Photovoltaic power generation intelligent early warning method and system

Abstract

The invention discloses a photovoltaic power generation intelligent early warning method and system, and the method comprises the steps: determining the multivariable time series data of a photovoltaic array, and the multivariable time series data comprises the electrical feature data and environment feature data of a photovoltaic array string; inputting the multivariable time series data into the fault prediction model, extracting space correlation feature vectors among the photovoltaic array group strings, and carrying out time series dependence coding on the space correlation feature vectors to obtain fault sensitive feature vectors, so as to predict the fault precursor of the photovoltaic array based on the fault sensitive feature vectors; and based on the constructed fault evolution knowledge graph, knowledge link reasoning is carried out on the predicted fault precursor to generate graded early warning information, and the photovoltaic array fault risk is intervened in advance according to the graded early warning information. According to the method, the multivariable time series data is collected, features are extracted and coded through the fault prediction model, and the fault evolution knowledge graph is combined for reasoning, so that early weak signal recognition is realized, the early warning timeliness and accuracy are improved, and the refined intelligent requirements are met.

Description

Technical Field

[0001] This application relates to the field of circuit technology, and more specifically, to a method and system for intelligent early warning of photovoltaic power generation. Background Technology

[0002] With the global energy structure transitioning towards clean energy, photovoltaic (PV) power generation, as an important form of renewable energy utilization, has seen its installed capacity grow rapidly. Large-scale PV power plants typically consist of numerous PV modules connected in series or parallel to form PV arrays. During long-term operation, affected by natural environmental factors (such as sunlight, temperature, and dust) and equipment aging, PV arrays are prone to various faults such as hot spots, aging wiring, and loose connectors. If these faults are not detected and addressed in a timely manner, they can not only lead to a decrease in power generation efficiency but also potentially cause safety hazards and economic losses. Therefore, effective fault early warning for PV arrays and early intervention in fault risks are of great significance for ensuring the safe and stable operation of PV power plants and improving power generation efficiency.

[0003] Currently, the fault early warning method for photovoltaic power generation systems uses threshold judgment for fault identification. For example, by setting thresholds for the normal range of current and voltage, an alarm is issued when the monitored value exceeds the threshold.

[0004] However, since the operating status of photovoltaic arrays is affected by a variety of factors, the precursory features before a fault often manifest as weak coordinated changes in multiple parameters. Existing methods rely on threshold judgment for analysis, which makes it difficult to effectively capture these early weak fault signals, resulting in insufficient early warning timeliness. This makes it difficult to improve the accuracy and reliability of fault warnings and meet the needs of photovoltaic power plants for refined and intelligent early warning. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a method and system for intelligent early warning of photovoltaic power generation, which can at least alleviate the aforementioned technical problems.

[0006] The technical solutions provided in this application are as follows: A method for intelligent early warning of photovoltaic power generation, comprising: Step 1: Determine the multivariate time-series data of the photovoltaic array, which includes the electrical characteristic data and environmental characteristic data of the photovoltaic array string; Step 2: Input multivariate time series data into the fault prediction model, extract the spatial correlation feature vector between photovoltaic array strings and perform time-dependent encoding on the spatial correlation feature vector to obtain the fault-sensitive feature vector, so as to predict the fault precursors of the photovoltaic array based on the fault-sensitive feature vector; Step 3: Based on the constructed fault evolution knowledge graph, perform knowledge linking reasoning on the predicted fault precursors to generate hierarchical early warning information and intervene in advance on the fault risk of photovoltaic arrays accordingly.

[0007] A photovoltaic power generation intelligent early warning system, comprising: A time-series data processing unit is used to determine the multivariate time-series data of the photovoltaic array, wherein the multivariate time-series data includes electrical characteristic data and environmental characteristic data of the photovoltaic array string; The prediction unit is used to input multivariate time series data into the fault prediction model, extract the spatial correlation feature vector between photovoltaic array strings and perform time-dependent encoding on the spatial correlation feature vector to obtain the fault-sensitive feature vector, so as to predict the fault precursors of the photovoltaic array based on the fault-sensitive feature vector. The intervention unit is used to perform knowledge linking reasoning on predicted fault precursors based on the constructed fault evolution knowledge graph, so as to generate hierarchical early warning information and intervene in the fault risk of photovoltaic array in advance.

[0008] This technical solution collects multivariate time-series data containing electrical and environmental features, extracts spatial correlation features between strings using a fault prediction model, and performs time-dependent encoding to obtain fault-sensitive feature vectors. This enables the capture of coordinated changes in multiple parameters and effectively identifies weak early fault signals. Based on the fault-sensitive feature vectors, fault precursors are predicted. A fault evolution knowledge graph is used to infer and generate tiered early warning information for early intervention. Compared to threshold judgment methods, this approach can detect fault signs earlier and improve the timeliness of early warnings. The fault prediction model comprehensively considers spatial correlation and time-dependent relationships, combining knowledge graph inference to reduce the limitations of single threshold judgments, improve the accuracy and reliability of early warnings, and meet the refined and intelligent early warning needs of photovoltaic power plants. Attached Figure Description

[0009] Figure 1 This is a flowchart illustrating the intelligent early warning method for photovoltaic power generation according to an embodiment of this application.

[0010] Figure 2 This is a flowchart illustrating the photovoltaic power generation intelligent early warning system according to an embodiment of this application. Detailed Implementation

[0011] like Figure 1 As shown in the figure, this application provides a method for intelligent early warning of photovoltaic power generation, which includes: Step 1: Determine the multivariate time-series data of the photovoltaic array, which includes the electrical characteristic data and environmental characteristic data of the photovoltaic array string; Step 2: Input multivariate time series data into the fault prediction model, extract the spatial correlation feature vector between photovoltaic array strings and perform time-dependent encoding on the spatial correlation feature vector to obtain the fault-sensitive feature vector, so as to predict the fault precursors of the photovoltaic array based on the fault-sensitive feature vector; Step 3: Based on the constructed fault evolution knowledge graph, perform knowledge linking reasoning on the predicted fault precursors to generate hierarchical early warning information and intervene in advance on the fault risk of photovoltaic arrays accordingly.

[0012] Optionally, step 1, determining the multivariate time-series data of the photovoltaic array, wherein the multivariate time-series data includes electrical characteristic data and environmental characteristic data of the photovoltaic array strings, specifically includes the following steps: Step 11: Obtain the spatiotemporally labeled data stream obtained by spatiotemporally co-sampling of the photovoltaic array using a multi-scale sensor network; Step 12: Perform multimodal feature extraction on the spatiotemporal labeled data stream to obtain multimodal spatiotemporal correlation features; Step 13: Perform spatiotemporal alignment and synchronization calibration on the multimodal spatiotemporal correlation features to generate multivariate time series data.

[0013] Optionally, step 11: Determine the spatiotemporally labeled data stream obtained by the multi-scale sensor network for spatiotemporal co-sampling of the photovoltaic array, specifically including the following steps: Step 111: Obtain the multidimensional heterogeneous data stream obtained by spatiotemporal collaborative sampling of the photovoltaic array using a multi-scale sensor network; Step 112: Perform spatiotemporal anchoring and feature enhancement processing on the multidimensional heterogeneous data stream to generate a spatiotemporally correlated feature stream; Step 113: Perform dynamic calibration and semantic annotation on the spatiotemporal correlation feature stream to obtain the spatiotemporal labeled data stream.

[0014] Preferably, step 111: Initiate the spatiotemporal collaborative sampling process of the multi-scale sensor network. Based on a preset benchmark sampling period (e.g., string electrical parameters every 10 minutes / time, module temperature every 30 minutes / time, environmental parameters every 5 minutes / time), trigger all sensors in the multi-scale sensor network to synchronously collect data to obtain the raw data stream. Specifically, the clocks of each acquisition node in the multi-scale sensor network are calibrated using a precise time protocol to ensure that the sampling time deviation of sensors at different locations is within 1ms. Real-time monitoring of key parameter changes in the raw data stream is also performed, such as when the string current fluctuation exceeds 8%, the temperature difference between modules is greater than 3℃, or the irradiance changes by more than 200W / m² within one hour. 2At that time, the sampling frequency of the sensor area and adjacent areas corresponding to the key parameter change is automatically increased to 5-10 times the normal frequency. At the same time, the trigger time and location information of the frequency adjustment event corresponding to the key parameter change are recorded to form a multi-dimensional heterogeneous data stream containing multi-scale monitoring data and different spatiotemporal resolutions.

[0015] Preferably, step 112: each data point in the multidimensional heterogeneous data stream is spatiotemporally anchored and bound with a precise spatiotemporal label. The time label uses a UTC timestamp accurate to milliseconds, and the spatial label uses a four-level encoding rule of "array area-row number-string number-module sequence number" to form a triplet data structure of "timestamp + spatial code + physical quantity value". Based on this triplet data structure, time-frequency features are extracted from electrical parameter data through wavelet transform, three-dimensional reconstruction is performed on temperature field data to obtain spatial distribution features, and environmental parameters are filtered to eliminate noise. The extracted and processed multi-type features are associated according to the spatiotemporal labels and integrated into a spatiotemporal associated feature stream containing spatiotemporal association information.

[0016] Preferably, step 113 involves: dynamically calibrating the spatiotemporal correlation feature stream; verifying the integrity of the feature stream through an edge computing gateway; removing data with missing spatiotemporal labels or abnormal physical quantity values; correcting data with disordered timestamps due to transmission delays by performing linear interpolation based on the time series of adjacent normal data; labeling each data point with potential fault semantic tags (such as "precursor to hot spots" or "connector aging") based on the calibrated spatiotemporal correlation feature stream and combining it with a historical fault case library, and calculating the fault confidence; and arranging the calibrated and labeled data in an orderly manner along the time axis to ultimately form a time-ordered spatiotemporally labeled data stream containing multi-scale monitoring information and with each data point carrying precise spatiotemporal and semantic labels.

[0017] Optionally, step 12: Perform multimodal feature extraction on the spatiotemporally labeled data stream to obtain multimodal spatiotemporal correlation features, specifically including the following steps: Step 121: Perform adaptive multi-resolution decomposition on the spatiotemporal labeled data stream to obtain time-frequency semantic logical units; Step 122: Perform topology-aware spatial correlation modeling on the time-frequency semantic logic unit to generate a spatiotemporal correlation tensor; Step 123: Perform multimodal attention fusion processing on the spatiotemporal correlation tensor to obtain multimodal spatiotemporal correlation features.

[0018] Preferably, step 121: Adaptive multi-resolution decomposition processing is performed on the spatiotemporal labeled data stream. First, the spatiotemporal labeled data stream is divided into data subsets according to data type (electrical parameters, temperature field, environmental parameters) to obtain electrical parameter subsets, temperature field data subsets, and environmental parameter subsets. For the electrical parameter subset, wavelet transform is used to perform time-frequency decomposition to obtain electrical parameter decomposition features. For example, the time domain signal is decomposed into sub-band signals of different frequencies through a sliding window. The window size is adaptively adjusted according to the fluctuation amplitude of current and voltage (when the fluctuation amplitude exceeds 10%, the window is reduced to 1 / 3 of the original size). The temperature field data subset is spatially multi-scale divided, for example, according to the hierarchical structure of 5×5 modular grid and 10×10 string region, spatial distribution features are extracted as temperature field decomposition features. For the environmental parameter subset, a variable step time window is used (the step size is 10 minutes when the irradiance is stable and shortened to 2 minutes when there are drastic changes). The time sequence is decomposed (minutes) to obtain the component decomposition features; the obtained electrical parameter decomposition features, temperature field decomposition features, and component decomposition features form multi-scale features and are bound to the corresponding spatiotemporal tags to generate time-frequency semantic logical units containing time-frequency features, spatial scale information, and semantic descriptions (such as "high-frequency current fluctuations" and "local high-temperature areas").

[0019] Preferably, step 122: topology-aware spatial association modeling is performed on the time-frequency semantic logic units. A graph structure is constructed based on the physical topology of the photovoltaic array (string connection relationship, module arrangement). Each time-frequency semantic logic unit is mapped to a node in the graph. The edge weights between nodes are dynamically calculated based on the string electrical distance (such as wire length, number of connected nodes) and the probability of historical fault co-occurrence. The graph structure is propagated through a graph convolutional network (GCN). Each layer of GCN captures local spatial associations within 1-3 hop neighborhoods (such as temperature coordination changes between adjacent strings) and global spatial associations across regions (such as the influence of irradiance differences between the array edge and the center strings) to obtain spatial association features and temporal sequence information. The spatial association features output by GCN are fused with the temporal sequence information to form a three-dimensional spatiotemporal association tensor containing timestamps, spatial coordinates, and multimodal association features.

[0020] Preferably, step 123 involves performing multimodal attention fusion processing on the three-dimensional spatiotemporal correlation tensor. For example, a three-branch attention mechanism is constructed: the temporal attention branch focuses on key periods of fault evolution (such as 1 hour before a sudden temperature rise), and calculates the temporal attention weights through the hidden states of gated recurrent units (GRUs); the spatial attention branch highlights fault-sensitive areas (such as clusters of historically frequent faults), and allocates spatial attention weights based on node degree centrality and the fault impact range; the modal attention branch dynamically adjusts the contribution of electrical parameters, temperature field, and environmental parameters (such as enhancing the weight of current parameters under low irradiance conditions) to obtain intermodal attention weights; the three-dimensional spatiotemporal correlation tensor is weighted and fused based on the attention weights of these three branches (temporal attention weight, spatial attention weight, and intermodal attention weight), and the modal differences are eliminated by residual connection and normalization of the weighted results, ultimately outputting multimodal spatiotemporal correlation features that fuse multidimensional correlation information.

[0021] Optionally, step 13: Perform spatiotemporal alignment and synchronization calibration on the multimodal spatiotemporal correlation features to generate multivariate time series data, specifically including the following steps: Step 131: Perform spatiotemporal alignment processing on the multimodal spatiotemporal correlation features with topological constraints to obtain the spatial calibration feature tensor; Step 132: Perform multimodal synchronization calibration on the spatial calibration feature tensor to generate a spatiotemporal synchronization feature matrix; Step 133: Perform semantic enhancement and standardization on the spatiotemporal synchronization feature matrix to generate multivariate time series data.

[0022] Preferably, step 131: Perform spatiotemporal alignment processing of multimodal spatiotemporal correlation features with topological constraints. First, construct a unified three-dimensional spatial coordinate system based on the physical topology of the photovoltaic array (string arrangement, electrical connection relationship). Map the feature data (electrical, temperature, environment) of different modes in the multimodal spatiotemporal correlation features to the corresponding positions in this coordinate system. Use a dynamic time warping algorithm to align the sampling time series of different sensors with the clock of the array main control unit as the reference, so that the multimodal feature data at the same time are accurately matched on the time axis. In this process, spatial topological constraints (such as the electrical distance threshold of adjacent strings and the temperature conduction range of modules) are introduced to correct the spatial coordinates of abnormal feature data that deviate from the topological relationship, so as to obtain a spatial calibration feature tensor with accurate spatial position and time series alignment.

[0023] Preferably, step 132: Perform multimodal synchronous calibration processing on the spatial calibration feature tensor, construct a cross-modal time offset model based on the physical correlation of multimodal data (such as the hysteresis relationship between irradiance change and current change), calculate the time response difference between different modal feature data and perform compensation calibration; fuse the calibrated feature data through the Kalman filter algorithm to dynamically estimate the real operating state of the photovoltaic array and correct the modal deviation caused by sensor measurement noise, and obtain the processed feature data; rearrange the processed feature data according to time series and spatial location to generate a spatiotemporal synchronous feature matrix with rows and columns corresponding to time points and spatial nodes, and elements being multimodal fused feature values.

[0024] Preferably, in step 133, when performing semantic enhancement and standardization on the spatiotemporal synchronization feature matrix, semantic labels are added to each feature value in the spatiotemporal synchronization feature matrix by combining the entity relationships in the fault evolution knowledge graph (such as "current abnormal fluctuation - hot spot association" and "temperature gradient mutation - connector fault association"). At the same time, the max-min standardization method is used to map feature data of different magnitudes along with semantic labels to the [0,1] interval to eliminate the difference in dimensions. Then, the key features in the spatiotemporal synchronization feature matrix are extracted by principal component analysis, retaining more than 95% of the feature information, and finally generating multivariate time series data with unified dimensions, semantic labels, and temporal coherence.

[0025] Optionally, step 2, inputting multivariate time-series data into the fault prediction model, extracting spatial correlation feature vectors between photovoltaic array strings and performing time-dependent encoding on the spatial correlation feature vectors to obtain fault-sensitive feature vectors, and predicting fault precursors of the photovoltaic array based on the fault-sensitive feature vectors, specifically includes the following steps: Step 21: Based on the spatial feature extraction network in the fault prediction model, perform feature extraction on the multivariate time series data to generate spatial correlation feature vectors between photovoltaic array strings; Step 22: Based on the temporal dependency coding network in the fault prediction model, perform temporal dependency coding on multiple pairs of spatially correlated feature vectors to obtain fault-sensitive feature vectors; Step 23: Based on the spatiotemporal evolution inference network in the fault prediction model, perform multi-scale temporal evolution and probabilistic inference on the fault-sensitive feature vector to predict the fault precursors of the photovoltaic array.

[0026] Optionally, step 21: Based on the spatial feature extraction network in the fault prediction model, feature extraction is performed on the multivariate time series data to generate spatial correlation feature vectors between photovoltaic array strings, specifically including the following steps: Step 211: Based on the topology-aware embedding layer, perform physical structure encoding and modality fusion processing on the multivariate time series data to obtain the array topology feature map; Step 212: Based on the spatial dependency propagation layer, perform fault propagation mode mining on the array topology feature map to generate a spatial correlation feature tensor; Step 213: Based on the fault-sensitive enhancement layer, the spatial correlation feature tensor is processed to highlight fault features and suppress noise, so as to obtain the spatial correlation feature vector between photovoltaic array strings.

[0027] Optionally, in step 211, the topology-aware embedding layer adopts a hybrid structure of "1D convolution + graph embedding convolution". First, for the electrical and environmental features in the multivariate time-series data, a 1D convolution kernel (size 3×1, stride 1) is used to extract time-series features. Each modality (current, voltage, temperature, irradiance) corresponds to an independent 1D convolution branch, and a preliminary time-series feature map is generated through 64 convolution kernels. Next, a graph embedding convolution module is introduced to transform the photovoltaic array topology into an adjacency matrix as the convolution kernel weights. A 2D graph convolution (kernel_size=3×3, padding=1) is used to encode the spatial coordinates of the string. Each string node interacts with the nodes within the adjacent 3×3 range to output a 128-dimensional spatial embedding vector. Finally, the time-series feature map and the spatial embedding vector are fused through 1×1 convolution kernels (number 128) to obtain the array topology feature map.

[0028] Optionally, in step 212, the spatial dependency propagation layer adopts a "hierarchical graph convolution" structure, which includes two convolutional branches: local and global. The local branch uses 2D graph convolution (kernel_size=3×3, stride=1, padding=1), and the kernel weights are dynamically adjusted according to the physical distance of the strings (adjacent strings have a weight coefficient of 0.8, and a string with an interval of 0.3). It captures local associations within the 1-2 hop neighborhood in the array topology feature map through 32 convolutional kernels to obtain local association features within the region. The global branch uses dilated graph convolution (kernel_size=5×5, dilation=2, padding=4), which expands the receptive field to cover the 3-5 hop neighborhood in the array topology feature map. 64 convolutional kernels extract global association features across regions. The local association features and global association features are fused through residual connections (1×1 convolutions are added after adjusting the dimension), and finally processed by the BatchNorm layer and the ReLU activation function to generate a spatial association feature tensor that preserves multi-level spatial dependencies.

[0029] Optionally, in step 213, the fault-sensitive enhancement layer uses "attention convolution + noise reduction convolution" as its core structure. First, the fault-sensitive attention module uses 1×1 convolutions (the number of which is the same as the dimension of the input features) to generate an attention weight matrix and fuses it with the spatial correlation feature tensor to obtain fault-sensitive features. Then, the fault feature regions in the fault-sensitive features are highlighted by the Sigmoid activation function. Next, 2D convolutions with adaptive thresholds (kernel_size=3×3, padding=1) are used to enhance the fault-sensitive features, where the convolution kernel weights are dynamically updated according to the historical fault feature distribution (the weights of the high-temperature regions corresponding to hot spot faults are increased by 2 times). Finally, 1×1 convolutions (number 64) are used to compress the dimensions of the enhanced fault-sensitive features, and a Dropout layer (rate=0.2) is used to suppress noise interference, outputting a highly discriminative spatial correlation feature vector, with each dimension corresponding to the enhanced fault-sensitive mode.

[0030] Optionally, step 22: Based on the temporal dependency coding network in the fault prediction model, perform temporal dependency coding on multiple pairs of spatially correlated feature vectors to obtain fault-sensitive feature vectors, specifically including the following steps: Step 221: Based on the multi-scale temporal feature extraction layer, perform temporal window decomposition and parallel feature extraction on multiple pairs of spatially correlated feature vectors to obtain a multi-scale temporal feature set; Step 222: Based on the attention-enhanced temporal encoder, perform temporal dependency enhancement and fault feature highlighting on the multi-scale temporal feature set to obtain the fault-sensitive feature vector.

[0031] Preferably, step 221: Based on the multi-scale temporal feature extraction layer, time window decomposition and parallel feature extraction processing are performed on multiple pairs of spatially correlated feature vectors. This layer adopts a "multi-window 1D convolution" structure, setting three time windows (short-term window with 5 time steps, medium-term window with 20 time steps, and long-term window with 60 time steps), corresponding to the rapid change period, stable evolution period, and long-term accumulation period of fault development, respectively; for each pair of spatially correlated feature vectors, sliding cutting is performed according to the three window lengths to form multiple sets of temporal segments; an independent 1D convolution branch is configured for each window (16 kernels with kernel_size=3 for the short-term window, 32 kernels with kernel_size=5 for the medium-term window, and 64 kernels with kernel_size=7 for the long-term window), and local temporal patterns (such as short-term current pulse, medium-term temperature rise, and long-term performance degradation) are extracted at different time scales respectively; the features output by each branch are concatenated in chronological order to obtain a multi-scale temporal feature set containing multi-scale temporal information.

[0032] Preferably, step 222: Based on the attention-enhanced temporal encoder, the multi-scale temporal feature set is subjected to temporal dependency enhancement and fault feature highlighting processing. The encoder consists of a bidirectional LSTM and a temporal attention mechanism. First, the multi-scale temporal feature set is input into the bidirectional LSTM network. The forward LSTM captures the temporal dependency from the past to the present (such as the influence of historical temperature changes on the current state), and the backward LSTM captures the potential trend from the present to the future (such as the subsequent faults predicted by the current current anomaly). The output is a hidden state sequence containing bidirectional temporal information. Then, a temporal attention module is introduced. The attention weight matrix is ​​obtained by training historical fault data. High weights (up to 0.8) are given to time points in the hidden state sequence that are strongly correlated with the fault (such as the features of the 2 hours before the hot spot appears), and low weights (down to 0.1) are given to normal fluctuation periods. The weighted hidden states are fused by max pooling and average pooling. The dimensions are compressed by 1×1 convolution (number 64), and finally, a fault-sensitive feature vector containing the temporal evolution law and fault-sensitive mode is obtained.

[0033] Optionally, step 23: Based on the spatiotemporal evolution inference network in the fault prediction model, perform multi-scale temporal evolution and probabilistic inference on the fault-sensitive feature vector to predict the fault precursors of the photovoltaic array, specifically including the following steps: Step 231: Based on the spatiotemporal evolution generation layer, perform multi-scale spatiotemporal evolution modeling on the fault-sensitive feature vector to obtain the fault evolution state sequence; Step 232: Based on the probabilistic reasoning enhancement layer, perform uncertainty quantification and evidence fusion processing on the fault evolution state sequence obtained in Step 231 to generate a multi-dimensional fault probability distribution; Step 233: Based on the decision optimization output layer, perform risk decision-making and early warning classification processing on the multi-dimensional fault probability distribution to predict the fault precursors of the photovoltaic array.

[0034] Preferably, step 231: Based on the spatiotemporal evolution generation layer, multi-scale spatiotemporal evolution modeling processing is performed on the fault-sensitive feature vector. This layer adopts a "spatiotemporal Transformer" structure, which includes two parallel modules: spatial evolution and temporal evolution. The spatial evolution module constructs a graph attention network (GAT) based on the photovoltaic array topology, mapping fault-sensitive feature vectors to graph node features. By calculating the attention weights between nodes (e.g., the influence weight of a hot spot component on adjacent components reaches 0.7), it captures the spatial propagation trend of faults and outputs spatial evolution features. The temporal evolution module sets up three time scale Transformer encoders (short-term 1-6 hours, medium-term 12-24 hours, and long-term 48-72 hours). Each encoder uses a different number of attention heads (short-term 8 heads, medium-term 16 heads, and long-term 32 heads) to capture the evolution law of faults in different time dimensions (e.g., rapid changes in arc faults and slow degradation of component aging). The spatial evolution features and the temporal evolution features of the three time scales are concatenated along the channel dimension and integrated into a feature sequence of a unified dimension through 1×1 convolution (number of 128), generating a fault evolution state sequence containing spatiotemporal dynamic changes.

[0035] Preferably, step 232: Based on the probabilistic reasoning enhancement layer, uncertainty quantification and evidence fusion processing are performed on the fault evolution state sequence. This layer consists of a Bayesian LSTM and an evidence fusion module. First, the fault evolution state sequence is input into the Bayesian LSTM network. By introducing a Gaussian prior distribution into the network weights, the probability distribution of the fault state at each time step and the corresponding uncertainty interval (e.g., "hot spot probability 68%±9% at t=24h") are output, realizing the quantification of random noise and measurement error during the evolution process. Then, the evidence fusion module is called to extract case evidence similar to the current state from the historical fault database (e.g., hot spot development cases under the same environmental conditions). The case evidence is fused with the probability distribution output by the Bayesian LSTM using DS evidence theory to correct the low-confidence probability values ​​(e.g., correcting the 45%±20% caused by isolated samples to 52%±8%). Finally, the fused probability distribution is split dimensionally according to the fault type (hot spot, arc, aging, etc.) to generate a multi-dimensional fault probability distribution containing three-dimensional information of spatial location, time node, and fault type.

[0036] Preferably, step 233: Based on the decision optimization output layer, risk decision-making and early warning classification processing are performed on the multi-dimensional fault probability distribution. This layer adopts a "risk matrix-reinforcement learning" hybrid structure, which includes two core modules: risk quantification and early warning decision-making. The risk quantification module takes a multi-dimensional fault probability distribution as input to construct a three-dimensional risk assessment matrix (the dimensions being fault occurrence probability, impact range, and severity, respectively). It calculates the comprehensive risk value corresponding to each probability distribution using weight coefficients trained from historical fault loss data (e.g., the severity weight for hot spot faults is 0.4), generating a risk value matrix. The early warning decision module introduces a reinforcement learning agent, taking the risk value matrix as state input and combining it with preset early warning grading rules (e.g., risk value < 0.3 corresponds to Level 1 warning, 0.3-0.6 corresponds to Levels 2-3, and > 0.6 corresponds to Levels 4-5). Through interactive learning with the environment (e.g., minimizing false alarm and false negative rates), it dynamically adjusts the grading thresholds and outputs a preliminary early warning level. The preliminary early warning level undergoes spatiotemporal consistency verification to ensure that changes in the early warning level of adjacent strings conform to the fault propagation law (e.g., the early warning level of the central string is higher than that of the edge strings). Finally, it generates a photovoltaic array fault precursor prediction result containing the specific fault location, early warning level, and confidence level.

[0037] Optionally, step 3, based on the constructed fault evolution knowledge graph, performs knowledge linking reasoning on the predicted fault precursors to generate hierarchical early warning information and intervene in advance on the fault risk of the photovoltaic array accordingly, specifically including the following steps: Step 31: Based on the fault evolution knowledge graph, perform semantic mapping and knowledge anchoring on the predicted fault precursor feature vectors to obtain graph-related fault instances. Step 32: Perform temporal evolution reasoning and propagation path deduction on the graph-related fault instances to obtain the fault development probability distribution; Step 33: Conduct multi-dimensional risk assessment and hierarchical decision processing on the probability distribution of fault development to generate hierarchical early warning information; Step 34: Optimize the intervention strategy and plan the execution path for the graded early warning information to generate early intervention for photovoltaic array failure risks.

[0038] Optionally, step 31, based on the fault evolution knowledge graph, performs semantic mapping and knowledge anchoring processing on the predicted fault precursor feature vectors to obtain graph-associated fault instances, specifically including the following steps: Step 311: For the predicted fault precursor feature vector, map it to the semantic space of the knowledge graph through the entity embedding model of the knowledge graph, calculate its semantic similarity with historical fault cases, and obtain multiple candidate fault instances with similarity greater than the set similarity threshold to form a candidate fault instance set. Step 312: For the set of candidate fault instances, extract the associated entities of each candidate instance based on the fault evolution knowledge graph, construct a local subgraph to generate a structured subgraph containing fault context information; Step 313: For the structured subgraph, apply the graph attention mechanism to weight the entities in each subgraph according to their importance, combine the current environmental parameters to perform evidence fusion, and output the graph association failure instance with the highest probability, which includes the confidence score of failure type, development stage and potential impact.

[0039] Preferably, step 311: Semantic mapping and preliminary screening are performed on the predicted fault precursor feature vectors. First, the predicted fault precursor feature vectors are input into the entity embedding model of the knowledge graph (this model is trained by fusing multi-dimensional data such as fault type, feature parameters, and environmental conditions), and the output is a precursor semantic embedding vector aligned with the semantic space of the knowledge graph. Based on the precursor semantic embedding vector, the semantic similarity between it and the entity embedding vectors of all historical fault cases in the knowledge graph is calculated (using the cosine similarity algorithm), and a semantic similarity scoring table containing the similarity value of each historical case is generated. Historical fault cases with scores higher than a preset threshold (such as 0.7) are selected from the semantic similarity scoring table, and these cases are integrated to form a candidate fault instance set.

[0040] Preferably, step 312: The candidate fault instance set is processed by extracting associated entities and constructing a subgraph. For each instance in the candidate fault instance set, based on the entity relationship network of the fault evolution knowledge graph (such as relationship types like "fault-component", "fault-environment", "fault-maintenance"), its associated entities (including environmental parameters at the time of the fault, the model of the component involved, historical maintenance records, etc.) are extracted to form a list of associated entities for each instance. Based on the hierarchical relationship between entities in the knowledge graph (such as the membership relationship of "string-module-component") and the association strength (such as co-occurrence frequency), entity relationship edges are constructed for the list of associated entities of each candidate instance, generating an initial local subgraph with the candidate instance as the core node. The initial local subgraph is structurally optimized, retaining entities directly related to fault evolution (such as removing irrelevant meteorological data), supplementing entity attribute information (such as component operating years, maintenance frequency), and generating a structured subgraph containing fault context information.

[0041] Preferably, step 313: The structured subgraph undergoes entity weighting and evidence fusion processing. A graph attention mechanism (GAT) is applied to score the importance of entity nodes in the structured subgraph. Weights are dynamically adjusted based on the entity's contribution to fault prediction (e.g., "temperature sensor data" has a weight of 0.6 for hot spot faults) and historical frequency of occurrence, generating a weighted structured subgraph. Current environmental parameters (e.g., real-time temperature, irradiance) are input into the weighted structured subgraph, and the attention mechanism strengthens the weights of entities matching the current operating conditions (e.g., the weight of the entity "poor heat dissipation" is increased by 30% in high-temperature environments), resulting in an environment-adapted weighted subgraph. Entity features in the environment-adapted weighted subgraph are pooled and fused to calculate the comprehensive matching probability of each candidate fault instance. The instance with the highest probability is selected as the output, and confidence scores (0-100%) for the corresponding fault type (e.g., "hot spot fault"), development stage (e.g., "initial formation stage"), and potential impact (e.g., "may spread to 3 clusters within 24 hours") are generated for each indicator. (The credibility of the graph) ultimately yields the graph association fault instances.

[0042] Optionally, step 32 involves performing temporal evolution reasoning and propagation path deduction on the graph-related fault instances to obtain the fault development probability distribution, specifically including the following steps: Step 321: For the fault instances associated with the graph, based on the fault evolution rules and combined with the current environmental parameters (temperature, irradiance), apply the probabilistic graphical model to deduce the possible evolution paths in the next 12 / 24 / 48 hours to obtain the fault development state sequence; Step 322: For the fault development state sequence, use a graph neural network to calculate the propagation speed and range of the fault in the array topology in each state, predict the number of affected strings and their severity, and generate a spatiotemporal fault propagation heatmap. Step 323: Apply Bayesian inference to the spatiotemporal propagation heatmap, fuse the prior distribution of historical fault data, calculate the probability of fault occurrence at each time point, and output the fault development probability distribution containing confidence intervals.

[0043] Preferably, step 321: Evolution path deduction processing is performed on the fault instances associated with the fault graph. First, the fault evolution rule base corresponding to the instance is extracted from the fault evolution knowledge graph (including causal rules such as "hot spot - temperature rise - power decay" and "connector aging - contact resistance increase - current fluctuation"). Each rule is accompanied by an environmental influence coefficient (e.g., the hot spot evolution rate coefficient is 1.5 under high temperature conditions). The current environmental parameters (real-time temperature, irradiance) are input into the rule base, and the applicable core evolution rules are selected through the rule matching algorithm (e.g., the "high temperature accelerates hot spot diffusion" rule is activated when the temperature is >35℃). Based on the selected rules, a dynamic Bayesian network is constructed as a probabilistic graphical model, and the state transition probabilities of three time nodes (12 / 24 / 48 hours) are set (e.g., the probability of hot spot spreading from a single component to a string within 12 hours is 30%). The fault development state sequence containing the fault state at each time node (e.g., "component-level hot spot" and "string-level hot spot") is deduced.

[0044] Preferably, step 322: Spatial propagation modeling is performed on the fault development state sequence. A graph neural network (GNN) is constructed based on the photovoltaic array topology. Each state in the fault development state sequence is used as the input feature of the GNN. Nodes represent photovoltaic modules, and edge weights are dynamically assigned according to the electrical connection strength between modules (such as conductor cross-sectional area and connection tightness). Through the message passing mechanism of the GNN, the probability of the fault propagating from the initial position to the surrounding modules in each state is calculated (e.g., the propagation probability of adjacent modules is 60%, and the propagation probability of a module separated by one is 20%), and the fault influence range matrix at each time node is obtained. This matrix is ​​mapped to the spatial coordinates of the photovoltaic array to generate a spatiotemporal fault propagation heatmap (color depth represents the probability of fault occurrence, and red areas represent high-risk areas). Each pixel in the heatmap is accompanied by an impact score (0-10 points, with higher scores indicating more severe impact).

[0045] Preferably, step 323: performing probability distribution calculation processing on the fault propagation heatmap. First, historical propagation data of similar faults are extracted from the historical fault database. A prior distribution model of the fault occurrence probability is generated through kernel density estimation (e.g., the mean probability distribution of hot spot faults within 24 hours is 55%). The spatiotemporal probability values ​​in the fault propagation heatmap are used as the likelihood function input into the Bayesian inference model. The prior distribution and the likelihood function are fused to calculate the posterior probability (e.g., the 24-hour hot spot probability of a certain area = prior 55% × likelihood 70% / normalization constant). Confidence interval estimation is performed on the posterior probability of each time node (12 / 24 / 48 hours) (a 95% confidence interval is generated using the bootstrap method). Finally, the fault development probability distribution containing time nodes, spatial locations, fault probabilities, and confidence intervals is output.

[0046] Optionally, step 33 involves conducting a multi-dimensional risk assessment and hierarchical decision-making process on the probability distribution of fault development to generate hierarchical early warning information, specifically including the following steps: Step 331: For the probability distribution of fault development, integrate the evaluation indicators of three dimensions: power generation loss model (based on power curve and historical data), repair cost model (spare parts + labor), and safety risk assessment (such as fire probability), and calculate the multidimensional risk score vector at each time point. Step 332: Apply hierarchical analysis to the multidimensional risk score vector to determine the weights of each dimension (e.g., economic loss weight 0.6, safety risk weight 0.3), and generate a comprehensive risk index time series by weighting and aggregating the multidimensional risk score vector according to the weights of each dimension; Step 333: For the comprehensive risk index time series, compare it with the preset risk threshold (e.g., Level 1: index < 0.2, Level 5: index > 0.8), determine the warning level for each time point, and generate a preliminary graded warning sequence accordingly; Step 334: Perform dynamic planning on the preliminary graded early warning sequence to optimize the early warning triggering timing, so as to balance the benefits of early intervention and the cost of false alarms, and output the final graded early warning information including fault location, development trend, risk level, and recommended intervention time.

[0047] Preferably, step 331: The multidimensional risk index calculation is performed on the fault development probability distribution. First, the power generation loss model is called to correlate the fault development probability distribution with the power curve of the photovoltaic array (fitted based on historical operating data). The power generation loss that may occur due to the fault at each time point is calculated (e.g., a 24-hour power loss of 200 kWh corresponds to a fault probability of 60% in a certain area), thus obtaining the power generation loss index. Next, the repair cost model is called to match the corresponding spare parts price (e.g., hot spot module replacement cost of 500 yuan / piece) and labor cost (e.g., 200 yuan / hour) according to the fault type (e.g., hot spot module replacement cost of 500 yuan / piece) and labor cost (e.g., 200 yuan / hour). The total repair cost is calculated based on the fault impact range, thus obtaining the cost loss index. Finally, the safety risk assessment model is called to calculate the safety risk index (0-10 points, with higher scores indicating higher risk) based on the fault type and historical accident data (e.g., the probability of a fire caused by a hot spot is 0.5%), combined with current environmental parameters (e.g., the fire risk coefficient increases by 1.2 times when the wind speed is >5). The power generation loss index, cost loss index, and safety risk index are integrated according to time points to generate a multidimensional risk score vector containing three dimensions.

[0048] Preferably, step 332: weight allocation and comprehensive calculation are performed on the multidimensional risk score vector. A judgment matrix is ​​constructed using the Analytic Hierarchy Process (AHP). The relative importance of each dimension is determined by expert scoring (e.g., economic loss weight 0.6, safety risk weight 0.3, other factors weight 0.1). After consistency testing (CR value < 0.1), the final dimension weights are obtained. The scores of each dimension in the multidimensional risk score vector are weighted according to their corresponding weights (e.g., score at a certain time node = power generation loss × 0.6 + cost loss × 0.1 + safety risk × 0.3) to obtain the comprehensive risk index at that time node. The comprehensive risk index time series containing time series is generated by calculating sequentially according to 12 / 24 / 48 hour time nodes. Each index value is accompanied by the contribution ratio of each dimension (e.g., power generation loss accounts for 70% in the index at a certain node).

[0049] Preferably, step 333 involves classifying the comprehensive risk index time series into warning levels, pre-setting five risk threshold levels (Level 1: index < 0.2, Level 2: 0.2-0.4, Level 3: 0.4-0.6, Level 4: 0.6-0.8, Level 5: > 0.8), with each threshold range corresponding to a clear risk description (e.g., Level 3 corresponds to "moderate risk, requiring planned maintenance"); comparing each index value in the comprehensive risk index time series with the threshold to determine the corresponding warning level (e.g., an index of 0.52 corresponds to Level 3), generating a preliminary graded warning sequence; smoothing the level changes between adjacent time nodes in the sequence (e.g., supplementing a 20-hour transitional warning when the level is Level 3 at 12 hours and Level 4 at 24 hours), ensuring that the level changes conform to the fault evolution law, and obtaining a preliminary graded warning sequence containing time nodes and risk levels.

[0050] Preferably, step 334: Optimize the triggering timing and integrate information for the preliminary graded early warning sequence, construct a cost-benefit model, calculate the comprehensive benefits of different early warning triggering times (early intervention benefit = avoided loss - intervention cost, false alarm cost = unnecessary maintenance expenditure), and use a dynamic programming algorithm to select the early warning triggering time point that maximizes benefits (e.g., triggering at 24 hours can save 5000 yuan more than triggering at 12 hours); based on the optimized triggering timing, integrate spatial location information in the fault development probability distribution (e.g., "Eastern District 3-group string"), development trends in the comprehensive risk index time series (e.g., "risk continues to rise in 24-48 hours"), risk levels in the preliminary graded early warning sequence (e.g., level 4), and suggested intervention times (e.g., "within 20 hours") to generate final graded early warning information containing multi-dimensional information, with each field in the information accompanied by a confidence score (e.g., location information confidence score 90%).

[0051] Optionally, step 34 involves optimizing the intervention strategy and planning the execution path for the graded early warning information to generate early intervention for photovoltaic array failure risks. This specifically includes the following steps: Step 341: For graded early warning information, call the handling plan library in the fault evolution knowledge graph, match basic intervention strategies according to the early warning level (e.g., remote monitoring for level 1-2, preventive maintenance for level 3-4, and emergency shutdown for level 5), and generate a set of strategy candidates; Step 342: For the strategy candidate set, combine real-time meteorological data (such as the probability of precipitation in the next 24 hours) and the power plant operation and maintenance schedule to calculate the comprehensive benefits of different strategies (minimize power generation loss and maintenance costs), and select the optimal intervention strategy; Step 343: For the optimal intervention strategy, associate it with the operation and maintenance knowledge base in the knowledge graph, automatically generate an execution guide containing a spare parts list, personnel qualification requirements, and operation steps, and synchronize it to the power plant management system to form a directly executable early intervention plan for fault risks.

[0052] Preferably, step 341: Perform basic strategy matching processing on the graded early warning information. First, analyze the core elements (risk level, fault type, and impact range) in the final graded early warning information, such as "Level 4 risk, hot spot fault, affecting 3 groups in the eastern area". Based on these elements, call the disposal plan library of the fault evolution knowledge graph. This library stores standardized strategies in a two-dimensional matrix of "early warning level - fault type" (e.g., matching "hourly data sampling + remote monitoring" for Level 1-2 risks, matching "component cooling + partial power outage inspection" for Level 3-4 risks, and matching "emergency shutdown + fire plan activation" for Level 5 risks). Perform preliminary screening on the matched basic strategies, eliminating plans that conflict with the current environment (e.g., excluding the "water cooling" strategy during heavy rain), and generate a strategy candidate set containing 3-5 optional plans. Each plan is accompanied by an explanation of applicable conditions (e.g., "component cooling requires ambient temperature < 30℃").

[0053] Preferably, step 342: Perform comprehensive benefit evaluation and optimal screening on the strategy candidate set, construct a dynamic benefit evaluation model, input real-time meteorological data (e.g., 80% probability of precipitation in the next 24 hours) and operation and maintenance schedule (e.g., available maintenance teams the next morning), calculate the execution feasibility score of each strategy (0-10 points, with outdoor operation score decreasing when the probability of precipitation is high); combine the cost parameters of each scheme in the strategy candidate set (e.g., 200kWh power generation loss and 500 yuan labor cost for "partial power outage inspection") and benefit parameters (e.g., 3000 yuan potential benefit from avoiding fault propagation), calculate the comprehensive benefit value (benefit = potential benefit - execution cost - feasibility penalty); sort the comprehensive benefit values ​​using a multi-objective optimization algorithm (balancing benefit maximization and execution complexity minimum), and select the scheme with the highest score as the optimal intervention strategy, such as "component replacement after the rain stops the next day (10:00-12:00)".

[0054] Preferably, step 343 involves generating and synchronizing the execution details of the optimal intervention strategy with the system. Based on the optimal intervention strategy, the operation and maintenance knowledge base of the knowledge graph is invoked to extract the associated standardized operation modules (such as "hot spot component replacement," which includes a safe power-off procedure, a component model matching table, and a tool list). Combining the spatial coordinates of the fault location (the 5th component in string 3 of the East Zone) and the topology (the direction of the series line), personalized execution steps are generated (such as "first disconnect the main switch of the string, then remove the 4 fixing screws of the faulty component"). Personnel qualification requirements (such as "must hold a high-voltage electrician's certificate") and safety precautions (such as "wear insulated gloves") are added. The generated execution guidelines are converted into structured data (such as JSON format) that can be recognized by the power plant management system and synchronized to the operation and maintenance scheduling module, spare parts management module, and personnel terminals to form a closed-loop intervention plan that includes "task allocation - spare parts preparation - operation steps - acceptance criteria," ensuring that on-site personnel can directly execute according to the guidelines.

[0055] like Figure 2 As shown in the illustration, this application also provides a photovoltaic power generation intelligent early warning system, which includes: A time-series data processing unit is used to determine the multivariate time-series data of the photovoltaic array, wherein the multivariate time-series data includes electrical characteristic data and environmental characteristic data of the photovoltaic array string; The prediction unit is used to input multivariate time series data into the fault prediction model, extract the spatial correlation feature vector between photovoltaic array strings and perform time-dependent encoding on the spatial correlation feature vector to obtain the fault-sensitive feature vector, so as to predict the fault precursors of the photovoltaic array based on the fault-sensitive feature vector. The intervention unit is used to perform knowledge linking reasoning on predicted fault precursors based on the constructed fault evolution knowledge graph, so as to generate hierarchical early warning information and intervene in the fault risk of photovoltaic array in advance.

[0056] The technical processing of each unit in the above system can be explained exemplarily by referring to the above. Figure 1 .

[0057] The above description is merely a preferred embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for intelligent early warning of photovoltaic power generation, characterized in that, include: Step 1: Determine the multivariate time-series data of the photovoltaic array, which includes the electrical characteristic data and environmental characteristic data of the photovoltaic array string; Step 2: Input multivariate time series data into the fault prediction model, extract the spatial correlation feature vector between photovoltaic array strings and perform time-dependent encoding on the spatial correlation feature vector to obtain the fault-sensitive feature vector, so as to predict the fault precursors of the photovoltaic array based on the fault-sensitive feature vector; Step 3: Based on the constructed fault evolution knowledge graph, perform knowledge linking reasoning on the predicted fault precursors to generate hierarchical early warning information and intervene in advance on the fault risk of photovoltaic arrays accordingly.

2. The method according to claim 1, characterized in that, Step 1: Determine the multivariate time-series data of the photovoltaic array. The multivariate time-series data includes the electrical characteristic data and environmental characteristic data of the photovoltaic array strings. Specifically, it includes the following steps: Step 11: Obtain the spatiotemporally labeled data stream obtained by spatiotemporally co-sampling of the photovoltaic array using a multi-scale sensor network; Step 12: Perform multimodal feature extraction on the spatiotemporal labeled data stream to obtain multimodal spatiotemporal correlation features; Step 13: Perform spatiotemporal alignment and synchronization calibration on the multimodal spatiotemporal correlation features to generate multivariate time series data.

3. The method according to claim 2, characterized in that, Step 11: Determine the spatiotemporally labeled data stream obtained by spatiotemporally co-sampling of the photovoltaic array using a multi-scale sensor network, specifically including the following steps: Step 111: Obtain the multidimensional heterogeneous data stream obtained by spatiotemporal collaborative sampling of the photovoltaic array using a multi-scale sensor network; Step 112: Perform spatiotemporal anchoring and feature enhancement processing on the multidimensional heterogeneous data stream to generate a spatiotemporally correlated feature stream; Step 113: Perform dynamic calibration and semantic annotation on the spatiotemporal correlation feature stream to obtain the spatiotemporal labeled data stream.

4. The method according to claim 2, characterized in that, Step 12: Perform multimodal feature extraction on the spatiotemporally labeled data stream to obtain multimodal spatiotemporal correlation features, specifically including the following steps: Step 121: Perform adaptive multi-resolution decomposition on the spatiotemporal labeled data stream to obtain time-frequency semantic logical units; Step 122: Perform topology-aware spatial correlation modeling on the time-frequency semantic logic unit to generate a spatiotemporal correlation tensor; Step 123: Perform multimodal attention fusion processing on the spatiotemporal correlation tensor to obtain multimodal spatiotemporal correlation features.

5. The method according to claim 2, characterized in that, Step 13: Perform spatiotemporal alignment and synchronization calibration on the multimodal spatiotemporal correlation features to generate multivariate time series data, specifically including the following steps: Step 131: Perform spatiotemporal alignment processing on the multimodal spatiotemporal correlation features with topological constraints to obtain the spatial calibration feature tensor; Step 132: Perform multimodal synchronization calibration on the spatial calibration feature tensor to generate a spatiotemporal synchronization feature matrix; Step 133: Perform semantic enhancement and standardization on the spatiotemporal synchronization feature matrix to generate multivariate time series data.

6. The method according to claim 1, characterized in that, Step 2: Input multivariate time-series data into the fault prediction model, extract spatial correlation feature vectors between photovoltaic array strings, and perform time-dependent encoding on the spatial correlation feature vectors to obtain fault-sensitive feature vectors. These fault-sensitive feature vectors are then used to predict the precursors of photovoltaic array faults. Specifically, this includes the following steps: Step 21: Based on the spatial feature extraction network in the fault prediction model, perform feature extraction on the multivariate time series data to generate spatial correlation feature vectors between photovoltaic array strings; Step 22: Based on the temporal dependency coding network in the fault prediction model, perform temporal dependency coding on multiple pairs of spatially correlated feature vectors to obtain fault-sensitive feature vectors; Step 23: Based on the spatiotemporal evolution inference network in the fault prediction model, perform multi-scale temporal evolution and probabilistic inference on the fault-sensitive feature vector to predict the fault precursors of the photovoltaic array.

7. The method according to claim 6, characterized in that, Step 21: Based on the spatial feature extraction network in the fault prediction model, feature extraction is performed on the multivariate time series data to generate spatial correlation feature vectors between photovoltaic array strings. This includes the following steps: Step 211: Based on the topology-aware embedding layer, perform physical structure encoding and modality fusion processing on the multivariate time series data to obtain the array topology feature map; Step 212: Based on the spatial dependency propagation layer, perform fault propagation mode mining on the array topology feature map to generate a spatial correlation feature tensor; Step 213: Based on the fault-sensitive enhancement layer, the spatial correlation feature tensor is processed to highlight fault features and suppress noise, so as to obtain the spatial correlation feature vector between photovoltaic array strings.

8. The method according to claim 6, characterized in that, Step 22: Based on the temporal dependency coding network in the fault prediction model, perform temporal dependency coding on multiple pairs of spatially correlated feature vectors to obtain fault-sensitive feature vectors. This includes the following steps: Step 221: Based on the multi-scale temporal feature extraction layer, perform temporal window decomposition and parallel feature extraction on multiple pairs of spatially correlated feature vectors to obtain a multi-scale temporal feature set; Step 222: Based on the attention-enhanced temporal encoder, perform temporal dependency enhancement and fault feature highlighting on the multi-scale temporal feature set to obtain the fault-sensitive feature vector.

9. The method according to claim 6, characterized in that, Step 23: Based on the spatiotemporal evolution inference network in the fault prediction model, perform multi-scale temporal evolution and probabilistic inference on the fault-sensitive feature vector to predict the fault precursors of the photovoltaic array. This includes the following steps: Step 231: Based on the spatiotemporal evolution generation layer, perform multi-scale spatiotemporal evolution modeling on the fault-sensitive feature vector to obtain the fault evolution state sequence; Step 232: Based on the probabilistic reasoning enhancement layer, perform uncertainty quantification and evidence fusion processing on the fault evolution state sequence obtained in Step 231 to generate a multi-dimensional fault probability distribution; Step 233: Based on the decision optimization output layer, perform risk decision-making and early warning classification processing on the multi-dimensional fault probability distribution to predict the fault precursors of the photovoltaic array.

10. A photovoltaic power generation intelligent early warning system, characterized in that, include: A time-series data processing unit is used to determine the multivariate time-series data of the photovoltaic array, wherein the multivariate time-series data includes electrical characteristic data and environmental characteristic data of the photovoltaic array string; The prediction unit is used to input multivariate time series data into the fault prediction model, extract the spatial correlation feature vector between photovoltaic array strings and perform time-dependent encoding on the spatial correlation feature vector to obtain the fault-sensitive feature vector, so as to predict the fault precursors of the photovoltaic array based on the fault-sensitive feature vector. The intervention unit is used to perform knowledge linking reasoning on predicted fault precursors based on the constructed fault evolution knowledge graph, so as to generate hierarchical early warning information and intervene in the fault risk of photovoltaic array in advance.

Images