Photovoltaic station intelligent operation and maintenance simulation method and system based on digital twinning

Abstract

The invention provides a photovoltaic station intelligent operation and maintenance simulation method and system based on digital twinning, and relates to the technical field of photovoltaic station operation and maintenance, and the method comprises the steps: constructing a digital twinning hierarchical structure through multi-modal detection signals through discrete wavelet packet transformation, tensor decomposition and a graph neural network, and determining structured data; extracting fault features by using a dynamic neural network, predicting a fault probability based on a Markov model, and triggering maintenance early warning; inputting the early warning signal into a maintenance strategy model based on multi-agent reinforcement learning, and generating a candidate maintenance time window set; the maintenance scheduling module optimizes sorting to obtain an optimal time window sequence, and the resource allocation module generates a personnel material allocation scheme by adopting a genetic algorithm; and iteratively optimizing the strategy parameters based on the maintenance cost, and outputting a final operation and maintenance implementation scheme.

Description

Technical Field

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

[0002] Photovoltaic power generation is an important component of clean energy, and its safe and stable operation is of great significance to the power system. As the scale of photovoltaic stations continues to expand and the number of equipment increases dramatically, traditional manual inspections and experience-based maintenance methods can no longer meet the needs of operation and maintenance. Digital twin technology provides new ideas for the intelligent operation and maintenance of photovoltaic stations by building digital mappings of physical entities, but the overall operation and maintenance efficiency and intelligence level still need to be improved.

[0003] There are some major problems with the existing photovoltaic field operation and maintenance methods: inconsistent data formats and sampling frequencies, and a lack of effective multi-source heterogeneous data fusion methods; fault diagnosis methods based on single features or simple rules are difficult to characterize the gradual process and multi-mode characteristics of equipment failures; maintenance decisions that rely on manual experience fail to fully consider maintenance resource constraints and equipment maintenance sequence constraints, 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 dynamic changes in equipment status and operating environment.

[0004] In summary, there is an urgent need for a photovoltaic station intelligent operation and maintenance simulation method based on digital twins, which realizes the deep fusion of multi-source data by constructing a digital twin hierarchical structure, accurately describes the evolution law of equipment failures using dynamic neural networks and Markov models, optimizes maintenance decisions and resource allocation based on multi-agent reinforcement learning and genetic algorithms, and finally forms a closed-loop optimized intelligent operation and maintenance system; it not only improves the accuracy and timeliness of fault diagnosis, but also can adaptively generate the optimal maintenance plan, significantly improves operation and maintenance efficiency, and reduces maintenance costs. The present invention can solve the problems in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a photovoltaic station intelligent operation and maintenance simulation method and system based on digital twins, which can solve the problems in the prior art.

[0006] According to a first aspect of the embodiments of the present invention,

[0007] A photovoltaic station intelligent operation and maintenance simulation method based on digital twin is provided, including:

[0008] A multimodal detection signal is obtained through a photovoltaic module state acquisition device. Based on the multimodal detection signal, a signal decomposition tree is constructed through discrete wavelet packet transform, feature vectors are extracted from high-frequency subbands and low-frequency subbands, and a multidimensional feature space is established. A tensor decomposition operation is performed on the multidimensional feature space to perform multimodal spatiotemporal alignment, and a feature fusion matrix is ​​generated. The feature fusion matrix is ​​input 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;

[0009] Based on the hierarchical structure of digital twins, structured data is determined; a dynamic neural network is constructed to extract fault features of structured data, and similarity scores between the fault features and historical faults are calculated. A Markov fault evolution model is established based on the similarity score sequence in a continuous time window to predict the probability of fault occurrence; the fault probability is compared with a preset threshold to trigger a maintenance warning signal;

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

[0011] In an optional embodiment,

[0012] Based on the multimodal detection signal, a signal decomposition tree is constructed by discrete wavelet packet transform, feature vectors are extracted from high-frequency sub-bands and low-frequency sub-bands, and a multidimensional feature space is established, including:

[0013] Performing a signal-to-noise ratio evaluation on the multimodal detection signal to obtain a signal-to-noise ratio evaluation result, and adaptively selecting an optimal wavelet basis function according to the signal-to-noise ratio evaluation result;

[0014] The multimodal detection signal is recursively decomposed at the first level by using the optimal wavelet basis function, and low-frequency decomposition and high-frequency decomposition are performed respectively by using the low-pass filter coefficient and the high-pass filter coefficient of the orthogonal mirror filter group to obtain a first-level low-frequency component and a first-level high-frequency component, and the first-level low-frequency component and the first-level high-frequency component are used as input signals of the second-level recursive decomposition respectively, and the same low-frequency decomposition and the same high-frequency decomposition are used to obtain a second-level low-frequency component and a second-level high-frequency component; it is repeated until n-level recursive decomposition is completed, and all components obtained by the recursive decomposition are connected according to a hierarchical relationship to form a multiscale signal decomposition tree, wherein the 2n-th power components of the last level are used as leaf node signals of the multiscale signal decomposition tree;

[0015] Calculating the information gain of each leaf node signal of the multi-scale signal decomposition tree, setting an adaptive threshold based on the information gain, and automatically selecting an optimal low-frequency subband signal and an optimal high-frequency subband signal from the leaf node signals according to the adaptive threshold;

[0016] Performing a square accumulation operation on the optimal high-frequency sub-band signal to obtain a high-frequency energy feature, and performing 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 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, based on the low-frequency normalized distribution, calculate the low-frequency entropy value feature;

[0018] Calculating the degree of discreteness of the optimal high-frequency sub-band signal to obtain a high-frequency standard deviation feature, and calculating the degree of discreteness of the optimal low-frequency sub-band signal to obtain a low-frequency standard deviation feature;

[0019] Based on the information gain, weighting 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 is performed on the feature importance to form a six-dimensional feature vector;

[0020] By calculating the vector mean and vector standard deviation of the six-dimensional feature vector, subtracting the vector mean from the six-dimensional feature vector and dividing the result by the vector standard deviation, a normalized feature vector is obtained, and the normalized feature vector is used as a feature point to establish a multidimensional feature space.

[0021] In an optional embodiment,

[0022] Performing a tensor decomposition operation on the multidimensional feature space to perform multimodal spatiotemporal alignment to generate a feature fusion matrix includes:

[0023] Performing time series division on feature vectors in a multidimensional feature space, reorganizing the feature vectors into a third-order tensor based on a sliding time window according to the dimensions of feature quantity, time window quantity, and modality quantity, and performing normalization processing on feature data, time data, and modality data in the third-order tensor to obtain a normalized third-order tensor;

[0024] Performing Tucker decomposition operation on the normalized third-order tensor, projecting it on the feature dimension, time dimension and modal dimension respectively, to generate a core tensor, a feature mode matrix, a time mode matrix and a modal mode matrix;

[0025] Reconstructing the normalized third-order tensor based on the core tensor, the characteristic mode matrix, the time mode matrix and the modal mode matrix, calculating the reconstruction error value, adjusting the target dimension of the core tensor according to the reconstruction error value, and repeatedly performing the Tucker decomposition operation until the reconstruction error value is less than a preset error threshold, thereby obtaining an optimal core tensor;

[0026] Mapping the feature vector to a unified feature space using the feature pattern matrix, aligning the feature vector to a unified time reference using the time pattern matrix, and mapping the feature vector to a unified representation space using the modal pattern matrix to obtain aligned feature data, time data, and modal data;

[0027] Performing a Kronecker product operation on the temporal pattern matrix and the modal pattern matrix to obtain a combined pattern matrix, and performing a transposition operation on the combined pattern matrix to obtain a transposed pattern matrix;

[0028] The feature pattern matrix is ​​multiplied by the optimal core tensor to obtain a feature tensor, and the feature tensor is multiplied by the transposed pattern matrix to generate a feature fusion matrix containing multimodal spatiotemporal alignment information.

[0029] In an optional embodiment,

[0030] Inputting the feature fusion matrix into the 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:

[0031] Mapping each photovoltaic module to a graph network node, constructing a graph network edge based on the physical connection relationship between each photovoltaic module, and determining the topological structure of the graph neural network model; assigning the feature fusion matrix to the corresponding graph network node according to the acquisition position, and determining the initial node feature of each graph network node;

[0032] Extracting sensor data features and time series correlation features from the initial node features, and organizing the sensor data features and the time series correlation features into a physical data layer;

[0033] For each of the graph network nodes, initial node features of adjacent nodes are collected based on the graph network edges, feature fusion weights are determined according to physical connection relationships, and neighbor feature aggregation values ​​are obtained through weighted average operations; based on the neighbor feature aggregation values, first-order correlation features of the graph network nodes are generated through linear transformation and nonlinear transformation, and the first-order correlation features are fused with the temporal correlation features of the physical data layer to construct a feature association layer;

[0034] Inputting the first-order correlation features of the feature correlation layer into a preset feature extraction network, extracting the deep correlation patterns of each photovoltaic module through multi-layer feature transformation, and generating high-order correlation features; inputting the high-order correlation features into a preset fully connected neural network, extracting the characterization features of the overall operating state of the photovoltaic module through multi-layer feature mapping, and constructing a state characterization layer;

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

[0036] In an optional embodiment,

[0037] Based on the hierarchical structure of digital twins, structured data is determined; a dynamic neural network is constructed to extract fault features of structured data, and similarity scores between the fault features and historical faults are calculated. A Markov fault evolution model is established based on the similarity score sequence in a continuous time window to predict the probability of fault occurrence; the fault probability is compared with a preset threshold, and a maintenance warning signal is triggered, including:

[0038] Constructing a dynamic neural network, dividing the dynamic neural network into an input layer, a dynamic neuron layer and an output layer according to a hierarchical structure, wherein 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 a first inter-layer connection weight, and the dynamic neuron layer is associated with the output layer through a second inter-layer connection weight;

[0039] Receiving structured data of a digital twin hierarchical structure, and inputting the structured data into an input layer of the dynamic neural network;

[0040] Calculating the membrane potential value of each dynamic neuron in the dynamic neuron layer, wherein the membrane potential value is obtained by adding the product of the membrane potential value at the previous moment and the membrane potential attenuation coefficient to the characteristic value of the structured data after being transferred through the first inter-layer connection weight, and then subtracting the dynamic threshold;

[0041] Determining the magnitude relationship between the membrane potential value and the dynamic threshold value, and generating a pulse output signal when the membrane potential value is greater than the dynamic threshold value;

[0042] The product of the dynamic threshold value at the previous moment and the threshold attenuation coefficient is added to the product of the pulse output signal at the previous moment and the threshold increment coefficient to obtain the dynamic threshold value at the current moment;

[0043] Processing the pulse output signal, transmitting the pulse output signal to the output layer through the second inter-layer connection weight and performing a weighted sum operation to generate a fault feature vector at the current moment;

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

[0045] Collecting the similarity scores within a continuous preset time period to form a similarity score sequence, dividing the similarity score intervals, using the similarity score intervals as state spaces, counting the number of interval transitions between adjacent moments in the similarity score sequence, and generating a state transition probability matrix;

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

[0047] Multiplying the initial probability distribution vector at the current moment by the state transition probability matrix to obtain a fault probability distribution vector at the next moment, and extracting 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] The fault occurrence probability is compared with a preset warning threshold, and when the fault occurrence probability is greater than the preset warning threshold, a maintenance warning signal is triggered.

[0049] In an optional embodiment,

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

[0051] Receive maintenance warning signals and build a maintenance strategy model based on multi-agent reinforcement learning, wherein 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] Obtaining equipment status parameters, maintenance resource parameters and production plan parameters of each node, calculating a status score according to the equipment status parameters, calculating a resource score according to the maintenance resource parameters, calculating an impact score according to the production plan parameters, and performing weighted calculation on the status score, the resource score and the impact score to obtain a comprehensive node score;

[0053] An exploration item is calculated based on the cumulative number of visits to the parent node, the number of visits to the current node, and a preset exploration coefficient, and the node comprehensive score is added to the exploration item to obtain a confidence upper bound value for each node. According to the confidence upper bound value, a node search path is selected, and a candidate maintenance time window set is generated based on the search path;

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

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

[0056] The maintenance tasks in the candidate maintenance time window set are prioritized according to the time margin to generate an optimal maintenance time window sequence.

[0057] In an optional embodiment,

[0058] Inputting the optimal time window sequence into the resource allocation module, and using the genetic algorithm to generate a personnel and material deployment plan includes:

[0059] Encode the maintenance personnel requirement quantity and material requirement quantity in the optimal time window sequence into maintenance personnel allocation matrix and material requirement matrix respectively to form genetic coding;

[0060] The resource usage cost, time cost and violation penalty cost are weighted and summed to obtain the total cost, and the fitness value of the genetic code is calculated based on the inverse of the total cost;

[0061] Randomly select three genetic codes from 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;

[0062] Based on the interval [0, 1], a first random number, a second random number and a third random number are generated respectively, and a crossover operation is performed when the first random number is less than a preset crossover probability; the second random number is multiplied by the length of the parent code and rounded down to obtain a crossover start position; the third random number is multiplied by the remaining code length and rounded down to obtain a crossover segment length; based on the gene segment determined by the crossover start position and the crossover segment length, the gene segment is converted into a range of [0, 1] according to interval mapping to obtain a converted gene value, an expansion factor that obeys a simulated binary distribution is generated, and the converted gene value is linearly combined with the expansion factor; the gene value after the linear combination is restored to the original value range according to the reverse interval mapping to generate a progeny code;

[0063] According to a preset mutation probability, random mutation is performed on the child code to output the mutated code; the actual workload of each type of maintenance resource in the mutated code is calculated respectively, and the actual workload is divided by the available capacity of the corresponding resource to obtain the resource utilization rate;

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

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

[0066] According to a second aspect of the embodiments of the present invention,

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

[0068] The first unit is used to obtain a multimodal detection signal through a photovoltaic module state acquisition device, construct a signal decomposition tree based on the multimodal detection signal through discrete wavelet packet transform, extract feature vectors from high-frequency subbands and low-frequency subbands, establish a multidimensional feature space, perform tensor decomposition operations on the multidimensional feature space to perform multimodal spatiotemporal 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;

[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 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 in a continuous time window, and predict the probability of fault occurrence; compare the fault probability with a preset threshold value to trigger a maintenance warning signal;

[0070] 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 candidate maintenance time window set to obtain the optimal time window sequence, input the optimal time window sequence into the resource allocation module, and use the genetic algorithm to generate a personnel and material deployment plan; calculate the maintenance cost of the personnel and material deployment plan, iteratively optimize the strategy parameters of the maintenance strategy model based on the maintenance cost, and output the final operation and maintenance implementation plan.

[0071] According to a third aspect of the embodiments of the present invention,

[0072] An electronic device is provided, comprising:

[0073] processor;

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

[0075] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0076] A fourth aspect of the embodiments of the present invention is:

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

[0078] In the embodiment of the present invention, a digital twin model is constructed by multimodal signal fusion and graph neural network, which can more comprehensively characterize the operating status of the photovoltaic station. Combined with the Markov fault evolution model, the probability of fault occurrence can be more accurately predicted, 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 allocation module, and finally obtain the optimal operation and maintenance implementation plan, thereby improving maintenance efficiency and reducing maintenance costs. The digital twin technology, artificial intelligence algorithm and operation and maintenance management are combined to realize the automation and intelligence of photovoltaic station operation and maintenance, reduce the need for manual intervention, and improve the level of operation and maintenance management. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 It is a flow chart of a photovoltaic station intelligent operation and maintenance simulation method based on digital twins according to an embodiment of the present invention;

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

[0081] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

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

[0083] Figure 1FIG. 1 is a flow chart of a photovoltaic station intelligent operation and maintenance simulation method based on digital twins according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0084] S101. Obtain a multimodal detection signal through a photovoltaic module state acquisition device, construct a signal decomposition tree based on the multimodal detection signal through discrete wavelet packet transform, extract feature vectors from high-frequency subbands and low-frequency subbands, establish a multidimensional feature space, perform tensor decomposition operations on the multidimensional feature space to perform multimodal spatiotemporal 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;

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

[0086] Then, the detection signal of each mode is decomposed by discrete wavelet packets. The signal is first decomposed into two sub-bands, low frequency and high frequency, and then each sub-band is further decomposed to form a multi-layer signal decomposition tree structure. Through this recursive decomposition method, the original signal is decomposed to the preset number of decomposition layers.

[0087] After obtaining the signal decomposition tree, extract features from high-frequency subbands and low-frequency subbands respectively. Extract statistical features such as mean, variance, and peak in the time domain; extract spectral features such as frequency center and bandwidth 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, the feature vectors of each modality are arranged in time order to construct a three-dimensional feature tensor, whose three dimensions correspond to the modality dimension, time dimension, and feature dimension. A tensor decomposition operation is performed on the feature tensor to decompose it into the sum of several rank-1 tensors, so as to align the data of different modalities in time and space dimensions and generate a feature fusion matrix.

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

[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 components are fully captured to improve the accuracy of signal analysis; a three-dimensional feature tensor is constructed and the multimodal data is aligned in time and space dimensions through tensor decomposition, effectively solving the problem of heterogeneous data fusion; the layered architecture of the graph neural network model is used to comprehensively describe the component status, and the layer-by-layer integration from physical data to state representation significantly improves the accuracy of operation status monitoring and evaluation; the digital twin model realizes real-time monitoring and diagnosis of the status of photovoltaic components, improves operation and maintenance efficiency and the reliability of component operation; the method can be extended to other photovoltaic field station components or energy management systems, and has universality and adaptability.

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

[0092] In an optional specific implementation, the monitoring data of the physical data layer, the correlation information of the feature correlation layer, and the state information of the state representation layer are first extracted from the digital twin hierarchical structure, and this information is normalized according to a unified data format to form a structured fault diagnosis data set.

[0093] Then, a dynamic neural network is constructed, which includes an input layer, a hidden layer, and an output layer. The input layer receives 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 fault features. The structured data is input into the network, and the fault features are extracted through forward propagation.

[0094] Then, the extracted fault features are compared with the fault features in the historical fault database. The cosine similarity calculation method is used to calculate the similarity score between the current fault feature and each historical fault feature. A similarity score is calculated for each time point, forming a similarity score sequence in a continuous time window.

[0095] Based on the similarity score sequence, a Markov fault evolution model is established. First, the fault state space is determined, and the similarity score is divided into multiple intervals corresponding to different fault states. Then, the state transition frequency between adjacent time points is counted, and the state transition probability matrix is ​​calculated. This matrix can be used to predict the fault state at the next moment, thereby obtaining the probability of the fault occurring.

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

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

[0098] S103. 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 candidate maintenance time window set, and obtain the optimal time window sequence. Input the optimal time window sequence into the resource allocation module, and use the genetic algorithm to generate a personnel and material deployment plan; calculate the maintenance cost of the personnel and material deployment plan, iteratively optimize the strategy parameters of the maintenance strategy model based on the maintenance cost, and output the final operation and maintenance implementation plan.

[0099] In an optional specific implementation, a maintenance strategy model based on multi-agent reinforcement learning is first constructed, which includes multiple agents, each of which is responsible for the maintenance decision of a type of equipment. The maintenance warning signal is used as the input state of the model, and each agent generates multiple maintenance time windows within a feasible time range according to its own strategy network to form a set of candidate maintenance time windows.

[0100] Then the maintenance scheduling module is started, which optimizes and sorts the candidate maintenance time window set by considering factors such as equipment importance, maintenance duration, and maintenance sequence constraints between equipment. First, the time windows that do not meet the maintenance sequence constraints are eliminated, and then the remaining time windows are weighted and sorted according to the equipment importance. Finally, the time windows are combined and optimized in combination with the maintenance duration to obtain the optimal time window sequence.

[0101] Then, 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 number of maintenance personnel and the required number of materials are encoded as chromosomes to construct the initial population. Then, through genetic operations such as fitness evaluation, selection, crossover and mutation, the chromosomes are continuously iterated and optimized, and finally a personnel and material allocation plan that meets resource constraints is generated.

[0102] The maintenance cost of the deployment plan is then calculated, including labor cost, material cost, time cost, and penalty cost for violations. 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 strategy of the agent is continuously optimized.

[0103] Finally, the optimized maintenance time window sequence and personnel and material deployment plan are combined to form the final operation and maintenance implementation plan. The plan includes detailed information such as specific maintenance time schedule, required number of maintenance personnel, material type and quantity, maintenance process, etc., which can directly guide the implementation of on-site maintenance work.

[0104] In this embodiment, based on the multi-agent reinforcement learning model, dynamic optimization of equipment maintenance decisions is achieved to ensure accurate maintenance strategies for different equipment; the importance of equipment, maintenance sequence constraints and maintenance period are comprehensively considered to optimize candidate maintenance time windows, generate optimal time arrangements and improve scheduling efficiency; genetic algorithms are used to optimize the allocation of personnel and materials to ensure maximum resource utilization while meeting maintenance requirements and constraints; maintenance costs are used as feedback signals to iteratively optimize the reinforcement learning strategy network to gradually reduce labor, material and time costs and improve economy; detailed time arrangements, resource requirements and maintenance procedures are provided to form directly executable operation and maintenance plans to improve the operability of on-site maintenance work; combined with optimized maintenance time and resource allocation, equipment downtime is reduced, operation and maintenance costs are reduced, and the overall operation efficiency and reliability of the system are improved.

[0105] In an optional implementation, based on the multimodal detection signal, a signal decomposition tree is constructed by discrete wavelet packet transform, feature vectors are extracted from high-frequency sub-bands and low-frequency sub-bands, and a multidimensional feature space is established, including:

[0106] Performing a signal-to-noise ratio evaluation on the multimodal detection signal to obtain a signal-to-noise ratio evaluation result, and adaptively selecting an optimal wavelet basis function according to the signal-to-noise ratio evaluation result;

[0107] The multimodal detection signal is recursively decomposed at the first level by using the optimal wavelet basis function, and low-frequency decomposition and high-frequency decomposition are performed respectively by using the low-pass filter coefficient and the high-pass filter coefficient of the orthogonal mirror filter group to obtain a first-level low-frequency component and a first-level high-frequency component, and the first-level low-frequency component and the first-level high-frequency component are used as input signals of the second-level recursive decomposition respectively, and the same low-frequency decomposition and the same high-frequency decomposition are used to obtain a second-level low-frequency component and a second-level high-frequency component; it is repeated until n-level recursive decomposition is completed, and all components obtained by the recursive decomposition are connected according to a hierarchical relationship to form a multiscale signal decomposition tree, wherein the 2n-th power components of the last level are used as leaf node signals of the multiscale signal decomposition tree;

[0108] Calculating the information gain of each leaf node signal of the multi-scale signal decomposition tree, setting an adaptive threshold based on the information gain, and automatically selecting an optimal low-frequency subband signal and an optimal high-frequency subband signal from the leaf node signals according to the adaptive threshold;

[0109] Performing a square accumulation operation on the optimal high-frequency sub-band signal to obtain a high-frequency energy feature, and performing a square accumulation operation on the optimal low-frequency sub-band signal to obtain a 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, based on the low-frequency normalized distribution, calculate the low-frequency entropy value feature;

[0111] Calculating the degree of discreteness of the optimal high-frequency sub-band signal to obtain a high-frequency standard deviation feature, and calculating the degree of discreteness of the optimal low-frequency sub-band signal to obtain a low-frequency standard deviation feature;

[0112] Based on the information gain, weighting 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 is performed on the feature importance to form a six-dimensional feature vector;

[0113] By calculating the vector mean and vector standard deviation of the six-dimensional feature vector, subtracting the vector mean from the six-dimensional feature vector and dividing the result by the vector standard deviation, a normalized feature vector is obtained, and the normalized feature vector is used as a feature point to establish a multidimensional feature space.

[0114] The n in the n-layer recursive decomposition specifically refers to the number of layers of wavelet decomposition, which determines the depth of the signal decomposition, that is, the degree of refinement of the 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 more complex, a higher number of decomposition layers (larger n) is required to capture more high-frequency and low-frequency information. If the signal is simpler 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 will be halved, and the resolution will be refined layer by layer. The number of decomposition layers n usually depends on the sampling frequency of the signal and the required time resolution; signal-to-noise ratio evaluation results. 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 too much decomposition may cause the high-frequency part to be submerged in the noise; based on the above logic, n is preferably between 3 and 8 layers.

[0115] In an optional specific implementation, first, the signal-to-noise ratio of the collected multimodal detection signal is evaluated. 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, the optimal wavelet basis function is adaptively selected. 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 purpose is to select the wavelet basis function that best suits the signal characteristics to improve the accuracy of feature extraction.

[0116] Next, the multimodal signal is recursively decomposed in multiple layers using the selected optimal wavelet basis function to construct a signal decomposition tree. For example, a three-layer recursive decomposition is performed on the signal. First, the original signal is decomposed into the first layer of low-frequency components and high-frequency components through a low-pass filter and a high-pass filter. Then, the first layer of low-frequency components and high-frequency components are decomposed again through the same low-pass and high-pass filters, respectively, to obtain the second layer of low-frequency and high-frequency components. This process is repeated until the three-layer decomposition is completed, and eight final components are obtained, which constitute the leaf nodes of the signal decomposition tree.

[0117] Then, the information gain of each leaf node signal in the signal decomposition tree is calculated. The information gain can measure the contribution 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. An adaptive threshold is set based on the calculated information gain, and the optimal low-frequency subband signal and the optimal high-frequency subband signal are selected 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 subband 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 squares of the amplitude 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 discreteness of the signal. For example, calculate the standard deviation of the amplitude of each subband signal to obtain the standard deviation feature value. Assume that 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 previously calculated information gain, the six extracted features (high-frequency energy feature, low-frequency energy feature, high-frequency entropy feature, low-frequency entropy feature, high-frequency standard deviation feature, and low-frequency standard deviation feature) are weighted to obtain a six-dimensional feature vector. For example, each feature value is multiplied by the information gain of the corresponding leaf node as a weight. Assuming that the information gains are 0.5 and 0.6 respectively, the weighted six-dimensional feature vector is [15, 103.2, 0.4, 0.54, 0.56, 0.67].

[0120] Finally, the six-dimensional eigenvector is normalized to construct a multidimensional feature space. For example, the mean and standard deviation of the six-dimensional eigenvector are calculated, and the normalized eigenvector is obtained by subtracting the mean from each eigenvalue and then dividing it by the standard deviation. The normalized eigenvector is used as a feature point, and all feature points constitute a multidimensional feature space. Assuming the mean is 31.63 and the standard deviation is 43.71, the normalized eigenvector is [-0.38, 1.64, -0.72, -0.70, -0.69, -0.68].

[0121] In this embodiment, by adaptively selecting wavelet basis functions and selecting the optimal sub-band based on information gain, key information in the signal can be better captured and the accuracy of feature extraction can be improved; by calculating multiple features and performing weighted combination, the impact of noise and interference on feature extraction can be reduced and the robustness of the features can be enhanced; by selecting the optimal sub-band through adaptive thresholds, the amount of data to be processed is reduced, the computational complexity is reduced, and the processing efficiency is improved.

[0122] In an optional implementation, performing a tensor decomposition operation on the multidimensional feature space to perform multimodal spatiotemporal alignment to generate a feature fusion matrix includes:

[0123] Performing time series division on feature vectors in a multidimensional feature space, reorganizing the feature vectors into a third-order tensor based on a sliding time window according to the dimensions of feature quantity, time window quantity, and modality quantity, and performing normalization processing on feature data, time data, and modality data in the third-order tensor to obtain a normalized third-order tensor;

[0124] Performing Tucker decomposition operation on the normalized third-order tensor, projecting it on the feature dimension, time dimension and modal dimension respectively, to generate a core tensor, a feature mode matrix, a time mode matrix and a modal mode matrix;

[0125] Reconstructing the normalized third-order tensor based on the core tensor, the characteristic mode matrix, the time mode matrix and the modal mode matrix, calculating the reconstruction error value, adjusting the target dimension of the core tensor according to the reconstruction error value, and repeatedly performing the Tucker decomposition operation until the reconstruction error value is less than a preset error threshold, thereby obtaining an optimal core tensor;

[0126] Mapping the feature vector to a unified feature space using the feature pattern matrix, aligning the feature vector to a unified time reference using the time pattern matrix, and mapping the feature vector to a unified representation space using the modal pattern matrix to obtain aligned feature data, time data, and modal data;

[0127] Performing a Kronecker product operation on the temporal pattern matrix and the modal pattern matrix to obtain a combined pattern matrix, and performing a transposition operation on the combined pattern matrix to obtain a transposed pattern matrix;

[0128] The feature pattern matrix is ​​multiplied by the optimal core tensor to obtain a feature tensor, and the feature tensor is multiplied by the transposed pattern matrix to generate a feature fusion matrix containing multimodal spatiotemporal 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 important information in the tensor, but also reduce the complexity of the data. Specifically, suppose there is a three-dimensional tensor (which can be regarded as a "multi-layer table"), and the three dimensions represent: feature dimension (such as the type of detection signal), time dimension (change at each time point), and modal dimension (such as different devices and sources of signals); Tucker decomposition will decompose the tensor into the following parts: a core tensor, which is a simplified version of the tensor and contains the most important information; three mode matrices, corresponding to features, time, and mode, respectively. These matrices describe how the data of each dimension is projected onto the core tensor. After decomposition, the original tensor can be reconstructed approximately using the core tensor and mode matrix. This method makes data analysis more efficient and facilitates the extraction of meaningful information from complex data.

[0130] The Kronecker product is a special multiplication between matrices that is used to generate a larger matrix. The corresponding rules are as follows: given two matrices A and B, the Kronecker product multiplies each element in matrix A by the entire matrix B; then, all 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 multidimensional feature space are divided into time series. Assume that there is a multimodal 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. The feature vectors of each modality are divided 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 the time window. Assuming that the data duration is 10 seconds and the time window length is 0.5 seconds, 20 time windows will be obtained.

[0132] Next, based on the sliding time window, the feature vector is reorganized into a third-order tensor according to the dimensions of the number of features, the number of time windows, and the number of modalities. Continuing with the above example, we will eventually get a third-order tensor with a dimension of [512 / 256 / 100, 20, 3], where the first dimension represents the number of features (depending on the modality), the second dimension represents the number of time windows, and the third dimension represents the number of modalities.

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

[0134] Perform Tucker decomposition on the normalized third-order tensor. Tucker decomposition decomposes the third-order tensor into a core tensor and three mode matrices (eigenmode matrix, time mode matrix, and modal mode matrix). These three mode matrices represent the projections on the feature dimension, time dimension, and modal dimension, respectively. Assume that the initial target dimension of the core tensor is 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 with 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] The target dimension of the core tensor is adjusted according to the reconstruction error value, and the Tucker decomposition operation is repeated 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 (for example, 0.01), the target dimension of the core tensor is adjusted (for example, increase or decrease the dimension) and the Tucker decomposition is repeated. Iterate continuously until the reconstruction error is less than the preset threshold to obtain the final core tensor and three pattern matrices.

[0137] The eigenvector is mapped to a unified eigenspace using the eigenmode matrix. The original eigenvector is multiplied by the eigenmode matrix to obtain the eigenvector mapped to the unified eigenspace. This allows features of different modes to be compared in the same space.

[0138] Use the time pattern matrix to align feature vectors to a unified time base. Apply the time pattern matrix to the time dimension to align features of different time windows to a unified time base and eliminate time deviations.

[0139] The modal pattern matrix is ​​used to map the feature vector to a unified representation space. The modal pattern matrix is ​​applied to the modal dimension to map the features of different modalities to a unified representation space, so that the information of different modalities can be integrated.

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

[0141] Perform a transposition operation on the combination pattern matrix to obtain a transposed pattern matrix. Perform a transposition operation on the combination pattern matrix.

[0142] Multiply the eigenmode matrix by the optimal core tensor to get the eigentensor. Multiply the eigenmode matrix by the resulting core tensor.

[0143] Multiply the feature tensor with the transposed pattern matrix to generate a feature fusion matrix containing multimodal spatiotemporal alignment information. Multiply the feature tensor with the transposed combined pattern matrix to obtain the final feature fusion matrix, which contains the fused feature representation of multimodal data under a unified spatiotemporal reference.

[0144] In this embodiment, by aligning the time and features of different modalities, the deviation between the data is eliminated, making the subsequent analysis more accurate and reliable; the features of different modalities are mapped to a unified representation space, so that the information of different modalities can be effectively integrated, thereby improving the richness and completeness of the feature representation; through the application of Tucker decomposition and pattern matrix, complex multimodal data is converted into a unified feature representation, simplifying the complexity of subsequent data processing.

[0145] In an optional implementation, the feature fusion matrix is ​​input 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, including:

[0146] Mapping each photovoltaic module to a graph network node, constructing a graph network edge based on the physical connection relationship between each photovoltaic module, and determining the topological structure of the graph neural network model; assigning the feature fusion matrix to the corresponding graph network node according to the acquisition position, and determining the initial node feature of each graph network node;

[0147] Extracting sensor data features and time series correlation features from the initial node features, and organizing the sensor data features and the time series correlation features into a physical data layer;

[0148] For each of the graph network nodes, initial node features of adjacent nodes are collected based on the graph network edges, feature fusion weights are determined according to physical connection relationships, and neighbor feature aggregation values ​​are obtained through weighted average operations; based on the neighbor feature aggregation values, first-order correlation features of the graph network nodes are generated through linear transformation and nonlinear transformation, and the first-order correlation features are fused with the temporal correlation features of the physical data layer to construct a feature association layer;

[0149] Inputting the first-order correlation features of the feature correlation layer into a preset feature extraction network, extracting the deep correlation patterns of each photovoltaic module through multi-layer feature transformation, and generating high-order correlation features; inputting the high-order correlation features into a preset fully connected neural network, extracting the characterization features of the overall operating state of the photovoltaic module through multi-layer feature mapping, and constructing a state characterization layer;

[0150] The state characterization layer integrates the original collected data corresponding to the physical data layer and the interactive relationship between the photovoltaic components corresponding to the feature association layer to form a digital twin hierarchical structure of the photovoltaic components.

[0151] In an optional specific implementation, first, a topological structure of a graph neural network is constructed. Each photovoltaic module is mapped as a node in the graph network. If there is a physical connection between two photovoltaic modules, such as through a junction box or an inverter, an edge is established between the two nodes. In this way, a graph network topological structure that reflects the physical connection relationship of the photovoltaic power station is formed. For example, a photovoltaic power station consisting of 100 photovoltaic modules can be mapped as 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 of the corresponding graph network node. For example, the seven feature values ​​of current, voltage, power, temperature, longitude, latitude, and installation angle can be combined into a 7-dimensional feature vector. Assuming that the collected data of a photovoltaic module is: current 10A, voltage 500V, power 5000W, temperature 30℃, longitude 120°, latitude 30°, installation angle 30°, then the initial node feature vector corresponding to the component is [10, 500, 5000, 30, 120, 30, 30].

[0153] Then, the digital twin hierarchical structure is constructed. First, the physical data layer is constructed. Sensor data features such as current, voltage, power, temperature, etc. are extracted from the initial node features, as well as features reflecting the timing relationship between photovoltaic components, such as historical power output data, etc. These features are organized into a physical data layer as the basis of the digital twin structure.

[0154] Secondly, construct a feature association layer. For each graph network node, collect the initial node features of its neighbor nodes. Determine the feature fusion weight according to the physical connection relationship between the nodes. For example, nodes that are closer and more tightly 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. The voltage features of the neighbor nodes are 500V, 480V, and 510V, respectively. The neighbor feature aggregation value is 500×0.5+480×0.3+510×0.2=496V. Then, perform linear and nonlinear transformations on the neighbor feature aggregation value, such as using the activation function ReLU, to generate the first-order association features of the node. The first-order association features are fused with the temporal association features of the physical data layer to form a feature association layer that reflects the interactive relationship between photovoltaic modules.

[0155] Finally, the state representation layer is constructed. The first-order correlation features of the feature association layer are input into a preset feature extraction network, such as a multi-layer convolutional neural network or a recurrent neural network. Through multi-layer feature transformation, the deep correlation patterns of each photovoltaic module are extracted to generate high-order correlation features. The high-order correlation features are input into a preset fully connected neural network, and the characterization features of the overall operating state of the photovoltaic module are extracted 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 association layer to form a digital twin hierarchical structure of the photovoltaic module, which fully reflects the operating state of the photovoltaic module.

[0156] In this embodiment, by fusing multi-source data and considering the correlation between photovoltaic components, the operating status 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 components can be quickly identified, and potential failure risks can be predicted, thereby 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 cleaning plans, adjusting component operating parameters, etc., to improve the operating efficiency and life of the photovoltaic power station.

[0157] In an optional implementation, based on the digital twin hierarchical structure, structured data is determined; a dynamic neural network is constructed to extract fault features of the structured data, similarity scores between the fault features and historical faults are calculated, and a Markov fault evolution model is established according to a similarity score sequence in a continuous time window to predict the probability of fault occurrence; the fault probability is compared with a preset threshold, and a maintenance warning signal is triggered, including:

[0158] Constructing a dynamic neural network, dividing the dynamic neural network into an input layer, a dynamic neuron layer and an output layer according to a hierarchical structure, wherein 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 a first inter-layer connection weight, and the dynamic neuron layer is associated with the output layer through a second inter-layer connection weight;

[0159] Receiving structured data of a digital twin hierarchical structure, and inputting the structured data into an input layer of the dynamic neural network;

[0160] Calculating the membrane potential value of each dynamic neuron in the dynamic neuron layer, wherein the membrane potential value is obtained by adding the product of the membrane potential value at the previous moment and the membrane potential attenuation coefficient to the characteristic value of the structured data after being transferred through the first inter-layer connection weight, and then subtracting the dynamic threshold;

[0161] Determining the magnitude relationship between the membrane potential value and the dynamic threshold value, and generating a pulse output signal when the membrane potential value is greater than the dynamic threshold value;

[0162] The product of the dynamic threshold value at the previous moment and the threshold attenuation coefficient is added to the product of the pulse output signal at the previous moment and the threshold increment coefficient to obtain the dynamic threshold value at the current moment;

[0163] Processing the pulse output signal, transmitting the pulse output signal to the output layer through the second inter-layer connection weight and performing a weighted sum operation to generate a fault feature vector at the current moment;

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

[0165] Collecting the similarity scores within a continuous preset time period to form a similarity score sequence, dividing the similarity score intervals, using the similarity score intervals as state spaces, counting the number of interval transitions between adjacent moments in the similarity score sequence, and generating a state transition probability matrix;

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

[0167] Multiplying the initial probability distribution vector at the current moment by the state transition probability matrix to obtain a fault probability distribution vector at the next moment, and extracting 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] The fault occurrence probability is compared with a preset warning threshold, and when the fault occurrence probability is greater than the preset warning threshold, a maintenance warning signal is triggered.

[0169] The membrane potential value specifically refers to a dynamic variable used to simulate the state of neurons in the neuron model. It is a quantitative expression of the "excitement level" of neurons, which determines whether the neurons will trigger pulse signals (i.e. "discharge"). In biology, the membrane potential value corresponds to the value of the potential difference between the inside and outside of 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, a digital twin hierarchical structure is established. The equipment system to be monitored, such as a wind turbine, is decomposed into multiple levels, such as blades, transmission system, generator, etc. Each level can be further decomposed into sub-components until the most basic sensor level. Each level and sub-component is associated with corresponding structured data, such as vibration data, temperature data, current data, etc., to form a complete digital twin hierarchical structure.

[0171] Next, a dynamic neural network is constructed. The network consists of an input layer, a dynamic neural layer, and an output layer. The number of neurons in the input layer corresponds to the dimensions of the structured data, such as the dimensions of wind turbine vibration, temperature, current, etc. The dynamic neural layer contains multiple dynamic neurons, each with a membrane potential value and a dynamic threshold. The input layer and the dynamic neural layer are associated with each other through the first layer of connection weights. The dynamic neural layer and the output layer are associated with each other through the second layer of connection weights.

[0172] Get structured data from the digital twin hierarchical structure. For example, get the vibration data, temperature data, and current data of the 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 adding the eigenvalue of the input data after passing through the first layer connection weight after attenuation of the membrane potential value at the previous moment, and then subtracting the dynamic threshold. For example, assuming that the membrane potential value at the previous moment is 0.5, the attenuation 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] Determine the relationship between the membrane potential value and the dynamic threshold. If the membrane potential value is greater than the dynamic threshold, the neuron generates a pulse output signal. For example, if the membrane potential value is 0.55 and the dynamic threshold is 0.1, a pulse output signal is generated.

[0175] Update the dynamic threshold. The dynamic threshold at the current moment is obtained by adding the product of the pulse output signal at the previous moment and the threshold increment coefficient after the dynamic threshold at the previous moment is attenuated. For example, assuming that the dynamic threshold at the previous moment is 0.1, the attenuation 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. The pulse output signals of all dynamic neurons are transmitted to the output layer through the second layer connection weights, and a weighted sum operation is performed to generate the fault feature vector at the current moment.

[0177] Calculate the similarity score between the current fault feature vector 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 the similarity score is 0.85.

[0178] Collect similarity scores in continuous time periods to form a similarity score sequence. For example, collect similarity scores in 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] The probability value corresponding to the interval where the current similarity score is located is set to 1, and the probability values ​​of other intervals are set to 0, to obtain the initial probability distribution vector at the current moment. For example, if the current similarity score is 0.85, which falls in the 0.8-1.0 interval, the probability value of this interval is 1, and the probability values ​​of other intervals are 0.

[0180] The initial probability distribution vector at the current moment is multiplied by the state transition probability matrix to obtain the fault probability distribution vector at the next moment. The probability value of the interval corresponding to the fault state is extracted from the fault probability distribution vector as the probability of the fault occurring.

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

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

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

[0184] Receive maintenance warning signals and build a maintenance strategy model based on multi-agent reinforcement learning, wherein 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;

[0185] Obtaining equipment status parameters, maintenance resource parameters and production plan parameters of each node, calculating a status score according to the equipment status parameters, calculating a resource score according to the maintenance resource parameters, calculating an impact score according to the production plan parameters, and performing weighted calculation on the status score, the resource score and the impact score to obtain a comprehensive node score;

[0186] An exploration item is calculated based on the cumulative number of visits to the parent node, the number of visits to the current node, and a preset exploration coefficient, and the node comprehensive score is added to the exploration item to obtain a confidence upper bound value for each node. According to the confidence upper bound value, a node search path is selected, and a candidate maintenance time window set is generated based on the search path;

[0187] The maintenance tasks in the candidate maintenance time window set are set as vertices, and the dependencies between the maintenance tasks are set as connecting edges to construct a maintenance task dependency graph; the earliest execution time of each maintenance task is recursively calculated along the forward path of the maintenance task dependency graph; and the latest completion time of each maintenance task is recursively calculated along the reverse path of the maintenance task dependency graph;

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

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

[0190] In an optional specific implementation, a maintenance warning signal is received and a maintenance strategy model is constructed. 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, we preset 12 noon every day in the next week as a possible maintenance time point, then these time points correspond to seven nodes on the tree respectively.

[0191] Get the relevant parameters of each node. For each maintenance time point (i.e. each node), we need to obtain three types of parameters: equipment status parameters, maintenance resource parameters, and production plan parameters. Equipment status parameters can include indicators such as equipment operation time, temperature, and vibration frequency. Maintenance resource parameters can include the number of available manpower, spare parts inventory, etc. Production plan parameters can include the scheduled number of production tasks, product delivery date, etc. For example, at the maintenance node on Monday, the equipment operation time is 100 hours, the temperature is 60 degrees Celsius, the vibration frequency is 10Hz, the available manpower is 5 people, the spare parts inventory is sufficient, and the scheduled number of production tasks is 1,000 pieces.

[0192] Calculate the score of each node. According to the obtained parameters, calculate the status score, resource score and impact score respectively. The status score reflects the health of the equipment, the resource score reflects the availability of maintenance resources, and the impact score reflects the 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, according to the parameters on Monday, the status score is calculated to be 80 points, the resource score is 90 points, and the impact score is 70 points.

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

[0194] Select a node search path and generate a set of candidate maintenance time windows. According to the confidence upper bound value of each node, select a search path from the root node to the leaf node. The nodes on the search path constitute the set of candidate maintenance time windows. For example, after searching, the path finally selected contains three nodes: Tuesday, Thursday, and Saturday. Then, 12 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 candidate maintenance time window set as a vertex of the graph, and set the dependencies between maintenance tasks as connecting edges. For example, if Tuesday's maintenance task must be completed before Thursday's maintenance task, then draw a directed edge from Tuesday's vertex to Thursday's vertex.

[0196] Calculate the earliest execution time and latest completion time of each maintenance task. 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. For example, assuming that Tuesday's task has no predecessor task, the earliest execution time is 12 noon on Tuesday; assuming that Thursday's task depends on Tuesday's task, and Tuesday's task duration is 2 hours, the earliest execution time of Thursday's task is 2 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 the earliest execution time minus the preset duration of the maintenance task. For example, if the latest completion time of Thursday's task is 8 am on Friday and the duration is 4 hours, the time margin is 20 hours.

[0198] Generate the optimal maintenance time window sequence. Prioritize each maintenance task according to its time margin. The smaller the time margin, the higher the priority. The sorted maintenance task sequence constitutes the optimal maintenance time window sequence. For example, assuming that the time margin for Tuesday's task is 24 hours, the time margin for Thursday's task is 20 hours, and the time margin for Saturday's task is 30 hours, then the optimal maintenance time window sequence is Thursday, Tuesday, and Saturday.

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

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

[0201] Encode the maintenance personnel requirement quantity and material requirement quantity in the optimal time window sequence into maintenance personnel allocation matrix and material requirement matrix respectively to form genetic coding;

[0202] The resource usage cost, time cost and violation penalty cost are weighted and summed to obtain the total cost, and the fitness value of the genetic code is calculated based on the inverse of the total cost;

[0203] Randomly select three genetic codes from 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], a first random number, a second random number and a third random number are generated respectively, and a crossover operation is performed when the first random number is less than a preset crossover probability; the second random number is multiplied by the length of the parent code and rounded down to obtain a crossover start position; the third random number is multiplied by the remaining code length and rounded down to obtain a crossover segment length; based on the gene segment determined by the crossover start position and the crossover segment length, the gene segment is converted into a range of [0, 1] according to interval mapping to obtain a converted gene value, an expansion factor that obeys a simulated binary distribution is generated, and the converted gene value is linearly combined with the expansion factor; the gene value after the linear combination is restored to the original value range according to the reverse interval mapping to generate a progeny code;

[0205] According to a preset mutation probability, random mutation is performed on the child code to output the mutated code; the actual workload of each type of maintenance resource in the mutated code is calculated respectively, and the actual workload is divided by the available capacity of the corresponding resource to obtain the resource utilization rate;

[0206] When the resource utilization rate is greater than a preset utilization rate threshold, the resource utilization rate is reduced by adjusting corresponding values ​​in the maintenance personnel allocation matrix and the material requirement matrix;

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

[0208] In an optional specific implementation, first, the required number of maintenance personnel and the required number of materials need to be encoded to form the coding structure required by the genetic algorithm. Specifically, the required number of maintenance personnel is represented as a matrix, each row of the matrix represents a time window, each column represents a type of maintenance personnel, and the value of the matrix element represents the required number of maintenance personnel of this type within the time window. Similarly, the required number of materials is also represented as a matrix, each row of the matrix represents a time window, each column represents a type of material, and the value of the matrix element represents the required number of materials of this type within the time window. Combining these two matrices together forms the coding of the genetic algorithm, which is called genetic coding. For example, assuming there are three time windows, two types of maintenance personnel and three types of materials, 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, you need to define the fitness function of the genetic algorithm. The fitness function is used to evaluate the pros and cons 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 inverse of the total cost is used as the fitness value of the genetic encoding. The higher the fitness value, the better the scheme corresponding to the encoding. For example, assuming that the resource usage cost weight is 0.5, the time cost weight is 0.3, and the violation penalty cost weight is 0.2, the resource usage cost of a certain scheme is 100, the time cost is 50, and the violation penalty cost is 20, then the total cost of the scheme 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, three genetic codes are first randomly selected from the population, their fitness values ​​are compared, and the code with the highest fitness value is selected as the parent code. Then, a crossover operation is performed. Three random numbers between 0 and 1 are generated. The first random number is used to determine whether to perform a crossover operation. If the first random number is less than the preset crossover probability, a crossover operation is performed. The second random number is used to determine the starting position of the crossover. The second random number is multiplied by the length of the parent code and rounded down to obtain the crossover starting position. The third random number is used to determine the length of the crossover segment. The third random number is multiplied by the remaining code length and rounded down to obtain the crossover segment length. The selected gene segment is converted to the range of 0 to 1 according to the interval mapping, and then an expansion factor that obeys the simulated binary distribution is generated. The converted gene value is linearly combined with the expansion factor to obtain a new gene value. Finally, the new gene value is restored to the original value range according to the reverse interval mapping to generate the offspring code. For example, assuming that the crossover probability is 0.8 and the first random number is 0.7, a crossover operation is performed. Assuming the parent code length is 10 and the second random number is 0.3, the crossover start position is 3. Assuming the third random number is 0.2, the crossover segment length is 2.

[0211] Next, perform the mutation operation. According to the preset mutation probability, randomly mutate the offspring code to generate the mutation code. For example, assuming the mutation probability is 0.1, there is a 10% probability of mutating a gene in the offspring code.

[0212] Then, calculate the resource utilization. Calculate the actual workload of each type of maintenance resource in the variant code, and divide the actual workload by the available capacity of the corresponding resource to get the resource utilization. For example, assuming that the actual workload of a certain type of maintenance personnel is 80 hours and the available capacity is 100 hours, the resource utilization of this type of maintenance personnel is 80 / 100=0.8.

[0213] Finally, adjust the resource utilization. If the resource utilization is greater than the preset utilization threshold, you need to adjust the corresponding values ​​in the maintenance personnel allocation matrix and the material requirement matrix to reduce the resource utilization. For example, if the utilization threshold is 0.9 and the resource utilization of a certain type of maintenance personnel is 0.95, you need to reduce the number of allocated maintenance personnel of this type, 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 deployment plan.

[0215] In this embodiment, by optimizing the personnel and material allocation plan, the resource utilization cost, time cost and penalty cost caused by violation of regulations can be effectively reduced; the allocation plan that meets the resource utilization constraints can be quickly generated, thereby improving the efficiency of personnel and material allocation; the utilization of various maintenance resources can be balanced to avoid excessive use or idleness of resources.

[0216] Figure 2 FIG. 1 is a schematic diagram of the structure of a photovoltaic station intelligent operation and maintenance simulation system based on digital twins according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0217] The first unit is used to obtain a multimodal detection signal through a photovoltaic module state acquisition device, construct a signal decomposition tree based on the multimodal detection signal through discrete wavelet packet transform, extract feature vectors from high-frequency subbands and low-frequency subbands, establish a multidimensional feature space, perform tensor decomposition operations on the multidimensional feature space to perform multimodal spatiotemporal 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;

[0218] 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 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 in a continuous time window, and predict the probability of fault occurrence; compare the fault probability with a preset threshold value 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 candidate maintenance time window set to obtain the optimal time window sequence, input the optimal time window sequence into the resource allocation module, and use the genetic algorithm to generate a personnel and material deployment plan; calculate the maintenance cost of the personnel and material deployment plan, iteratively optimize the strategy parameters of the maintenance strategy model based on the maintenance cost, and output the final operation and maintenance implementation plan.

[0220] According to a third aspect of the embodiments of the present invention,

[0221] An electronic device is provided, comprising:

[0222] processor;

[0223] a memory for storing processor-executable instructions;

[0224] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0225] A fourth aspect of the embodiments of the present invention is:

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

[0227] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[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 aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A photovoltaic station intelligent operation and maintenance simulation method based on digital twins, characterized in that: include: A multimodal detection signal is obtained through a photovoltaic module state acquisition device. Based on the multimodal detection signal, a signal decomposition tree is constructed through discrete wavelet packet transform, feature vectors are extracted from high-frequency subbands and low-frequency subbands, and a multidimensional feature space is established. A tensor decomposition operation is performed on the multidimensional feature space to perform multimodal spatiotemporal alignment, and a feature fusion matrix is ​​generated. The feature fusion matrix is ​​input 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; Determine structured data based on the digital twin hierarchical structure; Construct a dynamic neural network, extract fault features from structured data, calculate the similarity scores between the fault features and historical faults, establish a Markov fault evolution model based on the similarity score sequence in a continuous time window, and predict the probability of fault occurrence; Comparing the failure occurrence probability with a preset threshold value, and triggering a maintenance warning signal; The maintenance warning signal is input into a maintenance strategy model constructed based on a multi-agent reinforcement learning network to generate a set of candidate maintenance time windows, and the maintenance scheduling module is called to optimize and sort the candidate maintenance time window set to obtain an optimal time window sequence. The optimal time window sequence is input into a resource allocation module, and a genetic algorithm is used to generate a personnel and material deployment plan; the maintenance cost of the personnel and material deployment plan is calculated, and the strategy parameters of the maintenance strategy model are iteratively optimized based on the maintenance cost, and the final operation and maintenance implementation plan is output.

2. The method according to claim 1, characterized in that: Based on the multimodal detection signal, a signal decomposition tree is constructed by discrete wavelet packet transform, feature vectors are extracted from high-frequency sub-bands and low-frequency sub-bands, and a multidimensional feature space is established, including: Performing a signal-to-noise ratio evaluation on the multimodal detection signal to obtain a signal-to-noise ratio evaluation result, and adaptively selecting an optimal wavelet basis function according to the signal-to-noise ratio evaluation result; The multimodal detection signal is recursively decomposed at the first level by using the optimal wavelet basis function, and low-frequency decomposition and high-frequency decomposition are performed respectively by using the low-pass filter coefficient and the high-pass filter coefficient of the orthogonal mirror filter group to obtain a first-level low-frequency component and a first-level high-frequency component, and the first-level low-frequency component and the first-level high-frequency component are used as input signals of the second-level recursive decomposition respectively, and the same low-frequency decomposition and the same high-frequency decomposition are used to obtain a second-level low-frequency component and a second-level high-frequency component; it is repeated until n-level recursive decomposition is completed, and all components obtained by the recursive decomposition are connected according to a hierarchical relationship to form a multiscale signal decomposition tree, wherein the 2n-th power components of the last level are used as leaf node signals of the multiscale signal decomposition tree; Calculating the information gain of each leaf node signal of the multi-scale signal decomposition tree, setting an adaptive threshold based on the information gain, and automatically selecting an optimal low-frequency subband signal and an optimal high-frequency subband signal from the leaf node signal according to the adaptive threshold; Performing a square accumulation operation on the optimal high-frequency sub-band signal to obtain a high-frequency energy feature, and performing a square accumulation operation on the optimal low-frequency sub-band signal to obtain a 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, based on the low-frequency normalized distribution, calculate the low-frequency entropy value feature; Calculating the degree of discreteness of the optimal high-frequency sub-band signal to obtain a high-frequency standard deviation feature, and calculating the degree of discreteness of the optimal low-frequency sub-band signal to obtain a low-frequency standard deviation feature; Based on the information gain, weighting 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 is performed on the feature importance to form a six-dimensional feature vector; By calculating the vector mean and vector standard deviation of the six-dimensional feature vector, subtracting the vector mean from the six-dimensional feature vector and dividing the result by the vector standard deviation, a normalized feature vector is obtained, and the normalized feature vector is used as a feature point to establish a multidimensional feature space.

3. The method according to claim 2, characterized in that Performing a tensor decomposition operation on the multidimensional feature space to perform multimodal spatiotemporal alignment to generate a feature fusion matrix includes: Performing time series division on feature vectors in a multidimensional feature space, reorganizing the feature vectors into a third-order tensor based on a sliding time window according to the dimensions of feature quantity, time window quantity, and modality quantity, and performing normalization processing on feature data, time data, and modality data in the third-order tensor to obtain a normalized third-order tensor; Performing Tucker decomposition operation on the normalized third-order tensor, projecting it on the feature dimension, time dimension and modal dimension respectively, to generate a core tensor, a feature mode matrix, a time mode matrix and a modal mode matrix; Reconstructing the normalized third-order tensor based on the core tensor, the characteristic mode matrix, the time mode matrix and the modal mode matrix, calculating the reconstruction error value, adjusting the target dimension of the core tensor according to the reconstruction error value, and repeatedly performing the Tucker decomposition operation until the reconstruction error value is less than a preset error threshold, thereby obtaining an optimal core tensor; Mapping the feature vector to a unified feature space using the feature pattern matrix, aligning the feature vector to a unified time reference using the time pattern matrix, and mapping the feature vector to a unified representation space using the modal pattern matrix to obtain aligned feature data, time data, and modal data; Performing a Kronecker product operation on the temporal pattern matrix and the modal pattern matrix to obtain a combined pattern matrix, and performing a transposition operation on the combined pattern matrix to obtain a transposed pattern matrix; The feature pattern matrix is ​​multiplied by the optimal core tensor to obtain a feature tensor, and the feature tensor is multiplied by the transposed pattern matrix to generate a feature fusion matrix containing multimodal spatiotemporal alignment information.

4. The method according to claim 3, characterized in that Inputting the feature fusion matrix into the 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: Mapping each photovoltaic module to a graph network node, constructing a graph network edge based on the physical connection relationship between each photovoltaic module, and determining the topological structure of the graph neural network model; assigning the feature fusion matrix to the corresponding graph network node according to the acquisition position, and determining the initial node feature of each graph network node; Extracting sensor data features and time series correlation features from the initial node features, and organizing the sensor data features and the time series correlation features into a physical data layer; For each of the graph network nodes, initial node features of adjacent nodes are collected based on the graph network edges, feature fusion weights are determined according to physical connection relationships, and neighbor feature aggregation values ​​are obtained through weighted average operations; based on the neighbor feature aggregation values, first-order correlation features of the graph network nodes are generated through linear transformation and nonlinear transformation, and the first-order correlation features are fused with the temporal correlation features of the physical data layer to construct a feature association layer; Inputting the first-order correlation features of the feature correlation layer into a preset feature extraction network, extracting the deep correlation patterns of each photovoltaic module through multi-layer feature transformation, and generating high-order correlation features; inputting the high-order correlation features into a preset fully connected neural network, extracting the characterization features of the overall operating state of the photovoltaic module through multi-layer feature mapping, and constructing a state characterization layer; The state characterization layer integrates the original collected data corresponding to the physical data layer and the interactive relationship between the photovoltaic components corresponding to the feature association layer to form a digital twin hierarchical structure of the photovoltaic components.

5. The method according to claim 1, characterized in that Based on the hierarchical structure of digital twins, structured data is determined; a dynamic neural network is constructed to extract fault features of structured data, and similarity scores between the fault features and historical faults are calculated. A Markov fault evolution model is established based on the similarity score sequence in a continuous time window to predict the probability of fault occurrence; Comparing the failure occurrence probability with a preset threshold and triggering a maintenance warning signal includes: Constructing a dynamic neural network, dividing the dynamic neural network into an input layer, a dynamic neuron layer and an output layer according to a hierarchical structure, wherein 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 a first inter-layer connection weight, and the dynamic neuron layer is associated with the output layer through a second inter-layer connection weight; Receiving structured data of a digital twin hierarchical structure, and inputting the structured data into an input layer of the dynamic neural network; Calculating the membrane potential value of each dynamic neuron in the dynamic neuron layer, wherein the membrane potential value is obtained by adding the product of the membrane potential value at the previous moment and the membrane potential attenuation coefficient to the characteristic value of the structured data after being transferred through the first inter-layer connection weight, and then subtracting the dynamic threshold; Determining the magnitude relationship between the membrane potential value and the dynamic threshold value, and generating a pulse output signal when the membrane potential value is greater than the dynamic threshold value; The product of the dynamic threshold value at the previous moment and the threshold attenuation coefficient is added to the product of the pulse output signal at the previous moment and the threshold increment coefficient to obtain the dynamic threshold value at the current moment; Processing the pulse output signal, transmitting the pulse output signal to the output layer through the second inter-layer connection weight and performing a weighted sum operation to generate a fault feature vector at the current moment; Calculating the cosine similarity between the current moment fault feature vector and the pre-stored historical fault feature vector to obtain a current moment similarity score; Collecting the similarity scores within a continuous preset time period to form a similarity score sequence, dividing the similarity score intervals, using the similarity score intervals as state spaces, counting the number of interval transitions between adjacent moments in the similarity score sequence, and generating a state transition probability matrix; The probability value corresponding to the similarity score interval where the similarity score at the current moment is located is set to 1, and the probability values ​​corresponding to other similarity score intervals are set to 0, to obtain the initial probability distribution vector at the current moment; Multiplying the initial probability distribution vector at the current moment by the state transition probability matrix to obtain a fault probability distribution vector at the next moment, and extracting the probability value of the similarity score interval corresponding to the fault state from the fault probability distribution vector as the fault occurrence probability; The fault occurrence probability is compared with a preset warning threshold, and when the fault occurrence probability is greater than the preset warning threshold, a maintenance warning signal is triggered.

6. The method according to claim 1, characterized in that The maintenance warning signal is input into the maintenance strategy model constructed based on the multi-agent reinforcement learning network to generate a set of candidate maintenance time windows, and the maintenance scheduling module is called to optimize and sort the set of candidate maintenance time windows to obtain the optimal time window sequence including: Receive maintenance warning signals and build a maintenance strategy model based on multi-agent reinforcement learning, wherein 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; Obtaining equipment status parameters, maintenance resource parameters and production plan parameters of each node, calculating a status score according to the equipment status parameters, calculating a resource score according to the maintenance resource parameters, calculating an impact score according to the production plan parameters, and performing weighted calculation on the status score, the resource score and the impact score to obtain a comprehensive node score; An exploration item is calculated based on the cumulative number of visits to the parent node, the number of visits to the current node, and a preset exploration coefficient, and the node comprehensive score is added to the exploration item to obtain a confidence upper bound value for each node. According to the confidence upper bound value, a node search path is selected, and a candidate maintenance time window set is generated based on the search path; The maintenance tasks in the candidate maintenance time window set are set as vertices, and the dependencies between the maintenance tasks are set as connecting edges to construct a maintenance task dependency graph; the earliest execution time of each maintenance task is recursively calculated along the forward path of the maintenance task dependency graph; and the latest completion time of each maintenance task is recursively calculated along the reverse path of the maintenance task dependency graph; Calculating a time margin for each maintenance task, where the time margin is equal to the latest completion time minus the earliest execution time minus a preset duration of the maintenance task; The maintenance tasks in the candidate maintenance time window set are prioritized according to the time margin to generate an optimal maintenance time window sequence.

7. The method according to claim 1, characterized in that Inputting the optimal time window sequence into the resource allocation module, and using the genetic algorithm to generate a personnel and material deployment plan includes: Encode the maintenance personnel requirement quantity and material requirement quantity in the optimal time window sequence into maintenance personnel allocation matrix and material requirement matrix respectively to form genetic coding; The resource usage cost, time cost and violation penalty cost are weighted and summed to obtain the total cost, and the fitness value of the genetic code is calculated based on the inverse of the total cost; Randomly select three genetic codes from 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; Based on the interval [0, 1], a first random number, a second random number and a third random number are generated respectively, and a crossover operation is performed when the first random number is less than a preset crossover probability; the second random number is multiplied by the length of the parent code and rounded down to obtain a crossover start position; the third random number is multiplied by the remaining code length and rounded down to obtain a crossover segment length; based on the gene segment determined by the crossover start position and the crossover segment length, the gene segment is converted into a range of [0, 1] according to interval mapping to obtain a converted gene value, an expansion factor that obeys a simulated binary distribution is generated, and the converted gene value is linearly combined with the expansion factor; the gene value after the linear combination is restored to the original value range according to the reverse interval mapping to generate a progeny code; According to a preset mutation probability, random mutation is performed on the child code to output the mutated code; the actual workload of each type of maintenance resource in the mutated code is calculated respectively, and the actual workload is divided by the available capacity of the corresponding resource to obtain the resource utilization rate; When the resource utilization rate is greater than a preset utilization rate threshold, the resource utilization rate is reduced by adjusting corresponding values ​​in the maintenance personnel allocation matrix and the material requirement matrix; Repeat until the preset number of iterations is reached and output the personnel and material deployment plan.

8. A photovoltaic station intelligent operation and maintenance simulation system based on digital twin, used to implement the method described in any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain a multimodal detection signal through a photovoltaic module state acquisition device, construct a signal decomposition tree based on the multimodal detection signal through discrete wavelet packet transform, extract feature vectors from high-frequency subbands and low-frequency subbands, establish a multidimensional feature space, perform tensor decomposition operations on the multidimensional feature space to perform multimodal spatiotemporal 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 fault features from structured data, calculate the similarity scores between the fault features and historical faults, establish a Markov fault evolution model based on the similarity score sequence in a continuous time window, and predict the probability of fault occurrence; Comparing the failure occurrence probability with a preset threshold value, and triggering a maintenance warning signal; 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 candidate maintenance time window set to obtain the optimal time window sequence, input the optimal time window sequence into the resource allocation module, and use the genetic algorithm to generate a personnel and material deployment plan; calculate the maintenance cost of the personnel and material deployment plan, iteratively optimize the strategy parameters of the maintenance strategy model based on the maintenance cost, and output the final operation and maintenance implementation plan.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images