Fire risk dynamic assessment method and system based on big data collection for photovoltaic plant

Abstract

The application discloses a photovoltaic factory fire risk dynamic assessment method and system based on big data collection, and relates to the technical field of big data collection. The method comprises the following steps: deploying a multi-source sensing unit in a photovoltaic factory for collection, and establishing a multi-source heterogeneous data set; activating a risk scene self-calibration mechanism to process data, and generating a spatiotemporal alignment risk factor tensor; extracting a four-dimensional risk primitive based on the tensor, inputting the multi-layer tensor coupling model to calculate the interaction intensity, and outputting a risk superposition matrix; constructing a cross-time slice risk conduction graph, calculating a directed transmission entropy and a path, and establishing a risk trigger chain; and evaluating the fire risk according to the matrix and the trigger chain. The technical problems that the traditional static assessment method cannot effectively fuse multi-source heterogeneous data, resulting in distortion of the description of risk evolution, and ignoring the dynamic coupling effect and conduction path between risk factors, resulting in early warning delay, are solved, and the technical effects of improving the accuracy, real-time performance and early warning capability of the fire risk dynamic assessment are achieved.

Description

Technical Field

[0001] This invention relates to the field of big data acquisition technology, specifically to a method and system for dynamic assessment of fire risk in photovoltaic plant areas based on big data acquisition. Background Technology

[0002] With the rapid development of photovoltaic power generation technology and the continuous expansion of its application scale, the safe operation and maintenance of large photovoltaic power plants are facing increasingly severe challenges. In particular, due to the long-term exposure of photovoltaic modules to a complex and ever-changing external environment, coupled with the continuous high-load operation of electrical equipment, fire risk has become a significant hidden danger threatening the safety of power plant assets and the stable supply of power. Traditional safety monitoring methods mainly rely on periodic manual inspections and isolated sensor alarms. This approach is insufficient for real-time perception and in-depth fusion analysis of massive, multi-dimensional risk factors, and it cannot accurately depict the complex dynamic coupling relationships between different risk dimensions and their transmission patterns over time. Existing technologies are often limited to static judgments based on single parameter thresholds, lacking unified processing and joint mining of multi-source heterogeneous data such as thermal, electrical, structural, and environmental data. This results in delayed risk assessment, a high false alarm rate, and difficulty in locating potential risk sources and development paths, failing to provide effective decision support for fire prevention and emergency management. Summary of the Invention

[0003] This application provides a method and system for dynamic fire risk assessment in photovoltaic plant areas based on big data collection. It solves the technical problems of traditional static assessment methods, which lead to distortion in the characterization of risk evolution due to the failure to effectively integrate multi-source heterogeneous data, and the delay in early warning due to ignoring the dynamic coupling effect and transmission path between risk factors. It achieves the technical effect of accurately calculating the intensity of risk interaction through multi-source data fusion and tensor coupling analysis, and revealing the risk triggering chain by combining dynamic causal reasoning, thereby improving the accuracy, real-time performance and early warning capability of dynamic fire risk assessment.

[0004] The first aspect of this application provides a method for dynamic assessment of fire risk in photovoltaic power plant areas based on big data collection, the method comprising: Multi-source sensing units are deployed at multiple distributed nodes within the photovoltaic plant area to perform data acquisition and establish a multi-source heterogeneous dataset. These multi-source sensing units include electrical sensing units, infrared thermal imaging units, environmental monitoring units, and video acquisition units. A risk scenario self-calibration mechanism mapped to the distributed nodes is activated, and data processing of the multi-source heterogeneous dataset is performed to establish a spatiotemporally aligned risk factor tensor. Based on the risk factor tensor, four-dimensional risk primitives representing thermal, electrical, structural, and environmental factors are extracted. These four-dimensional risk primitives are input into a multi-layer tensor coupling analysis model to calculate the interaction strength between each risk dimension and output a risk superposition matrix. A risk transmission graph structure spanning multiple time slices is constructed based on the risk superposition matrix. A dynamic causal inference channel is used to calculate the directed transmission entropy and influence path between nodes within the risk transmission graph structure, constructing a risk triggering chain. A dynamic fire risk assessment of the photovoltaic plant area is performed based on the risk superposition matrix and the risk triggering chain.

[0005] A second aspect of this application provides a dynamic fire risk assessment system for photovoltaic plant areas based on big data collection, the system comprising: Data Acquisition Module: Deploys multi-source sensing units at multiple distributed nodes within the photovoltaic plant area to perform data acquisition and establish a multi-source heterogeneous dataset. The multi-source sensing units include electrical sensing units, infrared thermal imaging units, environmental monitoring units, and video acquisition units. Data Processing Module: Activates a risk scenario self-calibration mechanism mapped to the distributed nodes, performs data processing on the multi-source heterogeneous dataset, and establishes a spatiotemporally aligned risk factor tensor. Coupling Analysis Module: Extracts four-dimensional risk primitives (thermal, electrical, structural, and environmental) based on the risk factor tensor, inputs these primitives into a multi-layer tensor coupling analysis model, calculates the interaction strength between each risk dimension, and outputs a risk superposition matrix. Risk Transmission Analysis Module: Constructs a cross-time-slice risk transmission graph structure based on the risk superposition matrix, and uses a dynamic causal inference channel to calculate the directed transmission entropy and influence path between nodes within the risk transmission graph structure, constructing a risk triggering chain. Risk Assessment Module: Performs a dynamic fire risk assessment of the photovoltaic plant area based on the risk superposition matrix and the risk triggering chain.

[0006] One or more technical solutions provided in this application have at least the following technical effects or advantages: First, multiple distributed nodes are installed within the photovoltaic plant area, utilizing various sensors to collect different types of data in real time, establishing a multi-source heterogeneous dataset. Then, by activating a risk scenario self-calibration mechanism mapped to the nodes, this data is processed to generate a spatiotemporally aligned risk factor tensor for further analysis. Next, based on this risk factor tensor, risk primitives from four main risk dimensions are extracted and input into a multi-layer tensor coupling analysis model to calculate the interaction strength between dimensions, resulting in a risk superposition matrix. Then, this matrix is ​​used to construct a risk transmission graph across time slices, combining dynamic causal reasoning techniques to calculate the directed transmission entropy and influence paths between different nodes, thereby identifying potential risk triggering chains. Finally, by combining the risk superposition matrix and risk triggering chains, the fire risk of the photovoltaic plant area is dynamically assessed, aiding in real-time monitoring and improving risk prevention and control capabilities. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 A schematic diagram of the process for a dynamic assessment method for fire risk in photovoltaic plant areas based on big data collection, provided in an embodiment of this application.

[0009] Figure 2 A schematic diagram of the structure of a dynamic fire risk assessment system for photovoltaic plant areas based on big data collection, provided in an embodiment of this application.

[0010] Figure labeling: Data acquisition module 11, data processing module 12, coupling analysis module 13, risk transmission analysis module 14, risk assessment module 15. Detailed Implementation

[0011] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0012] Example 1, as Figure 1 As shown, this application provides a dynamic assessment method for fire risk in photovoltaic plant areas based on big data collection. The method includes: Multi-source sensing units are deployed at multiple distributed nodes within the photovoltaic plant area to perform data acquisition and establish a multi-source heterogeneous dataset. The multi-source sensing units include electrical sensing units, infrared thermal imaging units, environmental monitoring units, and video acquisition units.

[0013] In this embodiment, multiple distributed nodes within the photovoltaic plant are equipped with multi-source sensing units for comprehensive data acquisition. These distributed nodes typically include different areas of the photovoltaic power station, such as the solar panel array, electrical control center, and inverter area. The multi-source sensing units typically include electrical sensing units, infrared thermal imaging units, environmental monitoring units, and video acquisition units. The electrical sensing units are primarily responsible for monitoring electrical parameters such as current, voltage, and power within the photovoltaic plant, capturing abnormal conditions in the operation of electrical equipment in real time, such as current overload and voltage fluctuations. This electrical data helps detect potential faults or safety hazards in electrical equipment, especially providing timely warnings when short circuits, overheating, or fire risks occur. The infrared thermal imaging units monitor the temperature distribution of equipment and the environment using infrared detection technology, detecting areas with abnormally high temperatures, particularly in electrical equipment, terminals, and power generation panels. Heat accumulation in these areas may be a precursor to fire; the infrared thermal imaging units can quickly detect and report abnormal temperature information, playing a crucial role in early detection of fire hazards. The environmental monitoring unit includes sensors for environmental parameters such as temperature, humidity, and wind speed. Environmental factors, such as excessively high temperatures, high humidity, or strong winds, can all trigger fires. Real-time monitoring of these environmental variables helps assess the potential impact of the external environment on fire risk, thus providing early warnings. The video acquisition unit monitors the operating status of equipment within the plant area through real-time video surveillance, focusing on key areas with high population density. Video surveillance provides intuitive visual information, which, combined with data from other sensors, further verifies and confirms the occurrence of risk events, improving the accuracy and comprehensiveness of data analysis. These multi-source sensing units collect their own data and transmit it to a data storage and processing platform via wireless or wired networks for subsequent comprehensive analysis. This data collection forms a multi-source heterogeneous dataset that reflects multi-dimensional information about electrical equipment, environmental factors, and equipment status within the photovoltaic plant area, providing fundamental data support for subsequent dynamic fire risk assessment.

[0014] A risk scenario self-calibration mechanism is activated and mapped to distributed nodes, data processing of multi-source heterogeneous datasets is performed, and a risk factor tensor with spatiotemporal alignment is established.

[0015] In one embodiment, each distributed node corresponds to one or more potential risk scenarios. These risk scenarios are jointly determined by the specific equipment and environmental conditions monitored by each node. For example, in the electrical equipment area, there may be risks of electrical faults such as short circuits and overloads; in the inverter area, there may be risks of equipment overheating or fan failure; and in the environmental monitoring node, there may be risks of extreme weather or fire spread. The monitoring data of each distributed node must be associated with a specific risk scenario in order to dynamically adjust the risk assessment criteria according to the actual situation. To ensure that the data of each node is highly matched with its risk scenario, the system activates a risk scenario self-calibration mechanism that maps to the distributed nodes. This risk scenario self-calibration mechanism can bind the data collected by each sensing unit in each distributed node to a specific location identifier. For example, a voltage data point may come from a transformer in the photovoltaic plant area, or a temperature data point may come from the heat dissipation area of ​​the inverter. By assigning a location label to each data point, the risk scenario self-calibration mechanism ensures that the system can accurately identify the source of the data and specify the area and monitoring equipment in which each data item is located, as well as the normal operating range of the equipment in that area, such as the normal temperature range of the inverter heat dissipation area. In this way, the relationship between data and specific equipment and areas is clearly mapped. Subsequently, for the collected multi-source heterogeneous dataset, based on the location information of each data point in the dataset, the data is compared with the normal operating range of the corresponding device at that location, such as the normal value range of temperature, voltage, current, image color, and image brightness, to calculate the risk factor for each sensor unit's corresponding acquisition location. For example, when the voltage data deviates from the normal range, the difference between the real-time voltage and the maximum or minimum value of the normal voltage range is calculated, and then divided by the maximum or minimum value to obtain the risk factor. The maximum or minimum value depends on which side the voltage data deviates from. Next, the risk factors of each sensor unit in each distributed node are concatenated in chronological order to form multiple risk factor tensors. Each risk factor tensor represents the risk status of the monitoring location corresponding to a specific sensor unit at a distributed node. These tensors have the same time dimension, reflecting the risk changes of the device or area at a specific point in time. Then, these risk factor tensors are aligned not only in the time dimension but also in the spatial location; that is, the risk factor tensors corresponding to different sensor units are synchronized to the same spatial framework, thereby ensuring that all risk factor tensors of each node can correctly correspond to its area and device. These spatiotemporally aligned risk factor tensors can provide a comprehensive data foundation for subsequent risk assessments, enabling the system to consider the dynamic changes and interrelationships of different risk factors at various time periods and spatial nodes when analyzing fire risks in photovoltaic plant areas. This lays an important foundation for the accuracy and reliability of the entire dynamic fire risk assessment system.

[0016] Based on the risk factor tensor, a four-dimensional risk primitive of thermal-electrical-structural-environment is extracted. The four-dimensional risk primitive is then input into a multi-layer tensor coupling analysis model to calculate the interaction strength between each risk dimension and output a risk superposition matrix.

[0017] In one embodiment, after generating multiple risk factor tensors through a risk scenario self-calibration mechanism and spatiotemporal alignment, the system performs multi-dimensional deconstruction on each risk factor tensor. That is, risk factors are extracted from each risk factor tensor according to four dimensions: thermal, electrical, structural, and environmental. The extracted risk factors are then combined with the corresponding monitoring data to form risk primitives for each dimension, thereby constructing a four-dimensional risk primitive of thermal-electrical-structural-environmental. The thermal dimension reflects the temperature change of the equipment or area; the electrical dimension reflects the fluctuation of electrical parameters such as current and voltage of electrical equipment; the structural dimension reflects the mechanical state of the equipment or structure, such as vibration, displacement, and strain data; and the environmental dimension reflects the state of the environment surrounding the photovoltaic plant, such as wind speed, humidity, and air quality data. Subsequently, the extracted four-dimensional risk primitives are input into a multi-layer tensor coupling analysis model. This multi-layer tensor coupling analysis model calculates the interaction of the four risk dimensions through a local mutual interference layer, a modal fusion layer, a cross-scale collaborative layer, and a dynamic feedback layer, and outputs the final risk superposition matrix. This risk superposition matrix reflects the interaction between different risk dimensions and their superposition effect, providing a reliable basis for subsequent fire risk assessment and decision-making.

[0018] Furthermore, inputting the four-dimensional risk primitive into a multilayer tensor coupling analysis model includes: The multi-layer tensor coupling analysis model includes a local mutual disturbance layer, a modal fusion layer, a cross-scale coordination layer, and a dynamic feedback layer. After the four-dimensional risk primitive is input into the multi-layer tensor coupling analysis model: the local mutual disturbance layer is used to extract the transient rate of change of each risk dimension in the four-dimensional risk primitive, calculate the coupling degree of thermal gradient disturbance, current pulse, structural strain mutation, and environmental wind pressure, and generate a local coupling weight matrix; after the local coupling weight matrix is ​​sent to the modal fusion layer, the nonlinear cross-influence between the four modes of thermal-electrical-structural-environment is weighted and normalized based on the tensor fusion operator of mutual information gain to form a modal interaction tensor; after the modal interaction tensor is synchronized to the cross-scale coordination layer, dual decoupling and reconstruction of the time scale and spatial scale are performed to extract the slow trend coefficient and fast trigger coefficient, and a spatiotemporal coordination tensor is formed through the scale coordination operator. The spatiotemporal coordination tensor is used to describe the coupling evolution path of risk modes at different scales; the dynamic feedback layer is used to perform temporal recursion and coupling residual analysis of the spatiotemporal coordination tensor, perform dynamic iterative correction, and output a risk superposition matrix.

[0019] Preferably, the multi-layer tensor coupling analysis model for four-dimensional risk primitive input includes a local mutual interference layer, a modal fusion layer, a cross-scale coordination layer, and a dynamic feedback layer. The primary objective of the local mutual interference layer is to extract the transient rate of change of each risk dimension in the four-dimensional risk primitive and calculate the instantaneous coupling strength between dimensions. The core task of the modal fusion layer is to perform weighted fusion of the coupling weight matrices from the local mutual interference layer, handling the nonlinear cross-influence between different risk dimensions. The objective of the cross-scale coordination layer is to perform dual decoupling and reconstruction of the modal interaction tensor on both temporal and spatial scales, thereby extracting risk patterns at different scales. The main task of the dynamic feedback layer is to utilize the spatiotemporal coordination tensor for temporal recursion and coupled residual analysis, performing dynamic iterative correction to optimize the risk assessment results. After receiving the four-dimensional risk primitives, the multi-layer tensor coupling analysis model first analyzes each risk dimension of the four-dimensional risk primitives through a local inter-perturbation layer, extracting the transient rate of change of each risk dimension, i.e., the rate of change of each risk dimension in a short period of time. For example, for the thermal dimension, the thermal gradient perturbation rate is calculated by the ratio of the temperature difference between adjacent time points to the current temperature; for the electrical dimension, the current pulse intensity is quantified by the ratio of the current mutation amplitude to the reference current; for the structural dimension, the structural strain mutation rate is quantified by the ratio of the strain gradient change to the historical mean; and for the environmental dimension, the environmental wind pressure is calculated by the ratio of the wind speed mutation to the wind pressure threshold. Subsequently, based on the local inter-perturbation layer, the local coupling weights are calculated based on the transient rates of change of each risk dimension by calling the internal Pearson correlation coefficient and Granger causality test methods. These local coupling weights are then stored in the form of a matrix to construct the local coupling weight matrix. Next, the obtained local coupling weight matrix is ​​passed to the modal fusion layer. The modal fusion layer uses a tensor fusion operator based on mutual information gain to perform weighted normalization fusion of the nonlinear cross-influences between the four risk dimensions (thermal, electrical, structural, and environmental), forming a modal interaction tensor. This modal interaction tensor integrates the interaction information of the four risk modes, comprehensively reflecting the mutual influence between these risk factors. Mutual information gain is a method used to measure the correlation between different data. In this way, the modal fusion layer can identify and quantify the complex relationships and cross-influences between the four risk dimensions (thermal, electrical, structural, and environmental). Then, the obtained modal interaction tensor is synchronized to the cross-scale collaboration layer, which performs dual decoupling and reconstruction of the modal interaction tensor at both the temporal and spatial scales. By using multi-resolution analysis, the slow trend coefficient and fast trigger coefficient within different time windows are extracted. Then, a spatiotemporal co-operation tensor is formed through scale co-operation operators. This spatiotemporal co-operation tensor describes the coupled evolution path of each risk mode under different time and spatial scales, providing spatiotemporally consistent information for subsequent dynamic risk analysis, enabling the system to simultaneously consider the risk evolution of long-term trends and short-term emergencies.Finally, the dynamic feedback layer performs time series recursive analysis and coupled residual analysis on the spatiotemporal collaborative tensor, and then performs weighted correction on the spatiotemporal collaborative tensor based on the analysis results. This process is carried out iteratively to gradually optimize the prediction results of risk factors until a corrected risk superposition matrix is ​​obtained. This risk superposition matrix integrates the interaction between various dimensions and their effects over time, and can comprehensively and accurately assess the fire risk in the photovoltaic plant area, identify potential risk triggering factors, and dynamically monitor and warn of them.

[0020] Furthermore, after sending the local coupling weight matrix to the modal fusion layer, the process includes: Modal correlation encoding is performed on the local coupling weight matrix. Based on the encoding results, the mutual information entropy differences between thermal-electric, thermal-structure, thermal-environment, electrical-structure, electrical-environment, and structure-environment are calculated to establish a modal dependency graph. Based on the modal dependency graph, a tensor fusion operator driven by mutual information gain is used to perform modal weighting calculation, wherein the fusion weights are adaptively adjusted based on the modal dependency degree. During the fusion process, the coupling weights are dynamically reconstructed based on the risk signal strength, response delay, and phase shift through a nonlinear cross-attention mechanism. The modal interaction tensor is output based on the modal weighting calculation results.

[0021] Optionally, in the modal fusion layer of the multi-layer tensor coupling analysis model, the local coupling weight matrix is ​​first encoded. Z-score standardization is then used to convert the local coupling weights of the four-dimensional modes (thermal, electrical, structural, and environmental) into standardized correlation vectors, eliminating dimensional differences. Next, the mutual information entropy difference between thermal-electrical, thermal-structural, thermal-environmental, electrical-structural, electrical-environmental, and structural-environmental modes is calculated using the information entropy formula. This difference reflects the strength of nonlinear dependencies between modes; positive values ​​indicate synergistic enhancement effects, while negative values ​​indicate inhibition effects. Smaller entropy differences indicate stronger correlations, and vice versa. Subsequently, based on the mutual information entropy difference between each risk dimension, a modal dependency graph is constructed. Each node in the graph represents a risk dimension, and the edges between nodes represent the dependencies between dimensions. The weight of each edge is proportional to its mutual information entropy difference. This modal dependency graph provides structural information for subsequent modal weighting calculations. Finally, based on the results of the modal dependency graph, a mutual information gain-driven tensor fusion operator is used to perform weighted calculations of the impact on different risk dimensions. Specifically, based on the edge weights of the modal dependency graph, the mutual information gain of each modal pair is calculated by dividing it by the maximum edge weight among all modal pairs. This mutual information gain is used as the basis for the modal fusion weights. A weighted average is then used to generate the fusion weights for the four modes. When the edge weights change, i.e., after the modal dependency is updated, the fusion weights are recalculated. Next, the risk factors of the four modes, such as thermal gradient, current pulse, structural strain, and wind pressure, are multiplied by their corresponding weights, and a weighted outer product is calculated using the tensor fusion operator to complete the initial fusion. Then, risk signal features for each modality pair are extracted, including risk signal strength, response delay, and phase shift. Signal strength can be the amplitude of thermal gradient perturbation, the amplitude of current pulse, etc.; response delay can be the time difference between the thermal response and the current change; and phase shift can be the phase difference between the thermal signal and the current signal. These risk signal features are then analyzed using a nonlinear cross-attention mechanism. Specifically, these risk signal features are input into the attention network, the attention weights for each modality pair are calculated, and the weights are adjusted using a nonlinear activation function, such as ReLU or Sigmoid, to dynamically reconstruct the coupling weight matrix. Finally, this dynamically reconstructed coupling weight matrix is ​​fused with the previous local coupling weight matrix, the fused weights are recalculated, and a weighted outer product is calculated using a tensor fusion operator to construct the final modal interaction tensor. This modal interaction tensor integrates the interactions between different modalities and combines strength, delay, and phase features from various dimensions, providing a foundation for subsequent cross-scale collaborative analysis and further processing in the dynamic feedback layer.

[0022] Furthermore, synchronizing the modal interaction tensor to the cross-scale collaboration layer includes: The modal interaction tensor is decomposed into temporal and spatial scales, and multi-resolution wavelet decomposition and local clustering algorithms are used to extract the high-frequency fast triggering component and low-frequency slow trend component of the risk signal, respectively. The cooperative strength of the high-frequency fast triggering component and the low-frequency slow trend component under different time windows is calculated using a multi-scale coupling operator to establish a scale response matrix. Based on the scale response matrix and the modal interaction tensor, joint tensor regression reconstruction is performed to establish a spatiotemporally consistent cooperative tensor. The cooperative tensor retains both the global consistency of the slow trend and the local sensitivity to the fast disturbance. The cooperative tensor is processed by a recursive cross-scale transfer mechanism to generate a spatiotemporal cooperative tensor.

[0023] Optionally, in the cross-scale collaborative layer of the multi-layer tensor coupling analysis model, to analyze the performance of the modal interaction tensor at different time and spatial scales, it is first decomposed into temporal scales and partitioned into spatial scales. The purpose of this process is to decompose complex spatiotemporal data into different scales, enabling independent processing of different risk signals at different time and spatial scales. Temporal scale decomposition refers to dividing the data in the modal interaction tensor into different time periods based on its rate of change, focusing on the trend of signal change within different time periods. Spatial scale partitioning divides the data according to spatial location to process risk events occurring at different spatial locations separately. Subsequently, multi-resolution wavelet decomposition is used to decompose the modal interaction tensor into components of different frequencies in the time dimension. High-frequency components represent rapid changes in the signal, typically associated with rapidly changing risk events such as electrical faults and sudden temperature rises; low-frequency components represent slow-trend changes in the signal, typically associated with long-term risk factors such as environmental changes and equipment aging. Through wavelet decomposition, the system can accurately extract the risk features of both rapid and long-term changes. Then, local clustering algorithms, such as K-means, are used to partition the spatially located data, and risk signals in different regions are clustered according to Euclidean distance, automatically dividing the entire photovoltaic plant area into several clusters with different risk characteristics, distinguishing between rapid changes and slow-trend changes within each region. After partitioning, high-frequency and low-frequency components are labeled in each cluster, forming high-frequency rapid triggering components and low-frequency slow-trend components. Next, to further understand the relationship between components of different frequency bands, a multi-scale coupling operator is used to calculate the statistical correlation strength between high-frequency rapid triggering components and low-frequency slow-trend components extracted from the same spatial partition within a sliding time window, such as calculating their mutual information at different lag times. The calculation results for all spatial partitions and all modal pairs are summarized to form a scale response matrix. This scale response matrix quantitatively characterizes the synergistic strength of the mutual driving and amplification of rapid-change risk events and slow-change risk backgrounds throughout the entire plant area. Then, the obtained scale response matrix and modal interaction tensor are used as input, and a joint tensor regression reconstruction is performed using the ridge regression algorithm. This process is achieved by minimizing an error function, such as mean squared error. The reconstructed spatiotemporal consistent cooperative tensor retains both the global consistency of slowly changing trends, such as the overall temperature increase trend, and highlights the local sensitivity to rapidly changing perturbations, such as local short-circuit events. Finally, the system processes the cooperative tensor obtained from the above reconstruction using a recursive cross-scale transfer mechanism. This mechanism is usually implemented by a recurrent neural network, such as an RNN or LSTM unit, which connects the cooperative tensor at the current moment with its historical state sequence for temporal modeling.In this way, the system can learn and simulate the evolution of risk patterns at different time scales. For example, how a rapid voltage disturbance gradually evolves into a sustained thermal runaway risk under a slowly increasing temperature background. After iterative processing and information transmission by this recurrent neural network, a spatiotemporal co-evolutionary tensor is finally generated. This spatiotemporal co-evolutionary tensor is a dynamic risk state representation containing historical memory. It fully describes the complex coupling evolution path between risk modes at different time and spatial scales, providing the core basis for dynamic feedback and final decision-making in the next stage.

[0024] Furthermore, the temporal recursion and coupled residual analysis of the spatiotemporal co-operation tensor using the dynamic feedback layer includes: A time series recursive relationship is established using the historical risk superposition matrix and the spatiotemporal co-operation tensor. The residual vectors under thermal, electrical, structural, and environmental modes are calculated within a continuous time window. Coupled residual analysis is performed on the residual vectors to identify the deviation transmission effects of different risk modes. The spatiotemporal co-operation tensor is weighted and corrected according to the deviation transmission effects to output the risk superposition matrix.

[0025] Optionally, in the dynamic feedback layer of the multi-layer tensor coupling analysis model, the spatiotemporal co-occurrence tensor at the current moment is first associated with the historical risk superposition matrix of several consecutive time windows to establish a time-series recursive relationship. This time-series recursive relationship is usually implemented through a recurrent neural network, such as LSTM or temporal convolutional network. This model can learn and memorize the dynamic patterns of risk evolution. Using this recursive relationship, the system can predict the expected state of each risk mode at the current moment based on the historical risk superposition matrix, and compare the predicted value with the actual value reflected by the spatiotemporal co-occurrence tensor to calculate the difference between them, forming residual vectors for thermal, electrical, structural, and environmental modes. These residual vectors quantify the deviation between the current actual risk state and the expected state based on historical trends. Positive residuals indicate that the risk exceeds expectations, while negative residuals indicate that the risk is lower than expected. Subsequently, coupled residual analysis is performed on the obtained residual vectors of each mode. This analysis aims to reveal how the deviations between different risk modes influence and propagate each other. Specifically, the system employs a sliding time window, such as 5 minutes, to calculate the cross-correlation matrix between residual vectors. Then, it constructs a deviation propagation network based on a directed graph model, where nodes represent modal residuals and edge weights represent cross-correlation coefficients. The PageRank algorithm is used to calculate node importance scores, identifying key propagation paths and defining deviation propagation effects. For example, if the electrical-to-thermal path has the highest score, electrical faults are identified as the primary source of thermal risk deviation. Subsequently, based on the deviation propagation effects in the coupled residual analysis results, the weights of each risk dimension in the spatiotemporal co-operation tensor are adjusted. Modes exhibiting strong deviation propagation effects in the residual analysis are given higher weights to ensure a more accurate reflection of their dynamic changes. For instance, if the residuals of the thermal mode significantly impact the residuals of the electrical mode, it indicates that temperature change is a crucial factor in predicting electrical risks. In this case, the weight of the thermal mode can be increased to enhance sensitivity to its influence. This weighted adjustment dynamically and iteratively adjusts each dimension of the spatiotemporal co-operation tensor, enabling the model to better capture the interactions and deviation effects between different modes in subsequent risk assessments. After weighted correction, the final output is a corrected risk overlay matrix. This risk overlay matrix is ​​a multi-dimensional matrix that records the interaction strength between various risk dimensions. After weighted correction, each element in the matrix reflects the actual impact strength of each risk mode at the current moment and their interrelationships, providing accurate risk data for subsequent dynamic risk assessment, thereby enabling continuous monitoring and accurate assessment of risks.

[0026] Based on the risk superposition matrix, a risk transmission graph structure spanning time slices is constructed, and the directed transmission entropy and influence path calculation between nodes within the risk transmission graph structure are performed using a dynamic causal reasoning channel to construct a risk triggering chain.

[0027] In one embodiment, after obtaining the risk superposition matrix, the system constructs a cross-time-slice risk transmission graph structure based on this matrix, utilizing a sliding time window and the directed transmission entropy of any two risk nodes. The core of this risk transmission graph structure is to depict the transmission relationship between risk factors using data changes over time. Subsequently, the system uses a dynamic causal inference channel and algorithms such as weighted Bayesian causal structure learning to conduct in-depth analysis of the nodes in the constructed risk transmission graph. Simultaneously, it introduces temporal consistency constraints and uses a residual correction mechanism to eliminate pseudo-corresponding edges that are statistically significant but do not conform to physical common sense or temporal logic, forming a high-confidence risk transmission graph. Finally, based on this high-confidence risk transmission graph, a graph traversal algorithm is used to identify all directed paths with continuous causal relationships. Each path constitutes a clear risk triggering chain, intuitively demonstrating the entire chain evolution process from the initial trigger to the final fire risk, providing a decisive basis for accurately locating key risk sources and implementing chain-breaking interventions.

[0028] Furthermore, constructing a risk transmission graph structure across time slices based on the aforementioned risk overlay matrix includes: Based on the time evolution sequence of each risk dimension in the risk superposition matrix, a risk node set is constructed using a sliding time window; the directed transmission entropy of any two risk nodes in the risk node set is calculated to establish a risk transmission graph structure.

[0029] Preferably, for the time evolution sequence of each risk dimension in the risk superposition matrix, the system sets a fixed-length time window, slides across the entire time evolution sequence, and gradually extracts the data subset of each position within each time window. The data within each time window is considered a risk node, representing risk factor data at a specific point in time. Each risk node typically includes the corresponding dimension, spatial location, and risk data. By summarizing these risk nodes, a time-ordered set of risk nodes can be provided for subsequent analysis. Subsequently, for any two risk nodes in the risk node set, one is designated as the source node and the other as the target node, and the directed transfer entropy between them is calculated. This directed transfer entropy is an information-theoretic metric used to quantify the reduction in information about the future state of another node contained in the time series data of one node. During the calculation, statistical methods are used to analyze the information contribution of the current state of the source node to the future state of the target node. Specifically, the difference between the probability distribution of the next state of the target node when the current states of the source and target nodes are known and the probability distribution when only the current state of the target node is known is compared, and the information contained in this conditional probability change is calculated. Next, all possible state combinations are traversed and statistically analyzed. This data is then aggregated and averaged to obtain the directed transmission entropy value from the source node to the target node. By comparing the magnitudes of the bidirectional transmission entropy, the causal direction can be determined. If the transmission entropy from the source node to the target node is significantly greater than the reverse value, a unidirectional causal relationship from the source node to the target node is confirmed. This value serves as the weight of the edge in the risk transmission graph, used to quantify the intensity of risk transmission. Subsequently, based on the calculated directed transmission entropy, the system constructs a risk transmission graph structure. In this structure, each risk node is a node in the graph, and the edges between nodes represent the information flow and causal relationships between these risk factors. The weight of the edge is determined by the calculated directed transmission entropy; the higher the transmission entropy, the greater the edge weight, indicating a stronger influence of that risk factor on another factor. Through this risk transmission graph structure, the system can intuitively see how different risk factors interact and reveal possible risk propagation paths, providing a reliable basis for real-time early warning and decision-making.

[0030] Furthermore, constructing a risk trigger chain includes: Dynamic causal reasoning of the risk transmission graph structure is performed using a dynamic causal reasoning channel. The directed connection relationship between nodes is determined by a weighted Bayesian causal structure learning algorithm, and the graph structure is updated. The results of the graph structure update are used to optimize the temporal consistency constraint, and pseudo-associative edges are eliminated by a residual correction mechanism to construct a high-confidence risk transmission graph. Node paths with continuous triggering relationships are identified on the high-confidence risk transmission graph to establish a risk triggering chain.

[0031] Optionally, after constructing the risk transmission graph structure, the system performs causal inference on the graph structure through a dynamic causal inference channel. Dynamic causal inference aims to reveal the causal relationships between nodes in the graph and determine the direction of information flow by analyzing historical data. In the dynamic causal inference channel, based on the initial structure of the risk transmission graph, the system dynamically captures the node risk within each time slice using a sliding window, forming a spatiotemporal sequence data stream. Subsequently, the dynamic causal inference channel employs a weighted Bayesian causal structure learning algorithm, treating each node in the risk transmission graph structure as a random variable and using existing edge weights as prior knowledge. By calculating the posterior probability of the currently observed spatiotemporal sequence data stream under all possible directed acyclic graph structures, the most likely directed connections between nodes are determined. Through this weighted Bayesian causal structure learning algorithm, the likelihood of different causal structures can be evaluated, and graph structure updates can be performed accordingly—that is, strengthening connections highly supported by data and weakening or removing connections with lower probability, thereby obtaining a statistically more robust causal graph. Next, based on the updated graph structure, a physical-logical cleanup is performed. This process involves applying temporal consistency constraints, mandating that all directed edges point from nodes with earlier timestamps to nodes with later timestamps, strictly adhering to the causal law that "cause must precede effect," thus automatically eliminating any invalid connections that violate temporal order. Then, a residual correction mechanism is activated: a structural equation model is constructed based on the optimized causal graph, calculating the residual between the actual observed value of each node and the predicted values ​​based on all its "cause" nodes. If the residual of an edge consistently exceeds a threshold and its transmission entropy is lower than the neighborhood mean, it is considered a spurious association and deleted. After these two optimization steps, a high-confidence risk transmission graph can be constructed, where each directed edge possesses both statistical significance and physical temporal rationality. Then, based on the high-confidence risk transmission graph, the system runs path search algorithms in graph theory, such as depth-first search or shortest path algorithms, to identify node paths with continuous triggering relationships. These paths consist of a series of directed edges connected end-to-end, starting from one or more initial risk source nodes (e.g., "abnormal temperature in area A at time t1"), passing through several intermediate nodes for transmission and amplification (e.g., "causing a current pulse in area A at time t2," which in turn "triggers insulation degradation in area B at time t3"), and finally pointing to one or more target nodes with high fire risk. Each such complete causal sequence is established as a clear risk triggering chain. This risk triggering chain clearly reveals the entire process of risk from generation and transmission to final outbreak, providing the most direct decision-making basis for achieving accurate early warning and chain-breaking prevention and control.

[0032] Dynamic fire risk assessment of the photovoltaic plant area is performed based on the risk superposition matrix and the risk triggering chain.

[0033] In one embodiment, after calculating the risk superposition matrix and constructing the risk triggering chain, the system performs eigenvalue decomposition on the risk superposition matrix, extracts the energy distribution of the main risk modes, and fuses the multi-dimensional risk values ​​of each spatial unit using a weighted geometric average algorithm to generate a normalized comprehensive risk index in the range of 0-1, where a value above 0.8 indicates a high-risk state. Subsequently, the risk triggering chain, as the causal transmission framework, is mapped onto the spatiotemporal grid of the risk superposition matrix. Through precise correspondence between chain nodes and matrix elements, the starting point, path nodes, and endpoint of risk transmission are located. Then, the located starting point, path nodes, and endpoint are combined with the comprehensive risk index to construct a dynamic fire risk assessment result. This assessment result is presented in real-time through a 3D digital twin platform of the plant area, displaying the spatial distribution of risk in the form of a heat map and simulating the risk transmission path with flow animation, providing comprehensive support for operation and maintenance decisions.

[0034] Furthermore, the dynamic assessment of fire risk in photovoltaic plant areas based on the aforementioned risk superposition matrix and risk triggering chain also includes: The risk superposition matrix and the risk triggering chain are used to perform early warning level matching for independent regional early warning and propagation collaborative early warning, and a dual-source early warning signal is constructed; early warning issuance management is performed based on the dual-source early warning signal.

[0035] Preferably, during dynamic risk assessment, the system also performs level matching between independent area early warning and propagation-coordinated early warning. In this process, the real-time risk index of each independent area, such as a single inverter or string unit, is extracted based on the risk overlay matrix. When the index of any area exceeds a preset threshold, an independent area early warning signal is generated, with its level (e.g., low, medium, high) determined by the magnitude of the threshold exceedance. Simultaneously, the system analyzes risk triggering chains, identifies activated chains with high propagation probability, and generates propagation-coordinated early warning signals based on the current propagation stage, scope, and terminal risk level of the chain. By merging these two early warning signals, a unified dual-source early warning signal is constructed, encompassing both static risk status and dynamic propagation trends. Subsequently, based on the dual-source early warning signal, a tiered response protocol is immediately activated, and early warning information is simultaneously reported through multiple channels, including audible and visual alarms, monitoring center pop-ups, and mobile terminal push notifications. This provides enhanced information support for subsequent trend analysis and emergency response, thus forming a closed-loop management process from early warning generation to dynamic monitoring and adjustment.

[0036] Furthermore, the early warning dispatch management based on the dual-source early warning signal includes: The dual-source early warning signal is used to establish a regional collection and attention system; based on the regional collection and attention system, the collection strategy of multiple distributed nodes is reconstructed, and the attention monitoring and early warning are executed.

[0037] Optionally, after the dual-source early warning signal is generated, the system analyzes the signal to accurately locate one or more core areas that triggered the warning. For example, "Inverter No. 5 in Zone C" is marked because its independent risk index exceeds the limit, or all node areas involved in the "risk chain from Zone A to Zone B" are marked. The system identifies these areas as key areas of concern in the digital twin platform and registers their spatial coordinates, risk type, and early warning level information into a dynamic data acquisition task priority list. Subsequently, based on the location and risk characteristics of the key areas of concern, the system reconstructs the acquisition strategy of multiple distributed nodes and executes monitoring and early warning. For infrared thermal imaging units, the system can instruct its pan-tilt unit to immediately turn and align with the key areas of concern, switching the scanning mode from wide-area periodic inspection (e.g., once every 10 minutes) to continuous tracking scanning of specific areas (e.g., once per second), and activating a high-sensitivity temperature measurement mode. For electrical sensing units, the system can instruct them to increase the data sampling frequency from the conventional second level to the millisecond level to capture more transient current and voltage fluctuations. For the environmental monitoring unit, it can be instructed to increase the data reporting frequency of wind speed and temperature sensors located upwind of key areas of concern and locally, in order to accurately assess the impact of the environment on risk evolution. For the video acquisition unit, it can be instructed to adjust preset positions to continuously record video or capture high-definition images of key devices, focusing on identifying visual features such as smoke and open flames. Through this series of acquisition strategy reconstructions, the system can enhance its perception capability of high-risk areas during the warning period, providing an immediate chain of evidence to verify the accuracy of the warning, and providing data support for subsequent analysis of risk evolution trends and the formulation of precise response plans.

[0038] In summary, the embodiments of this application have at least the following technical effects: First, multi-source sensing units are deployed at multiple distributed nodes within the photovoltaic plant to perform data acquisition and establish a multi-source heterogeneous dataset. These multi-source sensing units include electrical sensing units, infrared thermal imaging units, environmental monitoring units, and video acquisition units. Next, a risk scenario self-calibration mechanism mapped to the distributed nodes is activated, and data processing of the multi-source heterogeneous dataset is performed to establish a spatiotemporally aligned risk factor tensor. Then, based on the risk factor tensor, four-dimensional risk primitives (thermal, electrical, structural, and environmental) are extracted and input into a multi-layer tensor coupling analysis model to calculate the interaction strength between each risk dimension, outputting a risk superposition matrix. Then, a risk transmission graph structure spanning time slices is constructed based on the risk superposition matrix, and the directed transmission entropy and influence path calculations between nodes within the risk transmission graph structure are performed using a dynamic causal inference channel to construct a risk triggering chain. Finally, a dynamic fire risk assessment of the photovoltaic plant is performed based on the risk superposition matrix and the risk triggering chain. This method solves the technical problems of traditional static assessment methods, which lead to distorted characterization of risk evolution due to the failure to effectively integrate multi-source heterogeneous data, and delayed early warning due to the neglect of dynamic coupling effects and transmission paths between risk factors. It achieves the technical effect of accurately calculating the intensity of risk interaction through multi-source data fusion and tensor coupling analysis, and revealing the risk triggering chain by combining dynamic causal reasoning, thereby improving the accuracy, real-time performance and early warning capability of dynamic fire risk assessment.

[0039] Example 2 is based on the same inventive concept as the photovoltaic plant fire risk dynamic assessment method based on big data collection in the previous examples, such as... Figure 2 As shown, this application provides a dynamic fire risk assessment system for photovoltaic plant areas based on big data collection. The system includes: Data Acquisition Module 11: Deploys multi-source sensing units at multiple distributed nodes within the photovoltaic plant area to perform data acquisition and establish a multi-source heterogeneous dataset. The multi-source sensing units include electrical sensing units, infrared thermal imaging units, environmental monitoring units, and video acquisition units. Data Processing Module 12: Activates the risk scenario self-calibration mechanism mapped to the distributed nodes, performs data processing on the multi-source heterogeneous dataset, and establishes a spatiotemporally aligned risk factor tensor. Coupling Analysis Module 13: Extracts four-dimensional risk primitives of thermal-electrical-structural-environment based on the risk factor tensor, inputs the four-dimensional risk primitives into a multi-layer tensor coupling analysis model, calculates the interaction strength between each risk dimension, and outputs a risk superposition matrix. Risk Transmission Analysis Module 14: Constructs a risk transmission graph structure across time slices based on the risk superposition matrix, and uses a dynamic causal inference channel to perform directed transmission entropy and influence path calculations between nodes within the risk transmission graph structure, constructing a risk triggering chain. Risk Assessment Module 15: Performs a dynamic assessment of the fire risk in the photovoltaic plant area based on the risk superposition matrix and the risk triggering chain.

[0040] Furthermore, the coupling analysis module 13 is used to perform the following methods: The multi-layer tensor coupling analysis model includes a local mutual disturbance layer, a modal fusion layer, a cross-scale coordination layer, and a dynamic feedback layer. After the four-dimensional risk primitive is input into the multi-layer tensor coupling analysis model: the local mutual disturbance layer is used to extract the transient rate of change of each risk dimension in the four-dimensional risk primitive, calculate the coupling degree of thermal gradient disturbance, current pulse, structural strain mutation, and environmental wind pressure, and generate a local coupling weight matrix; after the local coupling weight matrix is ​​sent to the modal fusion layer, the nonlinear cross-influence between the four modes of thermal-electrical-structural-environment is weighted and normalized based on the tensor fusion operator of mutual information gain to form a modal interaction tensor; after the modal interaction tensor is synchronized to the cross-scale coordination layer, dual decoupling and reconstruction of the time scale and spatial scale are performed to extract the slow trend coefficient and fast trigger coefficient, and a spatiotemporal coordination tensor is formed through the scale coordination operator. The spatiotemporal coordination tensor is used to describe the coupling evolution path of risk modes at different scales; the dynamic feedback layer is used to perform temporal recursion and coupling residual analysis of the spatiotemporal coordination tensor, perform dynamic iterative correction, and output a risk superposition matrix.

[0041] Furthermore, the coupling analysis module 13 is used to perform the following methods: Modal correlation encoding is performed on the local coupling weight matrix. Based on the encoding results, the mutual information entropy differences between thermal-electric, thermal-structure, thermal-environment, electrical-structure, electrical-environment, and structure-environment are calculated to establish a modal dependency graph. Based on the modal dependency graph, a tensor fusion operator driven by mutual information gain is used to perform modal weighting calculation, wherein the fusion weights are adaptively adjusted based on the modal dependency degree. During the fusion process, the coupling weights are dynamically reconstructed based on the risk signal strength, response delay, and phase shift through a nonlinear cross-attention mechanism. The modal interaction tensor is output based on the modal weighting calculation results.

[0042] Furthermore, the coupling analysis module 13 is used to perform the following methods: The modal interaction tensor is decomposed into temporal and spatial scales, and multi-resolution wavelet decomposition and local clustering algorithms are used to extract the high-frequency fast triggering component and low-frequency slow trend component of the risk signal, respectively. The cooperative strength of the high-frequency fast triggering component and the low-frequency slow trend component under different time windows is calculated using a multi-scale coupling operator to establish a scale response matrix. Based on the scale response matrix and the modal interaction tensor, joint tensor regression reconstruction is performed to establish a spatiotemporally consistent cooperative tensor. The cooperative tensor retains both the global consistency of the slow trend and the local sensitivity to the fast disturbance. The cooperative tensor is processed by a recursive cross-scale transfer mechanism to generate a spatiotemporal cooperative tensor.

[0043] Furthermore, the coupling analysis module 13 is used to perform the following methods: A time series recursive relationship is established using the historical risk superposition matrix and the spatiotemporal co-operation tensor. The residual vectors under thermal, electrical, structural, and environmental modes are calculated within a continuous time window. Coupled residual analysis is performed on the residual vectors to identify the deviation transmission effects of different risk modes. The spatiotemporal co-operation tensor is weighted and corrected according to the deviation transmission effects to output the risk superposition matrix.

[0044] Furthermore, the risk transmission analysis module 14 is used to perform the following methods: Based on the time evolution sequence of each risk dimension in the risk superposition matrix, a risk node set is constructed using a sliding time window; the directed transmission entropy of any two risk nodes in the risk node set is calculated to establish a risk transmission graph structure.

[0045] Furthermore, the risk transmission analysis module 14 is used to perform the following methods: Dynamic causal reasoning of the risk transmission graph structure is performed using a dynamic causal reasoning channel. The directed connection relationship between nodes is determined by a weighted Bayesian causal structure learning algorithm, and the graph structure is updated. The results of the graph structure update are used to optimize the temporal consistency constraint, and pseudo-associative edges are eliminated by a residual correction mechanism to construct a high-confidence risk transmission graph. Node paths with continuous triggering relationships are identified on the high-confidence risk transmission graph to establish a risk triggering chain.

[0046] Furthermore, the risk assessment module 15 is used to perform the following methods: The risk superposition matrix and the risk triggering chain are used to perform early warning level matching for independent regional early warning and propagation collaborative early warning, and a dual-source early warning signal is constructed; early warning issuance management is performed based on the dual-source early warning signal.

[0047] Furthermore, the risk assessment module 15 is used to perform the following methods: The dual-source early warning signal is used to establish a regional collection and attention system; based on the regional collection and attention system, the collection strategy of multiple distributed nodes is reconstructed, and the attention monitoring and early warning are executed.

[0048] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A dynamic fire risk assessment method for photovoltaic plant areas based on big data collection, characterized in that, The method includes: Multi-source sensing units are deployed at multiple distributed nodes within the photovoltaic plant area to perform data acquisition and establish a multi-source heterogeneous dataset. The multi-source sensing units include electrical sensing units, infrared thermal imaging units, environmental monitoring units, and video acquisition units. A risk scenario self-calibration mechanism is activated and mapped to distributed nodes, data processing of multi-source heterogeneous datasets is performed, and a risk factor tensor with spatiotemporal alignment is established. Based on the risk factor tensor, a four-dimensional risk element of thermal-electrical-structural-environment is extracted. The four-dimensional risk element is input into a multi-layer tensor coupling analysis model to calculate the interaction strength between each risk dimension and output the risk superposition matrix. Based on the risk superposition matrix, a risk transmission graph structure across time slices is constructed, and the directed transmission entropy and influence path calculation between nodes within the risk transmission graph structure are performed using a dynamic causal reasoning channel to construct a risk triggering chain. Dynamic fire risk assessment of the photovoltaic plant area is performed based on the risk superposition matrix and the risk triggering chain.

2. The method for dynamic assessment of fire risk in photovoltaic plant areas based on big data collection as described in claim 1, characterized in that, The four-dimensional risk primitive is input into a multilayer tensor coupling analysis model, including: The multi-layer tensor coupling analysis model includes a local mutual interference layer, a modal fusion layer, a cross-scale collaborative layer, and a dynamic feedback layer. After the four-dimensional risk primitive is input into the multilayer tensor coupling analysis model: The transient rate of change of each risk dimension in the four-dimensional risk primitive is extracted using the local mutual interference layer. The coupling degree of thermal gradient disturbance, current pulse, structural strain mutation and environmental wind pressure is calculated to generate a local coupling weight matrix. After the local coupling weight matrix is ​​sent to the modal fusion layer, the nonlinear cross-influence between the four modes of thermal-electrical-structural-environment is weighted and normalized based on the tensor fusion operator of mutual information gain to form a modal interaction tensor. After synchronizing the modal interaction tensor to the cross-scale collaboration layer, a dual decoupling and reconstruction of the temporal and spatial scales is performed to extract the slow trend coefficient and the fast trigger coefficient. A spatiotemporal collaboration tensor is formed through the scale collaboration operator. The spatiotemporal collaboration tensor is used to describe the coupling evolution path of risk modes at different scales. The dynamic feedback layer is used to perform temporal recursion and coupled residual analysis of the spatiotemporal co-operation tensor, execute dynamic iterative correction, and output a risk superposition matrix.

3. The method for dynamic assessment of fire risk in photovoltaic plant areas based on big data collection as described in claim 2, characterized in that, After sending the local coupling weight matrix to the modal fusion layer, the process includes: Modal correlation encoding is performed on the local coupling weight matrix. Based on the encoding results, the mutual information entropy difference between thermal-electric, thermal-structure, thermal-environment, electrical-structure, electrical-environment, and structure-environment is calculated respectively, and a modal dependency graph is established. Based on the modal dependency graph, a tensor fusion operator driven by mutual information gain is used to perform modal weighting calculation, wherein the fusion weights are adaptively adjusted based on the modal dependency degree; During the fusion process, the coupling weights are dynamically reconstructed based on the risk signal strength, response delay, and phase shift through a nonlinear cross-attention mechanism. Output the modal interaction tensor based on the modal weighted calculation results.

4. The method for dynamic assessment of fire risk in photovoltaic plant areas based on big data collection as described in claim 2, characterized in that, After synchronizing the modal interaction tensor to the cross-scale collaboration layer, the process includes: The modal interaction tensor is decomposed into temporal scales and partitioned into spatial scales, and multi-resolution wavelet decomposition and local clustering algorithms are used to extract the high-frequency fast triggering component and low-frequency slow trend component of the risk signal, respectively. The cooperative strength of high-frequency fast-triggering components and low-frequency slow-trend components under different time windows is calculated using multi-scale coupling operators to establish a scale response matrix. Joint tensor regression reconstruction is performed based on the scale response matrix and modal interaction tensor to establish a spatiotemporally consistent collaborative tensor, which simultaneously preserves the global consistency of slow-changing trends and the local sensitivity to fast-changing perturbations. The co-current tensor is processed by a recursive cross-scale transfer mechanism to generate a spatiotemporal co-current tensor.

5. The method for dynamic assessment of fire risk in photovoltaic plant areas based on big data collection as described in claim 2, characterized in that, Temporal recursion and coupled residual analysis of spatiotemporal co-operational tensors using the dynamic feedback layer includes: By using the historical risk superposition matrix and the spatiotemporal co-operation tensor to establish a time series recursive relationship, the residual vectors under thermal, electrical, structural and environmental modes are calculated within a continuous time window; Perform coupled residual analysis on the residual vector to identify the bias transmission effects of different risk modes; The spatiotemporal cooperative tensor is weighted and corrected based on the aforementioned deviation transmission effect, and a risk superposition matrix is ​​output.

6. The method for dynamic assessment of fire risk in photovoltaic plant areas based on big data collection as described in claim 1, characterized in that, Based on the risk overlay matrix, a risk transmission graph structure spanning time slices is constructed, including: Based on the time evolution sequence of each risk dimension in the risk superposition matrix, a risk node set is constructed using a sliding time window; Calculate the directed transmission entropy between any two risk nodes in the risk node set, and establish the risk transmission graph structure.

7. The method for dynamic assessment of fire risk in photovoltaic plant areas based on big data collection as described in claim 6, characterized in that, Constructing a risk trigger chain includes: Dynamic causal reasoning of the risk transmission graph structure is performed using a dynamic causal reasoning channel. The directed connection relationship between nodes is determined by a weighted Bayesian causal structure learning algorithm, and the graph structure is updated. The time-series consistency constraint optimization is performed using the graph structure update results, and pseudo-associative edges are eliminated through the residual correction mechanism to construct a high-confidence risk transmission graph; Identify node paths with continuous triggering relationships on the high-confidence risk transmission graph and establish a risk triggering chain.

8. The method for dynamic assessment of fire risk in photovoltaic plant areas based on big data collection as described in claim 1, characterized in that, Based on the aforementioned risk superposition matrix and the aforementioned risk triggering chain, the dynamic assessment of fire risk in the photovoltaic plant area also includes: By utilizing the risk superposition matrix and the risk triggering chain, the early warning level matching of independent regional early warning and propagation coordinated early warning is performed to construct a dual-source early warning signal; Early warning dispatch management is performed based on the dual-source early warning signals.

9. The method for dynamic assessment of fire risk in photovoltaic plant areas based on big data collection as described in claim 8, characterized in that, Based on the aforementioned dual-source early warning signals, early warning dispatch management is performed, including: The dual-source early warning signal is used to establish a regional collection and attention mechanism; Based on the data collection strategy of reconstructing multiple distributed nodes according to the region, the monitoring and early warning are executed.

10. A dynamic fire risk assessment system for photovoltaic plant areas based on big data collection, characterized in that, The system is used to implement the dynamic fire risk assessment method for photovoltaic plant areas based on big data collection as described in any one of claims 1-9, the system comprising: Data acquisition module: Multi-source sensing units are deployed at multiple distributed nodes within the photovoltaic plant area to perform data acquisition and establish a multi-source heterogeneous dataset. The multi-source sensing units include electrical sensing units, infrared thermal imaging units, environmental monitoring units, and video acquisition units. Data processing module: activates the risk scenario self-calibration mechanism mapped to distributed nodes, performs data processing on multi-source heterogeneous datasets, and establishes a spatiotemporally aligned risk factor tensor; Coupling Analysis Module: Based on the risk factor tensor, extract four-dimensional risk primitives of thermal-electrical-structural-environment, input the four-dimensional risk primitives into the multilayer tensor coupling analysis model, calculate the interaction strength between each risk dimension, and output the risk superposition matrix; Risk transmission analysis module: Constructs a risk transmission graph structure across time slices based on the risk superposition matrix, and uses a dynamic causal reasoning channel to calculate the directed transmission entropy and influence path between nodes within the risk transmission graph structure, thereby constructing a risk triggering chain; Risk assessment module: Performs dynamic fire risk assessment of the photovoltaic plant area based on the risk superposition matrix and the risk triggering chain.

Images