Health degree detection and early warning method and system for photovoltaic equipment

Abstract

The invention provides a health degree detection early warning method and system for photovoltaic equipment. The method comprises the following steps: acquiring electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters and second acquisition time information corresponding to the physical parameters of at least one kind of photovoltaic equipment; compressing the electrical parameter, the physical parameter, the first acquisition time information corresponding to the electrical parameter and the second acquisition time information corresponding to the physical parameter of the at least one photovoltaic device to obtain a pre-processing data packet of the photovoltaic device; inputting the preprocessed data packet into a photovoltaic equipment health degree detection model for photovoltaic equipment health degree detection processing to obtain a health degree detection result of the photovoltaic equipment; and generating a health degree early warning result of the photovoltaic equipment according to the health degree detection result and the health degree early warning threshold range, and outputting the health degree early warning result. According to the scheme of the invention, dynamic early warning of the health degree of the photovoltaic equipment is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of intelligent detection of photovoltaic equipment, and particularly relates to a health degree detection and early warning method and system for photovoltaic equipment. BACKGROUND

[0002] With the large-scale development of photovoltaic intelligent stations, the number of equipment has increased from thousands of units in a single station to tens of thousands of units, and the traditional operation and maintenance mode is facing three major bottlenecks: 1. Limited monitoring dimension: The existing early warning devices mainly collect electrical parameters such as voltage and current (accounting for 70%), and lack of physical state data such as temperature field distribution and mechanical vibration, which leads to 20% of hidden faults (such as component hidden cracks and inverter capacitor aging) that cannot be detected; 2. Response lag: The existing early warning devices rely on centralized analysis in the cloud, and after 50GB of data per day in a single station is transmitted to the cloud, the analysis delay is more than 30 minutes, and the fault detection time lags by an average of 24 hours; 3. Low early warning accuracy: The fixed threshold judgment (such as temperature > 85℃ alarm) is used, which does not combine with the equipment aging curve, and the false alarm rate is as high as 35%, and the invalid work rate of operation and maintenance personnel is more than 40%. The data of a million-kilowatt photovoltaic intelligent station shows that the non-planned downtime caused by insufficient monitoring and early warning is up to 80h per year, and the direct loss of power is more than 2 million kWh. SUMMARY

[0003] The technical problem to be solved by the present application is to provide a health degree detection and early warning method and system for photovoltaic equipment, which realizes dynamic and rapid early warning of photovoltaic equipment using multi-monitoring dimension data.

[0004] To solve the above technical problems, the technical solutions of the present application are as follows: In a first aspect, the present application provides a health degree detection and early warning method for photovoltaic equipment, comprising: obtaining electrical parameters, physical parameters, first collection time information corresponding to the electrical parameters, and second collection time information corresponding to the physical parameters of at least one photovoltaic equipment; compressing the electrical parameters, the physical parameters, the first collection time information corresponding to the electrical parameters, and the second collection time information corresponding to the physical parameters of the at least one photovoltaic equipment to obtain a preprocessed data packet of the photovoltaic equipment; inputting the preprocessed data packet into a photovoltaic equipment health degree detection model for photovoltaic equipment health degree detection processing to obtain a health degree detection result of the photovoltaic equipment; the photovoltaic equipment health degree detection model is trained according to at least one of historical electrical parameters, historical physical parameters, first historical collection time information corresponding to the historical electrical parameters, and second historical collection time information corresponding to the historical physical parameters of the at least one photovoltaic equipment; Based on the health status detection results and the health status warning threshold range, a health status warning result for the photovoltaic equipment is generated and output.

[0005] Optionally, the electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one photovoltaic device are compressed to obtain a preprocessed data package of the photovoltaic device, including: The electrical parameters of at least one photovoltaic device are preprocessed to obtain first preprocessed data; The physical parameters of at least one photovoltaic device are preprocessed to obtain second preprocessed data; The first acquisition time information corresponding to the electrical parameters is preprocessed to obtain the third preprocessed data; The second acquisition time information corresponding to the physical parameters is preprocessed to obtain the fourth preprocessed data; The first, second, third, and fourth preprocessed data are compressed to obtain a preprocessed data package for the photovoltaic equipment.

[0006] Optionally, the training process of the photovoltaic equipment health detection model includes: The cloud-edge collaborative platform trains a first preset network model based on the first historical electrical parameters and first historical physical parameters of at least one photovoltaic device to obtain a first network model; it then trains a second preset network model based on the first historical electrical parameters, first historical physical parameters, first historical acquisition time information corresponding to the first historical electrical parameters, and second historical acquisition time information corresponding to the first historical physical parameters of at least one photovoltaic device to obtain a second network model; finally, based on the first network model and the second network model, it obtains the photovoltaic device health detection model; or The edge computing gateway trains a third preset network model based on the second historical electrical parameters, the second historical physical parameters, the third historical acquisition time information corresponding to the second historical electrical parameters, and the fourth acquisition time information corresponding to the second historical physical parameters of at least one photovoltaic device, to obtain the photovoltaic device health detection model; the duration of the third historical acquisition time is less than the duration of the first historical acquisition time, and the duration of the fourth historical acquisition time is less than the duration of the second historical acquisition time.

[0007] Optionally, a first preset network model is trained based on first historical electrical parameters and first historical physical parameters of at least one photovoltaic device to obtain a first network model, including: Multiple first training subsets are determined based on the first historical electrical parameters and first historical physical parameters of the at least one photovoltaic device; Multiple first classifiers in the first preset network model are trained based on the multiple first training subsets to obtain multiple second classifiers, and the multiple second classifiers constitute the first network model.

[0008] Optionally, a second preset network model is trained based on at least one photovoltaic device's first historical electrical parameters, first historical physical parameters, first historical acquisition time information corresponding to the first historical electrical parameters, and second historical acquisition time information corresponding to the first historical physical parameters, to obtain a second network model, including: Based on the first historical acquisition time information corresponding to the first historical electrical parameter and the second historical acquisition time information corresponding to the first historical physical parameter, the first historical electrical parameter and the first historical physical parameter are aligned to determine multiple second training subsets; The second preset network model is trained based on the multiple second training subsets to obtain the second network model.

[0009] Optionally, the photovoltaic equipment health detection model is obtained based on the first network model and the second network model, including: The first dynamic weight and the second dynamic weight are determined according to the preset mapping table. The photovoltaic equipment health detection model is obtained based on the first dynamic weight, the second dynamic weight, the first network model, and the second network model.

[0010] Optionally, the preprocessed data packet is input into the photovoltaic equipment health detection model for photovoltaic equipment health detection processing to obtain the photovoltaic equipment health detection results, including: The preprocessed data packet is input into the first network model to obtain the output result of the first network model; The preprocessed data packet is input into the second network model to obtain the output result of the second network model; The target value is determined based on the output of the first network model and the output of the second network model. The first dynamic weight and the second dynamic weight are determined based on the target value; Based on the output results of the first network model, the output results of the second network model, the first dynamic weight, and the second dynamic weight, the health detection results of the photovoltaic equipment are obtained.

[0011] Optionally, based on the second historical electrical parameters, second historical physical parameters, third historical acquisition time information corresponding to the second historical electrical parameters, and fourth acquisition time information corresponding to the second historical physical parameters of at least one photovoltaic device, a third preset network model is trained to obtain the photovoltaic device health detection model, including: The second historical electrical parameters and the second historical physical parameters are aligned based on the third historical acquisition time information corresponding to the second historical electrical parameters and the fourth historical acquisition time information corresponding to the second historical physical parameters to determine multiple third training subsets. The third preset network model is trained based on the multiple third training subsets to obtain the third network model.

[0012] Optionally, The preprocessed data packet is input into the photovoltaic equipment health detection model for photovoltaic equipment health detection processing to obtain the photovoltaic equipment health detection results, including: The preprocessed data packet is input into the third network model for photovoltaic equipment health detection processing to obtain the photovoltaic equipment health detection results.

[0013] Secondly, embodiments of the present invention also provide a health detection and early warning system for photovoltaic equipment, comprising: A health prediction device is used to receive pre-processed data packets from photovoltaic (PV) devices sent by an edge computing gateway. These pre-processed data packets are obtained by the edge computing gateway through compression processing of electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one PV device. The pre-processed data packets are then input into a PV device health detection model for PV device health detection processing to obtain a PV device health detection result. The PV device health detection model is trained based on at least one of the electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one PV device. Based on the health detection result and a health warning threshold range, a PV device health warning result is generated and output.

[0014] The above-described solution of the present invention has at least the following beneficial effects: The above-described solution of the present invention obtains electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one photovoltaic device; compresses the electrical parameters, physical parameters, first acquisition time information, and second acquisition time information corresponding to the physical parameters of the at least one photovoltaic device to obtain a preprocessed data package of the photovoltaic device; inputs the preprocessed data package into a photovoltaic device health detection model for photovoltaic device health detection processing to obtain a photovoltaic device health detection result; the photovoltaic device health detection model is trained based on at least one of the historical electrical parameters, historical physical parameters, first historical acquisition time information corresponding to the historical electrical parameters, and second historical acquisition time information corresponding to the historical physical parameters of at least one photovoltaic device; and generates and outputs a photovoltaic device health warning result based on the health detection result and the health warning threshold range; thus realizing dynamic and rapid early warning of the photovoltaic device's health. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of an embodiment of the health detection and early warning method for photovoltaic equipment according to the present invention; Figure 2 This is a schematic diagram of a specific implementation example of the health detection and early warning method for photovoltaic equipment based on the present invention. Detailed Implementation

[0016] Exemplary embodiments of the invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the invention are shown in the drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this invention will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art.

[0017] like Figure 1 As shown, an embodiment of the present invention proposes a health detection and early warning method for photovoltaic equipment, comprising: Step 11: Obtain electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one photovoltaic device; Step 12: Compress the electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one photovoltaic device to obtain a preprocessed data package of the photovoltaic device; Step 13: Input the preprocessed data packet into the photovoltaic equipment health detection model for photovoltaic equipment health detection processing to obtain the photovoltaic equipment health detection result; the photovoltaic equipment health detection model is trained based on at least one of the following: historical electrical parameters, historical physical parameters, first historical acquisition time information corresponding to the historical electrical parameters, and second historical acquisition time information corresponding to the historical physical parameters of the photovoltaic equipment. Step 14: Based on the health detection results and the health warning threshold range, generate and output the health warning results of the photovoltaic equipment.

[0018] This invention constructs a full lifecycle health record for equipment by integrating multi-dimensional parameters (electrical parameters, physical parameters, and corresponding data collection timestamps) with big data analysis and AI models, achieving a health assessment accuracy of ≥92%. Based on the health level, it triggers multi-level early warnings and links with the operation and maintenance system to generate work orders. This achieves real-time perception of equipment status, dynamic health assessment, and accurate early warning, reducing fault detection time to 3-7 days compared to traditional methods, and lowering site operation and maintenance costs by 20%-30%. It is applicable to photovoltaic smart sites of all sizes.

[0019] Step 11 can be detected by a cluster of monitoring terminals installed on the photovoltaic equipment. The monitoring terminal cluster may include a sensing unit, a data processing unit, and a communication unit. The sensing unit is used to collect electrical parameters, such as the series current signal and surface temperature distribution data of the photovoltaic module, as well as physical data, such as the vibration signals of the inverter fan and transformer core of the photovoltaic module. The data processing unit is used to receive the series current signal, surface temperature distribution data, and vibration signal, convert them into readable parameters, and record the timestamps of parameter acquisition (i.e., the first acquisition time information corresponding to the electrical parameters and the second acquisition time information corresponding to the physical parameters). The communication unit is used to receive the parameters and timestamps from the data processing unit, encapsulate them into MODBUS-TCP data frames, and upload them to the edge computing gateway.

[0020] The sensing unit includes a Hall current sensor, an infrared temperature array sensor, and a microelectromechanical system (MEMS) vibration sensor. The Hall current sensor is configured to acquire the series current signal of the photovoltaic module, with a range of 0-50A and a linearity of 0.1%. The infrared temperature array sensor is configured to capture the temperature field distribution data of the photovoltaic module, with a resolution of 32×32 pixels and a temperature measurement resolution of 0.1℃. The MEMS vibration sensor is configured to monitor the vibration signals of the inverter fan and transformer core, with a bandwidth of 1-10kHz. The data processing unit uses an STM32L496 microcontroller with power consumption below 50mA and supports 16-bit AD sampling at a sampling rate of 1kHz. The Hall current sensor, infrared temperature array sensor, and MEMS vibration sensor are communicatively connected to the data processing unit, which in turn is communicatively connected to the communication unit. The communication unit integrates both Ethernet (100Mbps) and LoRa (868MHz, transmission distance 3km) interfaces.

[0021] In this embodiment, a distributed monitoring terminal cluster can be used. The monitoring terminal cluster may include photovoltaic module monitoring terminals, inverter monitoring terminals, and combiner box monitoring terminals, used to monitor each photovoltaic module, inverter, and combiner box of the photovoltaic equipment. Each monitoring terminal includes a sensing unit, a data processing unit, and a communication unit. The sensing unit may include an integrated current sensor (measurement range 0-50A, accuracy ±0.5%FS), a temperature sensor (-40℃-125℃, accuracy ±0.3℃), and a vibration sensor (range ±10g, frequency response 1-10kHz).

[0022] Each monitoring terminal can be installed using a strong magnetic base (neodymium iron boron material, 50N adsorption force), eliminating the need for drilling. The photovoltaic module monitoring terminal incorporates a high-energy lithium battery (5000mAh capacity, ≥6 months of battery life) and a solar charging module (≥22% conversion efficiency), adaptable to operating environments from -30℃ to 70℃. The inverter monitoring terminal can also be mounted via DIN rail (compatible with 35mm standard rails), supporting DC24V power supply (fluctuation range ±10%) and solar energy complementarity. The solar panels use monocrystalline silicon (22% efficiency), paired with a 5000mAh lithium-thionyl chloride battery, allowing for 15 days of continuous operation even in cloudy or rainy weather. Each terminal also includes a built-in protocol conversion module, supporting proprietary protocol parsing, and uniformly converting data output to the MODBUS-RTU / TCIP protocol, ensuring compatibility with existing SCADA systems at the site.

[0023] In some embodiments, step 12 involves compressing the electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one photovoltaic device to obtain a preprocessed data package for the photovoltaic device, including: Step 121: Preprocess the electrical parameters of at least one photovoltaic device to obtain first preprocessed data; Step 122: Preprocess the physical parameters of at least one photovoltaic device to obtain second preprocessed data; Step 123: Preprocess the first acquisition time information corresponding to the electrical parameters to obtain the third preprocessed data; Step 124: Preprocess the second acquisition time information corresponding to the physical parameters to obtain the fourth preprocessed data; Step 125: Compress the first preprocessed data, the second preprocessed data, the third preprocessed data, and the fourth preprocessed data to obtain the preprocessed data package of the photovoltaic device.

[0024] In this embodiment, the edge computing gateway is connected to the monitoring terminal cluster via an industrial Ethernet, and has a built-in ARM Cortex-A72 processor (2.0GHz) and 16GB storage unit. It deploys a lightweight health assessment model (second health assessment regression model) that can complete the health score (0-100 points) of a single device within 500ms.

[0025] In this embodiment, electrical parameters (such as voltage deviation rate and current harmonic distortion rate), physical parameters (such as temperature fluctuation rate and vibration effective value), and time parameters (such as running time and last maintenance interval) can be received through the edge computing gateway and then sent to the cloud-edge collaboration platform.

[0026] The above data undergoes outlier removal, missing value imputation, and normalization, including: (1) Remove outliers (based on the 3σ principle): (11) Calculate the mean (μ) and standard deviation (σ) of the entire data series.

[0027] (12) According to the normal distribution assumption, 99.7% of the data should fall within the interval (μ - 3σ, μ + 3σ).

[0028] (13) Traverse all data points and mark any value that is not in this range as an outlier (set to NaN) and remove it.

[0029] (2) After removing outliers, the empty positions (i.e., NaN) need to be filled with missing values ​​(using the moving average method, with a maximum filling time of 3 minutes): (21) Set a sliding window (e.g., a window size of 5 data points).

[0030] (22) Traverse the data and when a missing value is encountered, fill it with the average of the non-missing values ​​before and after it.

[0031] (3) After filling in the missing values, the data sequence needs to be normalized (mapped to the 0-1 interval): (31) Find the minimum value (min_val) and maximum value (max_val) of the entire data column.

[0032] (32) Apply the following formula to each data point for minimum-maximum normalization: Normalized value = (original value - min_val) / (max_val - min_val).

[0033] The data compression rate after the above processing reaches 70%.

[0034] For step 121, the electrical parameters of at least one photovoltaic device are preprocessed to obtain first preprocessed data. As an example, the electrical parameters are the DC power of the photovoltaic device (4500, 4520, 10, 4400, 4425). The mean (μ) of the DC power is calculated to be 4391 W and the standard deviation (σ) is 76.31 W. This should fall within the interval (4391 - 3*76.31, 4391 + 3*76.31) = (4162.07, 4619.93). Since 10 is not within this range, it is set to NaN, resulting in (4500, 4520, NaN, 4400, 4425). The missing values ​​4520 and 4400 are then imputed using the average of the non-missing values ​​before and after NaN (4520 + 4400) = 4460, resulting in (4500, 4520, 4460, 4400, 4425). Find the minimum value of the entire data column, which is 4400, and the maximum value, which is 4520. Normalize each data point to obtain [(4500-4400) / (4520-4400), (4520-4400) / (4520-4400), (4460-4400) / (4520-4400), (4400-4400) / (4520-4400), (4425-4400) / (4520-4400)], which gives (0.8, 1, 0.5, 0, 0.2).

[0035] For step 122, the physical parameters of at least one photovoltaic device are preprocessed to obtain second preprocessed data. As an example, the physical parameters are the temperature fluctuation rate of the photovoltaic device (0, 0.2, 0.5, 0.1, 0.2). The mean (μ) and standard deviation (σ) of the temperature fluctuation rate are calculated to be 0.2 and 0.161, respectively. These values ​​should fall within the range of (0.2 - 3*0.161, 0.2 + 3*0.161) = (-0.283, 0.683). Since these values ​​are within the range, no padding is required. Find the minimum value of 0 and the maximum value of 0.5 in the entire data column. Normalize each data point to get [(0-0) / (0.5-0), (0.2-0) / (0.5-0), (0.5-0) / (0.5-0), (0.1-0) / (0.5-0), (0.2-0) / (0.5-0)], which gives (0, 0.4, 1, 0.2, 0.4).

[0036] Steps 123 and 124 both involve time processing. Taking the preprocessing of the first acquisition time information corresponding to the electrical parameters to obtain the third preprocessed data as an example, the first acquisition time information is (00:10, 00:20, 00:30, 00:40, 00:50). The mean (μ) of the first acquisition time information is calculated to be 00:30 and the standard deviation (σ) is 14.14 minutes. It should fall within the interval (00:30 - 3*14.14, 00:30+3*14.14) = (11:48 the previous day, 01:12 the current day). Since it is within the range, no padding is needed. Find the minimum value of the entire data column and the maximum value of 00:10. Normalize each data point to obtain [(00:10-00:10) / (00:50-00:10), (00:20-00:10) / (00:50-00:10), (00:30-00:10) / (00:50-00:10), (00:40-00:10) / (00:50-00:10), (00:50-00:10) / (00:50-00:10)], which gives (0, 0.25, 0.5, 075, 1).

[0037] Then, the first, second, third, and fourth preprocessed data are compressed to obtain the preprocessed data packet for the photovoltaic device, thereby reducing the amount of data transmission.

[0038] The compression process includes: (1) Scan the sequence and store the unique values ​​that appear in a "dictionary". Taking the second preprocessed data (0, 0.4, 1, 0.2, 0.4) in step 122 above as an example, the unique values ​​are: 0, 0.4, 1, 0.2. Assign them indices: 0->00; 0.2->01; 0.4->10; 1->11; (2) Replace each value in the original sequence with its index in the dictionary. Encoded sequence: 00 10 1101 10; (3) Generate compressed data: The compressed data needs to contain two parts: Dictionary: {0, 0.2, 0.4, 1} or represented as (00:0, 01:0.2, 10:0.4, 11:1); Encoded stream: 0010110110; The compressed data is: Header: [0, 0.2, 0.4, 1] + Data: [00, 10, 11, 01, 10].

[0039] In the example above, the original data (0, 0.4, 1, 0.2, 0.4) contains 5 floating-point numbers. Assuming each occupies 4 bytes, the total is 20 bytes. The compressed data: the dictionary {0, 0.2, 0.4, 1} contains 4 floating-point numbers ≈ 16 bytes, and the encoded stream 0010110110 contains 5 values, each represented by 2 bits (0.25 bytes) ≈ 1.25 bytes, for a total of 17.25 bytes.

[0040] In some embodiments, step 13, the training process of the photovoltaic equipment health detection model includes: Step 131: The cloud-edge collaborative platform trains a first preset network model based on the first historical electrical parameters and first historical physical parameters of at least one photovoltaic device to obtain a first network model; it then trains a second preset network model based on the first historical electrical parameters, first historical physical parameters, first historical acquisition time information corresponding to the first historical electrical parameters, and second historical acquisition time information corresponding to the first historical physical parameters of at least one photovoltaic device to obtain a second network model; finally, based on the first network model and the second network model, it obtains the photovoltaic device health detection model; or Step 132: The edge computing gateway trains a third preset network model based on the second historical electrical parameters, the second historical physical parameters, the third historical acquisition time information corresponding to the second historical electrical parameters, and the fourth acquisition time information corresponding to the second historical physical parameters of at least one photovoltaic device, to obtain the photovoltaic device health detection model; the duration of the third historical acquisition time is less than the duration of the first historical acquisition time, and the duration of the fourth historical acquisition time is less than the duration of the second historical acquisition time.

[0041] In some embodiments, step 131 involves training a first preset network model based on first historical electrical parameters and first historical physical parameters of at least one photovoltaic device to obtain a first network model, including: Step 1311: Determine multiple first training subsets based on the first historical electrical parameters and first historical physical parameters of the at least one photovoltaic device; Step 1312: Train multiple first classifiers in the first preset network model according to the multiple first training subsets to obtain multiple second classifiers, and the multiple second classifiers constitute the first network model.

[0042] For step 1311, from the first historical electrical parameters (a1, a2, a3, a4, a5, a6, a7) and corresponding labels of at least one photovoltaic device, multiple random samples with replacement are drawn (the same sample may be drawn multiple times, while other samples may not be drawn at all). Each time, N samples are drawn, resulting in multiple first training subsets of electrical parameters. Each first training subset of electrical parameters contains N samples. For example, if N is 3, then the first training subsets of electrical parameters are (a1, a2, a3), (a2, a3, a4), (a1, a3, a4), and (a1, a2, a4). Approximately 37% of the samples will not be drawn in this process. These undrawn samples can be used as a validation set for the model, such as (a4, a5, a7), to evaluate model performance during training.

[0043] Similarly, multiple random samples with replacement are drawn from the first historical physical parameters (b1, b2, b3, b4, b5, b6, b7) of at least one photovoltaic device (the same sample may be drawn multiple times, while other samples may not be drawn at all). Each time, N samples are drawn, resulting in multiple first training subsets of physical parameters. Each first training subset of physical parameters contains N samples. For example, if N is 3, then the first training subsets of physical parameters are (b1, b2, b3), (b2, b3, b4), (b1, b3, b4), and (b1, b2, b4). Approximately 37% of the samples will not be drawn in this process. These undrawn samples can be used as a validation set for the model, such as (b4, b5, b7), to evaluate model performance during training.

[0044] The first training subset of electrical parameters is (a1, a2, a3), (a2, a3, a4), (a1, a3, a4) and (a1, a2, a4), and the first training subset of physical parameters is (b1, b2, b3), (b2, b3, b4), (b1, b3, b4) and (b1, b2, b4), forming multiple first training subsets [(a1, a2, a3), (a2, a3, a4), (a1, a3, a4), (a1, a2, a4), (b1, b2, b3), (b2, b3, b4), (b1, b3, b4), (b1, b2, b4)].

[0045] For step 1312, the multiple first classifiers in the first preset network model are multiple decision trees that have been constructed in advance.

[0046] For multiple first training subsets of electrical parameters, multiple decision trees are constructed. Each first training subset of electrical parameters is used to train one decision tree. The training process is as follows: find the optimal split point in the first training subset of electrical parameters to obtain the node of the decision tree. This results in multiple second classifiers for electrical parameters.

[0047] For multiple physical parameter first training subsets, multiple decision trees are constructed. Each physical parameter first training subset is used to train one decision tree. The training process is as follows: find the optimal split point in the physical parameter first training subset to obtain the node of the decision tree. This results in multiple physical parameter second classifiers.

[0048] Multiple electrical parameter second classifiers and multiple physical parameter second classifiers constitute multiple second classifiers, and multiple second classifiers constitute the first network model.

[0049] That is, each first training subset corresponds to training one decision tree. For example, the first training subset (a1, a2, a3) corresponds to training one decision tree, and the first training subset (b1, b2, b3) corresponds to training another decision tree, resulting in multiple trained decision trees. In other words, the first network model includes multiple trained decision trees.

[0050] The process of finding the optimal split point in the first training subset of electrical (or physical) parameters to obtain the nodes of the decision tree is as follows: Step 1: Sort each parameter in the first training subset of electrical (or physical) parameters, for example, sort (a1, a2, a3) to get (a1, a2, a3); Step 2: Take the average of two adjacent different parameters as the candidate split point. For example, take the average of a1 and a2, p1, as the candidate split point, and take the average of a2 and a3, p2, as the candidate split point. Step 3: Traverse all these candidate split points (e.g., traverse to obtain candidate split points p1 and p2). Based on the two adjacent different parameters corresponding to the candidate split points, separate the parameters in the first training subset of electrical (or physical) parameters to obtain two parameter sets (i.e., divide into one parameter set a1 and another parameter set a2 and a3 based on candidate split point p1; divide into one parameter set a1 and a2 and another parameter set a3 based on candidate split point p2). Calculate the purity index of the two parameter sets (e.g., Gini coefficient, MSE; calculate the Gini coefficient x1 for a1 and a2, and the Gini coefficient x2 for a2 and a3). Determine the candidate split point corresponding to the highest purity index value as the final split point (if x1 is greater than x2, then the data is divided into one parameter set a1 and another parameter set a2 and a3). Use the purity index x1 as the basis for splitting this node.

[0051] Repeat steps 1, 2, and 3 until a1, a2, and a3 are all leaf nodes. Use the historical health score (tag) of the photovoltaic equipment corresponding to a1, a2, and a3 as the value of the next leaf node of a1, a2, and a3.

[0052] In some embodiments, step 131 involves training a second preset network model based on first historical electrical parameters, first historical physical parameters, first historical acquisition time information corresponding to the first historical electrical parameters, and second historical acquisition time information corresponding to the first historical physical parameters of at least one photovoltaic device, thereby obtaining a second network model, including: Step 1313: Align the first historical electrical parameters and the first historical physical parameters according to the first historical acquisition time information corresponding to the first historical electrical parameters and the second historical acquisition time information corresponding to the first historical physical parameters, and determine multiple second training subsets; Step 1314: Train the second preset network model according to the multiple second training subsets to obtain the second network model.

[0053] For step 1313, the first historical electrical parameter can be (a1, a2, a3, a4, a5, a6, a7), and the first historical acquisition time information corresponding to the first historical electrical parameter is t = (1:10, 1:20, 1:20, 1:40, 1:50, 1:60, 1:70). The first historical physical parameter can be (b1, b2, b3, b4, b5, b6, b7), and the second historical acquisition time information corresponding to the first historical physical parameter is (1:10, 1:20, 1:20, 1:40, 1:53, 1:62, 1:81). Then, based on the first historical acquisition time information corresponding to the first historical electrical parameter and the second historical acquisition time information corresponding to the first historical physical parameter, the first historical electrical parameter and the first historical physical parameter are aligned (i.e., the first historical electrical parameter and the first historical physical parameter are combined into an array according to the same timestamp), and multiple data groups are determined as [(a1, b1, 1:10), (a2, b2, 1:20), (a3, b3, 1:10), (a4, b4, 1:40)]. The historical health scores corresponding to the times 1:10, 1:20, 1:10, and 1:40 are queried as f1, f2, f3, and f4, and multiple second training subsets are determined as [(f1, 1:10), (f2, 1:20), (f3, 1:10), (f4, 1:40)].

[0054] For step 1314, the second preset network model is trained based on the multiple second training subsets to obtain the second network model:

[0055] in, The input is the training data at time t. Here is the weight matrix for the forget gate. Let be the bias vector of the forget gate. This represents the output of the forget gate. () is a function that multiplies element by element, where t = 1, 2, 3, ..., T. Let... .

[0056]

[0057] in, The input is the training data at time t. Here is the weight matrix of the input gate. Let be the bias vector of the input gate. Indicates the degree of opening or closing of the input gate. () is a function that multiplies element by element, where t = 1, 2, 3, ..., T.

[0058]

[0059] in, The input is the training data at time t. This is the weight matrix for the candidate cell states. This is the bias vector for the candidate cell state. This indicates the status value to be added. This represents the hidden state at time t-1. () is an activation function that outputs values ​​between -1 and 1 and is used to generate new information.

[0060]

[0061] in, This represents the output of the forget gate. The cell state at time t-1. To indicate the degree of opening or closing of the input gate, This indicates the status value to be added. Let be the cell state at time t, where t = 1, 2, 3, ..., T. .

[0062]

[0063] in, The input is the training data at time t. Here is the weight matrix of the output gate. This is the bias vector for the output gate. This represents the hidden state at time t-1. Indicates the degree of opening or closing of the output gate. () is a function that multiplies element by element, where t = 1, 2, 3, ..., T.

[0064]

[0065] in, Let t represent the cell state at time t. Indicates the degree of opening or closing of the output gate. This represents the hidden state at time t.

[0066] Final hidden state Typically, the output is passed to a fully connected layer and an activation function (such as Softmax) to produce the predicted output. ,in, It is the weight matrix of the output layer. It is the bias vector of the output layer.

[0067] Using the above example, inputting [(f1, 1:10), (f2, 1:20), (f3, 1:10), (f4, 1:40)] into the formula will yield the predicted health score at time t=1:40. 4. Given that the true health score is = f4, calculated using the formula for the loss function. and The loss value is:

[0068] in, Represents the true value. L represents the predicted value. and The loss value.

[0069] Update the weight matrix of the forget gate based on the above loss values. and the bias vector of the forget gate :

[0070] in, This is the updated input gate weight matrix. This is the input gate weight matrix before the update. To calculate the loss L pair The partial derivatives, This represents the learning rate of the input gate.

[0071]

[0072] in, This is the updated forget gate bias vector. This is the forget gate bias vector before the update. To calculate the loss L pair The partial derivatives, This represents the learning rate of the forget gate.

[0073] Update the input gate weight matrix based on the aforementioned loss value. and the bias vector of the input gate :

[0074] in, This is the updated input gate weight matrix. This is the input gate weight matrix before the update. To calculate the loss L pair The partial derivatives, This represents the learning rate of the input gate.

[0075]

[0076] in, This is the updated input gate bias vector. The input gate bias vector before the update. To calculate the loss L pair The partial derivatives, This represents the learning rate of the input gate.

[0077] Update the weight matrix of the candidate cell states based on the aforementioned loss values. Bias vectors of candidate cell states :

[0078] in, This is the weight matrix for the updated candidate cell states. This is the weight matrix of the candidate cell states before the update. To calculate the loss L pair The partial derivatives, This represents the learning rate of the input gate.

[0079]

[0080] in, This is the bias vector for the updated candidate cell states. This is the bias vector of the candidate cell states before the update. To calculate the loss L pair The partial derivatives, The learning rate represents the candidate cell state.

[0081] Update the weight matrix of the output gate based on the above loss values. and the bias vector of the output gate :

[0082] in, The updated weight matrix of the output gate. The weight matrix of the output gate before the update. To calculate the loss L pair The partial derivatives, This represents the learning rate of the input gate.

[0083]

[0084] in, This is the updated output gate bias vector. This is the output gate bias vector before the update. To calculate the loss L pair The partial derivatives, This represents the learning rate of the output gate.

[0085] The bias vector and weight matrix are updated iteratively until the loss value reaches a preset range, thus obtaining the second network model. In some embodiments, step 131, obtaining the photovoltaic equipment health detection model based on the first network model and the second network model, includes: Step 1315: Determine the first dynamic weight and the second dynamic weight according to the preset mapping table; Step 1316: Based on the first dynamic weight, the second dynamic weight, the first network model, and the second network model, obtain the photovoltaic equipment health detection model.

[0086] For step 1315, using the above example, assuming the current time is t, the parameters at time t are input into the first network model to obtain the device health score predicted by the first network model at time t. The parameters at times t, t-1, t-2, t-3… are input into the second network model to obtain the predicted health score at time t. ,according to The device health score at time t is obtained based on the first network model. Calculate the Mean Absolute Error (MAE) (reflecting the reliability of time series predictions), and obtain the device health score at time t based on the first network model. The MAE calculation formula is as follows:

[0087] in, The mean absolute error, This represents the number of predicted values ​​from the output of the second network model. To obtain the device health score at time t based on the first network model, The predicted device health score at time t is obtained based on the second network model.

[0088] The default mapping table contains: When MAE ≤ 2.5% (high reliability): First dynamic weight w1 = 0.7, second dynamic weight w2 = 0.3 (focusing on time series trends); When 2.5% < MAE < 5% (medium reliability): w1 = 0.5, w2 = 0.5 (balancing the characteristics of both). When MAE ≥ 5% (low reliability): w1 = 0.3, w2 = 0.7 (focusing on static characteristics).

[0089] Step 1316: The static feature matrix of the first network model (such as current electrical parameters and physical state) and the second network model (such as 24-hour trend change) are weighted and summed according to the above weights to form a comprehensive feature matrix, which is then input into the final evaluation model to output a health score (0-100 points).

[0090] For step 1316, the health score output by the photovoltaic equipment health detection model is: Health Score = (First Dynamic Weight) The health score predicted by the first network model + the second dynamic weights The health score predicted by the second network model is 2.

[0091] By dynamically adjusting weights through MAE, the problem of high false alarm rate associated with traditional fixed thresholds is solved: when the second network model is reliable, the focus is on capturing latent faults (such as the vibration trend of capacitor aging); when the prediction error is large, the static parameters of the first network model are relied upon to ensure the accuracy of the current state assessment. This achieves a fault identification accuracy of ≥92% as stated in the document, with a false alarm rate reduced to below 8%, supporting the achievement of a 60% reduction in ineffective maintenance work.

[0092] In some embodiments, in step 13, the preprocessed data packet is input into the photovoltaic equipment health detection model for photovoltaic equipment health detection processing to obtain the photovoltaic equipment health detection result, including: Step 133: Input the preprocessed data packet into the first network model to obtain the output result of the first network model; Step 134: Input the preprocessed data packet into the second network model to obtain the output result of the second network model; Step 135: Determine the target value based on the output of the first network model and the output of the second network model; determine the first dynamic weight and the second dynamic weight based on the target value; Step 136: Based on the output results of the first network model, the output results of the second network model, the first dynamic weight, and the second dynamic weight, obtain the health detection results of the photovoltaic equipment.

[0093] In this embodiment, after receiving the preprocessed data packet, the cloud-edge collaborative platform needs to parse it to obtain the electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one photovoltaic device. The parsed data is then input into the second network model to obtain the health scores of one or more predicted photovoltaic devices. Based on the outputs of the first and second network models, the target value is determined, i.e., the mean absolute error of the output results of the first and second network models is determined.

[0094] For step 133, using the above example, assuming the current time is t, the parameters at time t are input into the first network model to obtain the device health score at time t predicted by the first network model. .

[0095] For step 134, using the above example, assuming the current time is t, input the parameters at times t, t-1, t-2, t-3... into the second network model to obtain the device health score predicted by the second network model at time t. .

[0096] For step 135, calculate the target value. :

[0097] in, The mean absolute error, This represents the number of predicted values ​​from the output of the second network model. To obtain the device health score at time t based on the first network model, The predicted device health score at time t is obtained based on the second network model.

[0098] The default mapping table contains: When MAE ≤ 2.5% (high reliability): First dynamic weight w1 = 0.7, second dynamic weight w2 = 0.3 (focusing on time series trends); When 2.5% < MAE < 5% (medium reliability): w1 = 0.5, w2 = 0.5 (balancing the characteristics of both). When MAE ≥ 5% (low reliability): w1 = 0.3, w2 = 0.7 (focusing on static characteristics).

[0099] Step 1316: The static feature matrix of the first network model (such as current electrical parameters and physical state) and the second network model (such as 24-hour trend change) are weighted and summed according to the above weights to form a comprehensive feature matrix, which is then input into the final evaluation model to output a health score (0-100 points).

[0100] For step 136, health score = (first dynamic weight) The health score predicted by the first network model is... +Second dynamic weight The health score predicted by the second network model ) / 2.

[0101] In this embodiment, the first model (processing static features, such as the correlation between temperature and faults) has an accuracy of 90%; the second model (processing temporal features, such as vibration trend prediction) has a mean absolute error (MAE) of less than 5%. Through a feature fusion algorithm (dynamic weight allocation), the overall accuracy is improved to 92%.

[0102] In some embodiments, step 132, training a third preset network model based on second historical electrical parameters, second historical physical parameters, third historical acquisition time information corresponding to the second historical electrical parameters, and fourth acquisition time information corresponding to the second historical physical parameters of at least one photovoltaic device, to obtain the photovoltaic device health detection model, includes: Step 1321: Align the second historical electrical parameters and the second historical physical parameters according to the third historical acquisition time information corresponding to the second historical electrical parameters and the fourth historical acquisition time information corresponding to the second historical physical parameters, and determine multiple third training subsets; Step 1322: Train the third preset network model according to the multiple third training subsets to obtain the third network model.

[0103] For step 1321, the second historical electrical parameters and the second historical physical parameters are aligned according to the third historical acquisition time information corresponding to the second historical electrical parameters and the fourth historical acquisition time information corresponding to the second historical physical parameters. The historical health score of the corresponding time is queried. That is, the second historical health score and the time point corresponding to the second historical health score together constitute the third training subset.

[0104] For step 1322, the third preset network model has the same network structure and training steps as the second preset network model. That is, the features of the third training subset are extracted from the forget gate, input gate, output gate, etc., to obtain the predicted value. The loss function is calculated based on the predicted value and the true value. The bias vectors of the forget gate, input gate, and output gate are updated based on the loss function to obtain the trained third network model.

[0105] The dataset used to train the third preset network model is smaller than that used to train the second preset network model. Therefore, to avoid overfitting, the number of training iterations can be reduced when training the third preset network model. Simultaneously, the accuracy requirement for the loss function is lowered. If the bias vectors of the forget gate, input gate, and output gate are less than a preset range, the bias vectors or their corresponding values ​​are directly set to 0. Thus, the parameters of the third preset network model are simpler, resulting in a more lightweight third network model.

[0106] In some embodiments, in step 13, the preprocessed data packet is input into the photovoltaic equipment health detection model for photovoltaic equipment health detection processing to obtain the photovoltaic equipment health detection result, including: Step 133: The preprocessed data packet is input into the trained third network model to perform photovoltaic equipment health detection processing, and the health detection result of the photovoltaic equipment is obtained.

[0107] The collected electrical and physical parameters are input into the third network model to obtain the predicted equipment health score. Using the above example, assuming the current time is t, the parameters at time t are input into the third network model to obtain the predicted equipment health score at time t. , This refers to the health test results of photovoltaic equipment.

[0108] In some embodiments, the health warning threshold range includes a reminder-level threshold range, a warning-level threshold range, and an emergency-level threshold range; step 14, generating a health warning result for the photovoltaic equipment based on the health detection result and the health warning threshold range, includes: Step 141: When the health detection result meets the alert level threshold range, the warning information is displayed on the operation and maintenance interface; Step 142: When the health detection result meets the warning level threshold range, push the warning information to the target device; Step 143: When the health detection result is within the range of the emergency level threshold, display the warning information on the operation and maintenance interface, push the warning information to the target device, and generate a maintenance work order.

[0109] This embodiment can deploy an early warning response module on a cloud-edge collaborative platform. The early warning response module has a built-in start-up timestamp for each photovoltaic device, and determines the operation time (duration) based on the current timestamp and the start-up timestamp. The early warning response module can be linked to the site's SCADA system. When an emergency-level early warning is triggered, it automatically pushes a work order containing the device location (GPS positioning accuracy ±5m), fault type, and recommended maintenance plan.

[0110] In this embodiment, the warning threshold is dynamically adjusted based on the equipment's operating time. For example, the warning threshold for new photovoltaic equipment (operating for <1 year) is 60 points, while for older equipment (operating for >5 years), it is adjusted to 65 points (early warning). The threshold can be manually calibrated through the platform (within ±5 points). Multiple channels are used for push notifications: Reminder-level warnings (80-60 points) are only displayed in the maintenance app; warning-level warnings (60-40 points) are pushed to the responsible person's mobile phone via SMS (including equipment number and health status); emergency-level warnings (<40 points) are triggered by: ① SMS + app push (resend if unread within 5 minutes); ② generating a work order through the maintenance system (including GPS coordinates of the fault location and recommended spare parts model); ③ linking with the site's SCADA system (open API interface). The work order status (not dispatched → processing → resolved → verified) is synchronized to the platform in real time. After resolution, the equipment health status is automatically tracked within 72 hours (it needs to rise back to above 60 points to be considered valid).

[0111] The above embodiments expand the monitoring dimensions from a single electrical parameter to multiple dimensions of "electrical + physical + temporal", increasing the detection rate of latent faults to over 90%; edge-side preprocessing reduces data transmission volume by 70%, shortens response time from 30 minutes to 500 ms, and provides early warning of faults 3-7 days in advance; dynamic threshold + AI fusion model improves early warning accuracy to 92%, reduces false alarm rate to below 8%, and reduces ineffective maintenance work by 60%; full life cycle health records extend the average lifespan of equipment by 1-2 years, and reduce annual power loss of a single site by more than 1.5 million kWh.

[0112] In an optional embodiment of the present invention, the method further includes: Step 15: Save the parsed data by downsampling at the granularity of day, week, and month; Step 16: Create a file for the photovoltaic device based on the downsampling and saved data.

[0113] In this embodiment, data storage can adopt an "edge caching + cloud archiving" architecture: 7 days of data (1-minute granularity) are stored at the edge, and 3 years of data are stored in the cloud using an HBase time-series database (100,000 write throughput / s). Historical data is automatically downsampled at the "day → week → month" granularity (preserving trend characteristics).

[0114] As an example, data can be sent to the cloud-edge collaboration platform at a 1-minute granularity via an edge computing gateway, and the cloud-edge collaboration platform stores the data synchronously. Assuming that the temperature data of a photovoltaic module is collected at a rate of 1 record per minute, a total of 24 × 60 = 1440 data records will be generated in one day (24 hours) (granularity: 1 minute), as shown in Table 1.

[0115] Table 1 Example (Partial Data):

[0116] Among them, daily granular data (raw storage, cloud-edge collaboration platform retains 1-minute granularity): The cloud archives raw 1-minute granular data by "day" dimension as the basic data layer. Data within a single day is completely retained for short-term traceability (such as detailed analysis on the day a fault occurred). For example, 1440 1-minute granularity temperature data records for August 1, 2024, are completely stored in HBase.

[0117] Among them, weekly granular data (downsampling, aggregating from daily data): using "week" as the unit, downsampling of 1-minute granular data for 7 days, retaining key trend features of each day (such as peak, trough, mean, median, etc.), and increasing the granularity from 1 minute to "1 aggregated value per day".

[0118] Calculation logic: From 1440 data points each day, four core features are extracted (covering trends): daily highest value (e.g., peak temperature of the day), daily lowest value (e.g., trough temperature of the day), daily average value (e.g., average temperature of the day), and daily median (to reduce interference from extreme values). Refer to Table 2 (Week 1, August 2024, 7 days): Table 2 Weekly Granularity Data

[0119] Storage requirements: Only 7 data entries are needed per week (granularity: 1 day), compared to the original 1-minute granularity (7 × 1440 = 10080 entries), the compression rate is approximately 99.3%, while preserving the daily temperature change trend.

[0120] Among them, monthly granular data (further downsampling, aggregating from weekly data): taking "month" as the unit, the weekly granular data for 4 weeks (about 30 days) is downsampled again to retain the core trend features of each week, or directly aggregated into monthly features, with the granularity improved to "1 point per week" or "1 key value per month".

[0121] Calculation logic (taking monthly aggregation as an example): Monthly highest value (the highest temperature among all daily high temperatures in the month), monthly lowest value (the lowest temperature among all daily low temperatures in the month), monthly average temperature (the average of all daily average temperatures in the month), and monthly trend slope (reflecting the overall upward / downward trend of temperature through linear fitting, such as the gradual increase in temperature during summer). Refer to Table 3 (August 2024, 4 weeks in total): Table 3 Monthly Granularity Data

[0122] The storage requirement is only 1 data record per month (or 4 weekly aggregate data records), which is about 96.7% higher than the weekly granularity (30 records), and retains the overall trend of monthly temperature (such as whether it is abnormally high temperature, whether it is on the rise, etc.).

[0123] The method of automatically downsampling and saving the parsed data at daily, weekly, and monthly granularities offers the following advantages: 1. It preserves trend characteristics: By extracting extreme values ​​(highest / lowest), mean, median, and trend slope, it replaces the original high-frequency data, ensuring that the "trend of change" in the data is not lost. 2. Granularity is progressively increased: From 1 minute to 1 day, and then to 1 week / 1 month, it reduces storage requirements while meeting the needs of different scenarios (such as short-term detailed analysis and long-term trend analysis). 3. It adapts to business scenarios: In photovoltaic equipment operation and maintenance, short-term fault analysis requires high-frequency data, while equipment aging trend assessment and annual energy efficiency analysis only require long-term trends; the downsampling data can fully support these needs.

[0124] For step 16, create a unique file for each photovoltaic device, including: ① basic information (model, installation date, supplier); ② operation and maintenance records (maintenance time, replaced parts); ③ health curve (trend over the past 1 month / 3 months / 1 year). Multi-dimensional queries are supported, including by device ID and location region.

[0125] In one application scenario, refer to Figure 2 A health monitoring system for photovoltaic equipment may include: a monitoring terminal cluster, an edge computing gateway, a cloud-edge collaboration platform, and an early warning response module; The monitoring terminal cluster is connected to each component of the photovoltaic equipment and is used to collect the electrical parameters, physical parameters and corresponding collection timestamps of the photovoltaic equipment; the electrical parameters include the current of the photovoltaic string and the physical parameters include at least the surface temperature field distribution data and mechanical vibration frequency of the photovoltaic module. The edge computing gateway is communicatively connected to the monitoring terminal cluster. The edge computing gateway is used to receive the electrical parameters, physical parameters, and timestamps, and to compress the electrical parameters, physical parameters, and timestamps to generate a first data packet. The cloud-edge collaboration platform is communicatively connected to the edge computing gateway. The cloud-edge collaboration platform is used to receive and parse the first data packet, extract feature data from the parsed data, input the feature data into a pre-trained first health assessment regression model, output a first health score, and send the first health score and the parsed timestamp to the early warning response module. The health assessment regression model is trained based on historical fault data. The early warning response module is communicatively connected to the cloud-edge collaborative platform. The early warning response module is used to receive the first health score and the timestamp, and based on a dynamic threshold mapping mechanism, determine multiple threshold ranges with different levels according to the timestamp, match the first health score with the multiple threshold ranges, and generate and send early warning information based on the matching results.

[0126] The system comprises a monitoring terminal cluster, an edge computing gateway, a cloud-edge collaboration platform, and an early warning and response module. The monitoring terminal cluster is equipped with multiple types of sensors, capable of collecting electrical parameters, temperature, vibration, and other status data of the core equipment of the photovoltaic system, with a sampling frequency of 1-10kHz. The edge computing gateway performs real-time preprocessing of the data, completing a preliminary health assessment through lightweight algorithms, achieving a data compression rate of ≥70%. The cloud-edge collaboration platform integrates big data analytics and AI models to construct a full lifecycle health record for the equipment, with a health assessment accuracy of ≥92%. The early warning and response module triggers multi-level warnings based on the health level and links with the operation and maintenance system to generate work orders. The device achieves real-time perception of equipment status, dynamic health assessment, and accurate early warning, reducing fault detection time to 3-7 days compared to traditional methods, lowering site operation and maintenance costs by 20%-30%, and is suitable for photovoltaic smart sites of all sizes.

[0127] The installation and deployment of monitoring terminal clusters, edge computing gateways, cloud-edge collaboration platforms, and early warning response modules can be referenced as follows: Monitoring terminal cluster installation: Component terminals should be magnetically attached to the center of the back panel (avoiding the junction box), ensuring the vibration sensor is in close contact with the metal frame; inverter terminals should be installed inside the cabinet (≥30cm from ventilation holes to avoid temperature interference). Configure the terminal ID, sampling frequency (default 10Hz), and communication parameters using a handheld debugger (supports Bluetooth connection).

[0128] Edge computing gateway deployment: Configure one gateway for every 50 terminals (maximum capacity 80 terminals), install it in the site control room (ambient temperature 15-30℃), and connect it to the core switch via fiber optic cable (ensuring bandwidth ≥100Mbps). Upon initial startup, the gateway automatically scans for terminals (batch addition is supported) and generates a device topology diagram.

[0129] Cloud-edge collaboration platform configuration / early warning response module deployment: The cloud platform is accessed through a browser. The administrator creates a site account (assigns regional permissions), imports the device ledger (batch import from Excel template), and completes the basic configuration (such as device type and initial value of early warning threshold).

[0130] Based on the above, the operational flow of the monitoring terminal cluster, edge computing gateway, cloud-edge collaboration platform, and early warning response module can be referenced as follows: Data Acquisition: The terminal collects data (current, temperature, etc.) at a frequency of 10Hz, and uploads it to the gateway (LoRa wireless or Ethernet wired) every 10 seconds. In abnormal conditions (such as excessive vibration), high-frequency sampling is automatically triggered (temporarily boosted to 100Hz for 30 seconds).

[0131] Edge analytics: The gateway aggregates terminal data every 1 minute, performing cleaning and health assessment. If the health of an inverter drops to 55 points (early warning level), the data is immediately cached and uploaded to the cloud; if it drops to 35 points (emergency level), a local audible and visual alarm is triggered simultaneously.

[0132] Cloud-based processing: After receiving the data, the platform's AI engine performs a secondary assessment (combining historical trends) and generates a health report. If an emergency alert is confirmed, an SMS message (including GPS location information) is sent to the operations team leader within 5 seconds.

[0133] Operation and maintenance closed loop: Maintenance personnel receive work orders via the APP, navigate to the equipment location (error < 5m), process the work, and upload on-site photos (automatically linked to the archive). The platform rechecks the health status after 72 hours (must be ≥ 60 points), otherwise a secondary warning is triggered.

[0134] In another application scenario, after deploying a monitoring terminal cluster, edge computing gateway, cloud-edge collaboration platform, and early warning response module in a 200MW smart photovoltaic power station: Inverter capacitor bulging fault: The vibration sensor detects a characteristic frequency of 1.2kHz (normal <0.5kHz) and provides a warning 5 days in advance to avoid downtime (a single inverter downtime results in a power loss of 800kWh). Hot spots on components: Infrared temperature array detected local temperatures 15°C higher than the surrounding area (normal ≤5°C), locating 23 components with microcracks (traditional inspection takes 1 week, this device detects them in 2 hours). Early warning accuracy: During the 3-month trial operation, there were 127 effective early warnings and 10 false alarms (false alarm rate of 7.8%), which is significantly lower than the traditional method (35%).

[0135] Terminal calibration: Calibrate the sensor every 6 months using a standard signal source (accuracy class 0.05) (e.g., if a 10A standard signal is injected, the terminal display deviation should be <0.05A). Model updates: The lightweight model of the edge computing gateway is upgraded quarterly via OTA (retaining local historical parameters); the first health assessment regression model deployed on the cloud-edge collaboration platform is iterated monthly with new fault data (automatically retaining the optimal version); Data backup: The cloud-edge collaboration platform automatically backs up data daily (off-site disaster recovery) and supports point-in-time recovery (accurate to the hour for the most recent 7 days).

[0136] It enables real-time acquisition of multiple parameters, rapid edge analysis, dynamic health assessment, and precise early warning for the health management of photovoltaic smart power station equipment.

[0137] The present invention also provides a schematic diagram of an embodiment of a health detection and early warning system for photovoltaic equipment, the system comprising: A health prediction device is used to receive pre-processed data packets from photovoltaic (PV) devices sent by an edge computing gateway. These pre-processed data packets are obtained by the edge computing gateway through compression processing of electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one PV device. The pre-processed data packets are then input into a PV device health detection model for PV device health detection processing to obtain a PV device health detection result. The PV device health detection model is trained based on at least one of the electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one PV device. Based on the health detection result and a health warning threshold range, a PV device health warning result is generated and output.

[0138] Optionally, the electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one photovoltaic device are compressed to obtain a preprocessed data package of the photovoltaic device, including: The electrical parameters of at least one photovoltaic device are preprocessed to obtain first preprocessed data; The physical parameters of at least one photovoltaic device are preprocessed to obtain second preprocessed data; The first acquisition time information corresponding to the electrical parameters is preprocessed to obtain the third preprocessed data; The second acquisition time information corresponding to the physical parameters is preprocessed to obtain the fourth preprocessed data; The first, second, third, and fourth preprocessed data are compressed to obtain a preprocessed data package for the photovoltaic equipment.

[0139] Optionally, the training process of the photovoltaic equipment health detection model includes: The cloud-edge collaborative platform trains a first preset network model based on the first historical electrical parameters and first historical physical parameters of at least one photovoltaic device to obtain a first network model; it then trains a second preset network model based on the first historical electrical parameters, first historical physical parameters, first historical acquisition time information corresponding to the first historical electrical parameters, and second historical acquisition time information corresponding to the first historical physical parameters of at least one photovoltaic device to obtain a second network model; finally, based on the first network model and the second network model, it obtains the photovoltaic device health detection model; or The edge computing gateway trains a third preset network model based on the second historical electrical parameters, the second historical physical parameters, the third historical acquisition time information corresponding to the second historical electrical parameters, and the fourth acquisition time information corresponding to the second historical physical parameters of at least one photovoltaic device, to obtain the photovoltaic device health detection model; the duration of the third historical acquisition time is less than the duration of the first historical acquisition time, and the duration of the fourth historical acquisition time is less than the duration of the second historical acquisition time.

[0140] Optionally, a first preset network model is trained based on first historical electrical parameters and first historical physical parameters of at least one photovoltaic device to obtain a first network model, including: Multiple first training subsets are determined based on the first historical electrical parameters and first historical physical parameters of the at least one photovoltaic device; Multiple first classifiers in the first preset network model are trained based on the multiple first training subsets to obtain multiple second classifiers, and the multiple second classifiers constitute the first network model.

[0141] Optionally, a second preset network model is trained based on at least one photovoltaic device's first historical electrical parameters, first historical physical parameters, first historical acquisition time information corresponding to the first historical electrical parameters, and second historical acquisition time information corresponding to the first historical physical parameters, to obtain a second network model, including: Based on the first historical acquisition time information corresponding to the first historical electrical parameter and the second historical acquisition time information corresponding to the first historical physical parameter, the first historical electrical parameter and the first historical physical parameter are aligned to determine multiple second training subsets; The second preset network model is trained based on the multiple second training subsets to obtain the second network model.

[0142] Optionally, the photovoltaic equipment health detection model is obtained based on the first network model and the second network model, including: The first dynamic weight and the second dynamic weight are determined according to the preset mapping table. The photovoltaic equipment health detection model is obtained based on the first dynamic weight, the second dynamic weight, the first network model, and the second network model.

[0143] Optionally, the preprocessed data packet is input into the photovoltaic equipment health detection model for photovoltaic equipment health detection processing to obtain the photovoltaic equipment health detection results, including: The preprocessed data packet is input into the first network model to obtain the output result of the first network model; The preprocessed data packet is input into the second network model to obtain the output result of the second network model; The target value is determined based on the output of the first network model and the output of the second network model. The first dynamic weight and the second dynamic weight are determined based on the target value; Based on the output results of the first network model, the output results of the second network model, the first dynamic weight, and the second dynamic weight, the health detection results of the photovoltaic equipment are obtained.

[0144] Optionally, based on the second historical electrical parameters, second historical physical parameters, third historical acquisition time information corresponding to the second historical electrical parameters, and fourth acquisition time information corresponding to the second historical physical parameters of at least one photovoltaic device, a third preset network model is trained to obtain the photovoltaic device health detection model, including: The second historical electrical parameters and the second historical physical parameters are aligned based on the third historical acquisition time information corresponding to the second historical electrical parameters and the fourth historical acquisition time information corresponding to the second historical physical parameters to determine multiple third training subsets. The third preset network model is trained based on the multiple third training subsets to obtain the third network model.

[0145] Optionally, the preprocessed data packet is input into the photovoltaic equipment health detection model for photovoltaic equipment health detection processing to obtain the photovoltaic equipment health detection results, including: The preprocessed data packet is input into the third network model for photovoltaic equipment health detection processing to obtain the photovoltaic equipment health detection results.

[0146] It should be noted that this device is the same as the method described above. All implementations in the above method embodiments are applicable to the embodiments of this device and can achieve the same technical effect.

[0147] The device embodiments described above are merely illustrative. The modules described as separate components may or may not be physically separate, and the components shown as modules may or may not be physical modules; that is, they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0148] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of various embodiments or some parts of embodiments.

[0149] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A method for detecting and warning the health status of photovoltaic equipment, characterized in that, include: Acquire electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one photovoltaic device; The electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one photovoltaic device are compressed to obtain a preprocessed data package of the photovoltaic device. The preprocessed data package is input into the photovoltaic equipment health detection model for photovoltaic equipment health detection processing to obtain the photovoltaic equipment health detection result; the photovoltaic equipment health detection model is trained based on at least one of the following: historical electrical parameters, historical physical parameters, first historical acquisition time information corresponding to the historical electrical parameters, and second historical acquisition time information corresponding to the historical physical parameters of the photovoltaic equipment. Based on the health status detection results and the health status warning threshold range, a health status warning result for the photovoltaic equipment is generated and output.

2. The method for health detection and early warning of photovoltaic equipment according to claim 1, characterized in that, The electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one photovoltaic device are compressed to obtain a preprocessed data package for the photovoltaic device, including: The electrical parameters of at least one photovoltaic device are preprocessed to obtain first preprocessed data; The physical parameters of at least one photovoltaic device are preprocessed to obtain second preprocessed data; The first acquisition time information corresponding to the electrical parameters is preprocessed to obtain the third preprocessed data; The second acquisition time information corresponding to the physical parameters is preprocessed to obtain the fourth preprocessed data; The first, second, third, and fourth preprocessed data are compressed to obtain a preprocessed data package for the photovoltaic equipment.

3. The method for health detection and early warning of photovoltaic equipment according to claim 1, characterized in that, The training process of the photovoltaic equipment health detection model includes: The cloud-edge collaborative platform trains a first preset network model based on the first historical electrical parameters and first historical physical parameters of at least one photovoltaic device to obtain a first network model; it then trains a second preset network model based on the first historical electrical parameters, first historical physical parameters, first historical acquisition time information corresponding to the first historical electrical parameters, and second historical acquisition time information corresponding to the first historical physical parameters of at least one photovoltaic device to obtain a second network model; finally, based on the first network model and the second network model, it obtains the photovoltaic device health detection model; or The edge computing gateway trains a third preset network model based on the second historical electrical parameters, the second historical physical parameters, the third historical acquisition time information corresponding to the second historical electrical parameters, and the fourth acquisition time information corresponding to the second historical physical parameters of at least one photovoltaic device, to obtain the photovoltaic device health detection model; the duration of the third historical acquisition time is less than the duration of the first historical acquisition time, and the duration of the fourth historical acquisition time is less than the duration of the second historical acquisition time.

4. The method for health detection and early warning of photovoltaic equipment according to claim 3, characterized in that, Based on the first historical electrical parameters and first historical physical parameters of at least one photovoltaic device, a first preset network model is trained to obtain a first network model, including: Multiple first training subsets are determined based on the first historical electrical parameters and first historical physical parameters of the at least one photovoltaic device; Multiple first classifiers in the first preset network model are trained based on the multiple first training subsets to obtain multiple second classifiers, and the multiple second classifiers constitute the first network model.

5. The method for health detection and early warning of photovoltaic equipment according to claim 3, characterized in that, Based on the first historical electrical parameters, first historical physical parameters, first historical acquisition time information corresponding to the first historical electrical parameters, and second historical acquisition time information corresponding to the first historical physical parameters of at least one photovoltaic device, a second preset network model is trained to obtain a second network model, including: Based on the first historical acquisition time information corresponding to the first historical electrical parameter and the second historical acquisition time information corresponding to the first historical physical parameter, the first historical electrical parameter and the first historical physical parameter are aligned to determine multiple second training subsets; The second preset network model is trained based on the multiple second training subsets to obtain the second network model.

6. The method for health detection and early warning of photovoltaic equipment according to claim 5, characterized in that, Based on the first network model and the second network model, the photovoltaic equipment health detection model is obtained, including: The first dynamic weight and the second dynamic weight are determined according to the preset mapping table. The photovoltaic equipment health detection model is obtained based on the first dynamic weight, the second dynamic weight, the first network model, and the second network model.

7. The method for health detection and early warning of photovoltaic equipment according to claim 6, characterized in that, The preprocessed data packet is input into the photovoltaic equipment health detection model for photovoltaic equipment health detection processing to obtain the photovoltaic equipment health detection results, including: The preprocessed data packet is input into the first network model to obtain the output result of the first network model; The preprocessed data packet is input into the second network model to obtain the output result of the second network model; The target value is determined based on the output of the first network model and the output of the second network model. The first dynamic weight and the second dynamic weight are determined based on the target value; Based on the output results of the first network model, the output results of the second network model, the first dynamic weight, and the second dynamic weight, the health detection results of the photovoltaic equipment are obtained.

8. The method for health detection and early warning of photovoltaic equipment according to claim 3, characterized in that, Based on the second historical electrical parameters, second historical physical parameters, third historical acquisition time information corresponding to the second historical electrical parameters, and fourth acquisition time information corresponding to the second historical physical parameters of at least one photovoltaic device, a third preset network model is trained to obtain the photovoltaic device health detection model, including: The second historical electrical parameters and the second historical physical parameters are aligned based on the third historical acquisition time information corresponding to the second historical electrical parameters and the fourth historical acquisition time information corresponding to the second historical physical parameters to determine multiple third training subsets. The third preset network model is trained based on the multiple third training subsets to obtain the third network model.

9. The method for health detection and early warning of photovoltaic equipment according to claim 8, characterized in that, The preprocessed data packet is input into the photovoltaic equipment health detection model for photovoltaic equipment health detection processing to obtain the photovoltaic equipment health detection results, including: The preprocessed data packet is input into the third network model for photovoltaic equipment health detection processing to obtain the photovoltaic equipment health detection results.

10. A health detection and early warning system for photovoltaic equipment, characterized in that, include: A health prediction device receives a pre-processed data packet from a photovoltaic (PV) device sent by an edge computing gateway. The pre-processed data packet is obtained by the edge computing gateway compressing electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one PV device. The pre-processed data packet is then input into a PV device health detection model for PV device health detection processing to obtain the PV device health detection result. The PV device health detection model is trained based on at least one of the electrical parameters, physical parameters, first acquisition time information corresponding to the electrical parameters, and second acquisition time information corresponding to the physical parameters of at least one PV device. Based on the health status detection results and the health status warning threshold range, a health status warning result for the photovoltaic equipment is generated and output.

Images