Photovoltaic equipment monitoring method and system based on cloud platform computing

Abstract

The present invention discloses a photovoltaic equipment monitoring method and system based on cloud platform computing, which relates to the technical field of photovoltaic equipment monitoring. The method collects operating data and environmental data of photovoltaic equipment, inputs sensor data and image features into a dual-branch network, generates a joint feature vector, trains a local model, aggregates sensor parameters by device type, aggregates image parameters by environmental classification, and fuses to generate a global model. The global model is used to analyze real-time data. The present invention fuses drone images and sensor data, accurately identifies complex faults through dynamic weight allocation, compensates for blind spots in single data source detection, combines dynamic threshold adjustment, and actively warns based on image features to achieve early fault prediction and rapid response. The monitoring strategy is dynamically adjusted using image risk indicators, reduces the frequency of invalid inspections, and balances power generation revenue and maintenance costs. The hierarchical federal aggregation strategy fuses multi-scenario features to improve the generalization ability of the global model.

Description

Technical Field

[0001] The present invention relates to the technical field of photovoltaic equipment monitoring, and in particular to a photovoltaic equipment monitoring method and system based on cloud platform computing. Background Art

[0002] As the global energy mix shifts toward cleaner, low-carbon energy, photovoltaic power generation, a core form of renewable energy, continues to expand its installed capacity. However, photovoltaic power stations are widely distributed across complex geographical environments, posing numerous challenges to equipment operation and maintenance.

[0003] At present, the Chinese invention patent application number CN202410446067.1 discloses a photovoltaic power generation equipment monitoring and management method and system, including: receiving input voltage and current information and output voltage and current information sent by the photovoltaic power generation equipment; determining the configured output power of the photovoltaic power generation equipment based on the output voltage and current information; obtaining the target output power based on the model information; determining the target maximum power point tracking algorithm information of the photovoltaic power generation equipment based on the configured output power and the target output power; generating a control instruction based on the target maximum power point tracking algorithm information and the input voltage and current information; returning the control instruction to the photovoltaic power generation equipment, and the control instruction is used to control the actual output power of the photovoltaic power generation equipment.

[0004] The above technologies have difficulty in deeply integrating drone images and sensor data, dynamically allocating feature weights, and hierarchically optimizing federated learning models, and are therefore ineffective in improving the accuracy of photovoltaic equipment anomaly detection and the real-time nature of maintenance decisions. Summary of the Invention

[0005] The technical problem solved by the present invention is that the existing technology is difficult to deeply integrate drone images and sensor data, dynamically allocate feature weights and hierarchically optimize the federated learning model, and is not effective in improving the accuracy of photovoltaic equipment anomaly detection and the real-time nature of maintenance decisions.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] A photovoltaic equipment monitoring method based on cloud platform computing includes the following steps:

[0008] Step S1, synchronously collect the operating data and environmental data of the photovoltaic equipment, obtain the multispectral image of the photovoltaic panel through the drone, align the spatiotemporal information and pre-process it;

[0009] Step S2: Input the sensor data and image features into the dual-branch network, dynamically assign weights to generate a joint feature vector, and train the local model;

[0010] Step S3: Aggregate sensor parameters by device type, perform aggregate calculations on local monitoring model parameters using a federated learning algorithm, and fuse them to generate a global model.

[0011] Step S4: Analyze real-time data using a global model, output performance anomaly indicators, adjust anomaly thresholds based on image feature dynamics and preset indicator thresholds, and trigger an early warning;

[0012] Step S5: Optimize the monitoring frequency based on the image risk index and issue it for execution;

[0013] The step S1 further includes:

[0014] Collect multispectral image data of the photovoltaic panel surface, align it with the pre-processed data in time and space, and generate a time-space correlation dataset;

[0015] In step S2, feature extraction is performed on the multispectral image data, and the output is image feature data. A dual-branch adaptive spatiotemporal fusion network is used to perform fusion modeling on the preprocessed data and the multispectral image data. The dual branches include an image branch and a sensor branch. The fusion modeling includes:

[0016] For the sensor branch, the time series features and spatial correlation features of the preprocessed data are extracted and output as sensor features;

[0017] For the image branch, the image features of the multispectral image data are extracted and the time dimension of the preprocessed data is aligned;

[0018] Dynamically assign weights of sensor features and image features through a cross-modal attention mechanism to generate a joint feature vector that is input into the local monitoring model;

[0019] The extraction of the image features includes:

[0020] Perform rasterization positioning and segmentation on multispectral image data to extract the local image of a single photovoltaic panel;

[0021] A lightweight convolutional neural network is used to extract the structured feature vectors of dirt coverage, crack detection confidence, and shadow occlusion area ratio of the local image;

[0022] The execution logic of the cross-modal attention mechanism is:

[0023] The weight of the image branch in the joint feature vector is defined according to the dirt coverage and crack severity index in the image features and the corresponding relationship between the preset image features and weights;

[0024] When the ratio of the historical power output in the preprocessed data to the shadow occlusion area in the image features exceeds the preset power threshold and occlusion area threshold at the same time, the fusion modeling is triggered;

[0025] The aggregate calculation of the federated learning algorithm also includes:

[0026] The sensor branch parameters are grouped and aggregated according to the photovoltaic device type and the image branch parameters are grouped and aggregated according to the actual environmental data;

[0027] On the cloud platform, the grouped and aggregated sensor branch parameters and image branch parameters are globally fused to generate a dual-branch collaborative global monitoring model.

[0028] The setting logic of the indicator threshold in step S403 is:

[0029] When the dirt coverage is higher than a preset dirt coverage threshold, increasing the preset indicator threshold by a first percentage;

[0030] When the crack detection confidence is higher than the preset confidence threshold, an early warning signal is triggered and the power threshold condition is ignored.

[0031] Preferably, the step S1 includes the following sub-steps:

[0032] Step S101, retrieving historical operating data and historical environmental data of the photovoltaic device from a database of the edge node, wherein the historical operating data includes historical voltage, historical current, historical power output, and historical temperature parameters of the photovoltaic device, and the historical environmental data includes historical light intensity and historical ambient temperature parameters;

[0033] Step S102 : performing edge preprocessing on the historical operation data and the historical environment data to obtain preprocessed data.

[0034] Preferably, the edge preprocessing includes:

[0035] Outliers and erroneous data are eliminated, interpolation algorithms are used to fill missing data, the collected data are smoothed and filtered to reduce sensor noise interference, and normalized to unify the dimensions.

[0036] Preferably, step S2 includes the following sub-steps:

[0037] Step S201: input the preprocessed data as a training set into the adaptive spatiotemporal neural network model to perform local model training to obtain the operating status characteristics of the photovoltaic equipment;

[0038] Step S202: Perform local adaptive spatiotemporal neural network model training to obtain local monitoring model parameters.

[0039] Preferably, step S3 includes the following sub-steps:

[0040] Step S301: obtaining local monitoring model parameters uploaded by each edge node based on the cloud platform;

[0041] Step S302: performing aggregate calculation on the local monitoring model parameters according to the federated learning algorithm to generate a global monitoring model.

[0042] Preferably, step S4 includes the following sub-steps:

[0043] Step S401, obtaining current actual operation data and actual environmental data of the photovoltaic equipment, inputting the actual operation data and actual environmental data into a global monitoring model, and obtaining predicted performance data;

[0044] Step S402, calculating the difference between the predicted performance data and the actual operating data, and outputting a performance anomaly index;

[0045] Step S403: Compare the performance abnormality index with the preset index threshold:

[0046] If the performance abnormality index is greater than the index threshold, it is determined that the photovoltaic equipment has a performance abnormality risk and an early warning signal is output;

[0047] If the performance abnormality index is less than or equal to the index threshold, the photovoltaic equipment performance is determined to be normal and the analysis result is output;

[0048] The aggregate calculation of the federated learning algorithm includes:

[0049] Each edge node trains a local adaptive spatiotemporal neural network model based on local preprocessed data to obtain local model parameters;

[0050] Based on the cloud platform, local model parameters uploaded by each edge node are aggregated and calculated to obtain global model parameter updates;

[0051] Sending the global model parameters to each edge node;

[0052] Receive the global model parameters and continue to update the model based on local data, iterating multiple rounds until the federated learning algorithm converges.

[0053] Preferably, step S5 includes the following sub-steps:

[0054] Step S501, selecting a monitoring strategy based on the analysis result, wherein the monitoring strategy includes maintaining the current monitoring frequency and increasing the monitoring frequency, and outputting the corresponding monitoring frequency according to the monitoring strategy;

[0055] Step S502, sending the monitoring frequency to the control terminal of the photovoltaic equipment, adjusting the monitoring parameters and performing monitoring;

[0056] The selection logic of the monitoring strategy is:

[0057] If the analysis result indicates that the performance of the photovoltaic equipment is normal, the monitoring strategy is set to maintain the current monitoring frequency and adopt a preset current monitoring frequency value;

[0058] If the analysis result indicates that there is a risk of abnormal performance of the photovoltaic equipment, the monitoring strategy is set to increase the monitoring frequency.

[0059] A photovoltaic equipment monitoring system based on cloud platform computing, including a data acquisition module, a learning and training module, a monitoring aggregation module, an equipment analysis module and a control execution module;

[0060] The data acquisition module is used to synchronously collect the operating data and environmental data of the photovoltaic equipment, obtain multispectral images of the photovoltaic panels through the drone, align the spatiotemporal information and pre-process them;

[0061] The learning and training module is used to input sensor data and image features into a dual-branch network, dynamically assign weights to generate a joint feature vector, and train a local model;

[0062] The monitoring aggregation module is used to aggregate sensor parameters by device type, aggregate image parameters by environment classification, and fuse them to generate a global model;

[0063] The device analysis module is used to analyze real-time data using a global model, dynamically adjust anomaly thresholds based on image features, and trigger warnings;

[0064] The control execution module is used to optimize the monitoring frequency according to the image risk index and issue it for execution.

[0065] The beneficial effects of the present invention are as follows: the present invention integrates drone images and sensor data, accurately identifies complex faults through dynamic weight allocation, compensates for the detection blind spots of a single data source, combines dynamic threshold adjustment, and actively warns based on image features to achieve early fault prediction and rapid response. It uses image risk indicators to dynamically adjust monitoring strategies, reduce the frequency of invalid inspections, and balance power generation revenue and maintenance costs. The hierarchical federal aggregation strategy integrates multi-scenario features to enhance the generalization capability of the global model. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] Figure 1 A flowchart of a photovoltaic equipment monitoring method based on cloud platform computing provided by one embodiment of the present invention;

[0067] Figure 2 A schematic diagram of the basic flow of a photovoltaic equipment monitoring system based on cloud platform computing is provided in accordance with one embodiment of the present invention. DETAILED DESCRIPTION

[0068] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings. It is obvious that the described embodiments are only part of the embodiments of the present invention, but not all of the embodiments.

[0069] Example 1, with reference to Figure 1 , provides a photovoltaic equipment monitoring method based on cloud platform computing, including the following steps:

[0070] Step S1: synchronously collect the operating data and environmental data of the photovoltaic equipment, obtain the multispectral image of the photovoltaic panel through the drone, align the spatiotemporal information and preprocess it.

[0071] In step S2, the sensor data and image features are input into the dual-branch network, weights are dynamically assigned to generate a joint feature vector, and the local model is trained.

[0072] In step S3, sensor parameters are aggregated by device type, local monitoring model parameters are aggregated and calculated according to the federated learning algorithm, and the global model is generated by fusion.

[0073] Step S4: Use the global model to analyze real-time data, output performance anomaly indicators, adjust the anomaly threshold based on the dynamic image features and the preset indicator threshold, and trigger an early warning.

[0074] Step S5: Optimize the monitoring frequency according to the image risk index and issue it for execution.

[0075] Step S1 includes the following sub-steps:

[0076] Step S101, retrieve historical operation data and historical environmental data of the photovoltaic equipment from the database of the edge node, the historical operation data includes historical voltage, historical current, historical power output and historical temperature parameters of the photovoltaic equipment, and the historical environmental data includes historical light intensity and historical ambient temperature parameters.

[0077] Step S101 retrieves operating data such as voltage, current, and power output to quantify the historical performance degradation trajectory of the device. Combined with environmental data such as light intensity and ambient temperature, a correlation between device performance and external conditions is established. Data is directly obtained from edge nodes to reduce cloud transmission delays, support preprocessing operations with high real-time requirements, reuse historical data to avoid repeated collection, and reduce data acquisition costs.

[0078] Step S102 : performing edge preprocessing on the historical operation data and the historical environment data to obtain preprocessed data.

[0079] Step S102 filters outliers caused by sensor failure or communication interference to avoid model overfitting noise, performs linear / spline interpolation on missing data to ensure time series continuity, uses moving average or Kalman filtering to suppress sensor noise, improves data signal-to-noise ratio, maps data of different dimensions such as voltage and current to the [0,1] interval, eliminates the impact of dimensional differences on model training, and the preprocessed data can be directly input into the neural network to accelerate model convergence.

[0080] Step S1 further includes:

[0081] Multispectral image data of the photovoltaic panel surface is collected and spatially aligned with the preprocessed data to generate a spatiotemporal correlation dataset.

[0082] Edge preprocessing includes:

[0083] Outliers and erroneous data are eliminated, interpolation algorithms are used to fill missing data, the collected data are smoothed and filtered to reduce sensor noise interference, and normalized to unify the dimensions.

[0084] The multispectral images of drones are used to capture physical defects that cannot be perceived by sensors, supplement equipment status information, and implement raster positioning and segmentation to achieve refined analysis of single photovoltaic panels, locate the specific location of the fault, synchronize the image acquisition time with the sensor data timestamp, and correlate the temporal relationship between power fluctuations and dirt accumulation. Through GPS coordinates or image positioning technology, the photovoltaic panel area in the image is matched with the spatial position of the sensor node, establishing a three-dimensional association between equipment, area, and environment, and generating a spatiotemporal correlation data set, providing a cross-modal fusion foundation for the subsequent dual-branch network.

[0085] Step S1 fuses sensor historical data with drone image data to cover the multi-dimensional causes of device performance degradation, eliminates the time offset and spatial misalignment of heterogeneous data through spatiotemporal alignment, ensures the accuracy of cross-modal feature association, eliminates noise and anomalies through edge preprocessing, improves data availability, and provides a reliable foundation for model training.

[0086] Step S2 includes the following sub-steps:

[0087] In step S201 , the pre-processed data is input as a training set into an adaptive spatiotemporal neural network model to perform local model training and obtain the operating status characteristics of the photovoltaic equipment.

[0088] Step S202: Perform local adaptive spatiotemporal neural network model training to obtain local monitoring model parameters.

[0089] Steps S201 and S202 identify abnormal power output trends through the temporal variation patterns of the adaptive spatiotemporal neural network model, use graph neural networks to model device spatial associations, locate fault diffusion paths, and generate local monitoring model parameters containing device operating status characteristics. Adaptive network structures are used to achieve low-latency reasoning at edge nodes to meet real-time monitoring needs.

[0090] In step S2, feature extraction is performed on the multispectral image data, and the output is image feature data. A dual-branch adaptive spatiotemporal fusion network is used to perform fusion modeling on the preprocessed data and the multispectral image data. The dual branches include an image branch and a sensor branch. The fusion modeling includes:

[0091] For the sensor branch, the time series features and spatial correlation features of the preprocessed data are extracted and output as sensor features.

[0092] For the image branch, image features of multispectral image data are extracted and the time dimension of the preprocessed data is aligned.

[0093] The weights of sensor features and image features are dynamically assigned through the cross-modal attention mechanism to generate a joint feature vector that is input into the local monitoring model.

[0094] Image feature extraction includes:

[0095] The multispectral image data is rasterized, positioned and segmented to extract the local image of a single photovoltaic panel.

[0096] A lightweight convolutional neural network is used to extract the structured feature vectors of dirt coverage, crack detection confidence and shadow occlusion area ratio of the local image.

[0097] Multispectral image feature extraction divides the entire photovoltaic array into single-board-level areas, locates the specific fault location, extracts structured feature vectors, compresses image information into a low-dimensional feature space, aligns image feature timestamps with sensor data, establishes a temporal association between power fluctuations and dirt accumulation, and provides a data foundation for cross-modal fusion.

[0098] The execution logic of the cross-modal attention mechanism is:

[0099] The weight of the image branch in the joint feature vector is defined according to the dirt coverage and crack severity indicators in the image features and the corresponding relationship between the preset image features and weights.

[0100] When the ratio of the historical power output in the preprocessed data to the shadow occlusion area in the image features exceeds the preset power threshold and occlusion area threshold at the same time, the fusion modeling is triggered.

[0101] The sensor branch extracts temporal and spatial features to generate a representation of the equipment's operating status;

[0102] The image branch outputs structured indicators such as dirt coverage and crack confidence to quantify invisible physical defects.

[0103] According to the correspondence between the dirt coverage and crack severity indicators in the image features and the preset image features and weights, the influence of visual features on the model output is strengthened;

[0104] When the ratio of the historical power output in the preprocessed data to the shadow occlusion area in the image features exceeds the preset power threshold and occlusion area threshold at the same time, cross-modal attention is activated to jointly analyze the causal relationship between occlusion and performance degradation.

[0105] Step S2 uses a dual-branch adaptive spatiotemporal fusion network and a dynamic weight allocation mechanism to achieve deep collaborative modeling of sensor data and drone image data, and combines federated learning to optimize model parameters, significantly improving the accuracy and environmental adaptability of photovoltaic equipment status monitoring.

[0106] Multimodal feature fusion breaks through the limitations of a single data source, combines sensor timing features with image visual features, and realizes macro and micro joint diagnosis of device performance degradation. It adaptively adjusts data weights through a cross-modal attention mechanism, prioritizes high-confidence features to correct sensor misjudgments, aggregates model parameters by device type and environment, solves data heterogeneity problems, and generates a global monitoring model with stronger generalization capabilities.

[0107] Step S3 includes the following sub-steps:

[0108] Step S301: Obtain local monitoring model parameters uploaded by each edge node based on the cloud platform.

[0109] In step S301, the edge node only uploads the model gradient or parameters, and the original sensor data and image data are retained locally to avoid the leakage of sensitive information. This meets the compliance requirements for local storage of photovoltaic power station data. The cloud platform unifies the parameter format, supports model access of heterogeneous edge devices, and filters abnormal updates through the parameter verification mechanism to ensure aggregation stability.

[0110] Step S302: performing aggregate calculation on the local monitoring model parameters according to the federated learning algorithm to generate a global monitoring model.

[0111] The aggregate calculations of the federated learning algorithm include:

[0112] Each edge node trains a local adaptive spatiotemporal neural network model based on local preprocessed data to obtain local model parameters.

[0113] Based on the cloud platform, the local model parameters uploaded by each edge node are aggregated and calculated to obtain the global model parameter update.

[0114] Send global model parameters to each edge node.

[0115] Receive the global model parameters and continue to update the model based on local data, iterating multiple rounds until the federated learning algorithm converges.

[0116] The aggregate calculation of the federated learning algorithm also includes:

[0117] The sensor branch parameters are grouped and aggregated according to the photovoltaic device type, and the image branch parameters are grouped and aggregated according to the actual environmental data.

[0118] The grouped and aggregated sensor branch parameters and image branch parameters are globally fused on the cloud platform to generate a dual-branch collaborative global monitoring model.

[0119] Through dynamic weight allocation, data heterogeneity conflicts are resolved, and sensor misjudgments are corrected by prioritizing image features.

[0120] Sub-step S302 aggregates parameters by photovoltaic device type, retains the differences in the response characteristics of different materials to temperature and light, aggregates parameters by environmental classification, and adapts to the differentiated image feature distribution. During global fusion, the branch weights are dynamically adjusted according to the device type and environmental classification to resolve data heterogeneity conflicts. When the local model of an edge node has higher accuracy in a specific scenario, its parameters are given higher weight in the aggregation. Through multiple rounds of iteration, the model converges to the global optimum.

[0121] Traditional federated learning suffers from reduced accuracy in heterogeneous device scenarios. However, this method controls the accuracy loss to a lower range through group aggregation and dynamic weighting. It uses asynchronous federated learning, allowing edge nodes to upload parameters at the local training speed, reducing the time spent waiting for synchronization. For edge nodes that are offline or have abnormal parameters, historical parameter weighting is used to fill in the gaps, ensuring the continuity of global model training.

[0122] Step S3 uses a federated learning framework to achieve collaborative model optimization across edge nodes. While protecting data privacy, it generates a global monitoring model that is both device type adaptable and environmentally adaptable, significantly improving the model's generalization capability and cross-scenario diagnostic accuracy. It avoids uploading raw data and only aggregates model parameters to meet the data isolation requirements of photovoltaic power stations. It retains device type and environmental differences through grouping and aggregation strategies, solves the problem of inconsistent data distribution, adjusts the parameter weights of sensor and image branches in global fusion, and enhances the contribution of high-value features to model output.

[0123] Step S4 includes the following sub-steps:

[0124] Step S401 : obtaining the actual operation data and actual environment data of the photovoltaic equipment, inputting the actual operation data and actual environment data into a global monitoring model, and obtaining predicted performance data.

[0125] Step S401 inputs the current sensor data and image features, and generates predicted performance data through a dual-branch global model. The global model has integrated multi-environmental data through federated learning, and automatically adjusts the weight of the impact of temperature on power in high temperature and high humidity environments to improve prediction accuracy.

[0126] Step S402: Calculate the difference between the predicted performance data and the actual operation data, and output a performance anomaly index.

[0127] Step S402 calculates the difference between the predicted performance data and the actual data, locates the source of performance degradation, combines the contribution of the sensor branch and the image branch, outputs a difference attribution report, and maps the difference value to a performance anomaly indicator with a unified dimension to facilitate cross-device and cross-scenario comparative analysis.

[0128] Step S403: Compare the performance abnormality index with the preset index threshold:

[0129] If the performance abnormality index is greater than the index threshold, it is determined that the photovoltaic equipment has a performance abnormality risk and an early warning signal is output.

[0130] If the performance abnormality index is less than or equal to the index threshold, the photovoltaic equipment performance is determined to be normal and the analysis result is output.

[0131] The logic for setting the indicator threshold in step S403 is:

[0132] When the dirt coverage is higher than a preset dirt coverage threshold, the preset indicator threshold is increased by a first percentage.

[0133] When the crack detection confidence is higher than the preset confidence threshold, an early warning signal is triggered and the power threshold condition is ignored.

[0134] Step S403: when the dirt coverage is higher than a preset dirt coverage threshold, the indicator threshold is increased by a first percentage from the reference value to prevent the progressive power reduction caused by dirt from being ignored;

[0135] When the crack detection confidence level is higher than the preset confidence threshold, an early warning is triggered directly to prevent the risk of fire caused by panel breakage.

[0136] The warning level is divided according to the abnormal index value, and the monitoring frequency adjustment strategy of step S5 is linked.

[0137] Traditional fixed thresholds can cause long-term power deviations in dusty environments due to dirt accumulation. Dynamic thresholds, on the other hand, are corrected by dirt coverage, reducing warning delays. The crack detection confidence level triggers an "unconditional warning," shortening the response time to panel breakage accidents and avoiding significant losses. For monocrystalline silicon equipment and thin-film equipment, the weights of indicators related to temperature and light are dynamically adjusted to improve anomaly detection accuracy.

[0138] Step S4 uses a global monitoring model and a dynamic threshold mechanism to achieve real-time and accurate performance assessment and risk prediction of photovoltaic equipment. It also dynamically corrects the warning logic based on drone image features, significantly reducing the missed alarm rate and false alarm rate.

[0139] Based on the global model output, the predicted performance data is quantified to quantify the real-time efficiency attenuation trend of the equipment. By calculating the difference between the predicted and actual data, a highly interpretable performance anomaly indicator is generated. The indicator threshold is adaptively corrected according to the image features to solve the failure problem of traditional fixed thresholds in complex scenarios. An "unconditional warning" is triggered for fatal defects such as cracks to avoid maintenance delays due to power fluctuations not reaching the threshold.

[0140] Step S5 includes the following sub-steps:

[0141] Step S501 : selecting a monitoring strategy according to the analysis result. The monitoring strategy includes maintaining the current monitoring frequency and increasing the monitoring frequency, and outputting the corresponding monitoring frequency according to the monitoring strategy.

[0142] Maintain low-frequency monitoring under normal conditions to reduce the communication pressure between edge nodes and cloud platforms. Under abnormal conditions, high-frequency monitoring is used to capture transient anomalies and shorten detection delays. Monitoring levels are divided according to performance anomaly indicator values ​​to match operation and maintenance needs of different urgency levels.

[0143] Step S502: Send the monitoring frequency to the control terminal of the photovoltaic device, adjust the monitoring parameters and perform monitoring.

[0144] The selection logic of the monitoring strategy is:

[0145] If the analysis results indicate that the performance of the photovoltaic equipment is normal, the monitoring strategy is set to maintain the current monitoring frequency and adopt the preset current monitoring frequency value.

[0146] If the analysis results indicate that there is a risk of abnormal performance of the photovoltaic equipment, the monitoring strategy is set to increase the monitoring frequency.

[0147] Step S502 sends the adjusted frequency parameters to the control end, dynamically modifies the sensor sampling interval and image acquisition period, and synchronously adjusts the monitoring threshold and data reporting granularity to further optimize bandwidth usage.

[0148] Step S5 achieves a precise balance between resource consumption and risk response by dynamically monitoring the frequency adjustment mechanism, optimizing operation and maintenance costs while ensuring equipment safety.

[0149] When the equipment is operating normally, the monitoring frequency is reduced to reduce data storage and computing load. When anomalies occur, the monitoring density is increased to shorten the detection time window of the fault deterioration cycle. The strategy is automatically adjusted based on real-time analysis results without manual intervention. Dynamic strategies are used to reduce ineffective monitoring, extend equipment life, and reduce power generation losses caused by unplanned downtime.

[0150] Example 2, reference Figure 2 , provides a photovoltaic equipment monitoring system based on cloud platform computing, including a data acquisition module, a learning and training module, a monitoring aggregation module, an equipment analysis module and a control execution module.

[0151] The data acquisition module is used to synchronously collect the operating data and environmental data of photovoltaic equipment, obtain multispectral images of photovoltaic panels through drones, align the spatiotemporal information and preprocess it.

[0152] The learning and training module is used to input sensor data and image features into the dual-branch network, dynamically assign weights to generate a joint feature vector, and train the local model.

[0153] The monitoring aggregation module is used to aggregate sensor parameters by device type, aggregate image parameters by environment classification, and fuse them to generate a global model.

[0154] The device analysis module is used to analyze real-time data using a global model, dynamically adjust abnormal thresholds based on image features, and trigger early warnings.

[0155] The control execution module is used to optimize the monitoring frequency according to the image risk indicators and issue execution.

[0156] By spatiotemporally aligning drone multispectral imagery with sensor data, a spatiotemporal correlation dataset covering equipment operating parameters and environmental data was constructed, overcoming the limitations of a single data source. Dual-branch network collaborative modeling: Joint modeling of the image and sensor branches enables correlation analysis from micro-defects to macro-performance. Dynamic weights are assigned to the image and sensor branches based on the real-time status of image features and sensor data. Abnormality thresholds are dynamically adjusted based on image risk indicators. Sensor parameters are aggregated by device type and image parameters by environmental classification, preserving data characteristics across different scenarios. Global fusion of these grouped parameters is performed on the cloud platform to generate a dual-branch collaborative model that accounts for differences in device type and environment. Compared to traditional federated learning, this approach reduces model bias. Based on model analysis results, the monitoring frequency of high-risk equipment is increased, while the sampling frequency of stable equipment is reduced, achieving efficient resource allocation. Joint early warning of image features and sensor data allows for early prediction of potential failures, avoiding unplanned downtime. Edge preprocessing and noise reduction: Outlier removal, interpolation, and normalization are performed at the edge nodes, reducing the computing load on the cloud platform and improving data quality. When local model parameters are aggregated on the cloud platform, there is no need to upload the original data, which meets the data privacy requirements of photovoltaic power stations and reduces communication bandwidth requirements.

[0157] This system deploys edge processing technology on photovoltaic equipment to clean, complete, filter, and normalize collected operational and environmental data, effectively improving data transmission efficiency and quality. It uses adaptive spatiotemporal neural network modeling to comprehensively analyze temporal changes and spatial correlations to achieve intelligent prediction of equipment performance. Based on a federated learning mechanism, each site shares model knowledge without uploading raw data, improving overall prediction accuracy. Combined with game optimization methods, it dynamically adjusts monitoring frequency when performance anomalies occur, achieving an optimal balance between operation and maintenance costs and power generation revenue. It also introduces remaining life prediction technology to generate maintenance plans in advance and intelligently optimize inspection windows. The overall system builds an integrated photovoltaic equipment monitoring system that integrates edge and cloud collaboration, dynamic learning, and intelligent decision-making, featuring high real-time performance, high reliability, and strong scalability.

[0158] Those skilled in the art will appreciate that embodiments of the present invention may be provided as methods, systems, or computer program products. Therefore, the present invention may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention may take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium may be implemented by any type of volatile or non-volatile storage device, or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0159] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A photovoltaic equipment monitoring method based on cloud platform computing, characterized in that: The steps include: Step S1, synchronously collect the operating data and environmental data of the photovoltaic equipment, obtain the multispectral image of the photovoltaic panel through the drone, align the spatiotemporal information and pre-process it; Step S2: Input the sensor data and image features into the dual-branch network, dynamically assign weights to generate a joint feature vector, and train the local model; Step S3: Aggregate sensor parameters by device type, perform aggregate calculations on local monitoring model parameters using a federated learning algorithm, and fuse them to generate a global model. Step S4: Analyze real-time data using a global model, output performance anomaly indicators, adjust anomaly thresholds based on image feature dynamics and preset indicator thresholds, and trigger an early warning; Step S5: Optimize the monitoring frequency based on the image risk index and issue it for execution; The step S1 further includes: Collect multispectral image data of the photovoltaic panel surface, align it with the pre-processed data in time and space, and generate a time-space correlation dataset; In step S2, feature extraction is performed on the multispectral image data, and the output is image feature data. A dual-branch adaptive spatiotemporal fusion network is used to perform fusion modeling on the preprocessed data and the multispectral image data. The dual branches include an image branch and a sensor branch. The fusion modeling includes: For the sensor branch, the time series features and spatial correlation features of the preprocessed data are extracted and output as sensor features; For the image branch, the image features of the multispectral image data are extracted and the time dimension of the preprocessed data is aligned; Dynamically assign weights of sensor features and image features through a cross-modal attention mechanism to generate a joint feature vector that is input into the local monitoring model; The extraction of the image features includes: Perform rasterization positioning and segmentation on multispectral image data to extract the local image of a single photovoltaic panel; A lightweight convolutional neural network is used to extract the structured feature vectors of dirt coverage, crack detection confidence, and shadow occlusion area ratio of the local image; The execution logic of the cross-modal attention mechanism is: The weight of the image branch in the joint feature vector is defined according to the dirt coverage and crack severity index in the image features and the corresponding relationship between the preset image features and weights; When the ratio of the historical power output in the preprocessed data to the shadow occlusion area in the image features exceeds the preset power threshold and occlusion area threshold at the same time, the fusion modeling is triggered; The aggregate calculation of the federated learning algorithm also includes: The sensor branch parameters are grouped and aggregated according to the photovoltaic device type and the image branch parameters are grouped and aggregated according to the actual environmental data; The grouped and aggregated sensor branch parameters and image branch parameters are globally fused on the cloud platform to generate a dual-branch collaborative global monitoring model.

2. A photovoltaic equipment monitoring method based on cloud platform computing according to claim 1, characterized in that: The step S1 includes the following sub-steps: Step S101, retrieving historical operating data and historical environmental data of the photovoltaic device from a database of the edge node, wherein the historical operating data includes historical voltage, historical current, historical power output, and historical temperature parameters of the photovoltaic device, and the historical environmental data includes historical light intensity and historical ambient temperature parameters; Step S102 : performing edge preprocessing on the historical operation data and the historical environment data to obtain preprocessed data.

3. A photovoltaic equipment monitoring method based on cloud platform computing according to claim 2, characterized in that: The edge preprocessing includes: Outliers and erroneous data are eliminated, interpolation algorithms are used to fill missing data, the collected data are smoothed and filtered to reduce sensor noise interference, and normalized to unify the dimensions.

4. A photovoltaic equipment monitoring method based on cloud platform computing according to claim 3, characterized in that: The step S2 includes the following sub-steps: Step S201: input the preprocessed data as a training set into the adaptive spatiotemporal neural network model to perform local model training to obtain the operating status characteristics of the photovoltaic equipment; Step S202: Perform local adaptive spatiotemporal neural network model training to obtain local monitoring model parameters.

5. A photovoltaic equipment monitoring method based on cloud platform computing according to claim 4, characterized in that: The step S3 includes the following sub-steps: Step S301: obtaining local monitoring model parameters uploaded by each edge node based on the cloud platform; Step S302: performing aggregate calculation on the local monitoring model parameters according to the federated learning algorithm to generate a global monitoring model.

6. A photovoltaic equipment monitoring method based on cloud platform computing according to claim 5, characterized in that: The step S4 includes the following sub-steps: Step S401, obtaining current actual operation data and actual environmental data of the photovoltaic equipment, inputting the actual operation data and actual environmental data into a global monitoring model, and obtaining predicted performance data; Step S402, calculating the difference between the predicted performance data and the actual operating data, and outputting a performance anomaly index; Step S403: Compare the performance abnormality index with the preset index threshold: If the performance abnormality index is greater than the index threshold, it is determined that the photovoltaic equipment has a performance abnormality risk and an early warning signal is output; If the performance abnormality index is less than or equal to the index threshold, the photovoltaic equipment performance is determined to be normal and the analysis result is output; The aggregate calculation of the federated learning algorithm includes: Each edge node trains a local adaptive spatiotemporal neural network model based on local preprocessed data to obtain local model parameters; Based on the cloud platform, local model parameters uploaded by each edge node are aggregated and calculated to obtain global model parameter updates; Sending the global model parameters to each edge node; Receive global model parameters and continue to update the model based on local data, iterating multiple rounds until the federated learning algorithm converges; The setting logic of the indicator threshold in step S403 is: When the dirt coverage is higher than a preset dirt coverage threshold, increasing the preset indicator threshold by a first percentage; When the crack detection confidence is higher than the preset confidence threshold, an early warning signal is triggered and the power threshold condition is ignored.

7. A photovoltaic equipment monitoring method based on cloud platform computing according to claim 6, characterized in that: The step S5 includes the following sub-steps: Step S501: selecting a monitoring strategy based on the analysis result, wherein the monitoring strategy includes maintaining the current monitoring frequency and increasing the monitoring frequency, and outputting the corresponding monitoring frequency according to the monitoring strategy; Step S502, sending the monitoring frequency to the control terminal of the photovoltaic equipment, adjusting the monitoring parameters and performing monitoring; The selection logic of the monitoring strategy is: If the analysis result indicates that the performance of the photovoltaic equipment is normal, the monitoring strategy is set to maintain the current monitoring frequency and adopt a preset current monitoring frequency value; If the analysis result indicates that there is a risk of abnormal performance of the photovoltaic equipment, the monitoring strategy is set to increase the monitoring frequency.

8. A photovoltaic equipment monitoring system based on cloud platform computing, which is applied to a photovoltaic equipment monitoring method based on cloud platform computing as described in any one of claims 1 to 7, characterized in that: It includes data acquisition module, learning and training module, monitoring aggregation module, equipment analysis module and control execution module; The data acquisition module is used to synchronously collect the operating data and environmental data of the photovoltaic equipment, obtain multispectral images of the photovoltaic panels through the drone, align the spatiotemporal information and pre-process them; The learning and training module is used to input sensor data and image features into a dual-branch network, dynamically assign weights to generate a joint feature vector, and train a local model; The monitoring aggregation module is used to aggregate sensor parameters by device type, aggregate image parameters by environment classification, and fuse them to generate a global model; The device analysis module is used to analyze real-time data using a global model, dynamically adjust anomaly thresholds based on image features, and trigger warnings; The control execution module is used to optimize the monitoring frequency according to the image risk index and issue it for execution.

Images