on the real-time data variance of the two features, a fusion weight w is adaptively calculated. Finally, the extracted dielectric and thermal features are weighted and fused according to the weight w to generate the final fused feature vector used as the input to the physically constrained neural network.

[0088] Specifically, cross-validation and information complementarity of dust data from different physical modes (electromagnetic properties of SAR, thermal properties of TIR) are used to overcome the limitations of a single mode. Furthermore, a dynamic weight (w) allocation mechanism based on feature variance (or signal-to-noise ratio) is designed, enabling the fusion process to adaptively adjust according to real-time operating conditions, rather than using fixed or simply learned weights.

[0089] Extracting dielectric constant from SAR data The spatial temperature gradient (∇T) and temporal temperature change rate (∂T / ∂t) features were extracted from TIR data. The dielectric constant was retrieved from SAR data using an improved Oh model. ( , , (Desert scene optimization parameters)

[0090] Thermal infrared data: Temperature field inversion error ≤ 0.5℃

[0091] Feature-level fusion rules:

[0092]

[0093]

[0094] Where S represents SAR characteristic data (dielectric constant correlation); T represents thermal infrared characteristic data (temperature field correlation); and Var(⋅) represents the data variance (characterizing the signal-to-noise ratio).

[0095] Dynamic weight fusion: Dielectric characteristics are calculated separately within a 5x5 sliding window. The variance of the data for thermal characteristics (taking ∇T as an example).

[0096] Fusion weight w: According to the formula This was calculated dynamically. In a simulated dust storm event in September 2025 (with increased humidity), Var(ε) increased significantly. It adaptively increases from 0.4 to 0.65.

[0097] Final fusion features .

[0098] Step 4: Real-time inversion based on Physically Constrained Neural Network (PhysCNN)

[0099] Network Structure: PhysCNN uses the U-Net architecture, with the encoder containing 5 convolutional blocks and the decoder containing 5 corresponding upsampling and convolutional blocks. All convolutional layers are followed by the ReLU activation function. Through channel pruning, the model's parameter count is compressed to 0.8M.

[0100] like Figure 4 As shown, during the offline training phase, the fused feature vectors are input into a lightweight PhysCNN network to generate a predicted dust thickness map. The total loss function consists of a weighted average of the data loss and the physical residual loss, which is optimized through backpropagation. After training, the model is deployed at the edge for online inference, accelerated by an inference engine, and outputs the dust thickness map in real time.

[0101] In the offline phase, training was performed using a hybrid dataset containing 10,000 sets of simulation data and 200 sets of field-measured data (calibrated using a laser thickness gauge). The total loss function... ,in For MAE loss, This represents the physical residual loss of the heat conduction equation in step two. The weighting factor λ is set to 0.1. The Adam optimizer is used with an initial learning rate of 0.001 and training for 200 epochs.

[0102] The specific loss function is designed as follows:

[0103]

[0104]

[0105] Adaptive weights:

[0106] Boundary conditions:

[0107] The residual terms of the partial differential equation for heat conduction are directly added to the loss function as differentiable penalty terms. This drives the optimization of network parameters through backpropagation, forcing the network output to satisfy a physically consistent solution.

[0108] Step 5: Closed-loop feedback and decision output, such as Figure 5 As shown, the real-time generated dust thickness map is sent to the decision-making module. On the one hand, if the dust thickness is determined to exceed a preset threshold, a cleaning command is generated and sent to the operation and maintenance system. On the other hand, the thickness map is correlated with the actual power generation data of the power plant. If a systematic deviation is found between the inversion results and the power loss, a self-calibration procedure is initiated to fine-tune the core parameters of the physical model, thereby forming a continuously optimized closed-loop control system.

[0109] Based on a preset dust thickness threshold (set to an average thickness exceeding 0.5 mm or a local thickness exceeding 1 mm), a cleaning instruction is automatically generated. The instruction may include the photovoltaic module number that needs cleaning, the coordinates of the dust accumulation area, and the suggested cleaning priority, and this information is sent to the photovoltaic power plant's operation and maintenance control system or intelligent cleaning robot.

[0110] The retrieved dust thickness results are correlated with the actual power generation data of the photovoltaic array. If a systematic deviation is found between the retrieved dust accumulation degree and the theoretical model of power loss, a self-calibration procedure is initiated to fine-tune the key parameters of the heat-dust coupling model in step two (such as the equivalent thermal conductivity of dust). This forms a data-driven online model optimization closed loop, continuously improving inversion accuracy.

[0111] Experimental verification and results

[0112] Experiment 1

[0113] This embodiment was deployed at a photovoltaic demonstration base in Midong District, Xinjiang, China, for a three-month test. During the test, the ambient temperature ranged from 3°C to 50°C, covering sunny days, cloudy days, and two minor sandstorms. By comparing the data with synchronous measurement data from a high-precision laser thickness gauge, the mean absolute error (MAE) of the method in this embodiment was 0.068 mm, the single inference delay was stable at 92 ms, and the average power consumption of the entire system was 4.8 W. The results show that this invention achieves the monitoring objectives of high precision, high real-time performance, and low power consumption.

[0114] Comparative Example 1

[0115] To further verify the scheme proposed in this embodiment, a comparative scheme is set up to verify the significant beneficial effects of the physical constraint loss function and multimodal dynamic weight fusion in this invention through the control variable method. This embodiment sets up two comparative schemes, both tested on the exact same hardware and software platform and dataset as Embodiment 1.

[0116] Comparison with Solution A: A purely data-driven solution without physical constraints

[0117] Setup: This approach completely removes the physical residual loss term Loss_phys from PhysCNN in Example 1. The training of the neural network is driven solely by the data loss term Loss_data, making it a traditional, purely data-driven deep learning model. All other conditions remain consistent with Example 1.

[0118] Results: On the same test set, the mean absolute error of the comparison scheme A increased to 0.21 mm. In particular, under a high-temperature condition at noon (module surface temperature exceeding 70°C), the model output multiple physically impossible negative thickness values ​​(e.g., -0.4 mm), showing a significant "physical deviation" phenomenon.

[0119] Comparison with Solution B: A multimodal fusion scheme using fixed weights

[0120] Setup: This scheme removes the dynamic weight calculation module from Example 1. In the feature fusion step, SAR features and thermal infrared features are simply added together with a fixed weight of 50%-50% (w=0.5). All other conditions remain consistent with Example 1.

[0121] Results: On the same test set, the mean absolute error of the comparison scheme B was 0.13 mm. During a dust storm with high humidity, the signal-to-noise ratio of the thermal infrared image decreased, while the sensitivity of the SAR signal should have increased. However, the fixed weights could not be adaptively adjusted, resulting in a peak instantaneous error of 0.45 mm during this period, which was much higher than the peak error of 0.11 mm of the dynamic weight scheme in Example 1 during the same period.

[0122] Comparison Conclusion

[0123] The experimental results are summarized in the table below:

[0124]

[0125] The experimental data from this comparative example clearly lead to the following conclusion:

[0126] The physical constraint loss function introduced in this invention not only improves the average accuracy by about 3 times compared to the pure data-driven method, but also fundamentally eliminates the absurd results that violate physical laws from the model output, significantly enhancing the robustness and credibility of the model.

[0127] The multimodal dynamic weight fusion mechanism designed in this invention improves the average accuracy by nearly 2 times compared to the fixed weight method, and can effectively suppress error fluctuations under severe weather conditions, significantly improving the system's adaptability in complex environments.

[0128] Example 2

[0129] This embodiment aims to verify the robustness of the invention under harsh conditions such as approaching sandstorms, low atmospheric visibility, and drastic changes in environmental parameters.

[0130] Specifically, the test parameters in this embodiment are as follows:

[0131] Test time: 16:30 on the same afternoon, the sandstorm was approaching, the sky was yellowish and the wind speed was increasing.

[0132] Environmental parameters: ambient temperature 28.5℃, relative humidity 35%, wind speed 8.5m / s, total solar irradiance 320 W / m².

[0133] The difference between this embodiment and Embodiment 1 is in step three (multimodal feature fusion):

[0134] Due to reduced light intensity and temperature difference, the signal-to-noise ratio of thermal infrared images decreases, and their characteristic variance is significantly reduced.

[0135] Meanwhile, the suspended dust particles in the air and the increased humidity enhance the contrast of the dielectric constant of the dust layer on the surface of the photovoltaic panel, making the SAR signal response more sensitive and its characteristic variance significantly increased.

[0136] The dynamic weight allocation module automatically adapts to this change, and the calculated fusion weight w is updated to 0.82, meaning that the system automatically shifts the focus of analysis to more reliable SAR data.

[0137] Results and Analysis

[0138] Accuracy: Under these harsh conditions, the MAE of the inversion result is 0.08 mm.

[0139] Conclusion: Compared with Example 1, despite the drastic deterioration of environmental conditions, the accuracy of the present invention only decreased slightly, demonstrating the effectiveness of the multimodal dynamic fusion mechanism and the high robustness of the present invention in complex environments.

[0140] Example 3:

[0141] This embodiment proposes a real-time photovoltaic dust thickness inversion system, including an edge computing device and a data acquisition device. The data acquisition device is used to collect multimodal data, including at least thermal infrared image data and synthetic aperture radar (SAR) data. The edge computing device is equipped with a feature extraction unit configured to extract thermal and dielectric features from the multimodal data based on a pre-built thermal-dust physical coupling model. A feature fusion unit is configured to dynamically weight and fuse the thermal and dielectric features to obtain fused features. The inversion unit includes a pre-trained physical constraint neural network for receiving the fused features and performing real-time inversion of dust thickness.

[0142] Specifically, the hardware environment in this embodiment includes: an NVIDIA Jetson AGXXavier edge computing device deployed next to the photovoltaic array. This device is connected to a Ku-band miniaturized synthetic aperture radar (SAR) sensor and a FLIR A65 thermal infrared camera, and simultaneously accesses meteorological station data from the site via the Modbus protocol. The software environment is an Ubuntu 20.04 operating system, and the algorithm is implemented based on the PyTorch 2.5 deep learning framework and the OpenCV 4.12.0 library.

[0143] The edge computing device synchronously triggers the SAR sensor and thermal infrared camera at a frequency of 1Hz to acquire thermal infrared images with a resolution of 640x512 pixels and the corresponding SAR backscattering coefficient matrix of the region. Simultaneously, it reads ambient temperature, relative humidity, wind speed, and total solar irradiance data from the weather station every minute. Using the spatial transformation matrix of the thermal infrared camera and SAR sensor, obtained beforehand through a checkerboard calibration method, an affine transformation is performed on each frame of SAR data to achieve pixel-level spatial alignment with the thermal infrared image.

[0144] The edge computing device then executes the processing method described in Example 1 to perform inference on the latest fused feature map once per minute, outputting a high-resolution dust thickness distribution map. The cleaning decision threshold is set to the average dust thickness. > 0.5mm. When the monitoring results meet this condition, the system sends a cleaning command to the operation and maintenance center via the MQTT protocol (corresponding to...). Figure 5The instruction (S531) includes the number of the photovoltaic module that needs cleaning, the coordinates of the dust accumulation area, and the suggested cleaning priority, and sends this information to the operation and maintenance control system of the photovoltaic power station or the intelligent cleaning robot.

[0145] Furthermore, considering the need for efficient operation of complex models on resource-constrained hardware, techniques such as model quantization (≤64 channels, ≤10 layers), channel pruning, and embedded inference engine adaptation (TensorRT acceleration configuration) were employed to compress and accelerate the Physically Constrained Neural Network (PhysCNN). This achieved a closed-loop process on edge devices, encompassing data acquisition, preprocessing, fusion, inversion, and decision-making, resulting in millisecond-level inference latency and watt-level power consumption.

[0146] Example 4:

[0147] This embodiment aims to illustrate that the present invention is not limited to a specific hardware platform or neural network structure. Specifically, the main difference between this embodiment and Embodiment 1 is:

[0148] Hardware platform: The Huawei Atlas 200 DK AI Developer Kit is used as the edge computing device, which is equipped with the Ascend 310 AI processor.

[0149] Neural Network Architecture: To adapt to the computational characteristics of the Ascend processor, the U-Net structure in Example 1 is replaced with a lighter MobileNetV2-based encoder-decoder structure. This network significantly reduces the number of parameters and computational cost through depthwise separable convolutions, further compressing the model parameter count to 0.5M.

[0150] Software framework: The algorithm is implemented based on the MindSpore 1.8 deep learning framework and uses Huawei's CANN (Compute Architecture for Neural Networks) for operator-level optimization.

[0151] Apart from the changes mentioned above, the other steps in this embodiment, including the acquisition and preprocessing of multimodal data, the construction of the physical coupling model, the dynamic weight fusion mechanism, and the application of the physical constraint loss function, are the same as those described in Embodiment 1.

[0152] The solution in this embodiment was deployed in the same test environment. The results show that the solution can also run stably, with a mean absolute error (MAE) of 0.075mm, a single inference latency of 115ms, and an average system power consumption of 4.2W.

[0153] This embodiment demonstrates that the core technical solution of the present invention (physical model constraint and multimodal dynamic fusion) has good platform independence and model independence, can be flexibly adapted to different hardware and network architectures, and shows broad engineering application potential.

[0154] The embodiments described above merely illustrate specific implementation methods of this application, and while the descriptions are detailed and specific, they should not be construed as limiting the scope of protection of this application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the technical solution of this application, and these modifications and improvements all fall within the scope of protection of this application.

Claims

1. A method for real-time inversion of photovoltaic dust thickness, characterized in that, Includes the following steps: Step 1: Acquire multimodal data, including at least thermal infrared image data and synthetic aperture radar data; Step 2: Based on the pre-built thermal-dust physics coupling model, extract thermal and dielectric features from the multimodal data; Step 3: Perform dynamic weighted fusion of the thermal and dielectric characteristics to obtain the fused characteristics; Step 4: Input the fused features into a pre-trained physical constraint neural network to perform real-time inversion of sand and dust thickness; Step 5: Output cleaning instructions and perform closed-loop feedback analysis based on the real-time inversion results; The multimodal data also includes ambient temperature, relative humidity, wind speed, and real-time total solar irradiance; The heat-dust physics coupling model described in step 2 is a three-dimensional unsteady-state heat conduction equation. The construction of the equation includes the following steps: Step 2.1: Abstract the photovoltaic module covered with sand and dust into a multi-layer composite structure. Based on the heat exchange process and Fourier's law of heat conduction, establish a partial differential equation describing the temperature field inside and on the surface of the composite structure as a function of time. Step 2.2: Quantify the relationship between dust thickness and thermal resistance in the equation; Step 2.3: Use ambient temperature, irradiance, and wind speed as dynamic boundary conditions, and introduce a photovoltaic panel tilt angle compensation factor to correct the solar irradiance component and convective heat transfer coefficient received at different tilt angles. Step 2 extracts multimodal features, including thermal features and dielectric features; The thermal features include spatial temperature gradient and temporal temperature change rate extracted from thermal infrared image sequences; The dielectric characteristics include the dielectric constant distribution map of the photovoltaic panel surface obtained by inverting the preprocessed SAR data through an improved dielectric model; The dynamic weight adaptive fusion described in step 5 includes: calculating the data variance of thermal and dielectric features within a local region, determining the feature weights of different modal data based on the data distribution of different modal features, and fusing multimodal data based on the feature weights to obtain a multimodal feature map.

2. The method for real-time inversion of photovoltaic dust thickness according to claim 1, characterized in that, The spatiotemporal registration preprocessing of the multimodal data includes: using the pixel coordinate system of the thermal infrared image as a reference, and utilizing pre-calibrated camera intrinsic and extrinsic parameters, registering the SAR data matrix onto the thermal infrared image to achieve sub-pixel-level spatial alignment; simultaneously, adding a synchronized timestamp to all data; subsequently, using the Lee filtering algorithm to suppress speckle noise on the registered SAR data, and using the median filtering algorithm to remove random thermal noise on the thermal infrared image; finally, normalizing all data to form a standardized multimodal input dataset.

3. The method for real-time inversion of photovoltaic dust thickness according to claim 1, characterized in that, The physical constraint neural network described in step 4 adopts an encoder-decoder structure. The encoder is responsible for extracting deep abstract features from the multimodal feature map fused in step 3, while the decoder is responsible for mapping the extracted features into a pixel-level dust thickness distribution map.

4. The method for real-time inversion of photovoltaic dust thickness according to claim 1, characterized in that, The training loss function of the physical constraint neural network described in step 4 includes: a data loss term used to measure the difference between the predicted value and the true value, and a residual term of the physical model equation.

5. The method for real-time inversion of photovoltaic dust thickness according to claim 1, characterized in that, Step 5 includes: Based on the preset dust thickness threshold, a cleaning instruction is automatically generated. The instruction includes the photovoltaic module number that needs to be cleaned, the coordinates of the dust accumulation area, and the suggested cleaning priority. The inverted dust thickness results are correlated with the actual power generation data of the photovoltaic array. If a systematic deviation is found between the inverted dust accumulation degree and the theoretical model of power loss, a self-calibration procedure is initiated to fine-tune the key parameters of the heat-dust coupling model in step 2.

6. A real-time photovoltaic dust thickness inversion system, characterized in that, include The data acquisition unit is used to acquire multimodal data, which includes at least thermal infrared image data and synthetic aperture radar data; The feature extraction unit is configured to extract thermal and dielectric features from the multimodal data based on a pre-built thermal-dust physics coupling model. A feature fusion unit is configured to perform dynamic weighted fusion of the thermal features and dielectric features to obtain fused features; The inversion unit includes a pre-trained physical constraint neural network for receiving the fused features and performing real-time inversion of the dust thickness. The multimodal data also includes ambient temperature, relative humidity, wind speed, and real-time total solar irradiance; The heat-dust physical coupling model is a three-dimensional unsteady-state heat conduction equation. The construction of the equation includes the following steps: The photovoltaic module covered with sand and dust is abstracted into a multi-layer composite structure. Based on the heat exchange process and Fourier's law of heat conduction, a partial differential equation is established to describe the temperature field inside and on the surface of the composite structure as a function of time. The relationship between dust thickness and thermal resistance is quantified in the equation; Ambient temperature, irradiance, and wind speed are used as dynamic boundary conditions. At the same time, a photovoltaic panel tilt angle compensation factor is introduced to correct the solar irradiance component and convective heat transfer coefficient received at different tilt angles. Extracting multimodal features includes thermal features and dielectric features; The thermal features include spatial temperature gradient and temporal temperature change rate extracted from thermal infrared image sequences; The dielectric characteristics include the dielectric constant distribution map of the photovoltaic panel surface obtained by inverting the preprocessed SAR data through an improved dielectric model; The dynamic weight adaptive fusion includes: calculating the data variance of thermal and dielectric features within a local region, determining the feature weights of different modal data based on the data distribution of different modal features, and fusing multimodal data based on the feature weights to obtain a multimodal feature map.

Images