Photovoltaic digital twin operation and maintenance platform for multi-source collaborative diagnosis

Abstract

This invention discloses a photovoltaic digital twin operation and maintenance platform for multi-source collaborative diagnosis, specifically relating to the field of photovoltaic operation and maintenance technology. It normalizes component health and environmental parameters, extracts time-frequency features, and performs adaptive filtering through edge nodes to generate a lightweight encrypted performance summary, which is then differentially compressed and synchronized to the cloud. The cloud performs secure federated fusion and dynamic reputation weighting on the summary, outputting a robust global performance vector. A collaborative optimization engine uses this vector and historical policies to invoke reinforcement learning to generate the original power allocation strategy, which is then corrected using a projection gradient descent mapping function before being distributed. Edge actuators load environmental snapshots for parallel simulation verification and risk assessment in the digital twin environment, effectively solving the problems of multi-source data isolation and privacy leakage. The invention proposes adaptive multi-resolution mesh, heterogeneous accelerated parallelism, and distributed message scheduling to achieve high efficiency and real-time performance in large-scale multi-physics coupled simulations.

Description

Technical Field

[0001] This invention relates to the field of photovoltaic operation and maintenance technology, and more specifically, to a photovoltaic digital twin operation and maintenance platform for multi-source collaborative diagnosis. Background Technology

[0002] In photovoltaic power plant operation and maintenance platforms, various field sensors and digital twin simulation models are used to collect real-time data on component operating status and environmental parameter vectors. These data are then collaboratively constructed in the cloud and at the edge to create a profile of the equipment's operation, recreating the actual operating conditions of each branch and array to support daily monitoring and fault early warning. The photovoltaic digital twin operation and maintenance platform needs to deeply integrate locally collected electrical, thermal, and environmental information with simulation results to achieve integrated perception and display of the multi-level, multi-physical field characteristics of the photovoltaic array, thereby providing a comprehensive on-site and virtual fusion perspective for operation and maintenance decisions.

[0003] In such scenarios, the health status of each device is often analyzed based solely on its own data, making it difficult to closely link local diagnostic results with overall power generation conversion efficiency and load allocation strategies. Due to the lack of a cross-device collaborative scheduling mechanism, when a component experiences performance degradation or abnormal fluctuations, the platform cannot automatically coordinate adjacent arrays or overall power allocation for compensation. This leads to a decrease in overall power generation conversion efficiency and puts some devices under overload pressure, increasing energy waste and the risk of cascading failures. Summary of the Invention

[0004] To overcome the aforementioned deficiencies in existing technologies, this invention provides a photovoltaic digital twin operation and maintenance platform for multi-source collaborative diagnosis. It constructs a multi-source data acquisition and lightweight encrypted digest mechanism, calibrates digital twin parameters online using an edge physical information neural network, and generates a global performance vector through federated learning aggregation. Combined with reinforcement learning-driven collaborative power allocation and digital twin simulation closed-loop verification and differential backhaul, it achieves a secure and efficient closed-loop diagnosis and optimization platform for the entire photovoltaic operation and maintenance process, thereby solving the problems mentioned in the background technology.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a photovoltaic digital twin operation and maintenance platform for multi-source collaborative diagnosis, comprising:

[0006] Step S11: Each edge node performs parallel normalization, feature extraction, and noise filtering on the component health and environmental parameter vectors locally to generate a performance summary, which is then synchronized to the cloud via differential compression.

[0007] Step S12: The cloud performs secure encryption and fusion of the performance summaries of each node, and outputs a global performance vector that reflects the overall collaborative state;

[0008] Step S13: Based on the global performance vector and historical running strategies, reinforcement learning is invoked to dynamically generate a collaborative power allocation scheme for each subarray; the original power allocation vector A of the policy network is invoked; physical constraint correction is performed on the original power allocation vector A through projected gradient descent to generate the adjusted power allocation vector A′;

[0009] Step S14: The edge actuator receives the cooperative power allocation scheme and loads the environmental state snapshot. It performs rapid simulation verification and risk assessment in the local digital twin environment. It simulates the output power, temperature rise and conversion efficiency in parallel through the electrical module, thermal module and optical module, and calculates the safety score and conversion efficiency gain by comparing the threshold set in real time. If the target is not met, the power allocation vector A′ is adjusted and the simulation is repeated. Otherwise, a verification pass flag is output.

[0010] Step S15: The verified power allocation command is sent from the local actuator to the component controller, and the adjusted performance differential data is sent back to the cloud to start the next round of federated optimization iteration.

[0011] Preferably, step S11 includes:

[0012] Normalization transformations are performed on the collected component health vector H and environmental parameter vector E respectively;

[0013] The time-domain statistical features and frequency-domain spectral features of the health vector H and the environmental parameter vector E are calculated in parallel, and the feature vector F is output.

[0014] To address the abnormal fluctuations in the feature vector F caused by environmental interference or sensing errors, a median filtering algorithm based on a sliding window is applied to generate a filtered feature vector N.

[0015] The filtered feature vector N is combined with the node identifier and timestamp, and a performance summary P is generated by differential compression.

[0016] The previously reported performance summary is differentially compressed at the byte level to generate an incremental summary D, which is then synchronized to the cloud via a secure channel.

[0017] Preferably, the component health vector H includes open-circuit voltage, short-circuit current, maximum power point voltage, maximum power point current, fill factor, conversion efficiency, series internal resistance, parallel internal resistance, and insulation resistance, and the environmental parameter vector E includes flat-plate irradiance, ambient temperature, component surface temperature, wind speed, and relative humidity.

[0018] Preferably, the cloud federation fusion step S12 includes: calling the BSDiff inverse differential algorithm to merge the current round incremental digest with the previous round full digest to recover the current performance digest P, and using the AES-128 key to decrypt the performance digest P, removing feature components that exceed the median ± three times MAD based on the median absolute deviation, and reconstructing the filtered feature vector N through wavelet threshold denoising interpolation.

[0019] Preferably, after the power allocation scheme A is output by the deep reinforcement learning policy network, the excess component (the component that exceeds the subarray carrying capacity or inverter capacity) is truncated and the gradient backoff is corrected by the constraint mapping function C based on projection gradient descent to generate an adjusted power allocation vector A′. The adjusted power allocation vector A′, along with the execution timestamp and the subarray identifier list, is encapsulated into a scheme package M. A SHA-256 hash is calculated and an AES-128HMAC signature is attached for integrity and source verification.

[0020] Preferably, the physical constraint mapping function C corrects the original power allocation vector A using projective gradient descent. The projective gradient descent includes: using the original power allocation vector A as the initial value, defining the correction objective as minimizing the distance between the adjusted vector and the original vector; in each iteration, determining the adjustment direction by comparing the difference between the previously adjusted vector and the original vector, and updating and generating a temporary vector Atemp in the reverse direction with a fixed step size; truncating each component of the temporary vector Atemp to the corresponding minimum safe output and maximum rated output intervals; calculating the deviation between the sum of each component of the truncated vector and the total available power, and distributing the deviation equally among the adjusted components; repeating the above update, truncation, and residual distribution until the adjusted vector converges or reaches the predetermined number of iterations.

[0021] Preferably, the output power, temperature rise, and conversion efficiency are calculated as follows:

[0022] After iterative convergence, the electrical module calculates the subarray output power using the maximum power point voltage Vmp and the maximum power point current Imp. The output power is equal to the product of the maximum power point voltage and the maximum power point current, multiplied by the series-parallel linkage coefficient. The thermal module calculates the component surface temperature Tmod for each spatial discrete grid and obtains the temperature rise based on the ambient temperature Tamb. The conversion efficiency η is calculated as the ratio of the output power Pout to the total incident power of the corresponding subarray. The output power Pout, temperature rise ΔT, and conversion efficiency η are output as the simulation result set Rsim to the subsequent risk assessment and fine-tuning stages.

[0023] Preferably, the risk assessment and fine-tuning process includes: real-time comparison of output power, temperature rise, and conversion efficiency with corresponding preset thresholds; marking risk events and recording deviations when limits are exceeded; organizing the simulation results into a simulation verification vector; evaluating the operating status through a safety score and conversion efficiency gain; the safety score is 1 minus the ratio of the sum of all deviations to the maximum allowable deviation; the conversion efficiency gain is the improvement ratio of the total output power of the photovoltaic array relative to the reference power; when the safety score is lower than the threshold or the efficiency gain does not meet expectations, the power allocation scheme is proportionally fine-tuned based on the deviation and the simulation is repeated; otherwise, the verification pass result is output.

[0024] Preferably, the spatial discrete grid is configured as follows: based on the online fault heatmap and the intensity of environmental disturbance, a multi-resolution grid is adaptively generated in the same simulation space, marked as fine grid areas and coarse grid areas (fine grid areas are used for fine calculations, and coarse grid areas are used to reduce the computational burden). The fine grid areas are mapped to the edge side according to the computational load, and the calculations of illumination and heat transfer are processed in parallel to ensure efficient completion of the core physical field simulation. Full simulation is triggered only for blocks where environmental or output power changes exceed the limits (referring to the union of fine and coarse grid areas); otherwise, the results of the previous simulation are reused and redundant calculations are reduced through differential updates. An index library of disturbance features and simulation scenarios is constructed, and cached results are quickly matched and loaded through similarity to respond to repetitive conditions in seconds. Distributed message exchange is used to realize dynamic task allocation and state synchronization between nodes, accelerating cloud-edge collaborative simulation. Specifically, the following steps are included:

[0025] Step S21: Based on the real-time fault heat map and the intensity of environmental disturbance, adaptively generate multi-resolution grid blocks and mark the fine grid area and the coarse grid area;

[0026] Step S22: Classify the grid blocks according to the computational load, and drive the edge GPU and FPGA to accelerate together. Prioritize mapping high-complexity blocks to the GPU, and allocate low-complexity blocks to the FPGA to perform optical illumination and thermal conduction calculations in parallel.

[0027] Step S23: Trigger full simulation only for blocks in the mesh where the environmental parameter vector or output power changes beyond the limit; otherwise, reuse the results of the previous simulation and reduce redundant calculations through differential updates.

[0028] Step S24: Maintain the perturbation feature and simulation scene cache index, load the hit cache according to the similarity matching result between the current perturbation feature and the cached scene, and write the preheating simulation result into the cache for later use when there is no hit.

[0029] Step S25: Exchange simulation status and output results between adjacent edge nodes through MPI messages, dynamically adjust the simulation task allocation of each node and keep the status synchronized to achieve cloud-edge distributed collaborative acceleration.

[0030] Preferably, the multi-resolution mesh blocks are generated after the edge nodes acquire temperature and output power data in real time, construct a fault heat distribution map and an environmental disturbance field intensity map, and synthesize a weighted field intensity map. The weighted field intensity map is divided into fine and coarse mesh regions through threshold segmentation and connected component analysis. For the fine mesh region, a quadtree recursively splits the mesh until the field intensity difference is lower than the refinement tolerance. For the coarse mesh region, meshes are divided at fixed intervals to output multi-resolution mesh blocks for subsequent simulation acceleration and differential updates. Specific operations include:

[0031] Field strength map construction: Each edge node acquires temperature and output power data from the sensor in real time, draws a fault heat distribution map and an environmental disturbance field strength map corresponding to the component array, and combines the two into a weighted field strength map to characterize the simulation priority of different locations;

[0032] Hotspot region identification: Threshold segmentation and connected component analysis are applied to the weighted field strength map to extract connected regions whose heat or disturbance intensity exceeds the preset threshold and mark them as fine grid regions. The remaining regions are marked as coarse grid regions to ensure that suspected fault and strong disturbance regions obtain higher simulation resolution.

[0033] Multi-resolution mesh generation: A quadtree recursive splitting algorithm is applied to each fine mesh region until the difference between the maximum and minimum field strength in the region is lower than the refinement tolerance to generate a high-density mesh; a fixed-interval mesh is used to divide each coarse mesh region to generate a low-density mesh, and finally a unified multi-resolution mesh block is output.

[0034] Preferably, for each grid block in the multi-resolution grid, the product of its number of cells and the average field strength gradient is calculated as a complexity index. Based on the complexity index, all grid blocks are divided into high-load sets and low-load sets. The high-load sets are submitted to the GPU for lighting calculation and heat conduction calculation pipelines through the CUDA interface and OpenCL interface, respectively, while the low-load sets are submitted to the FPGA for parallel processing, so as to achieve parallel acceleration of heterogeneous hardware.

[0035] Preferably, step S23 includes: comparing the latest environmental parameter vector with the output power Pout obtained from the previous simulation and their respective variation range for each grid block; dividing the over-limit blocks into a trigger set Fset and calling the simulation interface to perform simulation on them; dividing the non-over-limit blocks into a reuse set Rset and adjusting the previous simulation results based on their environmental and power deviation increments through a differential updater; and finally merging all grid block results to form a complete simulation result set.

[0036] The technical effects and advantages of this invention are as follows:

[0037] (1) This invention generates a lightweight differential compression performance summary and synchronizes it to the cloud by performing normalization, time-frequency feature extraction and adaptive filtering on component health vector and environmental parameter vector in parallel at the edge node, taking into account both data integrity and privacy protection; in the cloud, it adopts secure federated fusion and dynamic reputation weighting, injects differential privacy noise to output a robust global performance vector, provides a highly reliable input for collaborative optimization, and effectively solves the problems of multi-source data isolation and privacy leakage.

[0038] (2) This invention maps high-risk blocks to GPU and FPGA heterogeneous accelerated parallel computing by introducing multi-resolution grids and adaptive grid refinement and coarsening strategies. Combined with incremental triggering and differential update, principal component feature caching and scene matching, and distributed message scheduling, it achieves second-level response of large-scale multi-physics digital twin simulation. By using projection gradient descent constraint correction and local simulation closed-loop verification to start differential feedback, it ensures that the scheme meets the efficiency target and avoids the risk of exceeding the limit, effectively solving the performance bottleneck and real-time problem of large-scale coupled simulation. Attached Figure Description

[0039] Figure 1 This is a flowchart of the photovoltaic digital twin operation and maintenance process for multi-source collaborative diagnosis according to the present invention.

[0040] Figure 2 This is a flowchart of the simulation optimization process based on dynamic mesh partitioning of the present invention. Detailed Implementation

[0041] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0042] At the same time, it should be understood that, for ease of description, the dimensions of the various parts shown in the accompanying drawings are not drawn according to actual scale.

[0043] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the scope of this application and its application or use.

[0044] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0045] Example 1, see Figure 1 This invention provides a photovoltaic digital twin operation and maintenance flowchart for multi-source collaborative diagnosis, as shown in the following figure. Figure 1The photovoltaic digital twin operation and maintenance platform for multi-source collaborative diagnosis shown includes:

[0046] Step S11: Each edge node performs parallel normalization, feature extraction, and noise filtering on the component health and environmental parameter vectors locally to generate a performance summary, which is then synchronized to the cloud via differential compression.

[0047] The component health vector H and the environmental parameter vector E are the basis for the platform to quantitatively represent the components and their external operating conditions locally. Their sources, composition, and selection criteria are as follows:

[0048] First: Definition and composition of the component health vector H:

[0049] The module health vector H is a set of features that reflect the health status of a photovoltaic module during operation from two dimensions: electrical and physical performance. Its parameters may include:

[0050] Open-circuit voltage Voc and short-circuit current Isc: These reflect the basic response performance of photovoltaic polycrystalline / monocrystalline silicon cells by measuring the voltage-current terminals of the element under unobstructed, standard conditions.

[0051] Maximum power point voltage Vmp and maximum power point current Imp: obtained by an online I-V scanner or inverter MPP tracking function, directly related to component conversion efficiency and aging degree;

[0052] Fill factor: Calculated from Voc, Isc, Vmp, and Imp, it is used to indicate the variation of series / parallel resistance within the component;

[0053] Conversion efficiency η: The ratio of module output power to incident irradiation power, reflecting the degree of module degradation and pollution;

[0054] Series internal resistance Rs and parallel internal resistance Rsh: can be estimated online by I-V curve fitting (such as the IEC61853 method), and are used to diagnose contact corrosion or local shielding;

[0055] Insulation resistance Rins: Measured by an insulation monitoring system, it provides early warning of component aging and potential leakage risks.

[0056] Acquisition methods: Electrical parameters are obtained through periodic or real-time MPP tracking and rapid scanning using an inverter or dedicated I-V scanning equipment; resistance is calculated using an online I-V curve fitting algorithm or a small IV-Sweeper instrument, combined with a temperature correction model; and test voltage is injected periodically and leakage current is measured through insulation monitoring.

[0057] The selection of the component health vector H is based on the following: The above parameters are all from the IEC61724 (Guideline for Monitoring the Operation of Photovoltaic Systems) and IEC61853 (Component Performance Testing) standards, which can comprehensively reflect the component's electrical characteristics, internal damage and aging tendency, and are representative and reproducible.

[0058] Second: Definition and composition of environmental parameter vector E

[0059] Environmental parameter vector E is a quantitative environmental indicator describing the impact of external operating conditions on component performance. Its parameters may include:

[0060] Flat plate irradiance G (W / m 2 The main sensor or reference module irradiance meter measures in real time and directly determines the module's power generation potential;

[0061] Amb temperature: Measured using PT100 or thermocouple, it affects the module's internal resistance and conversion efficiency;

[0062] Component surface temperature Tmod: obtained by infrared thermometry or patch temperature sensor, used to compensate for temperature drift of electrical parameters;

[0063] Wind speed v (m / s): Measured by an anemometer, related to heat dissipation conversion efficiency and temperature rise;

[0064] Relative humidity RH (%): Measured by a humidity sensor, it can be used to assess potential condensation and surface contamination risks;

[0065] Other meteorological factors (such as haze index and rainfall): can be selected according to on-site needs and used to analyze short-term power generation fluctuations.

[0066] Acquisition method: The above-mentioned environmental sensors are deployed in the array area in the form of weather stations and connected to the edge nodes through wired or wireless gateways.

[0067] Selection criteria: Based on the recommended items for environmental monitoring in IEC 61724, and numerous empirical studies, it has been shown that irradiance, temperature, wind speed, and humidity are the main environmental factors affecting the power output and aging rate of components, and have high correlation and diagnostic value.

[0068] Third: The roles of component health vector H and environmental parameter vector E in diagnosis.

[0069] The component health vector H directly represents the electrical health status of the component itself and supports the identification of fault types such as aging, shading, hot spots, and contact degradation.

[0070] The environmental parameter vector E provides the external operating condition background, which is used to isolate environmental influences (such as temperature compensation and irradiation correction) and ensure accurate judgment of the causes of "abnormalities" in H.

[0071] By rigorously selecting, acquiring online, and integrating the aforementioned H and E parameters, we can analyze component performance fluctuations and potential faults to the greatest extent possible, thereby scientifically supporting the collaborative diagnosis and optimization decisions of the digital twin model.

[0072] Furthermore, the specific implementation of step S11 is as follows:

[0073] Each edge node performs normalization transformation on the collected component health vector H and environmental parameter vector E to ensure comparability of each channel dimension and eliminate the influence of units. The specific operation includes: using the Z-Score normalization method based on a sliding time window, with the window length set to the most recent minute or a fixed number of data records, first calculating the mean and standard deviation within the window, and then subtracting the mean and dividing by the standard deviation from the newly sampled component health vector H and environmental parameter vector E respectively to ensure that the offset influence is eliminated under the same units while taking into account temporal changes.

[0074] The health vector H and the environmental parameter vector E are computed in parallel in the time domain and in the frequency domain, respectively, to output a feature vector F, which is used to characterize the operation mode of the component and the changes in the environment. The specific operations include: the time domain statistical features include mean, variance, kurtosis and skewness, and the frequency domain spectral features are obtained by obtaining the main spectral peaks and their amplitude ratios through fast Fourier transform (FFT). The above features are concatenated in a fixed order to form a feature vector F to support subsequent consistency processing and model input reproduction.

[0075] To address the abnormal fluctuations in the feature vector F caused by environmental interference or sensing errors, a median filtering algorithm based on a sliding window is applied to generate a filtered feature vector N, thereby improving the robustness of subsequent encrypted digests. The specific operation includes a median sliding window filtering algorithm with fixed size and step size. The window length and step size are set to one-fifth and one-tenth of the length of the feature vector F, respectively. The window size is dynamically adjusted according to the local variance of the signal. When the variance exceeds a preset threshold, the filtering window is automatically increased to improve the suppression of sudden noise and the tracking ability of slow drift.

[0076] The filtered feature vector N is combined with the node identifier and timestamp, and a performance digest P is generated using the differential privacy encryption algorithm EncDP, balancing data privacy protection and global information integrity. The specific operations include: adding Laplace noise based on differential privacy to the filtered feature vector N when generating the performance digest P, with a privacy budget ε of 0.5; encrypting the combined data containing the node identifier and timestamp using the AES-128 symmetric encryption algorithm; and automatically rotating the key every 24 hours or when the node restarts to balance data security and privacy protection.

[0077] Compare the current performance summary P with the previously reported performance summary, perform incremental differential compression, output incremental summary D and synchronize it to the cloud to reduce bandwidth consumption and ensure summary timeliness.

[0078] Step S12: The cloud performs secure encryption and fusion of the performance summaries of each node, and outputs a global performance vector that reflects the overall collaborative state;

[0079] Furthermore, the specific implementation of step S12 is as follows:

[0080] The previous full performance summary corresponding to the node identifier is called, and the incremental summary is merged with the previous full performance summary using the BSDiff differential decompression algorithm DiffDecompress to restore the current performance summary P;

[0081] The specific operations include: using the AES-128 key specified in step S11 to decrypt the recovered current performance digest P, and after verifying the range of Laplace noise addition, outputting the filtered feature vector N; the specific operations include: after decryption, performing anomaly removal based on median absolute deviation (MAD) on the feature vector N: removing all components that exceed the median ± three times MAD, and applying local weighted median interpolation to reconstruct the gaps, and then performing wavelet threshold denoising on the reconstruction result to remove residual noise interference and ensure the reliability of the filtered feature vector N;

[0082] Based on the filtered feature vectors N and their statistical consistency in recent aggregations, an exponential decay function is applied to calculate node reputation weights to enhance trust in long-term stable nodes and suppress the impact of sudden abnormal nodes. Specific operations include: employing an explicit exponential decay reputation formula. in, Let be the reputation weight of the i-th node after the t-th round of aggregation, and its value ranges from [0, 1]. Let S be the reputation weight of node i after the (t-1)th round of aggregation, used to preserve the influence of historical reputation. simcos(·) is the cosine similarity function. ref It is a weighted average of the global feature vectors S from the most recent five rounds, with the historical window length fixed at five rounds, to quantify the historical consistency of nodes and ensure the reproducibility of Ri;

[0083] The exponential decay coefficient α is used to balance the impact of historical reputation and current similarity on Ri. The value of α ranges from 0.5 to 0.9, and the optimal range is obtained by performing 10-fold cross-validation on different power plant historical operation datasets. 1-α is used to represent the compensation weight assigned to the current similarity, which together with α constitutes the weight normalization.

[0084] In most real-world testing scenarios, a value of 0.7 achieves a balance between accuracy and robustness, avoiding over-reliance on history (too large α) while maintaining sensitivity to novel anomalies (too small α). Specifically, α can be adjusted based on the global performance fluctuation range ΔV. When ΔV is large, α can be reduced by 0.1 to accelerate the response to new data, while when ΔV is small, it can be increased by 0.1 to enhance stability.

[0085] All filtered feature vectors are weighted and averaged according to their corresponding weights, and abnormal vectors that deviate from the current round fusion result are removed to generate a robust global feature vector S.

[0086] Controllable Gaussian difference privacy noise is injected into the robust global feature vector to form a global performance vector V, which is then provided to the collaborative optimization engine as the core input.

[0087] Step S12 Summary: Through the synergistic effect of differential decompression, privacy decryption, dynamic weighting, and differential privacy noise injection, this step achieves high-fidelity restoration, robust fusion, and privacy protection of multi-source performance summaries, providing a global performance input that balances security and accuracy for subsequent collaborative optimization.

[0088] Step S13: Based on the global performance vector and historical running strategies, reinforcement learning is invoked to dynamically generate a collaborative power allocation scheme for each subarray; the original power allocation vector A of the policy network is invoked; physical constraint correction is performed on the original power allocation vector A through projected gradient descent to generate the adjusted power allocation vector A′;

[0089] Furthermore, the specific implementation of step S13 is as follows:

[0090] The historical execution strategy set Hset, composed of the most recent N rounds of verified allocation schemes, is read from the cloud strategy library to reflect the past collaborative adjustment effects. Specific operations include: the number of the most recent rounds is combined with a fixed value and dynamic adaptation: the default is to take the most recent five rounds (N=5), and when the global performance fluctuation exceeds the preset threshold, it is dynamically expanded to the most recent ten rounds; if the corresponding round scheme is missing or there is a version conflict in the strategy library, the latest complete scheme in the cloud strategy backup library is automatically called and the exception log is recorded to ensure the integrity and consistency of the historical execution strategy set Hset.

[0091] The global performance vector V is concatenated and normalized with the historical policy set Hset using the state mapping function φ to generate the reinforcement learning input state SRL = φ(V‖Hset). Specific operations include: the state mapping function... It consists of a two-layer fully connected network layer and a feature cross layer: First, the global performance vector V and each adjusted power allocation vector A′ in the historical scheme set Hset are independently Z-score normalized. Then, the interaction features are calculated through the feature cross layer. Finally, the output state vector SRL is fed into the two-layer ReLU activated fully connected network layer to solve the difference in scale and improve the state representation ability.

[0092] The trained deep reinforcement learning policy network is invoked to output the original power allocation vector A based on the input state SRL. The length of A is consistent with the number of subarrays M, representing the percentage of power proposed by each subarray. Specifically, the deep reinforcement learning policy network passes the input state vector SRL sequentially through a first and second fully connected network. ReLU activation is applied to the outputs of the first and second fully connected networks to introduce nonlinearity, and then the vector is fed into a third fully connected network. The network output is divided into two paths: one is the Actor branch, whose output serves as the unnormalized weights for the power proposals of each subarray. This weight vector is then input into a Softmax function to obtain the power proposals for all subarrays. The array's normalized power percentage is assigned to a vector A; another path serves as the Critic branch, whose output is the overall value estimate for that state, used only for backpropagation of errors during online updates; throughout the inference process, the input state is standardized before each layer according to the mean and standard deviation obtained during the training phase to ensure that the network has the same response scale to new operating conditions; the Softmax function can be implemented using the standard implementation in existing deep learning libraries without additional design; the training data comes from a mixed dataset of historical running records generated by simulation and real operation and maintenance logs, and the network parameters are updated online with priority experience replay after every thousand interactions to ensure that the model adapts to new operating conditions.

[0093] Using the constraint mapping function C, projection correction is performed on the components in the original power allocation vector A that may exceed the subarray carrying capacity limit or inverter capacity, generating an adjusted scheme A′ = C(A) that meets physical and safety constraints. Specifically, the constraint mapping function C is implemented through projection gradient descent, hard-truncating each component in the original power allocation vector A to the interval [minouti, maxouti], where minouti and maxouti correspond to the minimum safe output of the subarray and the rated output of the inverter, respectively. If there is a risk of temperature or current exceeding limits, the gradient is reversed to the safety boundary to ensure that the adjusted scheme A′ simultaneously meets electrical and thermal constraints.

[0094] The adjusted power allocation vector A′, along with the current round timestamp and subarray identifier list, is encapsulated into a scheme packet M and output to the edge actuator for rapid verification in the next step. After encapsulation, the scheme packet M calculates a SHA-256 hash value and attaches an HMAC digital signature based on an AES key so that the edge actuator can perform integrity and source verification before receiving it, preventing the allocation scheme from being tampered with or replayed during transmission.

[0095] Furthermore, in this scheme, the physical constraint correction of A by projected gradient descent includes:

[0096] Using the original power allocation vector A as the initial value, the correction objective is defined as minimizing the squared Euclidean distance error between the power allocation vector A′ and the original power allocation vector A. In each iteration, the gradient increment is calculated based on the difference between A′(k-1) and A, and a temporary vector Atemp is updated along the reverse gradient direction with a fixed step size.

[0097] Perform a truncation projection operation on each component of the temporary vector Atemp. If a component is less than the corresponding minimum safe output value amin, set it to amin; if it is greater than the corresponding maximum rated output value amax, set it to amax, and obtain the truncated vector.

[0098] Calculate the deviation power ΔP between the sum of all components of the cutoff vector and the total available power Ptotal of the system. Distribute the deviation power ΔP equally according to the number of subarrays and adjust each component to satisfy the equality constraint.

[0099] Repeat the gradient update, truncation projection and residual allocation process described above until the adjusted power allocation vector A′ converges or reaches the predetermined number of iterations, thereby generating a final allocation scheme that fits the original strategy and fully complies with electrical and thermodynamic physical constraints.

[0100] Step S14: The edge actuator receives the cooperative power allocation scheme and loads an environmental state snapshot. It performs rapid simulation verification and risk assessment in the local digital twin environment. It uses the electrical module, thermal module, and optical module to deduce the output power Pout, temperature rise ΔT, and conversion efficiency η in parallel. It compares the threshold set in real time to calculate the safety score and conversion efficiency gain. If the target is not met, it adjusts the power allocation vector A′ and re-simulates. Otherwise, it outputs a verification pass flag.

[0101] To explain, in this invention, the real-time multiphysics simulation (parallel simulation of output power Pout, temperature rise ΔT, and conversion efficiency η by electrical, thermal, and optical modules) adopts a unified spatial discrete grid and synchronous time-step execution framework:

[0102] The spatial discrete mesh is divided into, for example, 100×100 cells at equal intervals according to the component size. Each cell represents a physical area of ​​10 cm × 10 cm, and the thermal boundary conditions of the boundary cells are labeled.

[0103] The optical module calculates the incident irradiance matrix G for each cell on the grid.

[0104] The thermal module solves the discretized heat conduction equations with the same grid and simulation time step, and outputs the temperature field T for each element.

[0105] The electrical module combines the latest incident irradiance matrix G and temperature field T, and calculates the current density on each grid cell and extracts the voltage and current at the maximum power point through a single diode equivalent circuit model.

[0106] The three modules exchange incident irradiance matrix G, temperature field T, and current density data through shared memory or high-speed bus at each simulation time step, ensuring that the simulation is completed at the same time level, and terminating the iteration after the local residuals (such as mesh temperature change less than 0.5 degrees Celsius) meet the convergence condition.

[0107] In this embodiment of the invention, the simulation time step is specified as 0.1 seconds, the heat conduction equation and the optical irradiance update both adopt the explicit finite difference scheme, and the electrical module IV iteration adopts the Newton-Raphson method for convergence accuracy.

[0108] The heat conduction equation itself is a well-known technique in the fields of thermodynamics and heat transfer. However, to ensure sufficient disclosure in the embodiments, the following is added: The thermal module adopts a three-dimensional unsteady-state heat conduction equation, the standard expression of which is: the partial derivative of temperature with respect to time is equal to the thermal conductivity multiplied by the spatial Laplace operator acting on the temperature field, plus the heat source term per unit volume; the physical properties such as thermal conductivity, density and specific heat capacity in the equation are specified according to the component material handbook or IEC standard.

[0109] The single-diode equivalent circuit model and its maximum power point (MPP) extraction method are well-known technologies for photovoltaic system modeling. However, to ensure sufficient disclosure in the embodiments, the following supplementary information is provided: First, each grid cell is considered as a photovoltaic cell, using a standard single-diode equivalent circuit. Its equivalent parameters include four parts: photocurrent, dark current, series resistance, and parallel resistance. The photocurrent can be obtained through optical and thermal coupling simulation based on the average irradiance received on the cell surface and the module area. The dark current, series resistance, and parallel resistance can be obtained directly from typical values ​​in the test manuals provided by the module manufacturer (e.g., dark current is approximately 10). -10 Up to 10 -7(Ampere, series resistance in the range of 0.1 to 1 ohm, parallel resistance in the range of 50 to 1000 ohms); During simulation, the above four parameters are applied to the calculation process of the equivalent circuit text description: first calculate the photocurrent value, then obtain the temperature dependence term according to the dark current and temperature correction model, and then successively deduct the voltage drop of the series resistance and the leakage current of the parallel resistance to obtain the output current density of the unit at each voltage point.

[0110] To extract the maximum power point, a perturbation-observation method is used for voltage optimization: In the initial stage of the simulation, the battery terminal voltage is adjusted with small voltage increments (e.g., 0.1 volts), and the corresponding output power is measured. If the power changes in the same direction as the voltage, the adjustment continues in the same direction; otherwise, it is adjusted in the opposite direction. This process is repeated until the direction of power change is opposite to the direction of voltage change, at which point the maximum power point is considered reached. The voltage and current values ​​at this point are recorded as Vmp and Imp, respectively. This algorithm typically converges to the maximum power point in fewer than ten iterations, and can be completed under the conditions of a simulation step size of 0.1 volts and a sampling period of 0.5 seconds. This is achieved through the above equivalent circuit parameter assignment and perturbation-observation method.

[0111] After the simulation, the temperature rise ΔT is calculated based on the output power Pout of each subarray, the average temperature Tmod of the module and the ambient temperature Tamb, and the conversion efficiency η is obtained by dividing the output power Pout by the average irradiance and multiplying by the irradiated area. The safety score PSafe and the conversion efficiency gain Geff are calculated based on the real-time threshold set (including the maximum temperature rise threshold and the minimum conversion efficiency threshold).

[0112] If any score is lower than the preset standard, the power allocation vector A′ is adjusted by a fixed step size according to the score deviation and the synchronous simulation is repeated; otherwise, a verification pass flag is output.

[0113] Overview: After receiving the solution package M from the cloud, the edge actuator needs to quickly reproduce the actual equipment operating conditions in the local digital twin environment and perform simulation verification of the power allocation scheme A′ to ensure that the adjustment scheme can meet the power generation conversion efficiency target without introducing the risk of exceeding limits, thereby providing a reliable basis for local execution.

[0114] Furthermore, the specific implementation of step S14 is as follows:

[0115] Extract the power allocation vector A′, subarray identifier list, and timestamp from the scheme package M, align the runtime parameters of the digital twin model, and load the corresponding environmental state snapshot Ssnap to reproduce the initial simulation conditions Ssim. Specific operations include: deploying ground meteorological sensor arrays and multi-rotor UAVs equipped with optical irradiance and thermal imaging cameras to jointly collect data at each edge node; ground sensors record micro-meteorological data; UAVs fly at regular intervals to acquire irradiance and wind field distributions above the array; all data adopt a unified time synchronization protocol (NTP+PTP hybrid) and are buffered to the edge cache; and generate a unified incremental JSON format environmental state snapshot Ssnap by aligning with the UTC timestamp T using linear interpolation and time delay compensation algorithms.

[0116] The output ratio of each subarray in the power allocation vector A′ is mapped to the electrical, thermal and optical modules of the digital twin model. The parallel computing framework is used to simultaneously deduce the output power Pout, temperature rise ΔT and conversion efficiency η of each subarray under the current environment to generate the simulation result set Rsim.

[0117] The explanation is as follows: the electrical module, optical module and thermal module define a consistent interface data structure (including three-dimensional arrays of power, temperature field and irradiance field) and use the MPI message passing protocol to perform domain partitioning at the sub-array level. Each process is responsible for the simulation of one sub-array. The iterative convergence stops when the state residual of each module is less than its own preset iteration threshold (such as temperature residual below 0.5℃).

[0118] The electrical module, optical module, and thermal module are defined as follows:

[0119] The electrical module is based on a single (or multiple) diode equivalent circuit model. It takes the solar irradiance distribution and temperature distribution of each photovoltaic unit as input, solves the IV characteristic curve online, and extracts the maximum power point voltage Vmp and maximum power point current Imp. The optical module uses geometric optics or ray tracing algorithms, combined with solar elevation angle, module tilt angle and shading information, to calculate the incident irradiance matrix G of each unit on the three-dimensional mesh. The thermal module uses the finite difference or finite element method to solve the module heat transfer and internal conduction equations under the influence of irradiance, wind speed and ambient temperature, and outputs the temperature field T of the corresponding grid point.

[0120] The simulation is as follows:

[0121] The entire photovoltaic array is logically divided into several computational units according to subarrays. Each computational unit's MPI process is responsible for the data of one subarray, including its incident irradiance matrix G, temperature field T, and electrical parameters. In each iteration, the optical module first calculates its own process G and then exchanges boundary irradiance values ​​with adjacent processes via non-blocking MPIIsend / MPIIrecv. Then, the thermal module jointly calculates the temperature field T locally and with the received incident irradiance matrix G, and exchanges boundary temperatures to ensure the continuity of heat conduction. The electrical module reads the latest incident irradiance matrix and temperature field, calculates IV, and updates the input for the next round. The entire process is repeated until the residuals of the temperature field and irradiance field in the global maximum norm sense are both below 0.5℃ and 5W / m. 2 The iteration will terminate after reaching a preset threshold.

[0122] The calculation methods for output power, temperature rise, and conversion efficiency are as follows:

[0123] After iterative convergence, the electrical module calculates the subarray output power using the maximum power point voltage Vmp and the maximum power point current Imp. The output power is equal to the product of the maximum power point voltage and the maximum power point current, multiplied by the series-parallel connection coefficient (Pout = Vmp × Imp × series-parallel connection coefficient, which is used to adjust the total power ratio after multiple components are combined in series or parallel). The thermal module calculates the component surface temperature Tmod for each grid and calculates the temperature rise (ΔT = Tmod – Tamb) based on the ambient temperature Tamb. The conversion efficiency η is calculated as the ratio of the output power Pout to the total incident power of the corresponding subarray (total incident power is the product of the average irradiance and the irradiated area). The output power Pout, temperature rise ΔT, and conversion efficiency η are output as the simulation result set Rsim to the subsequent risk assessment and fine-tuning stages.

[0124] Furthermore, the risk assessment and fine-tuning process includes:

[0125] The output power Pout, temperature rise ΔT, and conversion efficiency η in the simulation result set Rsim are compared in real time with the predefined threshold set Glimit (including maximum temperature rise, minimum conversion efficiency, and maximum current limit). If any indicator exceeds the limit, a risk event is marked and the deviation δ is recorded for subsequent evaluation.

[0126] The simulation result set Rsim is organized into a simulation verification vector. Based on the risk assessment rule Rev, the safety score Psafe and the conversion efficiency gain Geff are calculated. The safety score is 1 minus the ratio of the sum of all deviations to the maximum allowable deviation (Psafe = 1 – (∑δ / Δmax) reflects the risk of exceeding limits, Σδ represents the cumulative deviation, and Δmax represents the maximum allowable deviation). The conversion efficiency gain is the improvement ratio of the total output power of the photovoltaic array relative to the reference power (Geff = (∑Pout – Pref) / Pref, ΣPout represents the total output power of the photovoltaic array, Pref represents the reference power, and in this embodiment of the invention, Pref is the historical thirty-day rolling average of output) reflecting the improvement of the conversion efficiency of the collaborative power allocation scheme.

[0127] If the safety score Psafe is lower than the safety threshold or the conversion efficiency gain Geff is lower than the expected gain, the adjustment mapping function Adj is called to proportionally fine-tune the power allocation vector A′ according to the deviation δ to generate a proportional fine-tuning scheme Aadj, and the proportional fine-tuning scheme Aadj is fed back to the repeated simulation; otherwise, the verification pass flag Flagpass and the final simulation verification result Rsim are output and the process proceeds to step S15.

[0128] Step S15: The verified power allocation command is sent from the local actuator to the component controller, and the adjusted performance differential data is sent back to the cloud to start the next round of federated optimization iteration.

[0129] Furthermore, the specific implementation of step S15 is as follows:

[0130] The edge actuator encapsulates the verified power allocation vector A′ with the electronic signature algorithm HMAC, along with the timestamp and subarray identifier list in the scheme package M, into an execution command package, and sends it to each component controller through the local gateway to generate the execution status and execution timestamp;

[0131] After receiving the execution command packet, each component controller feeds back the actual output power vector, actual environmental parameter vector, and status flag to the edge actuator in real time; the latter merges the actual output power vector and the actual environmental parameter vector into a sub-execution performance vector.

[0132] The edge actuator calls DiffCompress to perform byte-level difference between the current execution performance vector and the simulation verification vector, generating an execution difference summary to highlight the actual deviation and save bandwidth.

[0133] The execution differential digest, execution status, and execution timestamp are combined and an HMAC digital signature is generated using an AES-128 key. This is then packaged into a feedback packet to ensure data integrity and traceability.

[0134] The edge actuator reports the feedback packet to the cloud aggregation interface through a secure channel, and triggers the next round of federated learning and collaborative optimization loop after confirming the ACK, so as to complete the continuous adaptation of the model and policy.

[0135] In summary, Embodiment 1 of this invention constructs an end-to-end multi-source collaborative diagnostic closed loop: At edge nodes, component health vectors and environmental parameter vectors are processed in parallel. Encrypted performance summaries are generated through sliding window normalization, time-frequency feature extraction, and adaptive filtering, and synchronized to the cloud using incremental differential compression, balancing data integrity and privacy protection. The cloud performs secure federated fusion of the summaries from each node, using dynamic weighting and differential privacy noise injection to generate robust global performance vectors, providing highly reliable input to the collaborative optimization engine. Based on this, the collaborative optimization engine calls reinforcement learning models and combines them with physical constraint mapping to achieve refined power allocation for each subarray. Edge actuators conduct real-time simulation verification and risk assessment in the local digital twin environment and initiate the next round of federated iterations through differential feedback. This ensures secure and efficient sharing of multi-source data while overcoming the limitations of traditional isolated optimization, achieving synchronous closed-loop optimization of global and local, virtual and real data.

[0136] Background: In a photovoltaic digital twin operation and maintenance platform, the system relies on a physics engine to deeply integrate multi-source data such as light, temperature, and current collected by field sensors with a 3D simulation model. This allows for real-time reproduction of the optical irradiance distribution, heat conduction process, and electrical output characteristics of the components, supporting fault identification and operation and maintenance decisions by maintenance personnel. The platform collaboratively schedules simulation resources in the cloud and at the edge, comparing local simulation results with the virtual model to achieve seamless mapping between the digital twin and the field status. However, as the scale of photovoltaic arrays continues to expand, the physics engine must simultaneously handle optical irradiance propagation, thermal field diffusion, and electrical load calculations over a larger area, leading to a sharp increase in simulation workload and a significant extension of computation time. Because multi-physics coupled simulation requires data interaction and iteration of optical, thermal, and electrical modules to be completed in the same time step, the platform struggles to provide simulation results within the required response timeframe for operation and maintenance. This results in delays in fault warning and efficiency optimization verification, affecting maintenance timeliness and increasing equipment risk. Based on this, Embodiment 2 of the present invention is proposed.

[0137] Example 2, see Figure 2 The simulation optimization flowchart based on dynamic mesh partitioning is shown below. The difference between this embodiment and Embodiment 1 is that it also includes a method for setting the spatial discrete mesh, specifically including:

[0138] Based on online fault heatmaps and environmental disturbance intensity, multi-resolution meshes are adaptively generated in the same simulation space, labeled as fine-mesh and coarse-mesh regions (fine-mesh regions are used for detailed calculations, while coarse-mesh regions are used to reduce computational burden). Fine-mesh regions are mapped to the edge according to computational load, and lighting and heat transfer calculations are processed in parallel to ensure efficient completion of core physics field simulations. Full simulation is triggered only for blocks where environmental or output power changes exceed limits (referring to the union of fine-mesh and coarse-mesh regions); otherwise, the results of the previous simulation are reused, and redundant calculations are reduced through differential updates. An index library of disturbance features and simulation scenarios is constructed, and cached results are quickly matched and loaded using similarity to respond to recurring conditions within seconds. Distributed message exchange is used to achieve dynamic task allocation and state synchronization between nodes, accelerating cloud-edge collaborative simulation. Specifically, the following steps are included:

[0139] Step S21: Based on the real-time fault heat map and the intensity of environmental disturbance, adaptively generate multi-resolution grid blocks and mark the fine grid area and the coarse grid area;

[0140] Step S22: Classify the grid blocks according to the computational load, and drive the edge GPU and FPGA to accelerate together. Prioritize mapping high-complexity blocks to the GPU, and allocate low-complexity blocks to the FPGA to perform optical illumination and thermal conduction calculations in parallel.

[0141] Step S23: Trigger full simulation only for blocks in the mesh where the environmental parameter vector or output power changes beyond the limit; otherwise, reuse the results of the previous simulation and reduce redundant calculations through differential updates.

[0142] Step S24: Maintain the perturbation feature and simulation scene cache index, load the hit cache according to the similarity matching result between the current perturbation feature and the cached scene, and write the preheating simulation result into the cache for later use when there is no hit.

[0143] Step S25: Exchange simulation status and output results between adjacent edge nodes through MPI messages, dynamically adjust the simulation task allocation of each node and keep the status synchronized to achieve cloud-edge distributed collaborative acceleration.

[0144] Furthermore, after acquiring temperature and output power data in real time at the edge nodes, the multi-resolution mesh blocks are generated by constructing a fault heat distribution map and an environmental disturbance field intensity map, and synthesizing a weighted field intensity map. The weighted field intensity map is divided into fine and coarse mesh regions through threshold segmentation and connected component analysis. For the fine mesh region, a quadtree is recursively split until the field intensity difference is below the refinement tolerance. For the coarse mesh region, meshes are divided at fixed intervals to output multi-resolution mesh blocks for subsequent simulation acceleration and differential updates. Specific operations include:

[0145] Field strength map construction: Each edge node acquires temperature and output power data from the sensor in real time, draws a fault heat distribution map and an environmental disturbance field strength map corresponding to the component array, and combines the two into a weighted field strength map to characterize the simulation priority of different locations;

[0146] Among them, the fault heat map refers to a functional map on the component array plane, which takes each component or subarray position as a sampling point, and maps the difference between the surface temperature measured on site and the long-term temperature baseline deviation value of that position to a color image; its generation process is as follows: each edge node collects the component surface temperature in real time through infrared thermometry or patch temperature sensor, then subtracts the historical average temperature of the corresponding position to obtain the temperature deviation value, and then renders it on a two-dimensional coordinate system according to color levels. Areas with large adjacent deviations are displayed as hot spots to indicate fault or hot spot risk.

[0147] The environmental disturbance field intensity map is an image on the same plane that normalizes the environmental factors affecting the output of photovoltaic modules (including real-time irradiance fluctuations, wind speed changes, and humidity abrupt changes), and then combines them into a comprehensive disturbance intensity index according to a set weight. The magnitude of this index is represented by the color depth on the plane. The specific operation is as follows: each edge node obtains raw environmental data from the irradiance meter, anemometer, and humidity sensor. First, the absolute difference between each data and its historical mean is calculated. Then, different weights are assigned according to the relative importance of irradiance, wind speed, and humidity, and the results are summed. Finally, the disturbance field is rendered on the array plane to reflect the degree of environmental interference in different areas.

[0148] The fault heat map and the environmental disturbance field intensity map are first mapped pixel-by-pixel in the same grid coordinate system. The values ​​in the two maps are then standardized (e.g., the mean of temperature deviation and disturbance intensity is subtracted and then divided by the standard deviation). The fusion weight of the two maps is then dynamically determined based on experience or operating conditions (e.g., each is initially set to 50%, and then fine-tuned according to their respective variance ratios). Finally, the standardized fault heat values ​​and disturbance intensity values ​​are added point by point according to the weights to generate a weighted field intensity map in the same coordinate system, which is used for subsequent grid division and priority determination.

[0149] Hotspot region identification: Threshold segmentation and connected component analysis are applied to the weighted field strength map to extract connected regions whose heat or disturbance intensity exceeds the preset threshold and mark them as fine grid regions. The remaining regions are marked as coarse grid regions to ensure that suspected fault and strong disturbance regions obtain higher simulation resolution.

[0150] Multi-resolution mesh generation: A quadtree recursive splitting algorithm is applied to each fine mesh region until the difference between the maximum and minimum field strength in the region is lower than the refinement tolerance to generate a high-density mesh; a fixed-interval mesh is used to divide each coarse mesh region to generate a low-density mesh, and finally a unified multi-resolution mesh block is output.

[0151] In hotspot region identification, the threshold segmentation can be based on the mean of the field strength map plus one standard deviation as the initial segmentation threshold, and connected components with fewer than three grid cells are removed to avoid noise-induced misjudgments. In multi-resolution grid generation, the refinement tolerance is defined as one-tenth of the difference between the maximum and minimum field strength values ​​in the region, and the quadtree recursion depth does not exceed six levels. The fixed spacing of the coarse grid area is equal to four times the side length of the smallest unit of the fine grid, so as to form multi-resolution grid blocks where the coarse and fine grids can be smoothly transitioned.

[0152] Furthermore, the product of the number of cells corresponding to each grid block in the multi-resolution grid and the average field strength gradient is used as a complexity index. Based on the complexity index, all grid blocks are divided into high-load sets and low-load sets. The high-load sets are submitted to the GPU for lighting calculation and heat conduction calculation pipelines through the CUDA interface and OpenCL interface, respectively, while the low-load sets are submitted to the FPGA for parallel processing to achieve parallel acceleration of heterogeneous hardware. Specifically, based on the complexity index, all grid blocks are sorted in descending order, the first half of the blocks are taken as the high-load set, the remaining blocks are taken as the low-load set, and a higher priority scheduling tag is assigned to the high-load set.

[0153] The number of units corresponding to a grid block refers to the smallest discrete computing unit in the simulation space, which is physically mapped to the ground projection area occupied by a single component or group of components in the photovoltaic array. Each grid unit corresponds to the installation position of a component and the space of equal area directly above it. It can correspond one-to-one with a single component, or multiple component areas can be merged into one grid unit according to the simulation accuracy requirements. In this scheme, the grid unit size is set to be the same size as the component plate by default, so that the grid division is mapped one-to-one with the actual array layout, thereby ensuring that the simulation calculation results can accurately reflect the illumination, temperature and electrical status of each component.

[0154] The average field strength gradient refers to the following: based on the weighted field strength map, the field strength gradient magnitude of each grid cell is calculated using the two-dimensional central difference method. Then, the arithmetic mean of the gradient magnitudes of all cells in the block is taken to obtain the average field strength gradient of the block, which is used to reflect the degree of drastic change in field strength within the block.

[0155] The field strength gradient magnitude is calculated using a two-dimensional central difference operator. Before the calculation, Gaussian smoothing (standard deviation is one grid cell) is applied to the weighted field strength map to suppress noise. In each block, the arithmetic mean of the absolute values ​​of the gradients of all grid points in that block is used as the average gradient.

[0156] In this embodiment of the invention, the division of high-load and low-load blocks is based on the median of complexity metrics: the top 30% of blocks with metrics higher than the median are classified as high-load sets, and the rest are classified as low-load sets; the GPU calls a floating-point data pipeline based on CUDA 11.2, and uses page-locked memory on the host side to batch transfer 32-bit floating-point grid data stored in row-first order; the FPGA calls a streaming interface conforming to the OpenCL 2.2 specification to process data in parallel with the same data format; the host thread queries the CUDA event status and registers OpenCL callbacks every 100 milliseconds, and initiates up to three retries if no progress update is received within five seconds to ensure that the task is not interrupted.

[0157] Furthermore, step S23 includes: comparing the latest environmental parameter vector with the output power Pout obtained from the previous simulation and their respective variation range for each grid block; dividing the over-limit blocks into a trigger set Fset and calling the simulation interface to perform simulation on them; dividing the non-over-limit blocks into a reuse set Rset and adjusting the previous simulation results based on their environmental and power deviation increments through a differential updater; and finally merging all grid block results to form a complete simulation result set.

[0158] In this embodiment of the invention, the environmental change threshold (the range of change corresponding to the environmental parameter vector) is defined as twice the standard deviation of each component of the environmental parameter vector of the corresponding block and its historical rolling window. The power change threshold (the range of change corresponding to the output power) is defined as the percentage difference between the output power Pout of the previous simulation of the block and the historical mean. The threshold is dynamically updated using an exponentially weighted moving average method to ensure that the threshold is adaptively adjusted according to the long-term fluctuation trend. The differential updater adopts a linear incremental model, mapping the difference between the current latest environmental parameter vector and the previous environmental parameter vector, and the power difference between the output power Pout obtained from the previous simulation and the output power Pout obtained from the previous simulation, to the local solver with a fixed incremental coefficient. Point-to-point interpolation is used to adjust the previous simulation results. The updated simulation result set converges to the accuracy required for the current deviation in the error backoff loop, without the need for manual parameter tuning.

[0159] Furthermore, step S24 includes:

[0160] Perturbation feature vector generation: After merging the simulation result set Rsim, feature extraction is performed on the latest environmental parameter vector and output power vector Poutb. Principal component analysis is used to compress them into a fixed-dimensional perturbation feature vector Dfeat to ensure the consistency of subsequent matching and low-dimensional retrieval efficiency.

[0161] Similarity calculation and hit determination: Based on the cosine similarity function, the similarity value is calculated iteratively between the perturbation feature vector Dfeat and each existing scene feature in the cache index. If a certain similarity value is greater than the preset matching threshold, it is considered a hit and the scene identifier Sid is recorded, and the process proceeds to the next step; otherwise, the comparison continues.

[0162] Loading cached results: For scene identifier Sid, retrieve the corresponding complete simulation result set Rsim from the cache storage area and output it directly as the result of this simulation block to the merging queue, skipping the simulation or differential update process;

[0163] Scene cache writing and updating: When no match is found, the current perturbation feature vector Dfeat and the complete simulation result set Rsim are packaged into a new scene record, written to the cache index, and the oldest record is evicted according to the least recently used strategy to maintain cache capacity and hit efficiency.

[0164] In summary, Embodiment 2 of this invention synthesizes a weighted field strength map based on real-time fault heatmaps and environmental disturbance fields. It generates fine-mesh regions (high-risk) and fixed-spacing coarse-mesh regions (low-risk) through recursive quadtree splitting, achieving dynamic balance across multiple resolutions. GPUs are scheduled to handle high-load regions and FPGAs to handle low-load regions based on the complexity index of the mesh blocks (number of elements × average gradient). Fully coupled simulation is triggered only for environmental / power variation exceeding limits, while historical results are reused through differential updates for other blocks. PCA dimensionality reduction and cosine similarity matching are used to achieve second-level reuse of scene caches. Edge node tasks are dynamically scheduled and their states synchronized via MPI messages, constructing a cloud-edge collaborative acceleration architecture that overcomes simulation performance bottlenecks.

[0165] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A photovoltaic digital twin operation and maintenance platform for multi-source collaborative diagnosis, characterized in that, Comprise: Step S11: Each edge node locally performs normalization processing, feature extraction and noise filtering on the component health degree and environmental parameter vector in parallel to generate a performance summary, and synchronizes to the cloud through differential compression; Step S12: The cloud encrypts and fuses the performance summaries of each node to output a global performance vector reflecting the overall collaborative state; Step S13: Based on the global performance vector and historical running strategy, call reinforcement learning, call policy network to output the original power allocation vector A; perform physical constraint correction on the original power allocation vector A through projected gradient descent to generate the adjusted power allocation vector A', and generate the collaborative power allocation scheme package M containing the adjusted power allocation vector A'; Step S14: The edge executor receives the collaborative power allocation scheme package M, loads the environmental state snapshot, and carries out rapid simulation verification and risk assessment in the local digital twin environment; through the parallel deduction of electrical modules, thermal modules and optical modules, the output power, temperature rise and conversion efficiency are output, and the safety score and conversion efficiency gain are calculated by comparing with the threshold set in real time; if it does not meet the standard, adjust the power allocation vector A' and re-simulate, otherwise output the verification pass flag; Step S15: The verified power allocation instruction is issued by the local executor to the component controller, and the adjusted performance differential data is returned to the cloud to start the next round of federated optimization iteration.

2. The photovoltaic digital twin operation and maintenance platform for multi-source collaborative diagnosis according to claim 1, characterized in that, The step S11 comprises: Perform normalization transformation on the collected component health degree vector H and environmental parameter vector E respectively; Parallelly calculate the time domain statistical features and frequency domain spectral features of the health degree vector H and the environmental parameter vector E, and output the feature vector F; For abnormal fluctuations in the feature vector F caused by environmental interference or sensing errors, a median filtering algorithm based on a sliding window is applied to generate a filtered feature vector N; Combine the filtered feature vector N with the node identifier and timestamp to generate a performance summary P through differential compression; Byte-level differential compression is performed on the previous reported performance summary to generate an incremental summary D, which is synchronized to the cloud through a secure channel.

3. The photovoltaic digital twin operation and maintenance platform for multi-source collaborative diagnosis according to claim 1, characterized in that: The collaborative power allocation scheme package M is formed by encapsulating the adjusted power allocation vector A', the execution timestamp and the subarray identifier list, and after encapsulation, SHA-256 hash is calculated and AES-128 HMAC signature is attached for integrity and source verification; wherein the adjusted power allocation vector A' is obtained by performing physical constraint correction on the original power allocation vector A based on projected gradient descent.

4. The photovoltaic digital twin operation and maintenance platform for multi-source collaborative diagnosis according to claim 1, characterized in that, The physical constraint mapping function C corrects the original power allocation vector A through projected gradient descent, which comprises: Taking the original power allocation vector A as the initial value, define the correction target as minimizing the distance between the adjusted vector and the original vector; In each iteration, determine the adjustment direction by comparing the difference between the last adjusted vector and the original vector, and generate a temporary vector Atemp by updating in the opposite direction with a fixed step size; Truncate each component of temporary vector Atemp to the corresponding minimum safe output and maximum rated output interval; calculate the deviation between the sum of the truncated vector components and the total available power, and adjust each component by an equal amount of deviation; Repeat the above update, truncation and residual allocation until the adjusted vector converges or reaches a predetermined number of iterations.

5. The photovoltaic digital twin operation and maintenance platform for multi-source collaborative diagnosis according to claim 1, characterized in that, The calculation methods of the output power, temperature rise and conversion efficiency are as follows: After iteration convergence, the electrical module calculates the subarray output power at the maximum power point voltage Vmp and the maximum power point current Imp, and the output power is equal to the product of the maximum power point voltage and the maximum power point current, and then multiplied by the series-parallel connection coefficient; the thermal module calculates the component surface temperature Tmod of each spatially discrete grid, and calculates the temperature rise according to the ambient temperature Tamb; the conversion efficiency η is calculated according to the ratio of the output power Pout to the total incident power of the corresponding subarray, and the output power Pout, the temperature rise ΔT and the conversion efficiency η are output as the simulation result set Rsim to the subsequent risk assessment and fine-tuning link.

6. The photovoltaic digital twin operation and maintenance platform for multi-source collaborative diagnosis according to claim 5, characterized in that, The risk assessment and fine-tuning link includes: comparing the output power, temperature rise and conversion efficiency with the corresponding preset threshold in real time, and marking a risk event and recording the deviation amount when the threshold is exceeded; sorting the simulation result set into a simulation verification vector, evaluating the running state through a safety score and a conversion efficiency gain, the safety score being 1 minus the ratio of the sum of all deviation amounts to the maximum allowed deviation, and the conversion efficiency gain being the improvement ratio of the total output power of the photovoltaic array relative to the reference power; when the safety score is lower than the threshold or the efficiency gain does not meet the expectation, proportionally fine-tune the power distribution scheme based on the deviation amount and re-simulate, otherwise output the verification pass result.

7. The photovoltaic digital twin operation and maintenance platform for multi-source collaborative diagnosis according to claim 5, characterized in that, The spatially discrete grid is set as follows: according to the online fault heat map and the environmental disturbance intensity, a multi-resolution grid is adaptively generated in the same simulation space, marked as a fine grid area and a coarse grid area, the fine grid area is mapped to the edge side according to the calculation load, and the illumination and heat transfer calculation are processed in parallel to ensure efficient completion of the core physical field simulation; only the blocks with environmental or output power changes exceeding the threshold are triggered for full simulation, otherwise the last simulation result is reused and differential update is used to reduce redundant operations; an index library of disturbance features and simulation scenarios is constructed, and the cached results are quickly matched and loaded to respond to repeated working conditions in seconds; distributed message exchange is used to realize dynamic task allocation and state synchronization between nodes, accelerate cloud-edge collaborative simulation, and the specific steps are as follows: Step S21: adaptively generate a multi-resolution grid block based on the real-time fault heat map and the environmental disturbance intensity, and mark the fine grid area and the coarse grid area; Step S22: classify the grid blocks according to the calculation load, and drive the edge GPU and FPGA to cooperate to accelerate, preferentially map the high complexity blocks to the GPU, and distribute the low complexity blocks to the FPGA for parallel execution of optical illumination and heat conduction calculation; Step S23: only the blocks with environmental parameter vector or output power changes exceeding the threshold are triggered for full simulation, otherwise the last simulation result is reused and differential update is used to reduce redundant operations; Step S24: maintain the disturbance feature and simulation scene cache index, load the hit cache according to the similarity matching result of the current disturbance feature and the cache scene, and write the preheating simulation result into the cache for subsequent use when missing; Step S25: exchange the simulation state and output result between adjacent edge nodes through MPI message, dynamically adjust the simulation task allocation of each node and keep the state synchronization, so as to realize the cloud-edge distributed collaborative acceleration.

8. The multi-source collaborative diagnosis-oriented photovoltaic digital twin operation and maintenance platform according to claim 7, characterized in that, The multi-resolution grid block generation constructs the fault heat distribution map and the environmental disturbance field intensity map after the edge node acquires the temperature and output power data in real time, synthesizes the weighted field intensity map, divides the weighted field intensity map into fine grid area and coarse grid area through threshold segmentation and connected domain analysis, applies the quadtree recursive splitting to the fine grid area until the field intensity difference is lower than the refinement tolerance, and divides the coarse grid area according to the fixed interval, so as to output the multi-resolution grid block for subsequent simulation acceleration and differential update.

9. The multi-source collaborative diagnosis-oriented photovoltaic digital twin operation and maintenance platform according to claim 7, characterized in that, The product of the cell number and the average value of the field intensity gradient of each grid block in the multi-resolution grid is calculated as the complexity index, and all grid blocks are divided into high-load set and low-load set based on the complexity index.

10. The multi-source collaborative diagnosis-oriented photovoltaic digital twin operation and maintenance platform according to claim 7, characterized in that, The step S23 includes: comparing the latest environmental parameter vector with the output power Pout and the respective change range obtained by the previous simulation for each grid block, dividing the out-of-limit block into the trigger set Fset and calling the simulation interface for simulation, dividing the non-out-of-limit block into the reuse set Rset and adjusting the previous simulation result based on the environmental and power deviation increment through the differential updater, and finally merging all grid block results to form the complete simulation result set.

Images