Artificial intelligence-based high-altitude photovoltaic power station intelligent operation and maintenance and fault prediction method

Abstract

The application discloses a high-altitude photovoltaic power station intelligent operation and maintenance and fault prediction method based on artificial intelligence, belongs to the technical field of data processing, and comprises the following steps: step one, estimating the real illumination value of each component at any time; step two, constructing the theoretical power output of the component on the basis of combining temperature aging and installation altitude and other information; step three, determining whether an abnormal trend exists by comparing the deviation trend of the theoretical power and the actual power in a time window and combining the light fluctuation condition to generate an event trigger signal; step four, collecting key data to generate structured features after the event is triggered, and completing preliminary diagnosis based on rules; and step five, outputting the final diagnosis result by fusing the edge feature preliminary judgment result and the power deviation information on the center side. The scheme combines physical modeling and rule judgment, and constructs an intelligent diagnosis scheme suitable for the plateau power station environment.

Description

Technical Field

[0001] This invention belongs to the field of data processing technology, and in particular relates to an intelligent operation and maintenance and fault prediction method for high-altitude photovoltaic power plants based on artificial intelligence. Background Technology

[0002] High-altitude regions possess natural advantages for developing photovoltaic power generation, characterized by long hours of sunshine, strong radiation, and good air quality. Therefore, they are widely used in the construction of large-scale photovoltaic power plants in western my country. However, these regions also face technical challenges posed by extreme climates and operating environments. These include uneven sunlight due to topographical variations, illuminance disturbances caused by mountain shading and snow reflection, thermal stress changes due to large diurnal temperature differences, and reduced heat dissipation capacity due to thin air. These factors can lead to frequent and complex power fluctuations in photovoltaic modules during operation.

[0003] Currently, most mainstream intelligent operation and maintenance methods in the industry are based on centrally collected meteorological data and unified modeling methods. These methods cannot accurately reflect the differences in the environment of each component, often misjudging power changes caused by environmental disturbances as equipment failures, or failing to identify real faults in a timely manner. In addition, in remote plateau areas, communication network coverage is unstable, and it is difficult for power plant operation and maintenance personnel to enter and exit. Some components are left unmonitored for a long time. Existing artificial intelligence methods that rely on continuous communication and large-scale training data are difficult to deploy in such environments, resulting in poor diagnostic effects and delayed fault handling, which seriously affects the operational safety and power generation efficiency of power plants.

[0004] To address these issues, we propose an AI-based intelligent operation and maintenance and fault prediction method for high-altitude photovoltaic power plants. Summary of the Invention

[0005] The purpose of this invention is to solve the problems of poor diagnostic results and delayed fault handling in the prior art, and to propose an intelligent operation and maintenance and fault prediction method for high-altitude photovoltaic power plants based on artificial intelligence.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] AI-based intelligent operation and maintenance and fault prediction methods for high-altitude photovoltaic power plants include:

[0008] S1: Collect the first raw data, which includes digital elevation model data of the area where the photovoltaic power station is located, tilt angle and orientation information of the components, and total all-sky radiation data at the site level;

[0009] The first raw data is input into the illuminance estimation model; the illuminance estimation model includes an occlusion determination module and an illuminance calculation module;

[0010] The occlusion determination module determines whether each component is subject to terrain occlusion by calculating the relationship between the solar incidence angle and the terrain model.

[0011] The illuminance calculation module calculates the actual illuminance value that each component can receive within a preset time window;

[0012] The final output component is the illuminance estimate within a preset time window;

[0013] S2: Construct a power estimation model, which includes ideal power baseline estimation and aging and plateau micro-topography correction;

[0014] The ideal power baseline estimation is obtained by calculating the initial conversion efficiency of the module, the temperature loss coefficient, the temperature reference point, the current module temperature, and the estimated illuminance value.

[0015] The aging and plateau micro-topography correction is obtained by calculating the correction factor based on the component's historical aging factor, the local elevation of the component's installation point relative to the site, and the unified reference altitude of the site.

[0016] The theoretical power output is obtained by combining the ideal power baseline and the correction factor.

[0017] S3: Acquire actual power output; construct a deviation sequence, which is a set of deviation values ​​between the theoretical power output and the actual power output per unit time;

[0018] The illuminance fluctuation penalty term is designed, which is the standard deviation of the illuminance estimate of the component in each time window;

[0019] The deviation sequence is combined with the illuminance fluctuation penalty term to form a trend score;

[0020] When the trend score exceeds the preset threshold and the trigger signal is not activated, activate the trigger signal and proceed to step S4; otherwise, do not respond.

[0021] S4: Acquire the second raw data within the time window, which includes voltage, current, and temperature;

[0022] The structural feature vector is obtained by fusing voltage fluctuation amplitude, current fluctuation amplitude, temperature change standard deviation and relative power deviation.

[0023] The system uses preset rules to classify and determine the initial diagnostic labels.

[0024] S5: Standardize the structural feature vectors to construct diagnostic modeling vectors;

[0025] A fault discrimination model is constructed. The fault discrimination model is a shallow neural structure with input weight regularization, which includes a feedforward hidden layer with a width of 8 and an output layer. Softmax is used for multi-class classification output.

[0026] The diagnostic modeling vector is input into the fault discrimination model to output the probability scores of various faults; the one with the highest score is selected from all fault categories as the final fault judgment result.

[0027] Preferably, the calculation steps of the occlusion determination module in step S1 are as follows:

[0028] Calculate the current solar altitude angle and solar azimuth angle based on the component's location and time;

[0029] Project a straight line along the direction of sunlight at each component location to determine if there is terrain above the component's installation height that blocks the light; if so, set the shading mask value of that component to 1 at that time point, otherwise set it to 0.

[0030] Preferably, the calculation steps of the illuminance calculation module in step S1 are as follows:

[0031] The total radiation at the current time of a known site is used as the original benchmark for the total radiation energy received by the site per unit time and per unit area;

[0032] The original baseline is corrected by the occlusion mask value. When the occlusion mask value is the highest, the illumination correction coefficient is the lowest. The occlusion mask value varies between 0 and 1. The occlusion mask value is obtained from terrain modeling and the direction of solar incidence.

[0033] The illuminance estimate is obtained by adjusting the angle between the component's tilt angle and the current solar altitude angle.

[0034] Preferably, different weights are assigned to the sample deviation values ​​in the deviation sequence, and the weight decay rule adopts a decreasing coefficient, with the weight of the sample closer to the current time being higher.

[0035] Preferably, the preset rules in step S4 are as follows:

[0036] When voltage fluctuation exceeds the voltage fluctuation threshold, current fluctuation exceeds the current fluctuation threshold, and the temperature standard deviation is less than the first temperature difference threshold, while the relative power deviation is greater than the first power deviation threshold, the output is a voltage anomaly.

[0037] When the temperature standard deviation is greater than the second temperature difference threshold and the relative power deviation exceeds the second power deviation threshold, while the voltage and current fluctuations are less than the voltage fluctuation threshold and the current fluctuation threshold, respectively, the output is thermal runaway.

[0038] If all electrical indicators are less than the above thresholds, and the edge-side illumination disturbance exceeds the shading threshold, while the relative power deviation is between the first power deviation threshold and the second power deviation threshold, the output is shading.

[0039] If none of the above conditions are met, the output will be unknown.

[0040] Preferably, the preset rules also include:

[0041] When the absolute value of the power deviation is lower than the first power deviation threshold, and the voltage and current fluctuations are lower than the voltage fluctuation threshold and the current fluctuation threshold respectively, and the temperature standard deviation is lower than the first temperature difference threshold, the output is unknown.

[0042] Preferably, a low-amplitude deviation suppression term is introduced into the fault discrimination model, which applies a negative penalty when the actual deviation is lower than a threshold and is judged as a fault class.

[0043] An electronic device is characterized by comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the method described above.

[0044] In summary, the technical effects and advantages of this invention are as follows: This solution calculates the true illuminance value of each component at any given time using a digital elevation model and component installation parameters, avoiding the risk of misjudgment caused by the traditional uniform irradiance assumption; it constructs a power deviation trend analysis method based on a time window, introduces an illuminance fluctuation suppression mechanism, and triggers diagnostic events only when a continuous anomaly exists and cannot be explained by environmental changes; it generates preliminary fault types through rule-based judgment, reducing the edge computing burden and improving response time. Attached Figure Description

[0045] Figure 1 This is a flowchart of the steps in this invention. Detailed Implementation

[0046] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0047] like Figure 1 As shown, the intelligent operation and maintenance and fault prediction method for high-altitude photovoltaic power plants based on artificial intelligence includes:

[0048] S1: Collect the first raw data, which includes digital elevation model data of the area where the photovoltaic power station is located, tilt angle and orientation information of the components, and total all-sky radiation data at the site level;

[0049] The first raw data is input into the illuminance estimation model; the illuminance estimation model includes an occlusion determination module and an illuminance calculation module;

[0050] The occlusion determination module determines whether each component is subject to terrain occlusion by calculating the relationship between the solar incidence angle and the terrain model.

[0051] The illuminance calculation module calculates the actual illuminance value that each component can receive within a preset time window;

[0052] The final output component is the illuminance estimate within a preset time window;

[0053] S2: Construct a power estimation model, which includes ideal power baseline estimation and aging and plateau micro-topography correction;

[0054] The ideal power baseline estimation is obtained by calculating the initial conversion efficiency of the module, the temperature loss coefficient, the temperature reference point, the current module temperature, and the estimated illuminance.

[0055] The aging and plateau micro-topography correction is obtained by calculating the correction factor based on the component's historical aging factor, the local elevation of the component's installation point relative to the site, and the unified reference altitude of the site.

[0056] The theoretical power output is obtained by combining the ideal power baseline and the correction factor.

[0057] S3: Acquire actual power output; construct a deviation sequence, which is a set of deviation values ​​between the theoretical power output and the actual power output per unit time;

[0058] The illuminance fluctuation penalty term is designed, which is the standard deviation of the illuminance estimate of the component in each time window;

[0059] The deviation sequence is combined with the illuminance fluctuation penalty term to form a trend score;

[0060] When the trend score exceeds the preset threshold and the trigger signal is not activated, activate the trigger signal and proceed to step S4; otherwise, do not respond.

[0061] S4: Acquire the second raw data within the time window, which includes voltage, current, and temperature;

[0062] The structural feature vector is obtained by fusing voltage fluctuation amplitude, current fluctuation amplitude, temperature change standard deviation and relative power deviation.

[0063] The system uses preset rules to classify and determine the initial diagnostic labels.

[0064] S5: Standardize the structural feature vectors to construct diagnostic modeling vectors;

[0065] A fault discrimination model is constructed. The fault discrimination model is a shallow neural structure with input weight regularization, which includes a feedforward hidden layer with a width of 8 and an output layer. Softmax is used for multi-class classification output.

[0066] The diagnostic modeling vector is input into the fault discrimination model to output the probability scores of various faults; the one with the highest score is selected from all fault categories as the final fault judgment result.

[0067] The specific steps are as follows:

[0068] Step 1: Component Illuminance Estimation Model

[0069] This step aims to provide the subsequent fault prediction and intelligent operation and maintenance modules with an estimate of the actual irradiance received by each photovoltaic module at any time t. Unlike traditional methods that only use site-level total radiation data, this approach considers the typical illuminance unevenness in high-altitude environments, particularly the local illuminance differences caused by factors such as terrain obstruction, inconsistent slope aspect, and large variations in solar altitude angle. It constructs a component-level illuminance inversion method based on available parameters. Its core principle is to achieve spatial illuminance estimation that most closely approximates actual working conditions using minimal data sources, ensuring the model's deployability, stability, and adaptability to high-altitude scenarios.

[0070] The input for this step consists of three parts:

[0071] H(x,y) represents the digital elevation model (DEM) data of the area where the photovoltaic power station is located. A resolution of 1–5 meters is recommended. This data can be obtained by importing publicly available remote sensing imagery or from collaborating surveying and mapping units. For example, the 1-second resolution DEM data provided by the National Geomatics Center of China can meet the requirements of most medium-sized power stations. After importing, the site area is sliced ​​using bilinear interpolation to generate a matrix-form H(x,y) tensor.

[0072] θ(i) represents the tilt angle and orientation information of the i-th component, which is directly derived from the photovoltaic layout diagram. Typically, the tilt angles of each row of components are consistent, and mapping can be performed according to the layout index. If the construction unit provides BIM (Building Information Modeling) or design software (such as PVsyst) output files, this information can be directly exported; otherwise, a mapping table can be manually created using RTK and drawing numbers.

[0073] R(t) is the site-level total sky radiation data, with a recommended sampling frequency of no less than 10 minutes, provided by the site's existing micro-weather station. For example, mainstream domestic photovoltaic operation and maintenance systems generally deploy domestically produced radiometers (such as models CM21 and MS-802), which connect the radiation data to the main control system via a Modbus interface and provide a standard timestamp sequence R(t) through edge computing nodes for caching.

[0074] The illuminance estimation model consists of two parts: an occlusion determination module and an illuminance calculation module.

[0075] The first part is the occlusion detection module. It determines whether each component is under terrain occlusion by calculating the relationship between the solar incidence angle and the terrain model. First, based on the component position (derived from the layout diagram) and time t, it calculates the current solar altitude angle α. s (t) and solar azimuth angle φ s (t). These two angles can be calculated using a solar trajectory model, such as NREL's Solar Position Algorithm (SPA), which is used in most engineering systems. Then, simplified ray tracing is performed on the terrain matrix represented by H(x,y), i.e., a straight line is projected along the direction of sunlight incidence at each component location to determine if there is terrain above the component's installation height blocking the light. If so, the occlusion mask value for that component at that time point is set to 1; otherwise, it is 0. To avoid the binarization problem between "full occlusion" and "no occlusion," we use the angular proportion of the occluded area as the continuous value occlusion coefficient M. block (i,t), with a range of [0,1]. The occlusion ratio is calculated using the view domain analysis method, which estimates the effective blocking degree of the optical path by using the terrain height difference in the 8-neighbor or 16-neighbor directions.

[0076] The second part is the illuminance calculation module. This is based on the known total irradiance R(t) at the current time at the site, the tilt angle parameter θ(i) of the component, and the shading mask value M. block Given (i,t), the system will calculate the actual illuminance value that each component i can receive at time t. The extrapolation is performed. The extrapolation process follows the following physical logic: First, R(t) is used as the original baseline for the total radiant energy received by the site per unit time per unit area; then, considering whether the component is affected by terrain shading at that moment, the shading mask value M is used. block The illuminance is corrected using (i,t). The higher the shading ratio, the lower the illuminance correction factor. This mask value varies between 0 and 1 and is derived from terrain modeling and solar incidence direction analysis. Finally, the component's tilt angle θ(i) is used relative to the current solar altitude angle α. s The angle relationship between (t) is used to further adjust the illuminance estimation results. The tilt correction is usually calculated by obtaining the correction factor c through a table lookup method. θ(i) This factor represents the proportion of irradiance that the tilted component surface can effectively receive compared to a horizontal surface under the current tilt angle conditions. In practical applications, this value can also be determined by a simple approximation, such as using the cosine of the tilt angle and the solar altitude angle to ensure that the irradiated projected area is maximized when the component faces the sun. When the solar altitude angle is significantly lower than the component tilt angle, the correction factor approaches zero, reflecting that the effective radiation is close to ineffective at this time. Therefore, the system ultimately determines the effective radiation based on these three factors—the site's original total radiation R(t), the terrain shading mask M, and the total irradiance M. block (i,t) and component tilt correction factor c θ (i) — Calculate the effective illuminance actually received by each component at that moment. It is worth emphasizing that all correction factors are dimensionless proportional terms, and R(t) has been normalized beforehand to ensure the consistency of calculations within the model and the comparability of results.

[0077] In terms of deployment, M block (i,t) only needs to be updated once per quarter, c θ (i) Initialization can be performed uniformly during the initial website setup. During operation, only real-time input of R(t) is required for output. It has extremely low resource requirements and is suitable for deployment on edge controllers or central master stations.

[0078] Output: The illuminance estimate of component i at time t, in units of normalized effective irradiance, is used as the power estimation input for subsequent steps.

[0079] Step 2: Component power behavior estimation model:

[0080] The goal of this step is to obtain the estimated true illuminance of the components. Based on (step one), construct a system that can accurately output the theoretically expected power of each photovoltaic module at time t. The power estimation model is used as a benchmark for subsequent anomaly detection, trend analysis, and diagnostic judgment. Considering that the patent is applicable to photovoltaic power station scenarios in high-altitude areas, this model fully considers the special characteristics of the plateau operating environment during its design, such as the decrease in heat dissipation efficiency caused by thin air, the fluctuation of component performance caused by the extreme temperature difference between day and night, and the systematic imbalance caused by the micro-topographic differences within the same power station.

[0081] The calculation of the power estimation model is divided into two parts. The first part is the ideal power baseline estimation, which takes into account the effects of illuminance and temperature.

[0082]

[0083] in:

[0084] η0 represents the initial conversion efficiency of the component (e.g., 0.18), derived from the component's technical specifications.

[0085] β is the temperature loss coefficient, with a typical value of 0.004, which means that the power decreases by 0.4% for every 1°C increase;

[0086] T ref The temperature reference point is uniformly set to 25℃;

[0087] T(i,t) represents the current component temperature;

[0088] The actual illuminance of the component is calculated in step one.

[0089] This part of the output This is the ideal power reference that the module should have under ideal conditions, without taking into account the effects of aging and environmental differences.

[0090] The second part is aging and plateau micro-topography correction, used to further construct the final output. The complete power estimation expression is as follows:

[0091]

[0092] in:

[0093] α(i) represents the component's historical aging factor, ranging from [0, 0.3], obtained from historical operating data;

[0094] H alt (i) indicates the local elevation (in meters) of the component installation point relative to the site;

[0095] H ref A unified reference altitude (unit: meters) is provided for the entire site for normalization purposes;

[0096] γ is the micro-topography correction factor, with a typical value of 0.02 to 0.05;

[0097] δ is the adjustment index, which controls the intensity of nonlinear effects. A value of 2 to 3 is recommended.

[0098] The plateau correction factor This step is a unique innovation designed for patent scenarios. It utilizes the existing DEM data of the power plant for local adjustments, explicitly considering the comprehensive impact of plateau micro-topography on heat dissipation efficiency and component efficiency. This mechanism can explain the systematic power offset problem caused by different areas of the same site due to varying terrain, and has engineering value in sites with complex terrain distributions such as Tibet and Qinghai.

[0099] Output:

[0100] The theoretical power output of component i at time t serves as a reference benchmark for subsequent fault deviation analysis;

[0101] The calculation process for the entire step is as follows: First, calculate and T(i,t) to obtain Then, use α(i) and H alt (i) to correct it and output the final M_power$.

[0102] Step 3: Trend deviation analysis and event triggering mechanism:

[0103] The goal of this step is to establish a deviation analysis mechanism at the time series level between the actual power output P_actual(i,t) of the component and the predicted value of the power estimation model to identify potential fault trends and issue an event trigger signal E(i,t) for initiating edge - side diagnostic actions.

[0104] The details of this step are as follows:

[0105] The first stage is the construction of the deviation sequence, aiming to analyze the difference between the theoretical power output of each component i at any time t and the actually collected power P_actual(i,t). The difference between the two is defined as the deviation ΔP(i,t) per unit time, which is used to directly quantify the "operating deviation degree" of the component at that moment. When this deviation is positive for several consecutive times, it usually indicates that the actual output of the component is less than expected, which may be due to problems such as occlusion, local thermal degradation, or long - term aging; when the deviation is negative, it may be related to short - term anomalies such as model errors, system dynamic compensation, or albedo over - emission. To avoid over - reacting to the error at a single time point, this step constructs a cumulative deviation judgment logic within a sliding time window T w with a length of N. Within this window, the power deviation ΔP(i,t - k) at each moment is included in the judgment, but samples closer to the current moment are given higher weights. The attenuation rule of the weights can adopt a simple decreasing coefficient (for example, the weight of the most recent frame is 1, the previous frame is 1 / 2, the frame before that is 1 / 3, and so on), thus forming an enhanced perception ability for "recent consecutive anomalies".

[0106] However, in high - altitude environments, due to natural factors such as instantaneous cloud drift, mountain albedo, or snow cover changes, drastic fluctuations in illuminance are common. If not suppressed, these short - term non - structural disturbances are easily misjudged as component failures. Therefore, an illuminance fluctuation penalty mechanism is innovatively introduced. That is, in each time window T w the system calculates the standard deviation of the illuminance estimated value of this component This is used to reflect the environmental stability during the specified period. A large standard deviation indicates drastic changes in illumination during that period, suggesting that power deviations may be driven by environmental factors rather than component malfunctions. Ultimately, the system combines the weighted deviation and illuminance fluctuation intensity into a trend score S(i,t). When this score exceeds a certain threshold (configurable as a percentage of rated power, such as 5%), and the current trigger signal E(i,t) has not yet been activated, E(i,t) is set to 1, indicating entry into the edge diagnosis phase. It is important to note that the influence coefficient λ of the illuminance penalty term is an adjustable parameter that can be adjusted according to the environmental fluctuation characteristics of different power plant areas. For example, in the Shigatse region of Tibet, where frequent snow cover in winter enhances tolerance, λ can be increased; while in the Gonghe region of Qinghai, where the proportion of sunny days is high, this parameter can be appropriately reduced to improve sensitivity. The advantage of this method lies in its integration of structural trend judgment with scene adaptability adjustment mechanism. It does not rely on complex models, but only on sliding window statistics and light fluctuation characteristics to establish a dynamic triggering mechanism. It is extremely suitable for deployment in low-computing-power main control or edge terminals, serving as a key link in the pre-anomaly identification of the intelligent diagnostic system for plateau photovoltaic sites.

[0107] The second stage involves the execution of the trigger judgment mechanism. Based on the trend score S(i,t) calculated in the previous stage, the system compares this value with a set threshold θ. trigger A comparison is made to determine whether to generate a trigger signal E(i,t). Specifically, when the value of S(i,t) exceeds the threshold, the system will determine that the power behavior of component i at time t shows an abnormal trend, and the trigger flag E(i,t) will be set to 1, thereby activating the edge-side diagnostic module M to enter the working state; otherwise, E(i,t) remains 0, and the system does not take further action. In actual engineering, θ trigger The value of θ can be set according to the rated power of the module, and is commonly set to 3% to 5% of the rated power. For example, for a photovoltaic module with a rated power of 400W, θ trigger The deviation can be set to the range of 12W to 20W to filter out non-structural fluctuations caused by minor disturbances, modeling residuals, or sampling noise, thus avoiding false triggering. It is worth noting that in the trend judgment structure designed for high-altitude power plant environments in this patent, S(i,t) is already a comprehensive result of "weighted cumulative deviation minus the illuminance disturbance influence factor." This means that only when the component exhibits a significant deviation from the power estimation model for a continuous period, and this deviation cannot be explained by external factors such as drastic fluctuations in illumination, will S(i,t) exceed the threshold θ. triggerTherefore, this triggering mechanism is inherently designed to resist disturbances, exhibiting higher tolerance for non-faulty disturbances such as short-term shading and cloud changes, while also identifying trend-based problems such as potential thermal runaway and unstable weld points earlier. From an engineering deployment perspective, the generation logic of E(i,t) can be deployed as an independent module in the component cascade master control unit or edge RTU, requiring extremely low computational resources and not relying on large model inference or complex network structures. This mechanism only needs to cache and statistically analyze the power deviation and illuminance fluctuations at recent moments, meeting the operational requirements of low power consumption, low communication, and rapid on-site response at the edge. It is one of the core logical foundations for achieving "early intelligent anomaly identification" at plateau photovoltaic sites.

[0108] Final output of this step:

[0109] ΔP(i,t): Component power deviation value, providing characteristic data for edge diagnosis;

[0110] E(i,t): Event trigger signal, which determines whether to proceed to step four of the diagnostic process.

[0111] Step 4: Edge wake-up diagnosis and feature extraction (M-edge):

[0112] The goal of this step is to activate the lightweight diagnostic module M, deployed within a field edge device (such as a component group controller or smart combiner box), when the event trigger signal E(i,t) output in the previous step is 1. This module combines key electrical and temperature data within a time window to perform feature extraction and preliminary fault identification, generating a low-dimensional feature representation F(i,t) and a preliminary diagnostic label L to be uploaded. 局部 (i,t). This step connects the modeling logic and the system response in the patent structure, playing a crucial role: it receives the power deviation value ΔP(i,t) and event trigger flag E(i,t) output from step three, and serves as the input basis for the fusion judgment in step five, ensuring the system's rapid response to real faults and improving remote inference efficiency.

[0113] First, when E(i,t) = 1, edge M is activated and enters the data acquisition and processing flow. The data in the acquisition window includes:

[0114] V(i,τ): obtained from the component voltage acquisition module;

[0115] I(i,τ): Measured in real time by a DC bus current sampler;

[0116] T(i,t): Acquired by the backplane temperature sensor, already used in step two, and continues as the edge sampling channel.

[0117] The process then proceeds to the feature extraction module. To achieve the ability to represent complex fluctuation patterns on a resource-constrained embedded platform, this step designs a feature structure that integrates time-difference and normalized fluctuation descriptions, defined as follows:

[0118]

[0119] in:

[0120] ΔV = max(V(i,τ)) - min(V(i,τ)), which represents the maximum and minimum voltage fluctuation values.

[0121] The voltage mean, normalized fluctuation amplitude;

[0122] Similarly, ΔI and For current fluctuations and average values;

[0123] σ(T(i,τ)) is the standard deviation of temperature within the time window;

[0124] The relative power deviation intensity is considered in conjunction with the current severity of the fault.

[0125] This structure incorporates power deviation as a "response regularization term," integrating it with time-series fluctuation characteristics. Compared to traditional statistical features, this structure more effectively distinguishes between "structural fluctuations" (such as poor contact and electrical interference) and "environmental fluctuations" (such as sunlight disturbances), especially at high-altitude photovoltaic sites where the latter are more frequent and the former are more difficult to identify. Normalization avoids the feature scale mismatch problem caused by batch-to-batch differences in module electrical parameters.

[0126] After the event is triggered and feature extraction is completed, the system enters the classification and judgment stage. The goal is to preliminarily determine the typical fault type of the current component i based on the feature vector F(i,t) output by the edge module M, for further confirmation by the subsequent remote fusion model. F(i,t) includes four indicators: voltage fluctuation amplitude, current fluctuation amplitude, temperature change standard deviation, and relative power deviation. The system uses a rule-based classification method for judgment: when the voltage fluctuation exceeds 8 volts (i.e., the difference between the maximum and minimum voltage values ​​is greater than 8V), the current fluctuation exceeds 1.5 amps (i.e., the difference between the maximum and minimum voltage values ​​is greater than 1.5A), and the temperature standard deviation is less than 0.5℃, and the relative power deviation is also within acceptable limits. If the deviation is greater than 5%, the system judges it as "voltage anomaly"; if the temperature standard deviation is greater than 1.2℃ and the relative power deviation exceeds 7%, while the voltage and current fluctuations are less than the set thresholds (i.e., less than 5V and 1.0A), it is judged as "thermal runaway"; if all electrical indicators are less than the above thresholds, and the edge-side illumination disturbance is significant (provided by the shading module), while the relative power deviation is within the range of 3% to 6%, it is marked as "shading"; if none of the above conditions are met, or the feature combination does not form an identifiable pattern, the system classifies the sample as "unknown".

[0127] To further reduce the probability of false triggering caused by strong disturbances in high-altitude environments, such as shadow drift and cloud changes, the system introduces a "false triggering suppression mechanism" in its classification and judgment process. When the absolute value of the power deviation ΔP(i,t) is lower than... If the relative power deviation is less than 5%, and the voltage and current fluctuations are both below 3V and 0.8A, and the temperature standard deviation is below 0.5℃, the system will directly output "Unknown" as the classification result without proceeding to a specific fault type determination. The final diagnostic result is recorded as L. 局部 The value of (i,t) is limited to $\mathcal{C}={$shading, thermal runaway, voltage anomaly, unknown$}$, and is uploaded to the remote platform along with the feature vector F(i,t) through the edge communication module. This judgment logic has extremely low computational overhead, is entirely based on preset rules, is compatible with MCU or lightweight CPU environments, has good edge deployment capabilities and adaptability to high-altitude photovoltaic scenarios, and is the key execution unit in the edge response module of this patent.

[0128] The output includes:

[0129] F(i,t): Structured feature vector with a fixed dimension of 4, which facilitates compressed transmission;

[0130] L 局部 (i,t): Preliminary diagnostic label, used for remote fusion judgment and operation and maintenance suggestions in step five.

[0131] Step 5: Remote Integrated Decision-Making and Scheduling Response:

[0132] The task of this step is to construct a fusion judgment model on the remote center side, and to compare the structured feature vector F(i,t) output by the edge diagnosis module in the previous step (step four) with the preliminary diagnosis label L. 局部 Aggregate modeling is performed on (i,t), and the power deviation ΔP(i,t) output in the previous step is used as a fault intensity reference index. The final fault classification result L is then output at the remote center. 最终 (i,t).

[0133] First, the edge-uploaded features F(i,t) are standardized to construct a diagnostic modeling vector Z(i,t), which contains all directly computable input factors. The definition is as follows:

[0134]

[0135] in:

[0136] F(i,t) is the original feature vector uploaded by the edge module;

[0137] This is the one-hot encoding format for edge labels;

[0138] |ΔP(i,t)| is the absolute value of the deviation, which is used as the fault intensity input for the final discrimination model;

[0139] Z(i,t) is a dimensionless integer. The structured input vector can be constructed directly from edge information received from a remote center, without accessing the original sampling sequence.

[0140] Next, a fault discrimination model is constructed. This model is designed as a shallow neural structure with input weight regularization, containing a feedforward hidden layer of width 8 and an output layer, and uses softmax for multi-class classification. To improve the model's discrimination sensitivity in minor fault modes commonly found in high-altitude power plants (such as temporary occlusion and low-amplitude thermal runaway), this step specifically introduces a "low-amplitude bias suppression term," which makes it less likely for the model output to directly fall into the severe fault label when the bias is small. The structure is as follows:

[0141]

[0142] in:

[0143] P j (k,t) indicates that component i is classified as fault type j. The probability of;

[0144] W1, b1, W2, b2 are the weights and biases of the hidden layer and the output layer, respectively. The structure is fixed, and the model is deployed after being trained offline at a remote center.

[0145] σ(·) is the ReLU activation function;

[0146] γ and κ are parameters for adjusting the low deviation penalty term, for example, γ = 1.0, κ = 5.0;

[0147] To exclude options labeled "unknown", a deviation penalty is added to the three categories of "occlusion", "thermal runaway" and "electrical anomaly".

[0148] This structure introduces a nonlinear deviation suppression mechanism: when |ΔP(i,t)| is very small (i.e., the actual deviation is slight), e -κ·|ΔP(i,t)| When the value approaches 1, the model will be negatively penalized when it classifies a fault, automatically biasing towards the "unknown" class. As the bias increases, the suppression term decays rapidly, and the model's judgment will return to normal. This design addresses the practical problems of large fluctuations in illumination, sudden occlusion, and large differences in component response at high-altitude sites, effectively reducing the false alarm rate while maintaining fault identification sensitivity.

[0149] Ultimately, the system will score P based on the probability of various types of faults. j (i,t), select the highest-scoring fault category from all fault categories as the final fault determination result L for component i at time t. 最终 (i,t). Here, P j (i,t) represents the overall confidence level of a component being classified as a type j fault (such as "occlusion", "thermal runaway", "voltage anomaly" or "unknown"). This value is the sum of the feature vector F(i,t) received by the fusion center and the initial edge label L. 局部 After processing multi-dimensional information such as (i,t) and power deviation ΔP(i,t), the result is obtained through decision reasoning using the central fusion model M. The system iterates through the scores of all candidate fault categories and selects the category with the highest confidence as the final output label. This final judgment result L 最终 (i,t) will serve as the diagnostic output of the entire intelligent operation and maintenance process and will be incorporated into the subsequent scheduling and management process to support maintenance plans, component output limits, remote confirmation, or other intelligent response mechanisms.

[0150] The technical solutions in the above embodiments of this application have at least the following technical effects or advantages: This solution calculates the real illuminance value of each component at any time by using a digital elevation model and component installation parameters, avoiding the risk of misjudgment caused by the traditional uniform irradiance assumption; it constructs a power deviation trend analysis method based on a time window, introduces an illuminance fluctuation suppression mechanism, and triggers diagnostic events only when a continuous anomaly exists and cannot be explained by environmental changes; it generates preliminary fault types through rule-based judgment, reducing the edge computing burden and improving response time.

[0151] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent operation and maintenance and fault prediction of high-altitude photovoltaic power plants based on artificial intelligence, characterized in that: include: S1: Collect the first raw data, which includes digital elevation model data of the area where the photovoltaic power station is located, tilt angle and orientation information of the components, and total all-sky radiation data at the site level; The first raw data is input into the illuminance estimation model; the illuminance estimation model includes an occlusion determination module and an illuminance calculation module; The occlusion determination module determines whether each component is subject to terrain occlusion by calculating the relationship between the solar incidence angle and the terrain model. The illuminance calculation module calculates the actual illuminance value that each component can receive within a preset time window; The final output component is the illuminance estimate within a preset time window; S2: Construct a power estimation model, which includes ideal power baseline estimation and aging and plateau micro-topography correction; The ideal power baseline estimation is obtained by calculating the initial conversion efficiency of the module, the temperature loss coefficient, the temperature reference point, the current module temperature, and the estimated illuminance. The aging and plateau micro-topography correction is obtained by calculating the correction factor based on the component's historical aging factor, the local elevation of the component's installation point relative to the site, and the unified reference altitude of the site. The theoretical power output is obtained by combining the ideal power baseline and the correction factor. S3: Collect actual power output; Construct a deviation sequence, which is a set of deviation values ​​between the theoretical power output and the actual power output per unit time; The illuminance fluctuation penalty term is designed, which is the standard deviation of the illuminance estimate of the component in each time window; The deviation sequence is combined with the illuminance fluctuation penalty term to form a trend score; When the trend score exceeds the preset threshold and the trigger signal is not activated, activate the trigger signal and proceed to step S4; otherwise, do not respond. S4: Acquire the second raw data within the time window, which includes voltage, current, and temperature; The structural feature vector is obtained by fusing voltage fluctuation amplitude, current fluctuation amplitude, temperature change standard deviation and relative power deviation. The system uses preset rules to classify and determine the initial diagnostic labels. S5: Standardize the structural feature vectors to construct diagnostic modeling vectors; A fault discrimination model is constructed. The fault discrimination model is a shallow neural structure with input weight regularization, which includes a feedforward hidden layer with a width of 8 and an output layer. Softmax is used for multi-class classification output. The diagnostic modeling vector is input into the fault discrimination model to output probability scores for various types of faults; The highest-scoring fault category is selected from all fault categories as the final fault determination result.

2. The intelligent operation and maintenance and fault prediction method for high-altitude photovoltaic power plants based on artificial intelligence according to claim 1, characterized in that, The calculation steps of the occlusion determination module mentioned in step S1 are as follows: Calculate the current solar altitude angle and solar azimuth angle based on the component's location and time; Project a straight line along the direction of sunlight at each component location to determine if there is terrain above the component's installation height that blocks the light; if so, set the shading mask value of that component to 1 at that time point, otherwise set it to 0.

3. The intelligent operation and maintenance and fault prediction method for high-altitude photovoltaic power plants based on artificial intelligence according to claim 1, characterized in that, The calculation steps of the illuminance calculation module mentioned in step S1 are as follows: The total radiation at the current time of a known site is used as the original benchmark for the total radiation energy received by the site per unit time and per unit area; The original baseline is corrected by the occlusion mask value. When the occlusion mask value is the highest, the illumination correction coefficient is the lowest. The occlusion mask value varies between 0 and 1. The occlusion mask value is obtained from terrain modeling and the direction of solar incidence. The illuminance estimate is obtained by adjusting the angle between the component's tilt angle and the current solar altitude angle.

4. The intelligent operation and maintenance and fault prediction method for high-altitude photovoltaic power plants based on artificial intelligence according to claim 1, characterized in that, The deviation sequence assigns different weights to the sample deviation values, and the weight decay rule adopts a decreasing coefficient, with the weight of the sample closer to the current time being higher.

5. The intelligent operation and maintenance and fault prediction method for high-altitude photovoltaic power plants based on artificial intelligence according to claim 1, characterized in that, The preset rules mentioned in step S4 are as follows: When voltage fluctuation exceeds the voltage fluctuation threshold, current fluctuation exceeds the current fluctuation threshold, and the temperature standard deviation is less than the first temperature difference threshold, while the relative power deviation is greater than the first power deviation threshold, the output is a voltage anomaly. When the temperature standard deviation is greater than the second temperature difference threshold and the relative power deviation exceeds the second power deviation threshold, while the voltage and current fluctuations are less than the voltage fluctuation threshold and the current fluctuation threshold, respectively, the output is thermal runaway. If all electrical indicators are less than the above thresholds, and the edge-side illumination disturbance exceeds the shading threshold, while the relative power deviation is between the first power deviation threshold and the second power deviation threshold, the output is shading. If none of the above conditions are met, the output will be unknown.

6. The intelligent operation and maintenance and fault prediction method for high-altitude photovoltaic power plants based on artificial intelligence according to claim 1, characterized in that, The preset rules also include: When the absolute value of the power deviation is lower than the first power deviation threshold, and the voltage and current fluctuations are lower than the voltage fluctuation threshold and the current fluctuation threshold respectively, and the temperature standard deviation is lower than the first temperature difference threshold, the output is unknown.

7. The intelligent operation and maintenance and fault prediction method for high-altitude photovoltaic power plants based on artificial intelligence according to claim 1, characterized in that, The fault discrimination model introduces a low-amplitude deviation suppression term, which applies a negative penalty when the actual deviation is below a threshold and is classified as a fault.

8. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the program, it implements the method as described in any one of claims 1-7.

Images