Digital Twin-Based Intelligent Operation and Maintenance Simulation Method and System for Photovoltaic Power Stations

Abstract

The present invention provides a digital-twin-based intelligent operation and maintenance simulation method and system for photovoltaic power stations, which relates to the technical field of photovoltaic power station operation and maintenance. It includes constructing a digital-twin hierarchical structure through multi-modal detection signals, discrete wavelet packet transform, tensor decomposition, and graph neural networks to determine structured data; using a dynamic neural network to extract fault features, predicting fault probabilities based on a Markov model, and triggering maintenance warnings; inputting the warning signals into a maintenance strategy model based on multi-agent reinforcement learning to generate a set of candidate maintenance time windows; the maintenance scheduling module optimizes and sorts to obtain the optimal time window sequence, and the resource allocation module uses a genetic algorithm to generate a personnel and material allocation plan; iteratively optimizing the policy parameters based on the maintenance cost and outputting the final operation and maintenance implementation plan.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic power station operation and maintenance, and in particular to an intelligent operation and maintenance simulation method and system for photovoltaic power stations based on digital twins. Background Art

[0002] Photovoltaic power generation, as an important part of clean energy, its safe and stable operation is of great significance to the power system. With the continuous expansion of the scale of photovoltaic power stations and the sharp increase in the number of devices, the traditional manual inspection and experience-based maintenance methods are no longer sufficient to meet the operation and maintenance requirements. Digital twin technology provides a new idea for the intelligent operation and maintenance of photovoltaic power stations by constructing a digital mapping of physical entities, but the overall operation and maintenance efficiency and intelligent level still need to be improved.

[0003] The existing operation and maintenance methods for photovoltaic power stations mainly have some problems. The data formats and sampling frequencies are inconsistent, and there is a lack of effective multi-source heterogeneous data fusion methods; the fault diagnosis methods based on single features or simple rules are difficult to depict the gradual process and multi-mode characteristics of equipment failures; the maintenance decisions relying on manual experience do not fully consider the constraints of maintenance resources and the maintenance sequence between devices, resulting in low maintenance efficiency and high costs; there is a lack of a continuous optimization mechanism for maintenance strategies, making it difficult to adapt to the dynamic changes of equipment status and operating environment.

[0004] In summary, there is an urgent need for an intelligent operation and maintenance simulation method for photovoltaic power stations based on digital twins. By constructing a digital twin hierarchical structure, deep fusion of multi-source data can be achieved. Dynamic neural networks and Markov models are used to accurately describe the evolution law of equipment failures. Based on multi-agent reinforcement learning and genetic algorithms, maintenance decisions and resource allocation are optimized, and finally a closed-loop optimized intelligent operation and maintenance system is formed; not only can the accuracy and timeliness of fault diagnosis be improved, but also an optimal maintenance plan can be adaptively generated, significantly enhancing the operation and maintenance efficiency and reducing the maintenance cost. The present invention can solve the problems in the prior art. Summary of the Invention

[0005] Embodiments of the present invention provide an intelligent operation and maintenance simulation method and system for photovoltaic power stations based on digital twins, which can solve the problems in the prior art.

[0006] In the first aspect of the embodiments of the present invention,

[0007] An intelligent operation and maintenance simulation method for photovoltaic power stations based on digital twins is provided, including:

[0008] Obtain multi-modal detection signals through a photovoltaic module status acquisition device. Based on the multi-modal detection signals, construct a signal decomposition tree through discrete wavelet packet transform, extract feature vectors from high-frequency sub-bands and low-frequency sub-bands, establish a multi-dimensional feature space, perform tensor decomposition operations on the multi-dimensional feature space for multi-modal spatio-temporal alignment, generate a feature fusion matrix, and input the feature fusion matrix into a graph neural network model to construct a digital twin hierarchical structure including a physical data layer, a feature association layer, and a status representation layer;

[0009] Based on the digital twin hierarchical structure, determine structured data; construct a dynamic neural network, extract fault features of the structured data, calculate the similarity score between the fault features and historical faults, establish a Markov fault evolution model according to the similarity score sequence within a continuous time window, and predict the probability of fault occurrence; compare the probability of fault occurrence with a preset threshold to trigger a maintenance warning signal;

[0010] Input the maintenance warning signal into a maintenance strategy model constructed based on a multi-agent reinforcement learning network to generate a set of candidate maintenance time windows, call a maintenance scheduling module to optimize and sort the set of candidate maintenance time windows to obtain an optimal time window sequence, input the optimal time window sequence into a resource allocation module, and use a genetic algorithm to generate a personnel and material allocation plan; calculate the maintenance cost of the personnel and material allocation plan, and iteratively optimize the policy parameters of the maintenance strategy model based on the maintenance cost to output a final operation and maintenance implementation plan.

[0011] In an alternative embodiment,

[0012] Based on the multi-modal detection signals, constructing a signal decomposition tree through discrete wavelet packet transform and extracting feature vectors from high-frequency sub-bands and low-frequency sub-bands to establish a multi-dimensional feature space includes:

[0013] Evaluate the signal-to-noise ratio of the multi-modal detection signals to obtain a signal-to-noise ratio evaluation result, and adaptively select an optimal wavelet basis function according to the signal-to-noise ratio evaluation result;

[0014] Perform the first-layer recursive decomposition on the multi-modal detection signals through the optimal wavelet basis function, and perform low-frequency decomposition and high-frequency decomposition respectively through the low-pass filter coefficients and high-pass filter coefficients of the quadrature mirror filter bank to obtain the first-layer low-frequency component and the first-layer high-frequency component. Take the first-layer low-frequency component and the first-layer high-frequency component as the input signals for the second-layer recursive decomposition respectively, and use the same low-frequency decomposition and the same high-frequency decomposition to obtain the second-layer low-frequency component and the second-layer high-frequency component; repeat the iteration until the n-layer recursive decomposition is completed, and connect all the components obtained by the recursive decomposition according to the hierarchical relationship to form a multi-scale signal decomposition tree, where the 2 to the nth power of the last layer of components are used as the leaf node signals of the multi-scale signal decomposition tree;

[0015] Calculate the information gain of the signals of each leaf node of the multi-scale signal decomposition tree, set an adaptive threshold based on the information gain, and automatically select the optimal low-frequency sub-band signal and the optimal high-frequency sub-band signal from the leaf node signals according to the adaptive threshold;

[0016] Perform a square accumulation operation on the optimal high-frequency sub-band signal to obtain a high-frequency energy feature, and perform a square accumulation operation on the optimal low-frequency sub-band signal to obtain a low-frequency energy feature;

[0017] Calculate the energy distribution of the optimal high-frequency sub-band signal, and obtain a high-frequency normalized distribution through adaptive weight normalization. Calculate the energy distribution of the optimal low-frequency sub-band signal, and obtain a low-frequency normalized distribution through the adaptive weight normalization. Based on the high-frequency normalized distribution, calculate a high-frequency entropy value feature, and based on the low-frequency normalized distribution, calculate a low-frequency entropy value feature;

[0018] Calculate the degree of dispersion of the optimal high-frequency sub-band signal to obtain a high-frequency standard deviation feature, and calculate the degree of dispersion of the optimal low-frequency sub-band signal to obtain a low-frequency standard deviation feature;

[0019] Based on the information gain, perform feature importance weighting on the high-frequency energy feature, the low-frequency energy feature, the high-frequency entropy value feature, the low-frequency entropy value feature, the high-frequency standard deviation feature, and the low-frequency standard deviation feature to form a six-dimensional feature vector;

[0020] By calculating the vector mean and vector standard deviation of the six-dimensional feature vector, subtract the vector mean from the six-dimensional feature vector and then divide by the vector standard deviation to obtain a normalized feature vector. Use the normalized feature vector as a feature point to establish a multi-dimensional feature space.

[0021] In an alternative embodiment,

[0022] Perform a tensor decomposition operation on the multi-dimensional feature space for multi-modal spatio-temporal alignment to generate a feature fusion matrix including:

[0023] Perform a temporal partitioning on the feature vectors in the multi-dimensional feature space, and recombine the feature vectors into a third-order tensor based on the dimensions of the number of features, the number of time windows, and the number of modalities using a sliding time window. Perform a normalization process on the feature data, time data, and modality data in the third-order tensor to obtain a normalized third-order tensor;

[0024] Perform a Tucker decomposition operation on the normalized third-order tensor, project it respectively in the feature dimension, time dimension, and modality dimension to generate a core tensor, a feature mode matrix, a time mode matrix, and a modality mode matrix;

[0025] Reconstruct the normalized third-order tensor based on the core tensor, the feature pattern matrix, the time pattern matrix, and the modality pattern matrix, calculate the reconstruction error value, adjust the target dimension of the core tensor according to the reconstruction error value, and repeat the Tucker decomposition operation until the reconstruction error value is less than a preset error threshold to obtain the optimal core tensor;

[0026] Map the feature vectors to a unified feature space using the feature pattern matrix, align the feature vectors to a unified time reference using the time pattern matrix, and map the feature vectors to a unified representation space using the modality pattern matrix to obtain the aligned feature data, time data, and modality data;

[0027] Perform the Kronecker product operation on the time pattern matrix and the modality pattern matrix to obtain a combined pattern matrix, and perform a transpose operation on the combined pattern matrix to obtain a transposed pattern matrix;

[0028] Multiply the feature pattern matrix by the optimal core tensor to obtain a feature tensor, and multiply the feature tensor by the transposed pattern matrix to generate a feature fusion matrix containing multimodal spatio-temporal alignment information.

[0029] In an alternative embodiment,

[0030] Input the feature fusion matrix into a graph neural network model. Constructing a digital twin hierarchical structure including a physical data layer, a feature association layer, and a state representation layer includes:

[0031] Map each photovoltaic component to a graph network node, construct graph network edges based on the physical connection relationships between the photovoltaic components, and determine the topological structure of the graph neural network model; allocate the feature fusion matrix to the corresponding graph network nodes according to the acquisition locations to determine the initial node features of each graph network node;

[0032] Extract sensing data features and temporal correlation features from the initial node features, and organize the sensing data features and the temporal correlation features into a physical data layer;

[0033] For each graph network node, collect the initial node features of adjacent nodes based on the graph network edges, determine the feature fusion weights according to the physical connection relationships, and obtain the neighbor feature aggregation value through weighted average operation; based on the neighbor feature aggregation value, generate the first-order association features of the graph network node through linear transformation and non-linear transformation, and fuse the first-order association features with the temporal correlation features of the physical data layer to construct a feature association layer;

[0034] Input the first-order correlation features of the feature correlation layer into a preset feature extraction network, and through multi-layer feature transformation, extract the deep correlation patterns of each photovoltaic module to generate high-order correlation features; input the high-order correlation features into a preset fully connected neural network, and through multi-layer feature mapping, extract the characterization features of the overall operating state of the photovoltaic module to construct a state characterization layer;

[0035] The state characterization layer integrates the original collected data corresponding to the physical data layer and the interaction relationship between the photovoltaic modules corresponding to the feature correlation layer to form a digital twin hierarchical structure of the photovoltaic module.

[0036] In an alternative embodiment,

[0037] Based on the digital twin hierarchical structure, determine the structured data; construct a dynamic neural network, extract the fault features of the structured data, calculate the similarity score between the fault features and historical faults, establish a Markov fault evolution model according to the similarity score sequence within a continuous time window, and predict the probability of fault occurrence; compare the probability of fault occurrence with a preset threshold to trigger a maintenance warning signal, including:

[0038] Construct a dynamic neural network, divide the dynamic neural network into an input layer, a dynamic neuron layer, and an output layer according to the hierarchical structure, where each dynamic neuron in the dynamic neuron layer has an accumulated membrane potential value and an adaptive dynamic threshold, the input layer is associated with the dynamic neuron layer through the first inter-layer connection weight, and the dynamic neuron layer is associated with the output layer through the second inter-layer connection weight;

[0039] Receive the structured data of the digital twin hierarchical structure and input the structured data into the input layer of the dynamic neural network;

[0040] Calculate the membrane potential value of each dynamic neuron in the dynamic neuron layer, where the membrane potential value is obtained by adding the product of the previous moment's membrane potential value and the membrane potential decay coefficient to the eigenvalue after the structured data is transmitted through the first inter-layer connection weight, and then subtracting the dynamic threshold;

[0041] Judge the magnitude relationship between the membrane potential value and the dynamic threshold. When the membrane potential value is greater than the dynamic threshold, generate a pulse output signal;

[0042] Add the product of the previous moment's dynamic threshold and the threshold decay coefficient to the product of the previous moment's pulse output signal and the threshold increment coefficient to obtain the current moment's dynamic threshold;

[0043] Process the pulse output signal, transmit the pulse output signal through the second inter-layer connection weight to the output layer and perform a weighted summation operation to generate the current moment's fault feature vector;

[0044] Calculate the cosine similarity between the fault feature vector at the current moment and the pre-stored historical fault feature vectors to obtain the similarity score at the current moment;

[0045] Collect the similarity scores within a continuous preset time period to form a similarity score sequence, divide the similarity score intervals, use the similarity score intervals as the state space, count the number of interval transitions between adjacent moments in the similarity score sequence, and generate a state transition probability matrix;

[0046] Set the probability value corresponding to the similarity score interval where the similarity score at the current moment is located to 1, and set the probability values corresponding to other similarity score intervals to 0 to obtain the initial probability distribution vector at the current moment;

[0047] Multiply the initial probability distribution vector at the current moment by the state transition probability matrix to obtain the fault probability distribution vector at the next moment, and extract the probability value of the similarity score interval corresponding to the fault state from the fault probability distribution vector as the fault occurrence probability;

[0048] Compare the fault occurrence probability with a preset warning threshold, and trigger a maintenance warning signal when the fault occurrence probability is greater than the preset warning threshold.

[0049] In an alternative embodiment,

[0050] Input the maintenance warning signal into a maintenance strategy model constructed based on a multi-agent reinforcement learning network to generate a set of candidate maintenance time windows, and call the maintenance scheduling module to optimize and sort the set of candidate maintenance time windows to obtain an optimal time window sequence, including:

[0051] Receive the maintenance warning signal, construct a maintenance strategy model based on multi-agent reinforcement learning. The maintenance strategy model adopts a Monte Carlo tree structure, and uses each preset maintenance time point as a node of the Monte Carlo tree;

[0052] Obtain the device status parameters, maintenance resource parameters, and production plan parameters of each node, calculate the status score according to the device status parameters, calculate the resource score according to the maintenance resource parameters, calculate the impact score according to the production plan parameters, and perform weighted calculation on the status score, the resource score, and the impact score to obtain the comprehensive score of the node;

[0053] Calculate the exploration term based on the cumulative visit times of the parent node, the visit times of the current node, and a preset exploration coefficient, add the comprehensive score of the node to the exploration term to obtain the upper confidence bound value of each node, and select the node search path according to the upper confidence bound value, and generate a set of candidate maintenance time windows based on the search path;

[0054] Set the maintenance tasks in the candidate maintenance time window set as vertices, and set the dependency relationships between the maintenance tasks as connecting edges to construct a maintenance task dependency graph; recursively calculate the earliest execution time of each maintenance task along the forward path of the maintenance task dependency graph; recursively calculate the latest completion time of each maintenance task along the reverse path of the maintenance task dependency graph;

[0055] Calculate the time margin of each maintenance task, where the time margin is equal to the latest completion time minus the earliest execution time minus the preset duration of the maintenance task;

[0056] Perform priority sorting on the maintenance tasks in the candidate maintenance time window set according to the time margin to generate an optimal maintenance time window sequence.

[0057] In an alternative embodiment,

[0058] Input the optimal time window sequence into the resource configuration module, and use the genetic algorithm to generate a personnel and material allocation plan, including:

[0059] Encode the number of maintenance personnel requirements and the number of material requirements in the optimal time window sequence as a maintenance personnel allocation matrix and a material requirement matrix respectively to form a genetic code;

[0060] Perform weighted summation on the resource usage cost, time cost, and violation penalty cost to obtain the total cost, and calculate the fitness value of the genetic code based on the reciprocal of the total cost;

[0061] Randomly select three genetic codes within the genetic population, compare the fitness values of the three genetic codes, and mark the genetic code with the highest fitness value as the parental code;

[0062] Based on the interval [0, 1], generate the first random number, the second random number, and the third random number respectively. When the first random number is less than the preset crossover probability, enter the crossover operation; multiply the second random number by the length of the parental code and round down to obtain the crossover start position; multiply the third random number by the remaining code length and round down to obtain the crossover fragment length; based on the gene fragment determined by the crossover start position and the crossover fragment length, map the gene fragment to the range [0, 1] to obtain the transformed gene value, generate an expansion factor that follows the simulated binary distribution, and perform a linear combination of the transformed gene value and the expansion factor; restore the linearly combined gene value to the original value range according to the reverse interval mapping to generate the offspring code;

[0063] Perform random mutation on the offspring coding according to the preset mutation probability, and output the mutated coding; calculate the actual workload of various maintenance resources in the mutated coding respectively, and divide the actual workload by the available capacity of the corresponding resources to obtain the resource utilization rate;

[0064] When the resource utilization rate is greater than the preset utilization threshold, reduce the resource utilization rate by adjusting the corresponding values in the maintenance personnel allocation matrix and the material requirement matrix;

[0065] Repeat the execution until the preset number of iterations is reached, and output the personnel and material allocation plan.

[0066] In the second aspect of the embodiments of the present invention,

[0067] Provide an intelligent operation and maintenance simulation system for a photovoltaic power station based on digital twin, including:

[0068] The first unit is used to obtain multi-modal detection signals through a photovoltaic module state acquisition device, based on the multi-modal detection signals, construct a signal decomposition tree through discrete wavelet packet transform, extract feature vectors from high-frequency sub-bands and low-frequency sub-bands, establish a multi-dimensional feature space, perform tensor decomposition operation on the multi-dimensional feature space for multi-modal spatio-temporal alignment, generate a feature fusion matrix, and input the feature fusion matrix into a graph neural network model to construct a digital twin hierarchical structure including a physical data layer, a feature association layer, and a state representation layer;

[0069] The second unit is used to determine structured data based on the digital twin hierarchical structure; construct a dynamic neural network, extract fault features of the structured data, calculate the similarity score between the fault features and historical faults, establish a Markov fault evolution model according to the similarity score sequence within a continuous time window, and predict the probability of fault occurrence; compare the probability of fault occurrence with a preset threshold to trigger a maintenance warning signal;

[0070] The third unit is used to input the maintenance warning signal into a maintenance strategy model constructed based on a multi-agent reinforcement learning network to generate a set of candidate maintenance time windows, call a maintenance scheduling module to optimize and sort the set of candidate maintenance time windows to obtain an optimal time window sequence, input the optimal time window sequence into a resource configuration module, and use a genetic algorithm to generate a personnel and material allocation plan; calculate the maintenance cost of the personnel and material allocation plan, and iteratively optimize the policy parameters of the maintenance strategy model based on the maintenance cost to output a final operation and maintenance implementation plan.

[0071] In the third aspect of the embodiments of the present invention,

[0072] Provide an electronic device, including:

[0073] A processor;

[0074] A memory for storing processor-executable instructions;

[0075] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0076] In the fourth aspect of the embodiments of the present invention,

[0077] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0078] In the embodiments of the present invention, by fusing multi-modal signals and constructing a graph neural network to build a digital twin model, the operating state of the photovoltaic power station can be more comprehensively characterized. Combining with the Markov fault evolution model, the probability of fault occurrence can be predicted more accurately, so as to make maintenance preparations in advance and avoid losses caused by faults; the maintenance strategy model constructed based on the multi-agent reinforcement learning network can generate candidate maintenance time windows according to the predicted fault probability, and optimize them through the maintenance scheduling module and the resource configuration module, and finally obtain the optimal operation and maintenance implementation plan, thereby improving the maintenance efficiency and reducing the maintenance cost; combining digital twin technology, artificial intelligence algorithms and operation and maintenance management realizes the automation and intelligence of the operation and maintenance of the photovoltaic power station, reduces the need for manual intervention, and improves the level of operation and maintenance management. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is a schematic flowchart of the intelligent operation and maintenance simulation method for a photovoltaic power station based on digital twins according to the embodiments of the present invention;

[0080] Figure 2 It is a schematic structural diagram of the intelligent operation and maintenance simulation system for a photovoltaic power station based on digital twins according to the embodiments of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0081] To make the objectives, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0082] The technical solutions of the present invention will be described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be repeated in some embodiments.

[0083] Figure 1The figure is a schematic flowchart of the intelligent operation and maintenance simulation method for a photovoltaic power station based on digital twin according to an embodiment of the present invention. As Figure 1 shown, the method includes:

[0084] S101. Obtain multi-modal detection signals through a photovoltaic module status acquisition device. Based on the multi-modal detection signals, construct a signal decomposition tree through discrete wavelet packet transform, extract feature vectors from high-frequency sub-bands and low-frequency sub-bands, establish a multi-dimensional feature space, perform tensor decomposition operations on the multi-dimensional feature space for multi-modal spatio-temporal alignment, generate a feature fusion matrix, and input the feature fusion matrix into a graph neural network model to construct a digital twin hierarchical structure including a physical data layer, a feature association layer, and a state representation layer;

[0085] In an optional specific implementation manner, first collect multi-modal detection signals such as current, voltage, and temperature of the module through sensors installed on the photovoltaic module. These sensors include current sensors, voltage sensors, and temperature sensors, and collect the operation data of the module in real time.

[0086] Then, perform discrete wavelet packet decomposition on the detection signals of each modality separately. The signal is first decomposed into two sub-bands, a low-frequency sub-band and a high-frequency sub-band, and then each sub-band is further decomposed into two, forming a multi-layer signal decomposition tree structure. Through this recursive decomposition method, the original signal is decomposed to a preset decomposition level.

[0087] After obtaining the signal decomposition tree, extract features from the high-frequency sub-band and the low-frequency sub-band respectively. Extract statistical features such as mean, variance, and peak value in the time domain; extract spectral features such as frequency center and frequency band width in the frequency domain; extract time-frequency joint features such as energy distribution and energy entropy in the time-frequency domain. Combine these features to form a feature vector.

[0088] Then, arrange the feature vectors of each modality in chronological order to construct a three-dimensional feature tensor. The three dimensions of this tensor correspond to the modality dimension, the time dimension, and the feature dimension respectively. Perform tensor decomposition operations on this feature tensor, decompose it into the sum of several rank-1 tensors, realize the alignment of different modality data in the time and space dimensions, and generate a feature fusion matrix.

[0089] Finally, input the feature fusion matrix into a pre-trained graph neural network model. This model adopts a three-layer architecture: the bottom physical data layer stores the original monitoring data; the middle feature association layer describes the association relationship between different features through a graph structure; the top state representation layer fuses the information of the lower layer to represent the overall operation state of the component. Through this hierarchical structure, a digital twin model of the photovoltaic module is established to realize the real-time monitoring and evaluation of the component state.

[0090] In this embodiment, through discrete wavelet packet decomposition and multi-layer feature extraction, the time-domain, frequency-domain, and time-frequency domain characteristics of the operation of photovoltaic modules are comprehensively captured, improving the accuracy of signal analysis; a three-dimensional feature tensor is constructed and tensor decomposition is used to align multi-modal data in the time and space dimensions, effectively solving the problem of heterogeneous data fusion; the hierarchical architecture of the graph neural network model is used to comprehensively describe the component state, and the fusion is carried out layer by layer from physical data to state representation, significantly improving the accuracy of operation state monitoring and evaluation; the digital twin model realizes real-time monitoring and diagnosis of the state of photovoltaic modules, improving the operation and maintenance efficiency and the reliability of component operation; the method can be extended to other photovoltaic power station components or energy management systems, with generality and adaptability.

[0091] S102. Based on the digital twin hierarchical structure, determine the structured data; construct a dynamic neural network, extract the fault features of the structured data, calculate the similarity score between the fault features and historical faults, establish a Markov fault evolution model according to the similarity score sequence within a continuous time window, and predict the probability of fault occurrence; compare the probability of fault occurrence with a preset threshold to trigger a maintenance warning signal.

[0092] In an optional specific implementation manner, first, extract the monitoring data of the physical data layer, the association information of the feature association layer, and the state information of the state representation layer from the digital twin hierarchical structure, and normalize these information in accordance with a unified data format to form a structured fault diagnosis data set.

[0093] Then construct a dynamic neural network, which includes an input layer, a hidden layer, and an output layer. The input layer receives the structured data, the number of neurons in the hidden layer can be dynamically adjusted according to the feature dimension of the input data, and the output layer outputs the fault features. Input the structured data into this network, and extract the fault features through forward propagation.

[0094] Next, compare the extracted fault features with the fault features in the historical fault database. Use the cosine similarity calculation method to calculate the similarity score between the current fault feature and each historical fault feature. Calculate a similarity score for each time point, and form a similarity score sequence within a continuous time window.

[0095] Based on this similarity score sequence, establish a Markov fault evolution model. First, determine the fault state space, and divide the similarity score into multiple intervals corresponding to different fault states. Then count the state transition frequencies between adjacent time points and calculate the state transition probability matrix. Using this matrix, the fault state at the next moment can be predicted, thereby obtaining the probability of fault occurrence.

[0096] Finally, compare the predicted probability of fault occurrence with a pre-set warning threshold. When the probability of fault occurrence exceeds the warning threshold, the system will trigger a maintenance warning signal, indicating that maintenance intervention is required. This warning signal can include information such as the type of fault, the location of the fault, and the probability of the fault, providing a basis for subsequent maintenance decisions.

[0097] In this embodiment, through the digital twin hierarchical structure, multi-layer information is standardized and extracted to form a structured fault diagnosis data set, improving data quality and analysis efficiency; the dynamic neural network automatically adjusts the hidden layer structure according to the input features, enhancing the flexibility of the network and the ability to extract complex fault features; through cosine similarity calculation, the current fault features are accurately compared with historical data, improving the accuracy of fault type recognition; based on the Markov fault evolution model, the state transition probability matrix is used to predict the fault trend, realizing the dynamic tracking of potential faults; through the comparison of the probability of fault occurrence and the warning threshold, the maintenance warning signal is intelligently triggered, providing information on the type, location, and probability of the fault, facilitating scientific decision-making and timely maintenance; realizing the early detection and accurate positioning of faults, reducing the operation and maintenance costs of the photovoltaic system and improving the operation stability.

[0098] S103. Input the maintenance warning signal into a maintenance strategy model constructed based on a multi-agent reinforcement learning network to generate a set of candidate maintenance time windows. Call the maintenance scheduling module to optimize and sort the set of candidate maintenance time windows to obtain an optimal time window sequence. Input the optimal time window sequence into the resource allocation module and use a genetic algorithm to generate a personnel and material allocation plan; calculate the maintenance cost of the personnel and material allocation plan, and iteratively optimize the policy parameters of the maintenance strategy model based on the maintenance cost to output a final operation and maintenance implementation plan.

[0099] In an optional specific implementation manner, first construct a maintenance strategy model based on multi-agent reinforcement learning. This model contains multiple agents, and each agent is responsible for the maintenance decision of a type of equipment. Use the maintenance warning signal as the input state of this model. Each agent generates multiple maintenance time windows within a feasible time range according to its respective policy network to form a set of candidate maintenance time windows.

[0100] Then start the maintenance scheduling module. This module considers factors such as equipment importance, maintenance duration, and the constraint of the maintenance order between equipment to optimize and sort the set of candidate maintenance time windows. First, eliminate the time windows that do not meet the maintenance order constraint, then perform weighted sorting on the remaining time windows according to equipment importance, and finally combine and optimize the time windows considering the maintenance duration to obtain an optimal time window sequence.

[0101] Next, the optimal time window sequence is input into the resource allocation module, and the genetic algorithm is used to optimize the allocation of personnel and materials. First, the required numbers of maintenance personnel and materials are encoded as chromosomes to construct an initial population. Then, through genetic operations such as fitness evaluation, selection, crossover, and mutation, the chromosomes are continuously iteratively optimized, and finally, a personnel and material allocation plan that meets resource constraints is generated.

[0102] After that, the maintenance cost of this allocation plan is calculated, including labor cost, material cost, time cost, and violation penalty cost. The maintenance cost is used as a reward signal and fed back to the multi-agent reinforcement learning network to update the policy network parameters of each agent. Through multiple rounds of iterative training, the maintenance decision-making strategy of the agent is continuously optimized.

[0103] Finally, the optimized maintenance time window sequence and the personnel and material allocation plan are combined to form the final operation and maintenance implementation plan. This plan contains detailed information such as specific maintenance time arrangements, the required number of maintenance personnel, material types and quantities, and maintenance procedures, and can directly guide the on-site maintenance work.

[0104] In this embodiment, based on the multi-agent reinforcement learning model, the dynamic optimization of equipment maintenance decisions is realized to ensure accurate maintenance strategies for different equipment; considering the equipment importance, maintenance order constraints, and maintenance duration, the candidate maintenance time windows are optimized to generate the optimal time arrangement and improve the scheduling efficiency; the genetic algorithm is used to optimize the allocation of personnel and materials to ensure the maximization of resource utilization while meeting maintenance requirements and constraints; through the maintenance cost as a feedback signal, the reinforcement learning policy network is iteratively optimized to gradually reduce labor, material, and time costs and improve economy; detailed time arrangements, resource requirements, and maintenance procedures are provided to form an operation and maintenance plan that can be directly executed, improving the operability of on-site maintenance work; by combining the optimized maintenance time and resource allocation, the equipment downtime is reduced, the operation and maintenance costs are lowered, and the overall operation efficiency and reliability of the system are improved.

[0105] In an alternative implementation manner, based on the multi-modal detection signals, a signal decomposition tree is constructed through discrete wavelet packet transform, feature vectors are extracted from the high-frequency subbands and low-frequency subbands, and a multi-dimensional feature space is established, including:

[0106] The signal-to-noise ratio of the multi-modal detection signals is evaluated to obtain a signal-to-noise ratio evaluation result, and the optimal wavelet basis function is adaptively selected according to the signal-to-noise ratio evaluation result;

[0107] Perform the first - layer recursive decomposition on the multi - modal detection signal using the optimal wavelet basis function. Through the low - pass filter coefficients and high - pass filter coefficients of the quadrature mirror filter bank, perform low - frequency decomposition and high - frequency decomposition respectively to obtain the first - layer low - frequency component and the first - layer high - frequency component. Use the first - layer low - frequency component and the first - layer high - frequency component as the input signals for the second - layer recursive decomposition respectively, and perform the same low - frequency decomposition and the same high - frequency decomposition to obtain the second - layer low - frequency component and the second - layer high - frequency component; repeat the iteration until the n - layer recursive decomposition is completed. Connect all the components obtained from the recursive decomposition according to the hierarchical relationship to form a multi - scale signal decomposition tree, where the \(2^n\) components of the last layer are used as the leaf - node signals of the multi - scale signal decomposition tree;

[0108] Calculate the information gain of each leaf - node signal of the multi - scale signal decomposition tree, set an adaptive threshold based on the information gain, and automatically select the optimal low - frequency sub - band signal and the optimal high - frequency sub - band signal from the leaf - node signals according to the adaptive threshold;

[0109] Perform a square - accumulation operation on the optimal high - frequency sub - band signal to obtain the high - frequency energy feature, and perform a square - accumulation operation on the optimal low - frequency sub - band signal to obtain the low - frequency energy feature;

[0110] Calculate the energy distribution of the optimal high - frequency sub - band signal and obtain the high - frequency normalized distribution through adaptive weight normalization processing. Calculate the energy distribution of the optimal low - frequency sub - band signal and obtain the low - frequency normalized distribution through the adaptive weight normalization processing; based on the high - frequency normalized distribution, calculate the high - frequency entropy value feature, and based on the low - frequency normalized distribution, calculate the low - frequency entropy value feature;

[0111] Calculate the degree of dispersion of the optimal high - frequency sub - band signal to obtain the high - frequency standard - deviation feature, and calculate the degree of dispersion of the optimal low - frequency sub - band signal to obtain the low - frequency standard - deviation feature;

[0112] Based on the information gain, perform feature - importance weighting on the high - frequency energy feature, the low - frequency energy feature, the high - frequency entropy value feature, the low - frequency entropy value feature, the high - frequency standard - deviation feature, and the low - frequency standard - deviation feature to form a six - dimensional feature vector;

[0113] By calculating the vector mean and vector standard deviation of the six - dimensional feature vector, subtract the vector mean from the six - dimensional feature vector and then divide by the vector standard deviation to obtain a normalized feature vector. Use the normalized feature vector as a feature point to establish a multi - dimensional feature space.

[0114] The specific value of n in the n - layer recursive decomposition refers to the number of wavelet decomposition layers, which determines the depth to which the signal is decomposed, that is, the refinement degree of multi - scale analysis. The value of n is usually determined by the following factors: spectral range and complexity. If the spectral range of the signal is wide or its details are complex, a higher decomposition layer number (larger n) is required to capture more high - frequency and low - frequency information. If the signal is simple or the spectrum is concentrated, fewer decomposition layers (smaller n) are sufficient; sampling theorem and resolution requirements. For each recursive layer of wavelet decomposition, the frequency range is halved, and at the same time, the resolution is refined layer by layer. The decomposition layer number n usually depends on the sampling frequency of the signal and the required time resolution; the result of signal - to - noise ratio evaluation. The size of n can be dynamically adjusted according to the signal - to - noise ratio evaluation. A larger n - layer decomposition can better separate details and noise, but excessive decomposition may cause the high - frequency part to be submerged in noise. Based on the above logic, preferably, n is between 3 and 8 layers.

[0115] In an optional specific implementation, first, perform a signal - to - noise ratio evaluation on the collected multi - modal detection signal. For example, the signal - to - noise ratio can be evaluated by calculating the ratio of signal power to noise power. According to the signal - to - noise ratio evaluation result, adaptively select the optimal wavelet basis function. For example, if the signal - to - noise ratio is high, the Daubechies wavelet basis can be selected; if the signal - to - noise ratio is low, the Symlets wavelet basis can be selected. The aim is to select the wavelet basis function most suitable for the signal characteristics to improve the accuracy of feature extraction.

[0116] Next, use the selected optimal wavelet basis function to perform multi - layer recursive decomposition on the multi - modal signal to construct a signal decomposition tree. For example, perform three - layer recursive decomposition on the signal. First, decompose the original signal into the first - layer low - frequency component and high - frequency component through a low - pass filter and a high - pass filter. Then, pass the first - layer low - frequency component and high - frequency component through the same low - pass and high - pass filters respectively to obtain the second - layer low - frequency and high - frequency components. Repeat this process until three - layer decomposition is completed, obtaining eight final components, which form the leaf nodes of the signal decomposition tree.

[0117] Then, calculate the information gain of each leaf - node signal in the signal decomposition tree. The information gain can measure the contribution degree of each leaf - node signal to the information contained in the original signal. For example, the variance of each leaf - node signal can be calculated. The larger the variance, the higher the information gain. Set an adaptive threshold based on the calculated information gain, and select the optimal low - frequency sub - band signal and optimal high - frequency sub - band signal from the leaf - node signals according to the threshold. For example, the two leaf - node signals with the highest information gain can be selected as the optimal low - frequency and high - frequency sub - band signals.

[0118] Extract multiple features from the selected optimal low-frequency and high-frequency subband signals. First, calculate the energy features of the optimal low-frequency and high-frequency subband signals. For example, sum the squared amplitudes of each subband signal to obtain the energy feature value. Then, calculate the energy distribution of the optimal low-frequency and high-frequency subband signals and perform normalization processing through adaptive weights to obtain the normalized energy distribution. For example, divide the energy value of each subband signal by the total energy value to obtain the normalized energy distribution. Based on the normalized energy distribution, calculate the low-frequency and high-frequency entropy value features. The entropy value can reflect the complexity of the signal. Finally, calculate the standard deviation features of the optimal low-frequency and high-frequency subband signals. The standard deviation can reflect the dispersion degree of the signal. For example, calculate the standard deviation of the amplitude of each subband signal to obtain the standard deviation feature value. Assume the optimal low-frequency subband signal is [1, 2, 3, 4] and the optimal high-frequency subband signal is [5, 6, 7, 8]. Then the low-frequency energy feature is 30, the high-frequency energy feature is 174, the low-frequency standard deviation is 1.12, and the high-frequency standard deviation is 1.12. Assume that after normalization and calculation, the low-frequency entropy value is 0.8 and the high-frequency entropy value is 0.9.

[0119] Next, based on the information gain calculated previously, weight the six extracted features (high-frequency energy feature, low-frequency energy feature, high-frequency entropy value feature, low-frequency entropy value feature, high-frequency standard deviation feature, and low-frequency standard deviation feature) to obtain a six-dimensional feature vector. For example, multiply each feature value by the information gain of the corresponding leaf node as the weight. Assume the information gains are 0.5 and 0.6 respectively, then the weighted six-dimensional feature vector is [15, 103.2, 0.4, 0.54, 0.56, 0.67].

[0120] Finally, perform normalization processing on the six-dimensional feature vector to construct a multi-dimensional feature space. For example, calculate the mean and standard deviation of the six-dimensional feature vector, subtract the mean from each feature value and then divide by the standard deviation to obtain the normalized feature vector. Take the normalized feature vector as a feature point, and all feature points constitute the multi-dimensional feature space. Assume the mean is 31.63 and the standard deviation is 43.71, then the normalized feature vector is [-0.38, 1.64, -0.72, -0.70, -0.69, -0.68].

[0121] In this embodiment, by adaptively selecting the wavelet basis function and selecting the optimal subbands based on information gain, it is possible to better capture the key information in the signal and improve the accuracy of feature extraction; by calculating multiple features and performing weighted combination, it is possible to reduce the influence of noise and interference on feature extraction and enhance the robustness of the features; by adaptively selecting the optimal subbands through the threshold, the amount of data to be processed is reduced, the computational complexity is lowered, and the processing efficiency is improved.

[0122] In an alternative embodiment, a tensor decomposition operation is performed on the multi-dimensional feature space for multi-modal spatio-temporal alignment, and generating a feature fusion matrix includes:

[0123] Perform a temporal partition on the feature vectors in the multi-dimensional feature space, and based on a sliding time window, reorganize the feature vectors into a third-order tensor in the dimensions of the number of features, the number of time windows, and the number of modalities. Perform normalization processing on the feature data, time data, and modality data in the third-order tensor to obtain a normalized third-order tensor;

[0124] Perform a Tucker decomposition operation on the normalized third-order tensor, project it respectively in the feature dimension, time dimension, and modality dimension to generate a core tensor, a feature mode matrix, a time mode matrix, and a modality mode matrix;

[0125] Reconstruct the normalized third-order tensor based on the core tensor, the feature mode matrix, the time mode matrix, and the modality mode matrix, calculate the reconstruction error value, adjust the target dimension of the core tensor according to the reconstruction error value, and repeat the Tucker decomposition operation until the reconstruction error value is less than a preset error threshold to obtain an optimal core tensor;

[0126] Use the feature mode matrix to map the feature vectors to a unified feature space, use the time mode matrix to align the feature vectors to a unified time reference, and use the modality mode matrix to map the feature vectors to a unified representation space to obtain aligned feature data, time data, and modality data;

[0127] Perform a Kronecker product operation on the time mode matrix and the modality mode matrix to obtain a combined mode matrix, and perform a transpose operation on the combined mode matrix to obtain a transposed mode matrix;

[0128] Multiply the feature mode matrix by the optimal core tensor to obtain a feature tensor, and multiply the feature tensor by the transposed mode matrix to generate a feature fusion matrix containing multi-modal spatio-temporal alignment information.

[0129] The Tucker decomposition is a method used to simplify high-dimensional data (i.e., tensors), splitting a complex multi-dimensional tensor into a combination of a smaller core tensor and several mode matrices. It can not only retain the important information in the tensor but also reduce the complexity of the data. Specifically, assume there is a three-dimensional tensor (which can be regarded as a "multi-layer table"), and the three dimensions represent: the feature dimension (such as the type of detection signal), the time dimension (the changes at each time point), and the modality dimension (such as different devices, signal sources); the Tucker decomposition will decompose the tensor into the following parts: a core tensor, which is a streamlined version of the tensor and contains the most important information; three mode matrices, corresponding to features, time, and modality respectively. These matrices describe how the data in each dimension projects onto the core tensor. After decomposition, the original tensor can be approximately reconstructed using the core tensor and the mode matrices. This method makes data analysis more efficient and facilitates extracting meaningful information from complex data.

[0130] The Kronecker product is a special multiplication between matrices used to generate larger matrices. The corresponding rule is as follows: Given two matrices A and B, the Kronecker product multiplies each element in matrix A by the entire matrix B; then, all the small matrices are arranged according to the elements of A and spliced into a larger matrix.

[0131] In an optional specific implementation, first, the feature vectors in the multi-dimensional feature space are partitioned by time series. Assume there is a piece of multi-modal data containing video, audio, and text, and their feature vectors are extracted respectively. The dimension of the video feature vector is 512, the dimension of the audio feature vector is 256, and the dimension of the text feature vector is 100. For the feature vectors of each modality, they are partitioned according to a fixed time window length (for example, 0.5 seconds), and the feature vector sequence is converted into a series of feature vector segments within time windows. Assume the data duration is 10 seconds and the time window length is 0.5 seconds, then 20 time windows will be obtained.

[0132] Next, based on the sliding time window, the feature vectors are reorganized into a third-order tensor in the dimensions of the number of features, the number of time windows, and the number of modalities. Continuing with the above example, a third-order tensor with a dimension of [512 / 256 / 100, 20, 3] will be finally obtained, where the first dimension represents the number of features (different according to the modality), the second dimension represents the number of time windows, and the third dimension represents the number of modalities.

[0133] Then, normalization processing is performed on the feature data, time data, and modality data in the third-order tensor to obtain the normalized third-order tensor. For the feature dimension, zero-mean unit-variance normalization can be used; for the time dimension, min-max normalization can be used; for the modality dimension, since it is already a discrete modality index, no processing is required.

[0134] Perform Tucker decomposition operation on the normalized third-order tensor. Tucker decomposition decomposes the third-order tensor into a core tensor and three mode matrices (feature mode matrix, time mode matrix, and modality mode matrix). These three mode matrices respectively represent the projections in the feature dimension, time dimension, and modality dimension. Assume that the initial target dimensions of the core tensor are set to [256, 10, 2].

[0135] Reconstruct the normalized third-order tensor based on the core tensor and the three mode matrices, and calculate the reconstruction error value. Multiply the core tensor by the three mode matrices to obtain the reconstructed third-order tensor. Compare the reconstructed third-order tensor with the original third-order tensor and calculate the difference between them, such as the mean square error.

[0136] Adjust the target dimensions of the core tensor according to the reconstruction error value, and repeat the Tucker decomposition operation until the reconstruction error value is less than the preset error threshold to obtain the optimal core tensor. If the reconstruction error value is greater than the preset threshold (e.g., 0.01), then adjust the target dimensions of the core tensor (e.g., increase or decrease the dimensions), and repeat the Tucker decomposition. Keep iterating until the reconstruction error is less than the preset threshold to obtain the final core tensor and the three mode matrices.

[0137] Map the feature vectors to a unified feature space using the feature mode matrix. Multiply the original feature vectors by the feature mode matrix to obtain the feature vectors mapped to the unified feature space. This enables features of different modalities to be compared in the same space.

[0138] Align the feature vectors to a unified time reference using the time mode matrix. Apply the time mode matrix to the time dimension to align the features of different time windows to a unified time reference and eliminate temporal biases.

[0139] Map the feature vectors to a unified representation space using the modality mode matrix. Apply the modality mode matrix to the modality dimension to map the features of different modalities to a unified representation space, enabling information of different modalities to be fused together.

[0140] Perform the Kronecker product operation on the time mode matrix and the modality mode matrix to obtain the combined mode matrix. Perform the Kronecker product operation on the time mode matrix and the modality mode matrix to combine the time and modality information.

[0141] Transpose the combined mode matrix to obtain the transposed mode matrix. Perform a transpose operation on the combined mode matrix.

[0142] Multiply the feature mode matrix by the optimal core tensor to obtain the feature tensor. Multiply the feature mode matrix by the finally obtained core tensor.

[0143] Multiply the feature tensor by the transposed mode matrix to generate a feature fusion matrix containing multimodal spatio-temporal alignment information. Multiply the feature tensor by the transposed combined mode matrix to obtain the final feature fusion matrix, which contains the fused feature representation of multimodal data under a unified spatio-temporal benchmark.

[0144] In this embodiment, by aligning the time and features of different modalities, the deviation between data is eliminated, making subsequent analysis more accurate and reliable; mapping the features of different modalities to a unified representation space enables the effective fusion of information from different modalities, improving the richness and integrity of feature representation; through the application of Tucker decomposition and mode matrices, complex multimodal data is converted into a unified feature representation, simplifying the complexity of subsequent data processing.

[0145] In an alternative embodiment, inputting the feature fusion matrix into a graph neural network model to construct a digital twin hierarchical structure including a physical data layer, a feature association layer, and a state representation layer includes:

[0146] Map each photovoltaic component to a graph network node, construct graph network edges based on the physical connection relationships between the photovoltaic components, and determine the topological structure of the graph neural network model; allocate the feature fusion matrix to the corresponding graph network nodes according to the acquisition locations to determine the initial node features of each graph network node;

[0147] Extract sensing data features and temporal correlation features from the initial node features, and organize the sensing data features and the temporal correlation features into a physical data layer;

[0148] For each graph network node, collect the initial node features of adjacent nodes based on the graph network edges, determine the feature fusion weights according to the physical connection relationships, and obtain the neighbor feature aggregation value through weighted average operation; based on the neighbor feature aggregation value, generate the first-order association features of the graph network node through linear transformation and non-linear transformation, and fuse the first-order association features with the temporal correlation features of the physical data layer to construct a feature association layer;

[0149] Input the first-order association features of the feature association layer into a preset feature extraction network, extract the deep association patterns of each photovoltaic component through multi-layer feature transformation to generate high-order association features; input the high-order association features into a preset fully connected neural network, and extract the representation features of the overall operating state of the photovoltaic components through multi-layer feature mapping to construct a state representation layer;

[0150] The state representation layer integrates the original acquisition data corresponding to the physical data layer and the interaction relationships between the photovoltaic components corresponding to the feature association layer to form a digital twin hierarchical structure of the photovoltaic components.

[0151] In an alternative specific implementation, first, construct the topological structure of the graph neural network. Map each photovoltaic module to a node in the graph network. If there is a physical connection between two photovoltaic modules, such as being connected through a busbar box or an inverter, then establish an edge between these two nodes. In this way, a graph network topological structure reflecting the physical connection relationship of the photovoltaic power station is formed. For example, a photovoltaic power station composed of 100 photovoltaic modules can be mapped to a graph network containing 100 nodes and several edges.

[0152] Next, determine the initial node features of each graph network node. Collect various operating data of each photovoltaic module, such as current, voltage, power, temperature, etc., as well as static information such as the geographical location and installation angle of the module. Fuse these data into a feature vector as the initial node feature corresponding to the graph network node. For example, a 7-dimensional feature vector can be composed of 7 feature values such as current, voltage, power, temperature, longitude, latitude, and installation angle. Suppose the collected data of a certain photovoltaic module is: current 10A, voltage 500V, power 5000W, temperature 30°C, longitude 120°, latitude 30°, installation angle 30°, then the initial node feature vector corresponding to this module is [10, 500, 5000, 30, 120, 30, 30].

[0153] Then, construct the digital twin hierarchical structure. First, construct the physical data layer. Extract sensing data features from the initial node features, such as current, voltage, power, temperature, etc., as well as features reflecting the temporal relationship between photovoltaic modules, such as historical power output data, etc. Organize these features into the physical data layer as the basis of the digital twin structure.

[0154] Secondly, construct the feature association layer. For each graph network node, collect the initial node features of its neighbor nodes. Determine the feature fusion weights according to the physical connection relationship between the nodes. For example, nodes that are closer and more closely connected are given higher weights. Perform a weighted average operation on the features of the neighbor nodes to obtain the neighbor feature aggregation value. For example, a node has three neighbors with weights of 0.5, 0.3, and 0.2 respectively, and the voltage features of the neighbor nodes are 500V, 480V, and 510V respectively, then the neighbor feature aggregation value is 500×0.5 + 480×0.3 + 510×0.2 = 496V. Then, perform a linear transformation and a non-linear transformation on the neighbor feature aggregation value, such as using the activation function ReLU, to generate the first-order association feature of this node. Fuse the first-order association feature with the temporal association feature of the physical data layer to form the feature association layer, which reflects the interaction relationship between photovoltaic modules.

[0155] Finally, construct the state representation layer. Input the first-order correlation features of the feature correlation layer into a preset feature extraction network, such as a multi-layer convolutional neural network or a recurrent neural network. Through multi-layer feature transformation, extract the deep correlation patterns of each photovoltaic module to generate high-order correlation features. Input the high-order correlation features into a preset fully connected neural network, and extract the representation features of the overall operating state of the photovoltaic module through multi-layer feature mapping to construct the state representation layer. The state representation layer integrates the information of the physical data layer and the feature correlation layer to form a digital twin hierarchical structure of the photovoltaic module, comprehensively reflecting the operating state of the photovoltaic module.

[0156] In this embodiment, by fusing multi-source data and considering the correlation relationship between photovoltaic modules, the operating state of the photovoltaic power station can be more accurately reflected, avoiding errors caused by a single data source; by analyzing the state representation in the digital twin model, faulty modules can be quickly identified, and potential fault risks can be predicted, improving the efficiency of fault diagnosis; based on accurate state monitoring and fault diagnosis results, more reasonable operation and maintenance strategies can be formulated, such as optimizing the cleaning plan, adjusting the module operating parameters, etc., improving the operating efficiency and lifespan of the photovoltaic power station.

[0157] In an alternative embodiment, based on the digital twin hierarchical structure, determine structured data; construct a dynamic neural network, extract the fault features of the structured data, calculate the similarity score between the fault features and historical faults, and establish a Markov fault evolution model based on the similarity score sequence within a continuous time window to predict the probability of fault occurrence; compare the probability of fault occurrence with a preset threshold to trigger a maintenance warning signal, including:

[0158] Construct a dynamic neural network, divide the dynamic neural network into an input layer, a dynamic neuron layer, and an output layer according to the hierarchical structure, where each dynamic neuron in the dynamic neuron layer has an accumulated membrane potential value and an adaptive dynamic threshold. The input layer is connected to the dynamic neuron layer through the first inter-layer connection weight, and the dynamic neuron layer is connected to the output layer through the second inter-layer connection weight;

[0159] Receive the structured data of the digital twin hierarchical structure and input the structured data into the input layer of the dynamic neural network;

[0160] Calculate the membrane potential value of each dynamic neuron in the dynamic neuron layer, where the membrane potential value is obtained by adding the product of the previous moment's membrane potential value and the membrane potential decay coefficient to the eigenvalue transmitted by the structured data through the first inter-layer connection weight, and then subtracting the dynamic threshold;

[0161] Judge the magnitude relationship between the membrane potential value and the dynamic threshold. When the membrane potential value is greater than the dynamic threshold, generate a pulse output signal;

[0162] Add the product of the dynamic threshold at the previous moment and the threshold decay coefficient to the product of the pulse output signal at the previous moment and the threshold increment coefficient to obtain the dynamic threshold at the current moment;

[0163] Process the pulse output signal, transfer the pulse output signal through the second inter-layer connection weight to the output layer and perform weighted summation operation to generate the fault feature vector at the current moment;

[0164] Calculate the cosine similarity between the fault feature vector at the current moment and the pre-stored historical fault feature vector to obtain the similarity score at the current moment;

[0165] Collect the similarity scores within a continuous preset time period to form a similarity score sequence, divide the similarity score interval, use the similarity score interval as the state space, count the number of interval transitions between adjacent moments in the similarity score sequence, and generate a state transition probability matrix;

[0166] Set the probability value corresponding to the similarity score interval where the similarity score at the current moment is located to 1, and set the probability values corresponding to other similarity score intervals to 0 to obtain the initial probability distribution vector at the current moment;

[0167] Multiply the initial probability distribution vector at the current moment by the state transition probability matrix to obtain the fault probability distribution vector at the next moment, and extract the probability value of the similarity score interval corresponding to the fault state from the fault probability distribution vector as the fault occurrence probability;

[0168] Compare the fault occurrence probability with the preset warning threshold, and trigger a maintenance warning signal when the fault occurrence probability is greater than the preset warning threshold.

[0169] The membrane potential value specifically refers to a dynamic variable used to simulate the state of a neuron in a neuron model. It is a quantitative representation of the "excitation level" of a neuron and determines whether the neuron will trigger a pulse signal (i.e., "fire"). In biology, the membrane potential value corresponds to the numerical value of the potential difference across the neuron cell membrane, while in computer simulation, it is an abstract mathematical value used to describe the dynamic behavior of neurons.

[0170] In an optional specific implementation, first, establish a digital twin hierarchical structure. Decompose the device system to be monitored, such as a wind turbine, into multiple levels, such as blades, drive systems, generators, etc. Each level can be further decomposed into sub-components until the most basic sensor level. Each level and sub-component are associated with corresponding structured data, such as vibration data, temperature data, current data, etc., to form a complete digital twin hierarchical structure.

[0171] Next, construct a dynamic neural network. This network consists of an input layer, a dynamic neuron layer, and an output layer. The number of neurons in the input layer corresponds to the dimension of the structured data, such as the dimensions of data like wind turbine vibration, temperature, current, etc. The dynamic neuron layer contains multiple dynamic neurons, and each neuron has a membrane potential value and a dynamic threshold. An association is established between the input layer and the dynamic neuron layer through the first-layer connection weights. An association is established between the dynamic neuron layer and the output layer through the second-layer connection weights.

[0172] Obtain structured data from the digital twin hierarchical structure. For example, obtain vibration data, temperature data, and current data of wind turbine blades. Input this data into the input layer of the dynamic neural network.

[0173] Calculate the membrane potential value of each neuron in the dynamic neuron layer. The membrane potential value of each neuron is obtained by decaying the membrane potential value of the previous moment, adding the eigenvalue after the input data is transmitted through the first-layer connection weights, and then subtracting the dynamic threshold. For example, assume that the membrane potential value at the previous moment is 0.5, the decay coefficient is 0.9, the input eigenvalue is 0.2, and the dynamic threshold is 0.1. Then the membrane potential value at the current moment is (0.5×0.9)+0.2 - 0.1 = 0.55.

[0174] Judge the magnitude relationship between the membrane potential value and the dynamic threshold. If the membrane potential value is greater than the dynamic threshold, then this neuron generates a pulse output signal. For example, if the membrane potential value is 0.55 and the dynamic threshold is 0.1, then a pulse output signal is generated.

[0175] Update the dynamic threshold. The dynamic threshold at the current moment is obtained by decaying the dynamic threshold of the previous moment and adding the product of the pulse output signal of the previous moment and the threshold increment coefficient. For example, assume that the dynamic threshold at the previous moment is 0.1, the decay coefficient is 0.8, the pulse output signal is 1, and the threshold increment coefficient is 0.05. Then the dynamic threshold at the current moment is (0.1×0.8)+(1×0.05)=0.13.

[0176] Process the pulse output signal. Transmit the pulse output signals of all dynamic neurons to the output layer through the second-layer connection weights and perform a weighted summation operation to generate the fault feature vector at the current moment.

[0177] Calculate the similarity score between the fault feature vector at the current moment and the pre-stored historical fault feature vector. For example, use cosine similarity to calculate the similarity between the current fault feature vector and the feature vector of the historical fault "bearing wear", and obtain a similarity score of 0.85.

[0178] Collect the similarity scores within consecutive time periods to form a similarity score sequence. For example, collect the similarity scores within the past hour to form a data sequence containing 3600 similarity scores. Divide the similarity scores into multiple intervals, such as 0 - 0.2, 0.2 - 0.4, and so on, as the state space. Count the number of interval transitions between adjacent moments in the similarity score sequence to generate a state transition probability matrix.

[0179] Set the probability value corresponding to the interval where the similarity score at the current moment is located to 1, and set the probability values of other intervals to 0, to obtain the initial probability distribution vector at the current moment. For example, if the current similarity score is 0.85 and falls within the interval 0.8 - 1.0, then the probability value of this interval is 1, and the probability values of other intervals are 0.

[0180] Multiply the initial probability distribution vector at the current moment by the state transition probability matrix to obtain the failure probability distribution vector at the next moment. Extract the probability value of the interval corresponding to the failure state from the failure probability distribution vector as the probability of failure occurrence.

[0181] Compare the probability of failure occurrence with a preset warning threshold. If the probability of failure occurrence is greater than the warning threshold, trigger a maintenance warning signal. For example, if the probability of failure occurrence is 0.9 and the warning threshold is 0.8, then trigger a maintenance warning signal.

[0182] In this embodiment, by extracting fault features through a dynamic neural network and combining with a Markov model to predict the probability of failure, it is possible to more accurately predict the likelihood of equipment failure; it is possible to issue a warning signal before the failure occurs, thus providing sufficient time for maintenance personnel to perform preventive maintenance and avoid the occurrence of major failures; through early warning and preventive maintenance, the equipment downtime and maintenance costs can be reduced, and the operation efficiency and economic benefits of the equipment can be improved.

[0183] In an alternative embodiment, input the maintenance warning signal into a maintenance strategy model constructed based on a multi-agent reinforcement learning network to generate a set of candidate maintenance time windows, and call the maintenance scheduling module to optimize and sort the set of candidate maintenance time windows to obtain an optimal time window sequence, including:

[0184] Receive the maintenance warning signal, construct a maintenance strategy model based on multi-agent reinforcement learning. The maintenance strategy model adopts a Monte Carlo tree structure, and takes each preset maintenance time point as a node of the Monte Carlo tree;

[0185] Obtain the device status parameters, maintenance resource parameters, and production plan parameters of each node, calculate the status score according to the device status parameters, calculate the resource score according to the maintenance resource parameters, calculate the impact score according to the production plan parameters, and perform weighted calculation on the status score, the resource score, and the impact score to obtain the comprehensive score of the node;

[0186] Calculate the exploration item based on the cumulative access times of the parent node, the access times of the current node, and a preset exploration coefficient, add the comprehensive score of the node to the exploration item to obtain the upper confidence bound value of each node, and select the node search path according to the upper confidence bound value, and generate a candidate maintenance time window set based on the search path;

[0187] Set the maintenance tasks in the candidate maintenance time window set as vertices, set the dependency relationship between maintenance tasks as connection edges, and construct a maintenance task dependency graph; recursively calculate the earliest execution time of each maintenance task along the forward path of the maintenance task dependency graph; recursively calculate the latest completion time of each maintenance task along the reverse path of the maintenance task dependency graph;

[0188] Calculate the time margin of each maintenance task, where the time margin is equal to the latest completion time minus the earliest execution time minus the preset duration of the maintenance task;

[0189] Sort the maintenance tasks in the candidate maintenance time window set according to the time margin to generate an optimal maintenance time window sequence.

[0190] In an optional specific implementation manner, receive a maintenance warning signal and start constructing a maintenance strategy model. This model is based on multi-agent reinforcement learning and adopts a Monte Carlo tree structure. Each preset maintenance time point is set as a node of the Monte Carlo tree. For example, if we preset 12 noon every day in the next week as possible maintenance time points, then these time points respectively correspond to seven nodes on the tree.

[0191] Obtain the relevant parameters of each node. For each maintenance time point (i.e., each node), we need to obtain three types of parameters: device status parameters, maintenance resource parameters, and production plan parameters. The device status parameters can include indicators such as the running duration, temperature, and vibration frequency of the device. The maintenance resource parameters can include the available number of personnel, the inventory of spare parts, etc. The production plan parameters can include the number of scheduled production tasks, the product delivery date, etc. For example, at the maintenance node on Monday, the device running duration is 100 hours, the temperature is 60 degrees Celsius, the vibration frequency is 10 Hz, the available manpower is 5 people, the spare parts inventory is sufficient, and the number of scheduled production tasks is 1000 pieces.

[0192] Calculate the score for each node. Based on the obtained parameters, calculate the status score, resource score, and impact score respectively. The status score reflects the health status of the device, the resource score reflects the availability of maintenance resources, and the impact score reflects the degree of impact of maintenance on the production plan. The calculation method can be set according to the actual situation. For example, expert experience, historical data, or machine learning models can be used for evaluation. For example, based on the parameters on Monday, the calculated status score is 80 points, the resource score is 90 points, and the impact score is 70 points.

[0193] Calculate the upper confidence bound value of the node. First, calculate the exploration term based on the cumulative access times of the parent node, the access times of the current node, and a preset exploration coefficient. The exploration coefficient is used to balance exploration and exploitation and is usually set as a small positive number. Then, add the comprehensive score of the node (the weighted average of the status score, resource score, and impact score) to the exploration term to obtain the upper confidence bound value for each node. The upper confidence bound value is used to guide the selection of the search path, and the higher the value, the more potential the node has. For example, assume that the cumulative access times of the parent node of the Monday node is 100 times, the access times of the current node is 10 times, and the exploration coefficient is 0.1. Then the calculated exploration term is 3.0, the comprehensive score is 75 points, and the upper confidence bound value is 78.0.

[0194] Select the node search path and generate a set of candidate maintenance time windows. Based on the upper confidence bound value of each node, select the search path from the root node to the leaf node. The nodes on the search path form the set of candidate maintenance time windows. For example, after the search, the finally selected path contains three nodes: Tuesday, Thursday, and Saturday. Then 12:00 noon on these three days constitutes the set of candidate maintenance time windows.

[0195] Construct a maintenance task dependency graph. Set each maintenance task in the set of candidate maintenance time windows as a vertex of the graph, and set the dependency relationship between maintenance tasks as connecting edges. For example, if the maintenance task on Tuesday must be completed before the maintenance task on Thursday, then draw a directed edge from the vertex of Tuesday to the vertex of Thursday.

[0196] Calculate the earliest execution time and the latest completion time for each maintenance task. Along the forward path of the maintenance task dependency graph, recursively calculate the earliest execution time for each maintenance task. Along the reverse path of the maintenance task dependency graph, recursively calculate the latest completion time for each maintenance task. For example, assume that the task on Tuesday has no preceding tasks, then the earliest execution time is 12:00 noon on Tuesday; assume that the task on Thursday depends on the task on Tuesday, and the duration of the task on Tuesday is 2 hours, then the earliest execution time of the task on Thursday is 2:00 pm on Tuesday.

[0197] Calculate the time margin for each maintenance task. The time margin for each maintenance task is equal to its latest completion time minus its earliest execution time minus the preset duration of the maintenance task. For example, assume that the latest completion time of a Thursday task is 8:00 am on Friday and the duration is 4 hours, then the time margin is 20 hours.

[0198] Generate an optimal maintenance time window sequence. Sort the maintenance tasks according to their time margins, and the smaller the time margin, the higher the priority. The sorted sequence of maintenance tasks constitutes the optimal maintenance time window sequence. For example, assume that the time margin of a Tuesday task is 24 hours, the time margin of a Thursday task is 20 hours, and the time margin of a Saturday task is 30 hours, then the optimal maintenance time window sequence is Thursday, Tuesday, Saturday.

[0199] In this embodiment, through the intelligent maintenance strategy model and the optimized sorting algorithm, the best maintenance time window can be effectively selected to avoid unnecessary downtime, thereby improving the maintenance efficiency; by planning and optimizing the maintenance tasks in advance, resource waste can be reduced and the maintenance cost can be lowered; through timely maintenance, equipment failures can be prevented, the service life of the equipment can be extended, and the reliability of the equipment can be improved.

[0200] In an alternative embodiment, inputting the optimal time window sequence into the resource allocation module and using the genetic algorithm to generate a personnel and material allocation plan includes:

[0201] Encode the required number of maintenance personnel and the required quantity of materials in the optimal time window sequence as a maintenance personnel allocation matrix and a material requirement matrix respectively, and form a genetic code;

[0202] Perform a weighted summation of the resource usage cost, time cost, and violation penalty cost to obtain the total cost, and calculate the fitness value of the genetic code based on the reciprocal of the total cost;

[0203] Randomly select three genetic codes within the genetic population, compare the fitness values of the three genetic codes, and mark the genetic code with the highest fitness value as the parent code;

[0204] Based on the interval [0, 1], generate the first random number, the second random number, and the third random number respectively. When the first random number is less than the preset crossover probability, enter the crossover operation; multiply the second random number by the length of the parental encoding and round down to obtain the crossover start position; multiply the third random number by the remaining encoding length and round down to obtain the crossover segment length; based on the gene segment determined by the crossover start position and the crossover segment length, map the gene segment to the range [0, 1] according to the interval mapping to obtain the transformed gene value, generate an expansion factor that follows the simulated binary distribution, and perform a linear combination of the transformed gene value and the expansion factor; restore the linearly combined gene value to the original value range according to the reverse interval mapping to generate the offspring encoding.

[0205] Perform random mutation on the offspring encoding according to the preset mutation probability, and output the mutated encoding; calculate the actual workloads of various maintenance resources in the mutated encoding respectively, and divide the actual workloads by the available capacities of the corresponding resources to obtain the resource utilization rates.

[0206] When the resource utilization rate is greater than the preset utilization threshold, adjust the corresponding values in the maintenance personnel allocation matrix and the material requirement matrix to reduce the resource utilization rate.

[0207] Repeat the execution until the preset number of iterations is reached, and output the personnel and material allocation plan.

[0208] In an optional specific implementation manner, first, it is necessary to encode the number of maintenance personnel required and the number of materials required to form the encoding structure required by the genetic algorithm. Specifically, represent the number of maintenance personnel required as a matrix, where each row of the matrix represents a time window and each column represents a type of maintenance personnel, and the value of the matrix element represents the number of maintenance personnel of this type within this time window. Similarly, represent the number of materials required as a matrix, where each row of the matrix represents a time window and each column represents a type of material, and the value of the matrix element represents the number of materials of this type within this time window. Combine these two matrices together to form the encoding of the genetic algorithm, which is called the genetic encoding. For example, assume there are three time windows, two types of maintenance personnel, and three types of materials. Then the maintenance personnel requirement matrix can be represented as a 3x2 matrix, and the material requirement matrix can be represented as a 3x3 matrix.

[0209] Next, it is necessary to define the fitness function of the genetic algorithm. The fitness function is used to evaluate the quality of each genetic encoding. Here, the resource usage cost, time cost, and violation penalty cost are weighted and summed to obtain the total cost. Then, the reciprocal of the total cost is used as the fitness value of the genetic encoding. The higher the fitness value, the better the corresponding solution of the encoding. For example, assume that the weight of the resource usage cost is 0.5, the weight of the time cost is 0.3, and the weight of the violation penalty cost is 0.2. The resource usage cost of a certain solution is 100, the time cost is 50, and the violation penalty cost is 20. Then the total cost of this solution is 0.5×100 + 0.3×50 + 0.2×20 = 79, and its fitness value is 1 / 79.

[0210] Then, the iterative process of the genetic algorithm is carried out. In each iteration, first randomly select three genetic encodings from the population, compare their fitness values, and select the encoding with the highest fitness value as the parent encoding. Then, perform the crossover operation. Generate three random numbers between 0 and 1. The first random number is used to determine whether to perform the crossover operation. If the first random number is less than the preset crossover probability, the crossover operation is performed. The second random number is used to determine the starting position of the crossover. Multiply the second random number by the length of the parent encoding and round down to obtain the starting position of the crossover. The third random number is used to determine the length of the crossover segment. Multiply the third random number by the remaining encoding length and round down to obtain the length of the crossover segment. Map the selected gene segment to the range of 0 to 1 according to the interval mapping, then generate the expansion factor that follows the simulated binary distribution, and perform a linear combination of the transformed gene value and the expansion factor to obtain the new gene value. Finally, restore the new gene value to the original value range according to the reverse interval mapping to generate the offspring encoding. For example, assume that the crossover probability is 0.8 and the first random number is 0.7, then the crossover operation is performed. Assume that the length of the parent encoding is 10 and the second random number is 0.3, then the starting position of the crossover is 3. Assume that the third random number is 0.2, then the length of the crossover segment is 2.

[0211] Next, perform the mutation operation. According to the preset mutation probability, randomly mutate the offspring encoding to generate the mutated encoding. For example, assume that the mutation probability is 0.1, then there is a 10% probability of mutating a certain gene in the offspring encoding.

[0212] Then, calculate the resource utilization rate. Calculate the actual workload of various maintenance resources in the mutated encoding respectively, and divide the actual workload by the available capacity of the corresponding resource to obtain the resource utilization rate. For example, assume that the actual workload of a certain type of maintenance personnel is 80 hours and the available capacity is 100 hours, then the resource utilization rate of this type of maintenance personnel is 80 / 100 = 0.8.

[0213] Finally, resource utilization adjustment is performed. If the resource utilization rate is greater than the preset utilization threshold, it is necessary to adjust the corresponding values in the maintenance personnel allocation matrix and the material requirement matrix to reduce the resource utilization rate. For example, assuming the utilization threshold is 0.9 and the resource utilization rate of a certain type of maintenance personnel is 0.95, it is necessary to reduce the allocation quantity of this type of maintenance personnel or increase the available capacity of this type of maintenance personnel.

[0214] Repeat the above iterative process until the preset number of iterations is reached, and finally output the personnel and material allocation plan.

[0215] In this embodiment, by optimizing the personnel and material allocation plan, the resource usage cost, time cost, and penalty cost due to violation of regulations can be effectively reduced; a deployment plan that meets the resource utilization rate constraint can be quickly generated, improving the efficiency of personnel and material allocation; the utilization rate of various maintenance resources can be balanced, avoiding the situation of overuse or idleness of resources.

[0216] Figure 2 This is a schematic structural diagram of the digital twin-based intelligent operation and maintenance simulation system for a photovoltaic power station according to an embodiment of the present invention. As Figure 2 shown, the system includes:

[0217] A first unit for obtaining multi-modal detection signals through a photovoltaic module status acquisition device, based on the multi-modal detection signals, constructing a signal decomposition tree through discrete wavelet packet transform, extracting feature vectors from high-frequency sub-bands and low-frequency sub-bands, establishing a multi-dimensional feature space, performing tensor decomposition operations on the multi-dimensional feature space for multi-modal spatio-temporal alignment, generating a feature fusion matrix, and inputting the feature fusion matrix into a graph neural network model to construct a digital twin hierarchical structure including a physical data layer, a feature association layer, and a state representation layer;

[0218] A second unit for determining structured data based on the digital twin hierarchical structure; constructing a dynamic neural network, extracting fault features of the structured data, calculating the similarity score between the fault features and historical faults, establishing a Markov fault evolution model according to the similarity score sequence within a continuous time window, predicting the probability of fault occurrence; comparing the probability of fault occurrence with a preset threshold to trigger a maintenance warning signal;

[0219] The third unit is used to input the maintenance warning signal into the maintenance strategy model constructed based on the multi-agent reinforcement learning network, generate a set of candidate maintenance time windows, call the maintenance scheduling module to optimize and sort the set of candidate maintenance time windows to obtain an optimal time window sequence, input the optimal time window sequence into the resource allocation module, and use a genetic algorithm to generate a personnel and material allocation plan; calculate the maintenance cost of the personnel and material allocation plan, and iteratively optimize the policy parameters of the maintenance strategy model based on the maintenance cost, and output a final operation and maintenance implementation plan.

[0220] In the third aspect of the embodiments of the present invention,

[0221] A kind of electronic device is provided, including:

[0222] A processor;

[0223] A memory for storing instructions executable by the processor;

[0224] Wherein, the processor is configured to call the instructions stored in the memory to execute the method described above.

[0225] In the fourth aspect of the embodiments of the present invention,

[0226] A computer-readable storage medium is provided, on which computer program instructions are stored, and when the computer program instructions are executed by a processor, the method described above is implemented.

[0227] The present invention can be a method, a device, a system and / or a computer program product. The computer program product may include a computer-readable storage medium, on which computer-readable program instructions for executing various aspects of the present invention are uploaded.

[0228] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that: it is still possible to modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A digital-twin-based intelligent operation and maintenance simulation method for photovoltaic power stations, characterized in that Including: Obtain multi-modal detection signals through a photovoltaic module status acquisition device. Based on the multi-modal detection signals, construct a signal decomposition tree through discrete wavelet packet transform, extract feature vectors from high-frequency and low-frequency sub-bands, establish a multi-dimensional feature space, perform tensor decomposition operations on the multi-dimensional feature space for multi-modal spatio-temporal alignment, generate a feature fusion matrix, input the feature fusion matrix into a graph neural network model, and construct a digital twin hierarchical structure including a physical data layer, a feature association layer, and a state representation layer; Based on the digital twin hierarchical structure, determine structured data; Construct a dynamic neural network, extract fault features of the structured data, calculate the similarity score between the fault features and historical faults, establish a Markov fault evolution model based on the similarity score sequence within a continuous time window, and predict the probability of fault occurrence; Compare the probability of fault occurrence with a preset threshold to trigger a maintenance warning signal; Input the maintenance warning signal into a maintenance strategy model constructed based on a multi-agent reinforcement learning network to generate a set of candidate maintenance time windows, call a maintenance scheduling module to optimize and sort the set of candidate maintenance time windows to obtain an optimal time window sequence, input the optimal time window sequence into a resource allocation module, and use a genetic algorithm to generate a personnel and material allocation plan; calculate the maintenance cost of the personnel and material allocation plan, and iteratively optimize the policy parameters of the maintenance strategy model based on the maintenance cost to output a final operation and maintenance implementation plan; Based on the digital twin hierarchical structure, determine structured data; Construct a dynamic neural network, extract fault features of the structured data, calculate the similarity score between the fault features and historical faults, establish a Markov fault evolution model based on the similarity score sequence within a continuous time window, and predict the probability of fault occurrence; Comparing the probability of fault occurrence with a preset threshold to trigger a maintenance warning signal includes: Construct a dynamic neural network, divide the dynamic neural network into an input layer, a dynamic neuron layer, and an output layer according to a hierarchical structure, where each dynamic neuron in the dynamic neuron layer has an accumulated membrane potential value and an adaptive dynamic threshold. The input layer is connected to the dynamic neuron layer through the first inter-layer connection weights, and the dynamic neuron layer is connected to the output layer through the second inter-layer connection weights; Receive the structured data of the digital twin hierarchical structure and input the structured data into the input layer of the dynamic neural network; Calculate the membrane potential value of each dynamic neuron in the dynamic neuron layer, where the membrane potential value is obtained by adding the product of the previous moment's membrane potential value and the membrane potential decay coefficient to the eigenvalue after the structured data is transmitted through the first inter-layer connection weights, and then subtracting the dynamic threshold; Judge the magnitude relationship between the membrane potential value and the dynamic threshold. When the membrane potential value is greater than the dynamic threshold, generate a pulse output signal; Add the product of the previous moment's dynamic threshold and the threshold decay coefficient to the product of the previous moment's pulse output signal and the threshold increment coefficient to obtain the current moment's dynamic threshold; Process the pulse output signal, transmit the pulse output signal to the output layer through the second inter-layer connection weight, and perform weighted summation operation to generate a fault feature vector at the current moment; Calculate the cosine similarity between the fault feature vector at the current moment and the pre-stored historical fault feature vector to obtain the similarity score at the current moment; Collect the similarity scores within a continuous preset time period to form a similarity score sequence, divide the similarity score interval, use the similarity score interval as the state space, and count the number of interval transitions between adjacent moments in the similarity score sequence to generate a state transition probability matrix; Set the probability value corresponding to the similarity score interval where the similarity score at the current moment is located to 1, and set the probability values corresponding to other similarity score intervals to 0 to obtain the initial probability distribution vector at the current moment; Multiply the initial probability distribution vector at the current moment by the state transition probability matrix to obtain the fault probability distribution vector at the next moment, and extract the probability value of the similarity score interval corresponding to the fault state from the fault probability distribution vector as the fault occurrence probability; Compare the fault occurrence probability with a preset warning threshold, and trigger a maintenance warning signal when the fault occurrence probability is greater than the preset warning threshold.

2. The method according to claim 1, wherein Based on the multi-modal detection signal, construct a signal decomposition tree through discrete wavelet packet transform, extract feature vectors from high-frequency sub-bands and low-frequency sub-bands, and establish a multi-dimensional feature space including: Perform signal-to-noise ratio evaluation on the multi-modal detection signal to obtain a signal-to-noise ratio evaluation result, and adaptively select the optimal wavelet basis function according to the signal-to-noise ratio evaluation result; Perform the first-layer recursive decomposition on the multi-modal detection signal through the optimal wavelet basis function. Through the low-pass filter coefficients and high-pass filter coefficients of the quadrature mirror filter bank, perform low-frequency decomposition and high-frequency decomposition respectively to obtain the first-layer low-frequency component and the first-layer high-frequency component. Take the first-layer low-frequency component and the first-layer high-frequency component as the input signals for the second-layer recursive decomposition respectively, and use the same low-frequency decomposition and the same high-frequency decomposition to obtain the second-layer low-frequency component and the second-layer high-frequency component; repeat the iteration until the n-layer recursive decomposition is completed, and connect all the components obtained by the recursive decomposition according to the hierarchical relationship to form a multi-scale signal decomposition tree, where the 2 to the power of n components of the last layer are used as the leaf node signals of the multi-scale signal decomposition tree; Calculate the information gain of each leaf node signal of the multi-scale signal decomposition tree, set an adaptive threshold based on the information gain, and automatically select the optimal low-frequency sub-band signal and the optimal high-frequency sub-band signal from the leaf node signals according to the adaptive threshold; Perform square accumulation operation on the optimal high-frequency sub-band signal to obtain the high-frequency energy feature, and perform square accumulation operation on the optimal low-frequency sub-band signal to obtain the low-frequency energy feature; Calculate the energy distribution of the optimal high-frequency sub-band signal, and obtain the high-frequency normalized distribution through adaptive weight normalization processing. Calculate the energy distribution of the optimal low-frequency sub-band signal, and obtain the low-frequency normalized distribution through the adaptive weight normalization processing; Based on the high-frequency normalized distribution, calculate the high-frequency entropy value feature, and based on the low-frequency normalized distribution, calculate the low-frequency entropy value feature; Calculate the degree of dispersion of the optimal high-frequency sub-band signal to obtain the high-frequency standard deviation feature, and calculate the degree of dispersion of the optimal low-frequency sub-band signal to obtain the low-frequency standard deviation feature; Based on the information gain, perform feature importance weighting on the high-frequency energy feature, the low-frequency energy feature, the high-frequency entropy value feature, the low-frequency entropy value feature, the high-frequency standard deviation feature, and the low-frequency standard deviation feature to form a six-dimensional feature vector; By calculating the vector mean and vector standard deviation of the six-dimensional feature vector, subtract the vector mean from the six-dimensional feature vector and then divide by the vector standard deviation to obtain a normalized feature vector. Use the normalized feature vector as a feature point to establish a multi-dimensional feature space.

3. The method according to claim 2, wherein Perform tensor decomposition operation on the multi-dimensional feature space for multi-modal spatio-temporal alignment, and generate a feature fusion matrix including: Perform temporal partitioning on the feature vectors in the multi-dimensional feature space, and based on a sliding time window, reorganize the feature vectors into a third-order tensor according to the dimensions of the number of features, the number of time windows, and the number of modalities. Perform normalization processing on the feature data, time data, and modality data in the third-order tensor to obtain a normalized third-order tensor; Perform Tucker decomposition operation on the normalized third-order tensor, project it on the feature dimension, time dimension, and modality dimension respectively, and generate a core tensor, a feature mode matrix, a time mode matrix, and a modality mode matrix; Reconstruct the normalized third-order tensor based on the core tensor, the feature mode matrix, the time mode matrix, and the modality mode matrix, calculate the reconstruction error value, adjust the target dimension of the core tensor according to the reconstruction error value, and repeat the Tucker decomposition operation until the reconstruction error value is less than a preset error threshold to obtain an optimal core tensor; Use the feature mode matrix to map the feature vectors to a unified feature space, use the time mode matrix to align the feature vectors to a unified time reference, and use the modality mode matrix to map the feature vectors to a unified representation space to obtain aligned feature data, time data, and modality data; Perform Kronecker product operation on the time mode matrix and the modality mode matrix to obtain a combined mode matrix, and perform transpose operation on the combined mode matrix to obtain a transposed mode matrix; Multiply the feature mode matrix by the optimal core tensor to obtain a feature tensor, and multiply the feature tensor by the transposed mode matrix to generate a feature fusion matrix containing multi-modal spatio-temporal alignment information.

4. The method according to claim 3, wherein Input the feature fusion matrix into a graph neural network model, and construct a digital twin hierarchical structure including a physical data layer, a feature association layer, and a state representation layer: Map each photovoltaic module to a graph network node, construct graph network edges based on the physical connection relationships between the photovoltaic modules, and determine the topological structure of the graph neural network model; allocate the feature fusion matrix to the corresponding graph network nodes according to the acquisition locations to determine the initial node features of each graph network node; Extract the sensing data features and temporal correlation features from the initial node features, and organize the sensing data features and the temporal correlation features into a physical data layer; For each graph network node, collect the initial node features of adjacent nodes based on the graph network edges, determine the feature fusion weights according to the physical connection relationships, and obtain the neighbor feature aggregation value through weighted average operation; based on the neighbor feature aggregation value, generate the first-order correlation features of the graph network node through linear transformation and non-linear transformation, and fuse the first-order correlation features with the temporal correlation features of the physical data layer to construct a feature correlation layer; Input the first-order correlation features of the feature correlation layer into a preset feature extraction network, extract the deep correlation patterns of each photovoltaic module through multi-layer feature transformation, and generate high-order correlation features; input the high-order correlation features into a preset fully-connected neural network, and extract the characterization features of the overall operating state of the photovoltaic modules through multi-layer feature mapping to construct a state characterization layer; The state characterization layer integrates the original acquisition data corresponding to the physical data layer and the interaction relationships between the photovoltaic modules corresponding to the feature correlation layer to form a digital twin hierarchical structure of the photovoltaic modules.

5. The method according to claim 1, characterized in that, Input the maintenance warning signal into a maintenance strategy model constructed based on a multi-agent reinforcement learning network to generate a set of candidate maintenance time windows, and call the maintenance scheduling module to optimize and sort the set of candidate maintenance time windows to obtain an optimal time window sequence including: Receive a maintenance warning signal and construct a maintenance strategy model based on multi-agent reinforcement learning. The maintenance strategy model adopts a Monte Carlo tree structure, and uses each preset maintenance time point as a node of the Monte Carlo tree; Obtain the device status parameters, maintenance resource parameters, and production plan parameters of each node, calculate the status score according to the device status parameters, calculate the resource score according to the maintenance resource parameters, calculate the impact score according to the production plan parameters, and perform weighted calculation on the status score, the resource score, and the impact score to obtain the node comprehensive score; Calculate the exploration term based on the cumulative access times of the parent node, the current node access times, and a preset exploration coefficient, add the node comprehensive score to the exploration term to obtain the upper confidence bound value of each node, and select the node search path according to the upper confidence bound value to generate a set of candidate maintenance time windows based on the search path; Set the maintenance tasks in the set of candidate maintenance time windows as vertices, set the dependency relationships between the maintenance tasks as connecting edges, and construct a maintenance task dependency graph; recursively calculate the earliest execution time of each maintenance task along the forward path of the maintenance task dependency graph; recursively calculate the latest completion time of each maintenance task along the reverse path of the maintenance task dependency graph; Calculate the time margin for each maintenance task, where the time margin is equal to the latest completion time minus the earliest execution time minus the preset duration of the maintenance task; Perform priority sorting on the maintenance tasks in the candidate maintenance time window set according to the time margin to generate an optimal maintenance time window sequence.

6. The method according to claim 1, wherein Input the optimal time window sequence into the resource allocation module, and use the genetic algorithm to generate a personnel and material allocation plan, including: Encode the number of maintenance personnel required and the number of material requirements in the optimal time window sequence into a maintenance personnel allocation matrix and a material requirement matrix respectively to form a genetic code; Perform a weighted sum of the resource usage cost, time cost, and violation penalty cost to obtain the total cost, and calculate the fitness value of the genetic code based on the reciprocal of the total cost; Randomly select three genetic codes within the genetic population, compare the fitness values of the three genetic codes, and mark the genetic code with the highest fitness value as the parental code; Based on the interval [0, 1], generate the first random number, the second random number, and the third random number respectively. When the first random number is less than the preset crossover probability, enter the crossover operation; multiply the second random number by the length of the parental code and round down to obtain the crossover start position; multiply the third random number by the remaining code length and round down to obtain the crossover segment length; based on the gene segment determined by the crossover start position and the crossover segment length, map the gene segment to the range [0, 1] to obtain the transformed gene value, generate an expansion factor that follows the simulated binary distribution, and perform a linear combination of the transformed gene value and the expansion factor; restore the linearly combined gene value to the original value range according to the reverse interval mapping to generate the offspring code; Perform random mutation on the offspring code according to the preset mutation probability, and output the mutated code; calculate the actual workload of various maintenance resources in the mutated code respectively, and divide the actual workload by the available capacity of the corresponding resources to obtain the resource utilization rate; When the resource utilization rate is greater than the preset utilization threshold, adjust the corresponding values in the maintenance personnel allocation matrix and the material requirement matrix to reduce the resource utilization rate; Repeat the execution until the preset number of iterations is reached, and output the personnel and material allocation plan.

7. A digital twin-based intelligent operation and maintenance simulation system for photovoltaic power stations, which is used to implement the method described in any one of the preceding claims 1-6, characterized in that Including: The first unit is used to obtain multi-modal detection signals through a photovoltaic module state acquisition device, based on the multi-modal detection signals, construct a signal decomposition tree through discrete wavelet packet transform, extract feature vectors from high-frequency sub-bands and low-frequency sub-bands, establish a multi-dimensional feature space, perform tensor decomposition operations on the multi-dimensional feature space for multi-modal spatio-temporal alignment, generate a feature fusion matrix, input the feature fusion matrix into a graph neural network model, and construct a digital twin hierarchical structure including a physical data layer, a feature association layer, and a state representation layer; The second unit is used to determine structured data based on the digital twin hierarchical structure; Construct a dynamic neural network, extract the fault features of structured data, calculate the similarity score between the fault features and historical faults, establish a Markov fault evolution model based on the similarity score sequence within a continuous time window, and predict the probability of fault occurrence; Compare the probability of fault occurrence with a preset threshold to trigger a maintenance warning signal; The third unit is configured to input the maintenance warning signal into a maintenance strategy model constructed based on a multi-agent reinforcement learning network to generate a set of candidate maintenance time windows, call a maintenance scheduling module to optimize and sort the set of candidate maintenance time windows to obtain an optimal time window sequence, input the optimal time window sequence into a resource allocation module, and use a genetic algorithm to generate a personnel and material allocation plan; calculate the maintenance cost of the personnel and material allocation plan, iteratively optimize the policy parameters of the maintenance strategy model based on the maintenance cost, and output a final operation and maintenance implementation plan.

8. An electronic device, characterized in that, Comprising: A processor; A memory for storing instructions executable by the processor; Wherein, the processor is configured to call the instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, the method according to any one of claims 1 to 6 is implemented.

Images