A photovoltaic power generation monitoring method and system based on artificial intelligence

Abstract

This invention relates to the field of intelligent operation and maintenance and fault diagnosis technology for photovoltaic power plants, and particularly to a photovoltaic power generation monitoring method and system based on artificial intelligence. The method includes the following steps: parsing the time-of-use power generation data of the target photovoltaic modules, retrieving the total cumulative power generation for the entire day, performing normalization calculations for irradiance fluctuations, and constructing an actual power generation percentage sequence; performing time-domain cross-correlation matching between the actual power generation percentage sequence and a preset historical benchmark percentage sequence to determine the optimal matching time shift, and generating an aligned benchmark sequence based on the optimal matching time shift; calculating the negative distortion component sequence of the actual power generation percentage sequence relative to the aligned benchmark sequence. This invention, by normalizing power generation and using solar trajectory for spatiotemporal mapping, achieves remote inference of the physical location of interference sources and automatic classification of fault characteristics without relying on image equipment, effectively reducing false alarm rates and improving troubleshooting efficiency.

Description

Technical Field

[0001] This invention relates to the field of intelligent operation and maintenance and fault diagnosis technology for photovoltaic power plants, and in particular to a photovoltaic power generation monitoring method and system based on artificial intelligence. Background Technology

[0002] Existing photovoltaic (PV) power generation monitoring logic typically relies on directly comparing the absolute value of power or electricity generation with preset thresholds. Since the output power of PV modules is strongly positively correlated with solar irradiance, under conditions of drastic irradiance fluctuations such as cloudy or rainy weather, module power drops significantly, triggering inefficiency alarms. This makes it difficult for the system to distinguish between widespread inefficiency caused by weather conditions and isolated inefficiencies caused by equipment malfunctions. This strong dependence on weather conditions results in a very high false alarm rate during non-sunny periods, while during periods of high irradiance, minor equipment performance losses are easily overlooked, leading to missed alarms.

[0003] Existing photovoltaic power generation monitoring data can only reflect a numerical decline in module power generation performance, meaning it can only inform maintenance personnel that the equipment's power generation has decreased, but it cannot provide spatial information about the direction from which the interference is coming from. When modules are partially shaded (such as by trees or building shadows), the backend data cannot deduce the physical location of the shading source. This means that after receiving an alarm, maintenance personnel cannot make a preliminary judgment remotely and must go to the site to conduct a comprehensive, blind inspection of the surrounding environment, greatly increasing the time and manpower costs of on-site inspections.

[0004] Existing photovoltaic power generation monitoring methods often lack the ability to automatically classify the causes of power anomalies. External environmental factors (such as vegetation shading and dust accumulation) and internal equipment malfunctions (such as open circuits in diodes and microcracks) both manifest as a decrease in power output, and the system cannot automatically distinguish between the two. This leads to a lack of targeted maintenance dispatching, often resulting in situations where cleaning is used to solve equipment damage problems, or maintenance personnel are assigned to handle environmental shading issues, causing misallocation and waste of maintenance resources.

[0005] Existing technologies suffer from weak resistance to weather interference, lack of fault spatial location capabilities, and inability to automatically distinguish the nature of faults, problems that urgently need to be addressed. Summary of the Invention

[0006] Therefore, it is necessary to provide an artificial intelligence-based photovoltaic power generation monitoring method and system to solve at least one of the above-mentioned technical problems.

[0007] To achieve the above objectives, an artificial intelligence-based photovoltaic power generation monitoring method includes the following steps: Step S1: In response to the real-time operating data uploaded by the data acquisition terminal, the time-of-use power generation data of the target photovoltaic module is analyzed using the cloud monitoring server, and the total cumulative power generation for the whole day is retrieved through the time series database. Normalization calculation is performed for irradiance fluctuations to construct the actual power generation ratio sequence. Step S2: Perform time-domain cross-correlation matching between the actual power generation ratio sequence and the preset historical benchmark ratio sequence to determine the optimal matching time shift, and generate the aligned benchmark sequence based on the optimal matching time shift; calculate the negative distortion component sequence of the actual power generation ratio sequence relative to the aligned benchmark sequence. Step S3: Identify the effective concave time interval in the negative distortion component sequence; call the geographic information system engine to establish the temporal and spatial mapping relationship, calculate the relative incident azimuth of the component, and generate the anomaly source azimuth vector by combining the effective concave time interval; Step S4: Construct a neighborhood component set based on the azimuth vector of the anomaly source, calculate the spatiotemporal overlap of the anomaly features within the neighborhood component set, determine the anomaly type as either an environmental occlusion anomaly or a device-related anomaly based on the spatiotemporal overlap of the anomaly features, and generate the corresponding processing work order.

[0008] This invention constructs an actual power generation ratio sequence, eliminating the overall interference of meteorological conditions on monitoring indicators from the data source. This significantly reduces the false alarm rate during cloudy and rainy weather, improving the system's robustness under complex operating conditions. The translation correction mechanism effectively solves the benchmark misalignment problem caused by seasonal changes, ensuring the sensitivity of anomaly detection. Utilizing solar azimuth mapping technology, remote and accurate inference of the spatial location of interference sources is achieved, providing intuitive directional guidance for maintenance personnel. The spatiotemporal correlation judgment function can automatically distinguish between regional environmental interference and individual equipment failure, transforming maintenance work from blind troubleshooting to precise handling, significantly shortening the fault response cycle and reducing manpower maintenance costs.

[0009] Preferably, the present invention also provides an artificial intelligence-based photovoltaic power generation monitoring system for executing the artificial intelligence-based photovoltaic power generation monitoring method described above, the artificial intelligence-based photovoltaic power generation monitoring system comprising: The data normalization module is used to respond to the real-time operating data uploaded by the data acquisition terminal, use the cloud monitoring server to analyze the time-of-use power generation data of the target photovoltaic module, retrieve the total cumulative power generation of the day through the time series database, perform normalization calculation for irradiance fluctuations, and construct the actual power generation ratio sequence. The distortion extraction module is used to perform time-domain cross-correlation matching between the actual power generation ratio sequence and the preset historical benchmark ratio sequence, determine the optimal matching time shift, generate the aligned benchmark sequence based on the optimal matching time shift, and calculate the negative distortion component sequence of the actual power generation ratio sequence relative to the aligned benchmark sequence. The vector source tracing module is used to identify the effective concave time interval in the negative distortion component sequence; it calls the geographic information system engine to establish a mapping relationship between time and space, calculates the relative incident azimuth of the components, and generates the anomaly source azimuth vector by combining the effective concave time interval; The fault attribution module is used to construct a neighborhood component set based on the azimuth vector of the anomaly source, calculate the spatiotemporal overlap of anomaly features within the neighborhood component set, determine the anomaly type as environmental occlusion anomaly or device-related anomaly based on the spatiotemporal overlap of anomaly features, and generate the corresponding processing work order.

[0010] This invention provides an AI-based photovoltaic power generation monitoring system that achieves automated closed-loop monitoring through modular design. The data normalization module ensures the stability of all-weather monitoring, while the distortion extraction module improves the accuracy of identifying subtle performance degradation through time-domain alignment technology. The vector tracing module provides intuitive spatial coordinates for monitoring data, solving the problem of traditional monitoring's inability to determine location. The fault attribution module enables intelligent classification and distribution of maintenance instructions through group feature analysis. The entire system reduces reliance on human experience and external high-performance imaging hardware, ensuring the power generation efficiency of photovoltaic parks while achieving scientific and refined management of maintenance decisions. Attached Figure Description

[0011] Figure 1 A flowchart illustrating an artificial intelligence-based photovoltaic power generation monitoring method provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the anomaly attribution and persistence determination logic provided in an embodiment of the present invention. Detailed Implementation

[0012] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0013] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0014] It should be understood that the term “and / or” as used herein includes any and all combinations of one or more of the associated items listed.

[0015] In this embodiment of the invention, reference is made to Figure 1 The diagram shown is a flowchart illustrating the steps of the AI-based photovoltaic power generation monitoring method of the present invention. In this example, the AI-based photovoltaic power generation monitoring method includes the following steps: Step S1: In response to the real-time operating data uploaded by the data acquisition terminal, the time-of-use power generation data of the target photovoltaic module is analyzed using the cloud monitoring server, and the total cumulative power generation for the whole day is retrieved through the time series database. Normalization calculation is performed for irradiance fluctuations to construct the actual power generation ratio sequence. In this embodiment of the invention, the cloud monitoring server divides the acquired time-of-use power generation data of the day into multiple time slices with a fixed duration, calculates the proportion of power generation of each slice to the total power generation of the day, and offsets the influence of irradiance fluctuations on the values ​​through this proportionalization process. Finally, the data are arranged in chronological order to form an actual power generation proportion sequence that reflects the power generation distribution characteristics of the components.

[0016] Step S2: Perform time-domain cross-correlation matching between the actual power generation ratio sequence and the preset historical benchmark ratio sequence to determine the optimal matching time shift, and generate the aligned benchmark sequence based on the optimal matching time shift; calculate the negative distortion component sequence of the actual power generation ratio sequence relative to the aligned benchmark sequence. In this embodiment of the invention, the system slides and aligns the actual proportion sequence with the historical benchmark sequence, determines the translation amount by finding the number of steps with the smallest matching deviation, compensates for the displacement of the solar trajectory caused by seasonal changes, thereby generating an aligned benchmark template, and then uses subtraction to peel off the negative residual representing the damage to power generation performance, forming a negative distortion component sequence.

[0017] Step S3: Identify the effective concave time interval in the negative distortion component sequence; call the geographic information system engine to establish the temporal and spatial mapping relationship, calculate the relative incident azimuth of the component, and generate the anomaly source azimuth vector by combining the effective concave time interval; In this embodiment of the invention, the cloud monitoring server retrieves continuously exceeding abnormal segments in the distortion sequence to lock the effective depression interval. It uses a geographic information system engine to infer the azimuth angle of the sun based on the geographic coordinates, date, and time of the depression, calculates the incident angle of sunlight relative to the component surface, and combines the cumulative severity of the depression with the incident angle to construct an anomaly source azimuth vector with spatial pointing significance.

[0018] Step S4: Construct a neighborhood component set based on the azimuth vector of the anomaly source, calculate the spatiotemporal overlap of the anomaly features within the neighborhood component set, determine the anomaly type as environmental occlusion anomaly or device-related anomaly based on the spatiotemporal overlap of the anomaly features, and generate the corresponding processing work order; In this embodiment of the invention, the system retrieves the neighborhood set around the target component on the electronic map, compares whether the components in the neighborhood show similar vector features at the same time and in the same direction, and quantifies the abnormal clustering by calculating the spatiotemporal overlap ratio. If the overlap is high, it is determined to be external environmental occlusion and guided to clean it up; if the overlap is low, it is determined to be an individual component failure and guided to repair it.

[0019] Preferably, step S1 includes: The effective daytime illumination was divided into several time slices of equal length using a time series database; Calculate the integral value of the time-of-use power generation data of the target photovoltaic module in each time slice, and divide the integral value by the total cumulative power generation for the whole day to obtain the power generation ratio coefficient of the time slice. The power generation ratio coefficients corresponding to all time slices throughout the day are arranged in chronological order to form the actual power generation ratio sequence.

[0020] In one embodiment, the cloud monitoring server reads the sunrise and sunset times of the current monitoring day from a time-series database, and determines the time period between sunrise and sunset as the effective daylight hours. The effective daylight hours are then divided into several equal-length time slices at fixed 15-minute intervals. Taking a total effective daylight hours of 10 hours as an example, this results in 40 time slices, which are numbered sequentially from time slice 1 to time slice 40.

[0021] The system receives time-of-use (TOU) power generation data of the target photovoltaic modules at various sampling times within the current monitoring day from the data acquisition terminal via a communication network, and writes the received TOU power generation data into a time series database for storage. For each time slice, it retrieves all TOU power generation data from the time series database within the range of the start and end times of that time slice, and performs discrete integration operations on these power data along the time axis. (The last sentence appears to be incomplete and requires further context.) Taking a time slice as an example, suppose that the time slice contains a total of At each sampling time, the instantaneous power generation at the l-th sampling time is The time interval between adjacent sampling times is Then the first Power generation per time slice By , Until The power generation of each time slice is obtained by summing up the results. This discrete integration operation is then performed sequentially on all time slices to obtain the power generation for each time slice.

[0022] The total daily power generation is obtained by summing the power generation of each time slice across all time slices. Then, the power generation of each time slice is divided by the total daily power generation to obtain the power generation ratio coefficient for that time slice. The power generation ratio coefficient is a dimensionless value representing the proportion of power generation within that time slice to the total daily power generation. The sum of the power generation ratio coefficients for all time slices is always equal to 1.

[0023] The power generation percentage coefficients corresponding to all time slices are arranged sequentially from the 1st to the Mth time slice number to form the actual power generation percentage sequence. Each value in the actual power generation percentage sequence corresponds to a time slice, reflecting the proportion of that time slice in the total daily power generation. Since the power generation of each time period is divided by the same total daily power generation, when the overall solar radiation intensity rises or falls proportionally due to weather changes, the percentage values ​​of each time period remain unchanged, and the overall shape of the actual power generation percentage sequence is not affected by weather factors.

[0024] Preferably, the specific process of generating the aligned reference sequence in step S2 includes: Using a pre-set historical benchmark proportion sequence as a template, a sliding window matching is performed on the actual power generation proportion sequence on the time axis; Calculate the sequence similarity under different sliding steps, convert the sliding step number corresponding to the maximum similarity into time duration, and determine it as the optimal matching time shift amount; The time axis of the preset historical baseline proportion sequence is shifted and corrected according to the optimal matching time shift amount to obtain the aligned baseline sequence.

[0025] In one embodiment, a pre-set historical baseline percentage sequence for the target photovoltaic module is read from a time-series database. The historical baseline percentage sequence and the actual power generation percentage sequence have the same number of time slices, with the same-numbered time slices in both sequences corresponding to the same time period of the day. Because the solar trajectory shifts with the seasons, sunrise and sunset times differ in different months, causing the peak position of the actual power generation percentage sequence to shift along the time axis relative to the peak position of the historical baseline percentage sequence. To eliminate the impact of this seasonal shift, a time-axis translation correction is needed for the historical baseline percentage sequence to align its peak position with that of the actual power generation percentage sequence.

[0026] The search range for the translation amount is set. In this embodiment, the search range is set to shift 4 time slices to the left and 4 time slices to the right. That is, the value range of the candidate translation amount τ is the set of integers from negative 4 to positive 4, for a total of 9 candidate translation amounts. Shifting to the left indicates that the historical baseline proportion sequence moves as a whole towards an earlier time period, and shifting to the right indicates that the whole sequence moves towards a later time period.

[0027] For each of the nine candidate shifts, the following matching operation is performed: the historical baseline proportion sequence is shifted along the time axis by the number of time slices corresponding to the current candidate shift, resulting in a shifted temporary sequence. For time slices that exceed the sequence boundary after shifting, the values ​​at the sequence boundary are used to fill the gaps. Subsequently, the numerical difference between the actual power generation proportion sequence and the temporary sequence in the same numbered time slice is calculated for each time slice. The square of each difference is then summed over all time slices to obtain the matching deviation value corresponding to the candidate shift. The smaller the matching deviation value, the closer the shapes of the two sequences are under that shift, and the higher the degree of time axis alignment.

[0028] The matching deviation values ​​corresponding to all 9 candidate shifts are iterated through, and the candidate shift with the smallest matching deviation value is selected as the optimal matching time shift. In this embodiment, taking a time slice duration of 15 minutes as an example, if the optimal matching time shift is positive 2, it means that the historical baseline proportion sequence needs to be shifted to the right by 2 time slices, with a corresponding shift duration of 30 minutes.

[0029] A time-axis shift operation is performed on the historical baseline proportion sequence according to the optimal matching time shift, shifting the values ​​of each time slice in the historical baseline proportion sequence by the corresponding number of time slices to generate an aligned baseline sequence. The aligned baseline sequence and the actual power generation proportion sequence achieve peak position alignment on the time axis, eliminating the interference of solar trajectory offset caused by seasonal variations on subsequent deviation calculations. The aligned baseline sequence is temporarily stored in a time series database for use in subsequent calculations of negative distortion components.

[0030] Preferably, the specific process of calculating the negative distortion component sequence in step S2 includes: The difference between the actual power generation ratio sequence and the aligned baseline sequence at each time point is calculated to obtain the original residual sequence; Traverse the original residual sequence, retain data points with values ​​less than zero as valid negative distortion values, and set data points with values ​​greater than or equal to zero to zero. The negative distortion component sequence is composed of all valid negative distortion values ​​arranged in chronological order.

[0031] In one embodiment, the cloud monitoring server reads the actual power generation ratio sequence and the aligned baseline sequence from the time series database. Both sequences have the same number of time slices, and the time slice numbers correspond one-to-one. Starting from the first time slice, the value of the current time slice in the actual power generation ratio sequence is subtracted from the value of the time slice with the same number in the aligned baseline sequence to obtain the original residual value corresponding to that time slice. After performing the above subtraction operation on all time slices, all original residual values ​​are arranged in order of time slice number to form the original residual sequence. Each value in the original residual sequence reflects the degree of deviation of the actual power generation ratio of the target photovoltaic module within the corresponding time slice from the historical baseline ratio. A negative value indicates that the actual ratio for that period is lower than the baseline ratio, meaning that the power generation contribution share for that period has decreased; a positive value indicates that the actual ratio for that period is higher than the baseline ratio.

[0032] Each data point in the original residual sequence is individually sign-determined. For data points with values ​​less than zero, the value remains unchanged and is retained as a valid negative distortion value; for data points with values ​​greater than or equal to zero, the value is forcibly set to zero. The basis for this processing is that when the actual proportion of a time slice is lower than the baseline proportion, it indicates that the power generation capacity of that period has decreased relative to the normal state, which is an abnormal signal that requires attention; when the actual proportion of a time slice is higher than the baseline proportion, it is usually due to the compensatory increase in value caused by the normalization calculation after the proportion of other periods has decreased, and it does not reflect the enhanced power generation capacity of that period itself. Therefore, it does not have the value of anomaly diagnosis and is set to zero to eliminate interference.

[0033] After sign discrimination processing, all data points are rearranged according to the original time slice numbering order. Valid negative distortion values ​​retaining their original values ​​are distributed at their corresponding time slice positions, while data points set to zero fill the remaining time slice positions, thus forming a negative distortion component sequence. The negative distortion component sequence has the same time slice number and numbering correspondence as the actual power generation ratio sequence. Time slices with zero values ​​in the sequence indicate normal or high power generation during that period, while time slices with negative values ​​indicate a negative deviation in the power generation ratio during that period. The negative distortion component sequence is temporarily stored in the time series database for subsequent identification of periods of consolidation.

[0034] Preferably, the specific process of identifying the effective indentation time interval in step S3 includes: Determine whether there are any consecutive time slice values ​​below a preset anomaly threshold in the negative distortion component sequence. If so, lock the time range corresponding to a consecutive preset number or more time slices and mark it as the effective concave time period interval.

[0035] In one embodiment, a negative distortion component sequence is read from a time series database, and a preset anomaly detection threshold and a preset minimum number of consecutive slices are loaded. In this embodiment, the anomaly detection threshold is set to -0.03, and the minimum number of consecutive slices is set to 3. The anomaly detection threshold is a negative value, and its absolute value represents the maximum allowable decrease in the power generation ratio of a single time slice relative to the baseline ratio. When the negative distortion component value of a certain time slice is lower than this threshold, it indicates that the decrease in the power generation ratio during that period has exceeded the normal fluctuation range. The minimum number of consecutive slices is used to exclude isolated and occasional fluctuations caused by instantaneous disturbances (such as birds flying by or clouds drifting by quickly). Only when multiple consecutive adjacent time slices show an excessive decrease is it considered that the continuous power generation capacity is impaired with practical diagnostic significance.

[0036] Starting with the first time slice of the negative distortion component sequence, the values ​​of each time slice are read sequentially according to their time slice numbers, and compared with anomaly detection thresholds. If the value of the current time slice is lower than the anomaly detection threshold, the time slice is marked as a concave slice, and counting begins; if the value of the current time slice is greater than or equal to the anomaly detection threshold, the time slice is not marked as a concave slice, and the existing consecutive count is reset to zero. During this iterative process, the number of consecutive concave slices is continuously recorded.

[0037] When the number of consecutive concave slices reaches the minimum number of consecutive slices, the start time of the first concave slice in this consecutive concave slice segment is taken as the starting point of the interval. The process continues traversing until a non-concave slice is encountered or the end of the sequence is reached. The end time of the last concave slice is taken as the end time of the interval. The time range covered by the interval starting point to the interval ending point is marked as the effective concave time interval. In this embodiment, taking a time slice duration of 15 minutes as an example, if the 12th to 16th time slices are all marked as concave slices (5 consecutive time slices), then the effective concave time interval is from the start time of the 12th time slice to the end time of the 16th time slice, corresponding to a time span of 75 minutes.

[0038] If the negative distortion component sequence contains multiple non-adjacent consecutive depression segments, and the number of consecutive segments in each segment reaches the minimum number of consecutive slices, then each segment is marked as an independent valid depression time interval. If the entire negative distortion component sequence does not contain any segments with the number of consecutive depression slices reaching the minimum number of consecutive slices, then it is determined that the target photovoltaic module does not exhibit a valid depression on the current monitoring day, and no valid depression time interval is generated. The monitoring process for this module on the current monitoring day terminates here, and subsequent anomaly orientation mapping and spatiotemporal correlation classification steps are no longer executed. The identification results of the valid depression time interval are temporarily stored in the time series database for use in the subsequent generation of anomaly source orientation vectors.

[0039] Preferably, the specific process of generating the anomaly source azimuth vector in step S3 includes: Use a geographic information system engine to obtain the latitude and longitude coordinates of the photovoltaic power generation park and the current monitoring date; Calculate the solar azimuth angles corresponding to the start and end times within the effective concave time interval; Obtain the array installation orientation angle of the target photovoltaic module, calculate the difference between the solar azimuth angle and the array installation orientation angle, and obtain the relative incident azimuth angle of the module; The negative distortion values ​​within the effective indentation time interval are integrated to obtain the integral value representing the severity of the anomaly. The integral value is used as the magnitude of the vector, and the component's relative incident azimuth angle is used as the direction of the vector to construct the azimuth vector of the anomaly source.

[0040] In one embodiment, a geographic information system (GIS) engine is invoked to obtain the longitude and latitude coordinates of the photovoltaic power generation park's location, and the date information of the current monitoring day is read. The GIS engine pre-stores the park's geographical location data. The longitude and latitude coordinates are used to determine the park's precise location on the Earth's surface, and the date information is used to determine the Earth's orbital position relative to the sun on that day. These three parameters together constitute the input conditions for calculating the sun's position in the sky.

[0041] Based on the start and end times of the effective concave time period, the midpoint of this period is calculated. Using this midpoint as the time reference point for calculating the sun's position, and combining longitude, latitude, and the current monitoring date, the corresponding solar azimuth angle is calculated according to the solar position algorithm. The solar azimuth angle is measured clockwise with true north as the zero-degree reference, ranging from 0 to 360 degrees, representing the direction of the sun's projection on the horizontal plane. In this embodiment, if the effective concave time period is from 9:00 AM to 10:15 AM, the midpoint is 9:37 AM. The solar azimuth angle at 9:37 AM is calculated; in this embodiment, this value is 127 degrees, indicating that the sun is located in the east-southeast direction at this time.

[0042] The array installation orientation angle of the target photovoltaic module is read from the geographic information system engine. The array installation orientation angle uses the same angular reference as the solar azimuth angle, with true north as zero degrees, and is measured clockwise. It represents the direction in which the projection of the normal direction of the photovoltaic module's front face onto the horizontal plane points. In this embodiment, the target photovoltaic module is installed facing due south, and its array installation orientation angle is 180 degrees.

[0043] Subtracting the array mounting orientation angle from the solar azimuth angle yields the relative incident azimuth angle of the module. The relative incident azimuth angle represents the horizontal direction of sunlight incident in a local coordinate system based on the normal direction of the photovoltaic module's front surface. In this embodiment, the relative incident azimuth angle is 127 degrees minus 180 degrees, resulting in -53 degrees. A negative value indicates that sunlight enters from the left side of the module's front surface, i.e., from the left front of the module. When the module's power generation ratio in this direction shows a dip, it indicates that the interference factor causing the reduced power generation capacity is located in the direction from which the sunlight is incident, i.e., in the southeast direction slightly to the left front of the module.

[0044] The negative distortion values ​​corresponding to each time slice covered by the effective depression period interval are extracted from the negative distortion component sequence. The absolute values ​​of each negative distortion value are then summed to obtain the depression integral severity. The depression integral severity is a positive value; the larger the value, the more severe the cumulative deviation of the power generation ratio within the effective depression period interval, i.e., the more significant the impact of the anomaly on the module's power generation capacity. In this embodiment, the effective depression period interval covers 5 time slices, with negative distortion values ​​of -0.04, -0.06, -0.08, -0.05, and -0.03 for each time slice. The absolute values ​​are then summed to obtain a depression integral severity of 0.26.

[0045] The severity of the indentation integral is used as the magnitude of the vector, and the relative incident azimuth of the module is used as the direction angle of the vector. The magnitude and direction angle are used together to construct the anomaly source azimuth vector. The anomaly source azimuth vector is a two-dimensional vector that carries both anomaly severity information and anomaly source spatial direction information. Its direction points to the spatial orientation of the interfering factor causing the power generation indentation, and its magnitude reflects the cumulative impact of the interfering factor on the module's power generation capacity. The anomaly source azimuth vector is temporarily stored in a time series database for use in subsequent spatiotemporal correlation classification steps.

[0046] Please see Figure 2 This paper demonstrates the core decision-making process of anomaly attribution classification and persistent filtering in this invention, which is used to transform the detected anomaly source location vector into precise operation and maintenance instructions. The specific steps are as follows: 1. Enter the starting point: The process input uses the anomaly source orientation vector, which contains information on the spatial orientation and severity of the anomaly.

[0047] 2. Calculate the neighborhood overlap: Centered on the target component, the neighboring components within a preset radius are retrieved, and the temporal overlap and directional consistency of the anomalies of each neighboring component are calculated to obtain a spatiotemporal overlap index, which is used to distinguish the nature of the anomalies.

[0048] 3. Anomaly type attribution determination: High overlap: If most components in the neighborhood have anomalies in the same direction and at the same time, it is judged as an environmental shading anomaly (such as vegetation shading, dust accumulation, bird droppings and other regional environmental factors).

[0049] Low overlap: If only the target component has an anomaly and there are no synchronous anomalies in neighboring components, it is determined to be an anomaly of the device itself (such as component microcracks, diode failure, or other individual device failures).

[0050] 4. Historical continuity verification: For the two types of abnormal results mentioned above, further query the historical abnormal records of the target component within a preset observation period (e.g., 7 days): Historical records show continuous occurrences: If an anomaly occurs consistently over multiple days and is located in the same location, it is marked as a persistent fault, and a persistent fault work order is generated and pushed to maintenance personnel for priority handling.

[0051] Historical records: If the anomaly only occurs on the same day, it is marked as an occasional disturbance (such as temporary bird obstruction or transient electrical disturbance), stored in the observation queue, and no work order is assigned for the time being. A decision will be made after further verification.

[0052] Preferably, the specific process of calculating the spatiotemporal overlap of anomalous features in step S4 includes: Based on the electronic map of the photovoltaic power generation park, with the target photovoltaic module as the geometric center, search for other modules within a preset physical distance radius to form a neighborhood module set; Iterate through the neighbor components in the neighborhood component set and check whether the neighbor components also have anomaly source orientation vectors on the same monitoring day; If they exist, count the number of neighboring components whose azimuth vectors overlap in both time period and direction, calculate the proportion of this number to the total number of neighboring components, and obtain the spatiotemporal overlap of the anomaly features.

[0053] In one embodiment, a geographic information system (GIS) engine is invoked to obtain an electronic map of the photovoltaic (PV) power generation park. Using the physical coordinates of the anomaly-causing PV module on the electronic map as the geometric center, and a preset physical distance radius of 15 meters, the device numbers of all PV modules within this radius are retrieved and extracted. All PV modules within this radius, excluding the target PV module, constitute a neighborhood module set. In this embodiment, the total number of modules within the neighborhood module set is counted and denoted as . .

[0054] Iterate through each neighboring component in the neighborhood component set, retrieving its operating record for the current monitoring day from the time-series database based on its device number. Check if each neighboring component also has an anomaly source azimuth vector for the current monitoring day. If no anomaly source azimuth vector exists, the neighboring component is marked as a normal component; if an anomaly source azimuth vector exists, it is determined that the neighboring component also exhibits a power generation ratio dip, and the effective dip time interval and relative incident azimuth angle of the component are further extracted.

[0055] For neighboring modules with anomaly source azimuth vectors, their effective concave time period intervals are compared with those of the target photovoltaic module to determine if there is any time overlap. If at least one overlapping time slice exists, it is considered a time period overlap. Subsequently, the absolute value of the difference between the relative incident azimuth angle of the neighboring module and that of the target photovoltaic module is calculated. In this embodiment, a directional consistency deviation threshold of 10 degrees is set. If the absolute value of the above difference is less than or equal to 10 degrees, it is considered a directional overlap.

[0056] The number of neighboring components in the neighborhood component set that simultaneously satisfy both time-time overlap and directional overlap is denoted as the number of spatiotemporally consistent components. The number of spatiotemporally consistent components. Divide by the total number of neighborhood component sets The spatiotemporal overlap of anomalous features is obtained. This anomalous feature spatiotemporal overlap is a ratio between 0 and 1. A higher value indicates that most components surrounding the target component experienced power generation performance degradation at the same time and location, consistent with external environmental shading. A lower value indicates that the anomaly occurs only in the target component, with surrounding components operating normally, consistent with a hardware failure of the component itself. The anomalous feature spatiotemporal overlap is written into the time series database for subsequent anomaly type determination.

[0057] Preferably, the specific process of generating a processing work order in step S4 includes: If the spatiotemporal overlap of abnormal features is higher than the preset environmental clustering threshold, it is determined to be an environmental occlusion anomaly, and an environmental cleanup work order containing cleanup location guidance is generated. If the spatiotemporal overlap of abnormal features is lower than the preset individual discrete threshold, it is determined to be an equipment-based abnormality, and an equipment maintenance work order containing the maintenance object guidance is generated.

[0058] In one embodiment, a preset environmental clustering threshold and a preset individual discrete threshold are read from a time-series database. In this embodiment, the environmental clustering threshold is set to 0.6, and the individual discrete threshold is set to 0.2. The calculated spatiotemporal overlap of the abnormal features is then compared numerically with the two thresholds mentioned above.

[0059] If the spatiotemporal overlap of the abnormal characteristics is higher than 0.6, the anomaly is determined to be an environmental shading anomaly caused by external environmental obstruction. The logic behind this determination is that when more than 60% of the neighboring modules experience reduced power generation capacity at the same time and location, it indicates that a common external obstruction is affecting the entire area. An environmental cleanup work order is automatically generated. The work order specifies the target photovoltaic module and the set of device numbers for all spatiotemporally consistent modules, and uses the module's relative incident azimuth angle as the cleanup direction guide. The work order indicates the specific deviation angle of the interference source relative to the module array, guiding maintenance personnel to perform vegetation trimming, removal of building debris, or cleaning of accumulated dust at that specific location.

[0060] If the spatiotemporal overlap of the abnormal characteristics is less than 0.2, the anomaly is determined to be a device-related anomaly caused by a hardware defect in the component itself. The logic behind this determination is that when the vast majority of surrounding components are operating normally, and only the target component experiences a significant drop in power generation, the possibility of external environmental influences is ruled out. An equipment maintenance work order is automatically generated. The work order specifies the target photovoltaic module's device number, physical coordinates, and the effective time period of the fault. The work order also specifies the exact start and end times of the anomaly, guiding maintenance personnel to perform infrared scanning of cell microcracks, bypass diode continuity testing, or junction box electrical connection checks on that specific component.

[0061] If the spatiotemporal overlap of abnormal features is between 0.2 and 0.6, the abnormal state is marked as an observational state, and no processing work order is generated temporarily. Further analysis is conducted through the anomaly persistence assessment steps on subsequent dates. The generated environmental cleanup work order or equipment maintenance work order is pushed to the mobile terminals of maintenance personnel via the communication network, and the detailed content of the work order is synchronously written into the time series database for archiving. All data processing processes do not involve manual prediction; the cloud monitoring server automatically completes fault classification and command issuance based on the set quantitative indicators.

[0062] Preferably, the method further includes determining the continuity of historical data, specifically including: Query the historical anomaly source location vector records of the target photovoltaic module within the past preset observation period; If the historical anomaly source azimuth vector records appear continuously within the preset observation period, a persistent fault marker will be added to the processing work order. If the historical anomaly source azimuth vector record only appears on the current day, the anomaly will be marked as an occasional disturbance and stored in the observation queue, and the work order generation operation will not be executed temporarily.

[0063] In one embodiment, a preset observation period is read from a time-series database. In this embodiment, the preset observation period is set to 7 consecutive monitoring days. The cloud monitoring server retrieves all historical anomaly source azimuth vector records stored for the target photovoltaic module within the 7 historical monitoring days prior to the current monitoring day.

[0064] An existence retrieval is performed on the retrieved records. If, within a preset observation period, the frequency of occurrence of historical anomaly source azimuth vector records generated by the target photovoltaic module reaches a preset frequency threshold (in this embodiment, this frequency threshold is set to 5 times), and the range of the module's relative incident azimuth angle in each record (the difference between the maximum and minimum values) is less than 15 degrees, then the anomaly is determined to meet the persistent characteristic. At this time, a persistent fault marker is added to the attribute field of the generated environmental cleanup work order or equipment maintenance work order. This marker indicates that the anomaly source location is fixed, and that appropriate action must be taken for fixed obstructions or permanent hardware damage.

[0065] If the cloud monitoring server finds that the anomaly's location vector is only recorded on the current monitoring day, and not on any of the other six historical monitoring days within the preset observation period, then the anomaly is determined to be an intermittent interference. Intermittent interference is usually caused by temporary moving objects obstructing the view or by momentary electrical disturbances without a time pattern. In this case, the issuance of a work order for handling this anomaly is blocked, the device number of the target photovoltaic module and the current anomaly characteristic data are marked as intermittent interference, and written to a preset observation queue in the time series database. Work order generation is not performed on this type of anomaly to prevent ineffective maintenance work caused by temporary fluctuations.

[0066] The components in the observation queue are tracked on subsequent monitoring days. If no anomaly source orientation vector is generated again within the next three consecutive monitoring days, the cloud monitoring server clears the anomaly record for that component from the observation queue; if the anomaly is triggered again during the observation period, the persistence determination is re-executed. The entire process is automatically run by the cloud monitoring server based on historical data logic in the time series database, ensuring the accuracy of work order issuance.

[0067] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0068] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A photovoltaic power generation monitoring method based on artificial intelligence, characterized in that, The method is applied to a photovoltaic power generation monitoring unit, which includes a data acquisition terminal and a cloud monitoring server connected via a communication network. The data acquisition terminal is connected to a photovoltaic module array and environmental monitoring sensors, and the cloud monitoring server integrates a time-series database and a geographic information system engine. The method includes the following steps: Step S1: In response to the real-time operating data uploaded by the data acquisition terminal, the time-of-use power generation data of the target photovoltaic module is analyzed using the cloud monitoring server, and the total cumulative power generation for the whole day is retrieved through the time series database. Normalization calculation is performed for irradiance fluctuations to construct the actual power generation ratio sequence. Step S2: Perform time-domain cross-correlation matching between the actual power generation ratio sequence and the preset historical benchmark ratio sequence to determine the optimal matching time shift, and generate the aligned benchmark sequence based on the optimal matching time shift; calculate the negative distortion component sequence of the actual power generation ratio sequence relative to the aligned benchmark sequence. Step S3: Identify the effective concave time interval in the negative distortion component sequence; call the geographic information system engine to establish the temporal and spatial mapping relationship, calculate the relative incident azimuth of the component, and generate the anomaly source azimuth vector by combining the effective concave time interval; Step S4: Construct a neighborhood component set based on the azimuth vector of the anomaly source, calculate the spatiotemporal overlap of the anomaly features within the neighborhood component set, determine the anomaly type as either an environmental occlusion anomaly or a device-related anomaly based on the spatiotemporal overlap of the anomaly features, and generate the corresponding processing work order.

2. The photovoltaic power generation monitoring method based on artificial intelligence according to claim 1, characterized in that, Step S1 includes: The effective daytime illumination was divided into several time slices of equal length using a time series database; Calculate the integral value of the time-of-use power generation data of the target photovoltaic module in each time slice, and divide the integral value by the total cumulative power generation for the whole day to obtain the power generation ratio coefficient of the time slice. The power generation ratio coefficients corresponding to all time slices throughout the day are arranged in chronological order to form the actual power generation ratio sequence.

3. The photovoltaic power generation monitoring method based on artificial intelligence according to claim 1, characterized in that, The specific process of generating the aligned reference sequence in step S2 includes: Using a pre-set historical benchmark proportion sequence as a template, a sliding window matching is performed on the actual power generation proportion sequence on the time axis; Calculate the sequence similarity under different sliding steps, convert the sliding step number corresponding to the maximum similarity into time duration, and determine it as the optimal matching time shift amount; The time axis of the preset historical baseline proportion sequence is shifted and corrected according to the optimal matching time shift amount to obtain the aligned baseline sequence.

4. The photovoltaic power generation monitoring method based on artificial intelligence according to claim 1, characterized in that, The specific process of calculating the negative distortion component sequence in step S2 includes: The difference between the actual power generation ratio sequence and the aligned baseline sequence at each time point is calculated to obtain the original residual sequence; Traverse the original residual sequence, retain data points with values ​​less than zero as valid negative distortion values, and set data points with values ​​greater than or equal to zero to zero. The negative distortion component sequence is composed of all valid negative distortion values ​​arranged in chronological order.

5. The photovoltaic power generation monitoring method based on artificial intelligence according to claim 1, characterized in that, The specific process of identifying the effective indentation time interval in step S3 includes: Determine whether there are any consecutive time slice values ​​below a preset anomaly threshold in the negative distortion component sequence. If so, lock the time range corresponding to a consecutive preset number or more time slices and mark it as the effective concave time period interval.

6. The photovoltaic power generation monitoring method based on artificial intelligence according to claim 1, characterized in that, The specific process of generating the anomaly source azimuth vector in step S3 includes: Use a geographic information system engine to obtain the latitude and longitude coordinates of the photovoltaic power generation park and the current monitoring date; Calculate the solar azimuth angles corresponding to the start and end times within the effective concave time interval; Obtain the array installation orientation angle of the target photovoltaic module, calculate the difference between the solar azimuth angle and the array installation orientation angle, and obtain the relative incident azimuth angle of the module; The negative distortion values ​​within the effective indentation time interval are integrated to obtain the integral value representing the severity of the anomaly. The integral value is used as the magnitude of the vector, and the component's relative incident azimuth angle is used as the direction of the vector to construct the azimuth vector of the anomaly source.

7. The photovoltaic power generation monitoring method based on artificial intelligence according to claim 1, characterized in that, The specific process of calculating the spatiotemporal overlap of anomalous features in step S4 includes: Based on the electronic map of the photovoltaic power generation park, with the target photovoltaic module as the geometric center, search for other modules within a preset physical distance radius to form a neighborhood module set; Iterate through the neighbor components in the neighborhood component set and check whether the neighbor components also have anomaly source orientation vectors on the same monitoring day; If they exist, count the number of neighboring components whose azimuth vectors overlap in both time period and direction, calculate the proportion of this number to the total number of neighboring components, and obtain the spatiotemporal overlap of the anomaly features.

8. The photovoltaic power generation monitoring method based on artificial intelligence according to claim 1, characterized in that, The specific process of generating a processing work order in step S4 includes: If the spatiotemporal overlap of abnormal features is higher than the preset environmental clustering threshold, it is determined to be an environmental occlusion anomaly, and an environmental cleanup work order containing cleanup location guidance is generated. If the spatiotemporal overlap of abnormal features is lower than the preset individual discrete threshold, it is determined to be an equipment-based abnormality, and an equipment maintenance work order containing the maintenance object guidance is generated.

9. The photovoltaic power generation monitoring method based on artificial intelligence according to claim 1, characterized in that, The method also includes continuous assessment of historical data, specifically including: Query the historical anomaly source location vector records of the target photovoltaic module within the past preset observation period; If the historical anomaly source azimuth vector records appear continuously within the preset observation period, a persistent fault marker will be added to the processing work order. If the historical anomaly source azimuth vector record only appears on the current day, the anomaly will be marked as an occasional disturbance and stored in the observation queue, and the work order generation operation will not be executed temporarily.

10. A photovoltaic power generation monitoring system based on artificial intelligence, characterized in that, For executing the AI-based photovoltaic power generation monitoring method as described in claim 1, the AI-based photovoltaic power generation monitoring system includes: The data normalization module is used to respond to the real-time operating data uploaded by the data acquisition terminal, use the cloud monitoring server to analyze the time-of-use power generation data of the target photovoltaic module, retrieve the total cumulative power generation of the day through the time series database, perform normalization calculation for irradiance fluctuations, and construct the actual power generation ratio sequence. The distortion extraction module is used to perform time-domain cross-correlation matching between the actual power generation ratio sequence and the preset historical benchmark ratio sequence, determine the optimal matching time shift, generate the aligned benchmark sequence based on the optimal matching time shift, and calculate the negative distortion component sequence of the actual power generation ratio sequence relative to the aligned benchmark sequence. The vector source tracing module is used to identify the effective concave time interval in the negative distortion component sequence; it calls the geographic information system engine to establish a mapping relationship between time and space, calculates the relative incident azimuth of the components, and generates the anomaly source azimuth vector by combining the effective concave time interval; The fault attribution module is used to construct a neighborhood component set based on the azimuth vector of the anomaly source, calculate the spatiotemporal overlap of anomaly features within the neighborhood component set, determine the anomaly type as environmental occlusion anomaly or device-related anomaly based on the spatiotemporal overlap of anomaly features, and generate the corresponding processing work order.

Images