Power terminal monitoring method and system based on artificial intelligence

Abstract

The invention is suitable for the field of power terminal monitoring, and provides a power terminal monitoring method and system based on artificial intelligence, and the method comprises the steps: collecting and transmitting a monitoring data set of a photovoltaic module collected by a multi-mode sensing array; preprocessing the received monitoring data to generate a standardized feature data set; calling a targeted adaptive analysis model trained for the characteristics of the photovoltaic module; after the standardized feature data set is input into a targeted adaptive analysis model, the early hidden fault probability, the performance degradation rate prediction value and the real-time health degree score of the photovoltaic module are calculated and output through spatio-temporal feature fusion; the terminal control module is linked to execute local quick response and remote operation and maintenance scheduling; the monitoring dimension of a photovoltaic module is complemented through multi-modal sensing, the state of the module is accurately analyzed by means of a self-adaptive model fusing individual and group characteristics, timely response to operation and maintenance and continuous optimization of the model are realized in combination with graded early warning and closed-loop feedback, and the problem of monitoring defects of a traditional power terminal is effectively solved.

Description

Technical Field

[0001] This invention belongs to the field of power terminal monitoring, and particularly relates to a power terminal monitoring method and system based on artificial intelligence. Background Technology

[0002] With the global trend towards a cleaner and lower-carbon energy structure, photovoltaic power plants, as the core carrier of new energy power generation, have seen their installed capacity grow rapidly. As the core power generation unit of a photovoltaic power plant, the operating status of photovoltaic modules directly determines the power plant's power generation efficiency, the equipment's life-cycle revenue, and its safety and stability. If modules experience hidden faults or performance degradation, it can not only lead to a significant decrease in power generation but may also cause serious safety accidents such as hot spots and fires.

[0003] Currently, the industry generally relies on power terminals to monitor the operation of photovoltaic modules. Existing power terminals can collect and transmit basic parameters such as module output voltage, current, power, ambient temperature, and light intensity, and can achieve simple fault alarms by setting fixed thresholds, providing basic data support for site operation and maintenance.

[0004] However, existing photovoltaic (PV) module power terminal monitoring methods still have significant technical shortcomings, making it difficult to meet the needs of refined operation and maintenance at PV sites. On the one hand, traditional monitoring relies on threshold judgment modes, which are insufficient for identifying early-stage latent faults in PV modules. These faults initially have minimal impact on electrical parameters, far from reaching the threshold alarm conditions, making them prone to missed detections. Simultaneously, environmental fluctuations may cause parameters to temporarily exceed thresholds, triggering numerous false alarms, ultimately leading to low fault diagnosis efficiency and increased risk of fault escalation. To address these technical issues and improve the overall operational efficiency of PV terminals, there is an urgent need to develop an artificial intelligence-based power terminal monitoring method and system. Summary of the Invention

[0005] The purpose of this invention is to provide an artificial intelligence-based power terminal monitoring method and system, which aims to solve the problems mentioned in the background art.

[0006] This invention is implemented as follows: On one hand, an artificial intelligence-based method for monitoring power terminals, the method comprising: Collect and transmit monitoring datasets of photovoltaic modules acquired by a multimodal sensor array; The monitoring data includes dynamic response data of electrical parameters of photovoltaic modules under different light intensities and temperature conditions, spatiotemporal characteristic data of thermal distribution on the module surface, and imaging data reflecting the physical state of the module. The received monitoring data is preprocessed to generate a standardized feature dataset; A targeted adaptive analysis model trained for the characteristics of photovoltaic modules is invoked, which integrates the individual aging baseline of photovoltaic modules with the group performance degradation law; After inputting the standardized feature dataset into the targeted adaptive analysis model, the model calculates and outputs the early latent failure probability, performance degradation rate prediction, and real-time health score of the photovoltaic module through spatiotemporal feature fusion. Based on the early probability of hidden faults, the predicted value of performance degradation rate, and the real-time health score, a graded early warning information is generated, which includes the fault development trend and the scope of impact. The terminal control module is then linked to perform local rapid response and remote operation and maintenance scheduling.

[0007] As a further aspect of the present invention, the acquisition and transmission of the monitoring dataset of the photovoltaic module collected by the multimodal sensing array specifically includes: It is equipped with a multimodal sensor array, including an electrical parameter sensor, an infrared thermal imager, an electroluminescent imaging device, and an environmental sensor; Set a dynamic sampling frequency. High-frequency sampling is triggered when the light intensity is greater than the intensity threshold, and a normal frequency is used in a stable state. Synchronously collect dynamic response data of electrical parameters, and record the complete waveform of current and voltage characteristic curves and power fluctuation data over time under different combinations of light and temperature. Collect spatiotemporal characteristic data of surface thermal distribution of photovoltaic modules; The spatiotemporal characteristic data of surface thermal distribution include the temperature value and area change of hot spot region and the temperature gradient difference between adjacent components; An encrypted transmission channel is established, and a data compression algorithm is used to transmit the monitoring dataset, while also attaching timestamps and sensor number information.

[0008] As a further aspect of the present invention, the invocation of the targeted adaptive analysis model trained for the characteristics of photovoltaic modules specifically includes: A basic framework for loading a targeted adaptive analysis model is provided, which includes a spatial feature extraction layer fused with a convolutional neural network, a temporal feature analysis layer fused with a long short-term memory network, and an attention fusion layer. Establish an individual aging baseline for photovoltaic modules, and calculate initial performance parameters and monthly degradation coefficients based on the factory parameters and previous operating data of the photovoltaic modules; Establish a database of group performance degradation patterns, collect operating data of several photovoltaic modules in the same batch, and extract common degradation characteristics in different environmental regions through clustering algorithms; The model's basic parameters are updated in real time, and the normal operating data from the previous period is retrieved at preset time points to perform incremental training on the model.

[0009] As a further embodiment of the present invention, the standardized feature dataset is divided into two channels according to the time window and the spatial domain to generate a spatiotemporal feature matrix; By fusing convolutional neural network layers to analyze spatial features, the hidden crack risk index, hot spot aggregation degree, and spatial distribution uniformity of electrical parameters are output. By analyzing the temporal characteristics through the long short-term memory network layer, the power decay acceleration, voltage fluctuation frequency, and IV curve shape change trend were calculated. An attention mechanism is activated to perform feature fusion, assigning preset weights to key features such as the temperature gradient at the hot spot edge and the power loss corresponding to microcracks. Output three-dimensional evaluation results: The three-dimensional evaluation results include: early latent failure probability, predicted future periodic performance degradation rate, and real-time health score on a 0-100 scale.

[0010] As a further aspect of the present invention, the step of generating graded early warning information containing fault development trends and impact range projections based on early latent fault probability, performance degradation rate prediction, and real-time health score, and linking the terminal control module to perform local rapid response and remote operation and maintenance scheduling specifically includes: Set tiered early warning thresholds: Level 1, Level 2, and Level 3. Generate a visual early warning report, including a fault development trend curve, an impact range projection diagram, and an estimated economic loss; The system executes a local rapid response: a Level 1 warning triggers component power limiting, a Level 2 warning starts the cooling fan, and a Level 3 warning illuminates the local warning light. Push remote operation and maintenance scheduling information; Level 1 warning generates emergency repair work orders within an hour; Level 2 warning generates maintenance work orders within a day. Establish a closed-loop feedback mechanism to record the changes in component status after the early warning response, and store the processing results as label data in the model training library for the next round of model optimization.

[0011] As a further aspect of the present invention, another option is an artificial intelligence-based power terminal monitoring system, the system comprising: The data acquisition and transmission module is used to acquire and transmit the monitoring dataset of the photovoltaic module collected by the multimodal sensor array; The preprocessing module is used to perform preprocessing on the received monitoring data; The targeted adaptive analysis model module is used to call the targeted adaptive analysis model trained for the characteristics of photovoltaic modules; The calculation output module is used to input the standardized feature dataset into the targeted adaptive analysis model, and then calculate and output the early latent failure probability, performance degradation rate prediction value and real-time health score of the photovoltaic module through spatiotemporal feature fusion. The generation module is used to generate graded early warning information that includes the fault development trend and the scope of impact, based on the early hidden fault probability, the predicted value of the performance degradation rate and the real-time health score. The execution module is used to coordinate with the terminal control module to perform local rapid response and remote operation and maintenance scheduling.

[0012] As a further aspect of the present invention, the targeted adaptive analysis model module specifically includes: Basic framework unit, used as the basic framework for loading the targeted adaptive analysis model; The first calculation unit is used to establish the individual aging baseline of photovoltaic modules and calculate the initial performance parameters and monthly degradation coefficient based on the photovoltaic module's factory parameters and previous operating data. The extraction unit is used to construct the individual aging baseline of photovoltaic modules and calculate the initial performance parameters and monthly degradation coefficient based on the photovoltaic modules' factory parameters and previous operating data. The incremental training unit is used to update the basic parameters of the model in real time and to call the normal operating data of the previous period at a preset time point to perform incremental training on the model.

[0013] As a further aspect of the present invention, the calculation output module specifically includes: The generation unit is used to divide the standardized feature dataset into two channels according to the time window and the spatial domain to generate a spatiotemporal feature matrix; The first output unit is used to analyze spatial features by fusing convolutional neural network layers and output the hidden crack risk index, hot spot aggregation degree and spatial distribution uniformity of electrical parameters. The second computing unit is used to analyze the timing characteristics through the long short-term memory network layer and calculate the power decay acceleration, voltage fluctuation frequency and the changing trend of the current and voltage characteristic curves. The feature fusion unit is used to initiate the attention mechanism to perform feature fusion, and assigns a preset multiple weight to key features such as the temperature gradient at the hot spot edge and the power loss corresponding to microcracks. The second output unit is used to output the three-dimensional evaluation results.

[0014] This invention provides an artificial intelligence-based power terminal monitoring method and system. By supplementing the monitoring dimensions of photovoltaic modules through multimodal sensing, it accurately analyzes the status of modules by relying on an adaptive model that integrates individual and group characteristics, and achieves timely operation and maintenance response and continuous model optimization by combining hierarchical early warning and closed-loop feedback. This effectively solves the shortcomings of traditional power terminal monitoring, ensures the safe operation of modules, and improves operation and maintenance efficiency. Attached Figure Description

[0015] Figure 1 This is the main flowchart of an artificial intelligence-based power terminal monitoring method.

[0016] Figure 2 This is a flowchart illustrating the process of calling a targeted adaptive analysis model trained on the characteristics of photovoltaic modules in an artificial intelligence-based power terminal monitoring method.

[0017] Figure 3 This is a flowchart illustrating how a standardized feature dataset is input into a targeted adaptive analysis model and then the model is used to calculate and output the early latent fault probability, performance degradation rate prediction, and real-time health score of photovoltaic modules through spatiotemporal feature fusion in an artificial intelligence-based power terminal monitoring method.

[0018] Figure 4 This is a main structure diagram of an artificial intelligence-based power terminal monitoring system.

[0019] Figure 5 This is a structural block diagram of a targeted adaptive analysis model module in an artificial intelligence-based power terminal monitoring system.

[0020] Figure 6 This is a structural block diagram of the calculation output module in an artificial intelligence-based power terminal monitoring system. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] The specific implementation of the present invention will be described in detail below with reference to specific embodiments.

[0023] The present invention provides an artificial intelligence-based power terminal monitoring method and system, which solves the technical problems in the background art.

[0024] like Figure 1 The diagram shown is a main flowchart of an artificial intelligence-based power terminal monitoring method according to an embodiment of the present invention. The artificial intelligence-based power terminal monitoring method includes: Step S100: Collect and transmit the monitoring dataset of the photovoltaic module acquired by the multimodal sensor array; The monitoring data includes dynamic response data of electrical parameters of photovoltaic modules under different light intensities and temperature conditions, spatiotemporal characteristic data of thermal distribution on the module surface, and imaging data reflecting the physical state of the modules; Step S200: Perform preprocessing on the received monitoring data to generate a standardized feature dataset; Step S300: Call the targeted adaptive analysis model trained for the characteristics of photovoltaic modules, wherein the targeted adaptive analysis model integrates the individual aging baseline of photovoltaic modules and the group performance degradation law; Step S400: After inputting the standardized feature dataset into the targeted adaptive analysis model, the model calculates and outputs the early latent failure probability, performance degradation rate prediction value, and real-time health score of the photovoltaic module through spatiotemporal feature fusion. Step S500: Based on the early latent fault probability, performance degradation rate prediction value and real-time health score, generate graded early warning information including fault development trend and impact range projection, and link the terminal control module to perform local rapid response and remote operation and maintenance scheduling. In this embodiment, a multimodal sensor array is first configured: a 0.1-class Hall current sensor is selected as the electrical parameter sensor, and a FLIR infrared thermal imager is used. The A325sc uses an electroluminescence imaging device with a wavelength of 900-1100nm and an environmental sensor sampling rate of 1Hz to collect current-voltage characteristic curves, thermal distribution, and electroluminescence image data under different illumination and temperature conditions. In the preprocessing stage, a temperature compensation algorithm is used to filter noise for silicon-based components, and the current-voltage characteristic curves are mapped to STC conditions for dynamic normalization. Hidden crack length and hotspot shape factors are extracted from the electroluminescence images to generate a standardized dataset of 28-dimensional electrical parameters, 15-dimensional thermal distribution, and 12-dimensional imaging. Subsequently, a targeted adaptive model is invoked, using a CNN, a 2-layer LSTM, and an attention layer framework, fusing the aging baseline of a single component over the past 12 months with the group patterns of 1000 components in the same batch. The dataset is then input into the model according to a 2-hour time window and a 6×6 spatial partition. The attention mechanism weights the hotspot gradient and microcrack power loss, outputting PID, microcrack failure probability, 90-day decay rate, and a 0-100 health score. Finally, warnings are issued based on level 1, 2, and 3 thresholds, with local power limiting and heat dissipation activation, and remote push notifications for 1 hour and 24 hours. Hourly work orders record results and feed them back into the model.

[0025] In a preferred embodiment of the present invention, the acquisition and transmission of the monitoring dataset of the photovoltaic module collected by the multimodal sensor array specifically includes: It is equipped with a multimodal sensor array, including an electrical parameter sensor, an infrared thermal imager, an electroluminescent imaging device, and an environmental sensor; Set a dynamic sampling frequency. High-frequency sampling is triggered when the light intensity is greater than the intensity threshold, and a normal frequency is used in a stable state. Synchronously collect dynamic response data of electrical parameters, and record the complete waveform of current and voltage characteristic curves and power fluctuation data over time under different combinations of light and temperature. Collect spatiotemporal characteristic data of surface thermal distribution of photovoltaic modules; The spatiotemporal characteristic data of surface thermal distribution include the temperature value and area change of hot spot regions, as well as the temperature gradient differences between adjacent components; An encrypted transmission channel was established, and the monitoring dataset was transmitted using a data compression algorithm, while also attaching timestamps and sensor number information. In this embodiment, the sensor array is first configured according to the monitoring requirements: the electrical parameter sensor can be the CSM025 series, the infrared thermal imager is FLIR A325sc, the electroluminescent imaging device is PV EL-100, and the environmental sensor is SHT30; when the light intensity suddenly increases from 500W / m² to 800W / m², a high-frequency acquisition is triggered once every 10 seconds, and when it stabilizes, the frequency is switched to a regular frequency once every 30 seconds to balance data integrity and redundancy; when synchronously acquiring electrical parameters, five groups of light intensity-temperature combinations are set, and each group measures the current-voltage characteristic curve and power fluctuation; when acquiring heat distribution data, the hot spot area is counted using the pixel counting method, and a 2℃ threshold is set for the temperature gradient between adjacent components; the transmission link uses SSL / TLS 1.3 encryption, and the data packet is appended with a millisecond-level timestamp and the codes for the site, array, string, component, and sensor. like Figure 2 As shown, in a preferred embodiment of the present invention, the step of calling the targeted adaptive analysis model trained for the characteristics of photovoltaic modules includes: Step S301: Load the basic framework of the targeted adaptive analysis model, which includes a fusion convolutional neural network spatial feature extraction layer, a long short-term memory network temporal feature analysis layer, and an attention fusion layer; Step S302: Construct an individual aging baseline for photovoltaic modules. Based on the factory parameters of the photovoltaic modules and the previous operating data, calculate the initial performance parameters and monthly degradation coefficient. Step S303: Establish a database of group performance degradation patterns, collect operating data of several photovoltaic modules in the same batch, and extract common degradation characteristics of different environmental regions through clustering algorithms; Step S304: Update the basic parameters of the model in real time, and call up the normal operating data of the previous period at preset time points to perform incremental training on the model; In this embodiment, ResNet18 is used as the CNN spatial feature extraction layer, with two LSTM layers (128 hidden units per layer to capture the temporal trend of electrical parameters) and a Scaled Dot-Product Attention layer. The initial parameters are derived from training with 100,000 sets of multi-brand component failure samples. When constructing the individual aging baseline, the monthly decay coefficient is calculated by linear fitting using the component's factory parameters and data from the previous 12 months. When establishing the group pattern database, data from more than 500 components in the same batch are collected and classified into three categories according to the environment: high temperature (≥35℃), high humidity (RH≥80%), and high radiation (≥1600kWh / m²). K-Means clustering (K=3) is used to extract commonalities. When updating the model parameters, normal data from the previous 24 hours is called up from 2-4 am every day (the station's off-peak time), and incremental training is performed using the SGD optimizer, which takes ≤30 minutes. At the same time, real-time updates ensure the long-term accuracy of the model and avoid model failure due to component aging.

[0026] like Figure 3 As shown, in a preferred embodiment of the present invention, the step of inputting the standardized feature dataset into the targeted adaptive analysis model and then calculating and outputting the early latent failure probability, performance degradation rate prediction, and real-time health score of the photovoltaic module through spatiotemporal feature fusion specifically includes: Step S401: Divide the standardized feature dataset into two channels according to the time window and the spatial domain to generate a spatiotemporal feature matrix; Step S402: By fusing convolutional neural network layers to analyze spatial features, output the hidden crack risk index, hot spot aggregation degree and spatial distribution uniformity of electrical parameters; Step S403: Analyze the timing characteristics through the long short-term memory network layer, and calculate the power decay acceleration, voltage fluctuation frequency, and IV curve shape change trend; Step S404: Activate the attention mechanism to perform feature fusion, and assign preset multiple weights to key features such as the temperature gradient at the hot spot edge and the power loss corresponding to microcracks; Step S405: Output the three-dimensional evaluation results; The three-dimensional evaluation results include: early latent failure probability, predicted future periodic performance degradation rate, and real-time health score on a 0-100 scale. It should be understood that the standardized dataset is first divided into two channels: a 2-hour time window is set to balance data volume and timeliness, covering short-term operating condition fluctuations; the spatial domain is divided into 6×6 partitions based on the component's 1640×992mm size, forming a time-space matrix; when the CNN layer analyzes spatial features, the hidden crack index is crack length / component diagonal, the hot spot aggregation degree is hot spot number / 36 partitions, and the electrical parameter uniformity is power standard deviation / average power; when the LSTM layer analyzes time-series features, the calculation process for power decay acceleration is as follows: , At 1-hour intervals, the frequency of voltage fluctuations exceeding ±2% of the rated value is statistically analyzed (≥3 abnormal occurrences), and the trend of current and voltage curves is compared with the weekly deviation (>3% abnormality). An attention mechanism is used to weight hotspot gradient (1.6) and microcrack power loss (1.8). The output consists of three dimensions: failure probability, decay over the next 90 days, and health score of 0-100. This result serves as the basis for graded early warning. Spatiotemporal fusion ensures accuracy, and the feature calculation logic corresponds with the collected thermal distribution and electrical parameter data, ensuring the continuity of the technical chain.

[0027] In a preferred embodiment of the present invention, the step of generating graded early warning information containing fault development trends and impact range projections based on early latent fault probability, performance degradation rate prediction, and real-time health score, and linking the terminal control module to perform local rapid response and remote operation and maintenance scheduling specifically includes: Set tiered early warning thresholds: Level 1, Level 2, and Level 3. Generate a visual early warning report, including a fault development trend curve, an impact range projection diagram, and an estimated economic loss; The system executes a local rapid response: a Level 1 warning triggers component power limiting, a Level 2 warning starts the cooling fan, and a Level 3 warning illuminates the local warning light. Push remote operation and maintenance scheduling information; Level 1 warning generates emergency repair work orders within an hour; Level 2 warning generates maintenance work orders within a day. Establish a closed-loop feedback mechanism to record the changes in component status after the early warning response, and store the processing results as label data in the model training library for the next round of model optimization. In this embodiment, when setting graded early warning thresholds, Level 1 (health level < 50 / failure probability > 80%, severe fault), Level 2 (50-70 / 50%-80%, moderate fault), and Level 3 (70-85 / 30%-50%, minor fault) balance sensitivity and false alarm rate. When generating a visualization report, a line chart is used to display the 7-day fault probability trend, a heat map of the risk of three adjacent components, and economic losses are calculated based on the current electricity price. Local response: Level 1 limits power to 60% of the rated value (to prevent overheating), Level 2 starts the cooling fan (2000r / min) at 40℃, and Level 3 uses bright lights (red / yellow / blue). Remote dispatch uses 5G slices to transmit data, with Level 1 work orders within 1 hour and Level 2 work orders within 24 hours, including component number, fault type, and handling suggestions. The closed-loop feedback records response time, handling results, performance recovery rate, and tagged data are stored in the model library for incremental model training.

[0028] like Figure 4 As shown, in another preferred embodiment of the present invention, an artificial intelligence-based power terminal monitoring system is provided, the system comprising: The data acquisition and transmission module 100 is used to acquire and transmit the monitoring dataset of the photovoltaic module acquired by the multimodal sensor array; The preprocessing module 200 is used to perform preprocessing on the received monitoring data; The targeted adaptive analysis model module 300 is used to call the targeted adaptive analysis model trained for the characteristics of photovoltaic modules; The calculation output module 400 is used to input the standardized feature dataset into the targeted adaptive analysis model, and then calculate and output the early latent failure probability, performance degradation rate prediction value and real-time health score of the photovoltaic module through spatiotemporal feature fusion. The generation module 500 is used to generate graded early warning information that includes the fault development trend and the scope of impact, based on the early hidden fault probability, the predicted value of the performance degradation rate and the real-time health score. The execution module 600 is used to coordinate with the terminal control module to perform local rapid response and remote operation and maintenance scheduling.

[0029] In this embodiment, the acquisition and transmission module 100 acquires and transmits the monitoring dataset of the photovoltaic module collected by the multimodal sensor array. The preprocessing module 200 performs preprocessing on the received monitoring data. The targeted adaptive analysis model module 300 calls the targeted adaptive analysis model trained for the characteristics of the photovoltaic module. The calculation and output module 400 inputs the standardized feature dataset into the targeted adaptive analysis model and calculates and outputs the early latent fault probability, performance degradation rate prediction value, and real-time health score of the photovoltaic module through spatiotemporal feature fusion. Based on the early latent fault probability, performance degradation rate prediction value, and real-time health score, the generation module 500 generates graded early warning information containing fault development trends and impact range inferences. The execution module 600 links with the terminal control module to perform local rapid response and remote operation and maintenance scheduling.

[0030] like Figure 5 As shown, in another preferred embodiment of the present invention, the targeted adaptive analysis model module 300 specifically includes: Basic framework unit 301 is used to load the basic framework of the targeted adaptive analysis model; The first calculation unit 302 is used to establish an individual aging baseline for photovoltaic modules and calculate initial performance parameters and monthly degradation coefficients based on the factory parameters and previous operating data of the photovoltaic modules. Extraction unit 303 is used to construct an individual aging baseline for photovoltaic modules and calculate initial performance parameters and monthly degradation coefficients based on the factory parameters and previous operating data of photovoltaic modules. Incremental training unit 304 is used to update the model's basic parameters in real time and to incrementally train the model by calling up the normal operating data from the previous cycle at preset time points. In this embodiment, when applied, the basic framework unit 301 loads the basic framework of the targeted adaptive analysis model, the first calculation unit 302 establishes the individual aging baseline of the photovoltaic module, and calculates the initial performance parameters and monthly degradation coefficient based on the photovoltaic module's factory parameters and previous operating data. The extraction unit 303 constructs the individual aging baseline of the photovoltaic module, and calculates the initial performance parameters and monthly degradation coefficient based on the photovoltaic module's factory parameters and previous operating data. The incremental training unit 304 updates the model's basic parameters in real time, and calls the normal operating data of the previous period at a preset time point to perform incremental training on the model.

[0031] like Figure 6 As shown, in another preferred embodiment of the present invention, the calculation output module 400 specifically includes: The generation unit 401 is used to divide the standardized feature dataset into two channels according to the time window and the spatial domain to generate a spatiotemporal feature matrix. The first output unit 402 is used to analyze spatial features by fusing convolutional neural network layers and output the hidden crack risk index, hot spot aggregation degree and uniformity of spatial distribution of electrical parameters. The second computing unit 403 is used to analyze the timing characteristics through the long short-term memory network layer and calculate the power decay acceleration, voltage fluctuation frequency and current-voltage characteristic curve shape change trend. The feature fusion unit 404 is used to initiate the attention mechanism to perform feature fusion, and assigns a preset multiple weight to the key features of the hot spot edge temperature gradient and the power loss corresponding to the microcracks. The second output unit 405 is used to output the three-dimensional evaluation results.

[0032] In this embodiment, the generation unit 401 divides the standardized feature dataset into two channels according to the time window and the spatial domain to generate a spatiotemporal feature matrix. The first output unit 402 is used to analyze the spatial features by fusing convolutional neural network layers and output the hidden crack risk index, hot spot aggregation degree and the spatial distribution uniformity of electrical parameters. The second calculation unit 402 analyzes the temporal features by using a long short-term memory network layer and calculates the power decay acceleration, voltage fluctuation frequency and the changing trend of the current-voltage characteristic curve. The feature fusion unit 404 starts the attention mechanism to perform feature fusion and assigns a preset multiple weight to the key features of the hot spot edge temperature gradient and the power loss corresponding to the microcrack. The second output unit 405 outputs the three-dimensional evaluation results.

[0033] The above embodiments of the present invention provide an artificial intelligence-based power terminal monitoring method and an artificial intelligence-based power terminal monitoring system. Firstly, a multimodal sensor array is configured: a 0.1-class Hall current sensor is selected as the electrical parameter sensor, and a FLIR infrared thermal imager is used. The A325sc uses an electroluminescence imaging device with a wavelength of 900-1100nm and an environmental sensor sampling rate of 1Hz to collect current-voltage characteristic curves, thermal distribution, and electroluminescence image data under different illumination and temperature conditions. In the preprocessing stage, a temperature compensation algorithm is used to filter noise for silicon-based components, and the current-voltage characteristic curves are mapped to STC conditions for dynamic normalization. Hidden crack length and hotspot shape factors are extracted from the electroluminescence images to generate a standardized dataset of 28-dimensional electrical parameters, 15-dimensional thermal distribution, and 12-dimensional imaging. Subsequently, a targeted adaptive model is invoked, using a CNN, a 2-layer LSTM, and an attention layer framework, fusing the aging baseline of a single component over the past 12 months with the group patterns of 1000 components in the same batch. The dataset is then input into the model according to a 2-hour time window and a 6×6 spatial partition. The attention mechanism weights the hotspot gradient and microcrack power loss, outputting PID, microcrack failure probability, 90-day decay rate, and a 0-100 health score. Finally, warnings are issued based on level 1, 2, and 3 thresholds, with local power limiting and heat dissipation activation, and remote push notifications for 1 hour and 24 hours. Hourly work orders are used to record results and feed them back into the model. First, the sensor array is configured according to monitoring requirements: the electrical parameter sensor can be the CSM025 series, the infrared thermal imager is FLIR A325sc, the electroluminescent imaging device is PV EL-100, and the environmental sensor is SHT30. If the illumination suddenly increases from 500W / m² to 800W / m², a high-frequency acquisition is triggered once every 10 seconds; when stable, a regular frequency of once every 30 seconds is used to balance data integrity and redundancy. When synchronously acquiring electrical parameters, five sets of illumination-temperature combinations are set, with each set measuring current-voltage characteristic curves and power fluctuations. When acquiring thermal distribution data, the hotspot area is counted using pixel counting, and a 2℃ threshold is set for the temperature gradient between adjacent components. SSL / TLS 1 is used for transmission.3. Encryption: Data packets are appended with millisecond-level timestamps and codes for the site, array, string, component, and sensor. ResNet18 is used as the CNN spatial feature extraction layer, along with two LSTM layers (128 hidden units per layer to capture the temporal trend of electrical parameters), and a Scaled Dot-Product Attention layer. Initial parameters are derived from training with 100,000 sets of multi-brand component failure samples. When constructing individual aging baselines, the monthly decay coefficient is calculated using linear fitting with the component's factory parameters and data from the previous 12 months. When establishing the group pattern database, data from over 500 components in the same batch are collected and categorized into three types based on environment: high temperature (≥35℃), high humidity (RH≥80%), and high radiation (≥1600kWh / m²). K-Means clustering (K=3) is used to extract commonalities. When updating model parameters, normal data from the previous 24 hours is retrieved daily between 2-4 AM (the site's lowest point), and incremental training is performed using the SGD optimizer, with a training time ≤ [time missing]. For 30 minutes, real-time updates are performed to ensure the long-term accuracy of the model and prevent model failure due to component aging. First, the standardized dataset is divided into two channels: a 2-hour time window to balance data volume and timeliness, covering short-term operational fluctuations; and a 6×6 partition based on the component's 1640×992mm size, forming a time-space matrix. When the CNN layer analyzes spatial features, the hidden crack index is calculated as crack length / component diagonal, hot spot aggregation as hot spot number / 36 partitions, and electrical parameter uniformity as power standard deviation / average power. When the LSTM layer analyzes temporal features, the calculation process for power decay acceleration is as follows: , At 1-hour intervals, the frequency of voltage fluctuations is counted for the number of times the voltage exceeds ±2% of the rated value (≥3 times abnormal), and the trend of current and voltage curves is compared with the weekly deviation (>3% abnormal); the attention mechanism is used to weight the hot spot gradient (1.6) and microcrack power loss (1.8); the output three-dimensional results are: failure probability, decay in the next 90 days, and health score of 0-100. The results are directly used as the basis for graded early warning. Spatiotemporal fusion ensures accuracy, and the feature calculation logic corresponds with the collected thermal distribution and electrical parameter data, ensuring the continuity of the technical link. When setting graded early warning thresholds, Level 1 (health level < 50 / fault probability > 80%, severe fault), Level 2 (50-70 / 50%-80%, moderate fault), and Level 3 (70-85 / 30%-50%, minor fault) are used to balance sensitivity and false alarm rate. When generating a visualization report, a line chart is used to show the fault probability trend over 7 days, the risk of three adjacent components is shown on the heat map, and economic losses are calculated based on the current electricity price. Local response: Level 1 limits power to 60% of the rated value (to prevent overheating), Level 2 starts the cooling fan (2000r / min) at 40℃, and Level 3 uses bright lights (red / yellow / blue). Remote dispatch uses 5G slices to transmit data. Level 1 work orders are issued within 1 hour, and Level 2 work orders are issued within 24 hours, including component number, fault type, and handling suggestions. The closed-loop feedback records response time, handling results, and performance recovery rate, and the tagged data are stored in the model library for incremental training of the model.

[0034] In order for the above methods and systems to operate smoothly, the system may include more or fewer components than those described above, or combine certain components, or different components, in addition to the various modules mentioned above. For example, it may include input / output devices, network access devices, buses, processors, and memory.

[0035] The processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (OPGs) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor can be a microprocessor or any conventional processor. This processor is the control center of the system, connecting various parts via various interfaces and lines.

[0036] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0037] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

[0038] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A power terminal monitoring method based on artificial intelligence, characterized in that, The method includes: Collect and transmit monitoring datasets of photovoltaic modules acquired by a multimodal sensor array; The monitoring data includes dynamic response data of electrical parameters of photovoltaic modules under different light intensities and temperature conditions, spatiotemporal characteristic data of thermal distribution on the module surface, and imaging data reflecting the physical state of the module. The received monitoring data is preprocessed to generate a standardized feature dataset; A targeted adaptive analysis model trained for the characteristics of photovoltaic modules is invoked, which integrates the individual aging baseline of photovoltaic modules with the group performance degradation law; After inputting the standardized feature dataset into the targeted adaptive analysis model, the model calculates and outputs the early latent failure probability, performance degradation rate prediction, and real-time health score of the photovoltaic module through spatiotemporal feature fusion. Based on the early probability of hidden faults, the predicted value of performance degradation rate, and the real-time health score, a graded early warning information is generated, which includes the fault development trend and the scope of impact. The terminal control module is then linked to perform local rapid response and remote operation and maintenance scheduling.

2. The power terminal monitoring method based on artificial intelligence according to claim 1, characterized in that, The specific data sets collected and transmitted from the multimodal sensor array for monitoring photovoltaic modules include: It is equipped with a multimodal sensor array, including an electrical parameter sensor, an infrared thermal imager, an electroluminescent imaging device, and an environmental sensor; Set a dynamic sampling frequency. High-frequency sampling is triggered when the light intensity is greater than the intensity threshold, and a normal frequency is used in a stable state. Synchronously collect dynamic response data of electrical parameters, and record the complete waveform of current and voltage characteristic curves and power fluctuation data over time under different combinations of light and temperature. Collect spatiotemporal characteristic data of surface thermal distribution of photovoltaic modules; The spatiotemporal characteristic data of surface thermal distribution include the temperature value and area change of hot spot region and the temperature gradient difference between adjacent components; An encrypted transmission channel is established, and a data compression algorithm is used to transmit the monitoring dataset, while also attaching timestamps and sensor number information.

3. The power terminal monitoring method based on artificial intelligence according to claim 1, characterized in that, The specific steps of calling the targeted adaptive analysis model trained on the characteristics of photovoltaic modules include: A basic framework for loading a targeted adaptive analysis model is provided, which includes a spatial feature extraction layer fused with a convolutional neural network, a temporal feature analysis layer fused with a long short-term memory network, and an attention fusion layer. Establish an individual aging baseline for photovoltaic modules, and calculate initial performance parameters and monthly degradation coefficients based on the factory parameters and previous operating data of the photovoltaic modules; Establish a database of group performance degradation patterns, collect operating data of several photovoltaic modules in the same batch, and extract common degradation characteristics in different environmental regions through clustering algorithms; The model's basic parameters are updated in real time, and the normal operating data from the previous period is retrieved at preset time points to perform incremental training on the model.

4. The power terminal monitoring method based on artificial intelligence according to claim 1, characterized in that, The step of inputting the standardized feature dataset into the targeted adaptive analysis model and then calculating and outputting the early latent failure probability, performance degradation rate prediction, and real-time health score of the photovoltaic module through spatiotemporal feature fusion specifically includes: The standardized feature dataset is divided into two channels according to time window and spatial domain to generate a spatiotemporal feature matrix; By fusing convolutional neural network layers to analyze spatial features, the hidden crack risk index, hot spot aggregation degree, and spatial distribution uniformity of electrical parameters are output. By analyzing the temporal characteristics through the long short-term memory network layer, the power decay acceleration, voltage fluctuation frequency, and IV curve shape change trend were calculated. An attention mechanism is activated to perform feature fusion, assigning preset weights to key features such as the temperature gradient at the hot spot edge and the power loss corresponding to microcracks. Output three-dimensional evaluation results: The three-dimensional evaluation results include: early latent failure probability, predicted future periodic performance degradation rate, and real-time health score on a 0-100 scale.

5. The power terminal monitoring method based on artificial intelligence according to claim 1, characterized in that, The process of generating tiered early warning information based on early latent fault probability, performance degradation rate prediction, and real-time health score, including fault development trends and impact range projections, and linking the terminal control module to perform local rapid response and remote operation and maintenance scheduling, specifically includes: Set tiered early warning thresholds: Level 1, Level 2, and Level 3. Generate a visual early warning report, including a fault development trend curve, an impact range projection diagram, and an estimated economic loss; The system executes a local rapid response: a Level 1 warning triggers component power limiting, a Level 2 warning starts the cooling fan, and a Level 3 warning illuminates the local warning light. Push remote operation and maintenance scheduling information; Level 1 warning generates emergency repair work orders within an hour; Level 2 warning generates maintenance work orders within a day. Establish a closed-loop feedback mechanism to record the changes in component status after the early warning response, and store the processing results as label data in the model training library for the next round of model optimization.

6. An artificial intelligence-based power terminal monitoring system, characterized in that, The system, employing the artificial intelligence-based power terminal monitoring method as described in any one of claims 1-5, comprises: The data acquisition and transmission module is used to acquire and transmit the monitoring dataset of the photovoltaic module collected by the multimodal sensor array; The preprocessing module is used to perform preprocessing on the received monitoring data; The targeted adaptive analysis model module is used to call the targeted adaptive analysis model trained for the characteristics of photovoltaic modules; The calculation output module is used to input the standardized feature dataset into the targeted adaptive analysis model, and then calculate and output the early latent failure probability, performance degradation rate prediction value and real-time health score of the photovoltaic module through spatiotemporal feature fusion. The generation module is used to generate graded early warning information that includes the fault development trend and the scope of impact, based on the early hidden fault probability, the predicted value of the performance degradation rate and the real-time health score. The execution module is used to coordinate with the terminal control module to perform local rapid response and remote operation and maintenance scheduling.

7. The artificial intelligence-based power terminal monitoring system according to claim 6, characterized in that, The targeted adaptive analysis model module specifically includes: Basic framework unit, used as the basic framework for loading the targeted adaptive analysis model; The first calculation unit is used to establish the individual aging baseline of photovoltaic modules and calculate the initial performance parameters and monthly degradation coefficient based on the factory parameters and previous operating data of photovoltaic modules. The extraction unit is used to construct the individual aging baseline of photovoltaic modules and calculate the initial performance parameters and monthly degradation coefficient based on the photovoltaic modules' factory parameters and previous operating data. The incremental training unit is used to update the basic parameters of the model in real time and to call the normal operating data of the previous period at a preset time point to perform incremental training on the model.

8. The artificial intelligence-based power terminal monitoring system according to claim 6, characterized in that, The calculation output module specifically includes: The generation unit is used to divide the standardized feature dataset into two channels according to the time window and the spatial domain to generate a spatiotemporal feature matrix; The first output unit is used to analyze spatial features by fusing convolutional neural network layers and output the hidden crack risk index, hot spot aggregation degree and spatial distribution uniformity of electrical parameters. The second computing unit is used to analyze the timing characteristics through the long short-term memory network layer and calculate the power decay acceleration, voltage fluctuation frequency and the changing trend of the current and voltage characteristic curves. The feature fusion unit is used to initiate the attention mechanism to perform feature fusion, and assigns a preset multiple weight to key features such as the temperature gradient at the hot spot edge and the power loss corresponding to microcracks. The second output unit is used to output the three-dimensional evaluation results.

Images