Flexible photovoltaic intelligent monitoring and management method, system and method based on Internet of Things

Abstract

The invention relates to the technical field of photovoltaic power generation, in particular to a flexible photovoltaic intelligent monitoring and management system and method based on the Internet of Things, multi-source heterogeneous data are comprehensively collected through deployed multiple types of Internet of Things sensor nodes, the data are uploaded to a cloud platform after being cleaned and standardized through edge nodes, a big data processing architecture integrated with flow and batch is adopted, and the intelligent monitoring and management system and method based on the Internet of Things are established. The method comprises the following steps: performing real-time analysis and state judgment on a real-time data stream, performing deep batch processing and feature mining on historical data, extracting high-order features such as a performance attenuation trend and an abnormal mode, fusing real-time and historical features, and realizing comprehensive scoring of a health state of a component and accurate prediction of residual life by utilizing a machine learning model. And based on an evaluation result and a preset knowledge base, automatically generating a differentiated precise operation and maintenance instruction, and issuing and executing the differentiated precise operation and maintenance instruction to form closed-loop management. According to the invention, the monitoring depth and breadth of the flexible photovoltaic system are effectively improved, the conversion from passive alarm to active predictive maintenance is realized, and the operation reliability of the system is significantly enhanced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of photovoltaic power generation, in particular to a flexible photovoltaic intelligent monitoring and management method system and method based on Internet of Things. BACKGROUND

[0002] At present, photovoltaic power generation technology is developing rapidly towards lightweight, flexibility and building integration. Flexible photovoltaic modules have the advantages of being bendable, lightweight and suitable for a wide range of applications, and are increasingly widely used on irregular surfaces such as roofs, roofs and curved buildings. However, its flexible nature also brings new challenges: under complex and variable environmental stress, the module is more prone to physical deformation, material fatigue and accelerated performance degradation, which are not common problems for traditional rigid modules, which puts higher requirements on the perception dimension and analysis depth of the monitoring system.

[0003] In the field of monitoring and management technology, traditional monitoring systems are mostly limited to the collection and threshold alarm of basic electrical parameters such as voltage, current and power. Their data processing methods are relatively simple and lack deep fusion and intelligent analysis of multi-source heterogeneous data. In the face of the massive time series data, environmental data and mechanical deformation data generated by flexible photovoltaic systems, traditional methods cannot effectively mine the value of the data and cannot achieve the transition from "passive alarm" to "active prediction". There are obvious deficiencies in device health status evaluation, performance degradation trend prediction and fine operation and maintenance guidance.

[0004] In addition, existing solutions usually regard data collection, processing and decision support as relatively independent links, and fail to form an intelligent system that is optimized in a closed loop. The overall value of the data has not been fully tapped, resulting in operation and maintenance strategies that are often lagging behind the occurrence of faults and lack of foresight. Therefore, the industry urgently needs an innovative method that can deeply integrate Internet of Things perception and big data analysis technology, specifically for the characteristics of flexible photovoltaic systems, to achieve intelligent monitoring and precise management throughout the life cycle. SUMMARY

[0005] To achieve the above-mentioned purpose, the present application provides the following technical solutions: According to the first aspect of the present application, the present application claims a flexible photovoltaic intelligent monitoring and management method based on Internet of Things, comprising: S1, collecting multi-source photovoltaic data in real time through various sensor nodes deployed in the flexible photovoltaic array, and uploading the multi-source photovoltaic data to a cloud platform after cleaning and standardization by an edge computing node through an Internet of Things gateway; S2, the cloud platform receives the multi-source photovoltaic data through a distributed message queue, and uses a stream processing engine to perform real-time analysis, correlation and calculation, generating real-time performance indicators at the module and array levels, and comparing them with preset thresholds to determine instantaneous abnormalities; S3, persist the processed multi-source photovoltaic data to a distributed data lake, periodically perform batch processing analysis on the full volume of historical data in the data lake, and extract high-order features representing component performance degradation, failure modes, and environmental correlations; S4, input the real-time performance indicators and high-order features into a pre-trained ensemble learning model for inference, and output the overall health status score and remaining useful life prediction value of each flexible photovoltaic component and array; S5, dynamically display the health status score and prediction value through a visualization interface, and provide an administrator with a comprehensive view of the operating status; S6, based on the health status score, prediction value, and pre-defined operation and maintenance strategy knowledge base, automatically generate differentiated operation and maintenance instructions, and issue the operation and maintenance instructions to corresponding operation and maintenance personnel or control systems.

[0006] Further, in step S1, further comprising: The edge computing node removes outliers based on the 3σ criterion or the interquartile range method when performing data cleaning and standardization, and uniformly interpolates sensor data of different sampling frequencies to the same timestamp.

[0007] Further, in step S3, further comprising: S3.1, extracting a historical time series data set of the target flexible photovoltaic component in at least one complete operating cycle from the distributed data lake, and dividing it into a training set and a validation set according to time sequence; S3.2, normalizing the power output data in the training set and the validation set to standard test conditions based on physical formulas or neural network models to eliminate the interference of environmental fluctuations on analysis; S3.3, using seasonal trend decomposition or variational mode decomposition algorithm to separate the long-term trend item, seasonal periodic item and residual item from the normalized power time series data; S3.4, nonlinear fitting the long-term trend item obtained by decomposition, calculating the slope, curvature and decay percentage relative to the initial value as the core features representing slow performance degradation; S3.5, analyzing the statistical characteristics of the residual term, calculating kurtosis, skewness, and root mean square value, and using isolation forest or one-class SVM algorithm to identify outliers as potential indication features of sudden failures; S3.6, calculating the time lag cross-correlation between power data and concurrent environmental data, and extracting the maximum correlation coefficient and the corresponding time lag value as the environmental correlation feature.

[0008] Further, in step S4, further comprising: S4.1 For the flexible photovoltaic module to be evaluated, construct a multi-dimensional feature vector, including the real-time performance index, higher-order features, and the cumulative runtime and historical fault record statistics of the module; S4.2, The multidimensional feature vector is input into a pre-trained health assessment model, which outputs a continuous health score between 0 (complete failure) and 1 (novel) based on the integration of gradient boosting decision trees or deep neural networks. S4.3, based on the predefined range of the health score, divide the component status into multiple levels such as "healthy", "attention", "abnormal" and "severe"; S4.4 Input the historical feature data sequence of the entire life cycle of the component into the remaining lifetime prediction model based on temporal convolutional network or long short-term memory network to predict the health score decay trajectory at future time points, and extrapolate it to the time point when the health score is lower than the failure threshold. The difference between the time point and the current time is the predicted remaining effective lifetime.

[0009] Furthermore, the health assessment model and remaining life prediction model are updated using an online incremental learning approach, continuously fine-tuning the model parameters using newly generated validated data, and a model performance monitoring mechanism is provided to trigger a full retraining of the model when the performance drops to a threshold.

[0010] Furthermore, step S6 also includes: The predefined operation and maintenance strategy knowledge base is a knowledge graph based on a graph structure. Nodes include device type, fault mode, symptom characteristics, and operation and maintenance actions. Edges define the relationships and triggering conditions between nodes. The generation of operation and maintenance instructions is accomplished by reasoning and matching on the graph.

[0011] Furthermore, step S6 also includes: The generated operation and maintenance instructions include economic optimization suggestions based on predictive maintenance strategies. These suggestions integrate the predicted remaining lifespan, the cost of performing maintenance, electricity price signals, and power generation forecasts to recommend the maintenance time window with the highest return on investment.

[0012] Furthermore, step S1 also includes: The sensor nodes are networked using a wireless mesh network and powered by energy harvesting from the attached flexible photovoltaic modules.

[0013] Furthermore, the method also includes: S7. Collect the execution results of operation and maintenance instructions and subsequent equipment status data, establish an early warning-feedback closed loop to evaluate the accuracy of early warning and the effectiveness of operation and maintenance, and use the evaluation results to iteratively optimize the preset threshold in step S2, the feature extraction algorithm in step S3, and the evaluation model in step S4. The evaluation of early warning accuracy and operational effectiveness in step S7 is achieved by calculating precision, recall, and F1-score. The optimization process uses an automated machine learning framework to search and tune the model hyperparameters.

[0014] According to a second aspect of the present invention, the present invention claims protection for a flexible photovoltaic intelligent monitoring and management system based on the Internet of Things, comprising: One or more processors; A memory that stores one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the IoT-based flexible photovoltaic intelligent monitoring and management method.

[0015] This invention relates to the field of photovoltaic power generation technology, specifically a flexible photovoltaic intelligent monitoring and management system and method based on the Internet of Things (IoT). It comprehensively collects multi-source heterogeneous data through the deployment of multiple types of IoT sensor nodes. After cleaning and standardization at edge nodes, the data is uploaded to a cloud platform. Employing a stream-batch integrated big data processing architecture, it performs instant analysis and status assessment of real-time data streams, and deep batch processing and feature mining of historical data to extract high-order features such as performance degradation trends and anomaly patterns. By fusing real-time and historical features, and utilizing machine learning models, it achieves a comprehensive score for component health status and accurate prediction of remaining lifespan. Based on the assessment results and a pre-built knowledge base, it automatically generates differentiated and precise operation and maintenance instructions and issues them for execution, forming a closed-loop management system. This invention effectively improves the depth and breadth of monitoring for flexible photovoltaic systems, realizing a shift from passive alarms to proactive predictive maintenance, and significantly enhancing the reliability of system operation. Attached Figure Description

[0016] Figure 1 A flowchart illustrating the workflow of a flexible photovoltaic intelligent monitoring and management method based on the Internet of Things, as claimed in the embodiments of this application; Figure 2 A second flowchart illustrating a flexible photovoltaic intelligent monitoring and management method based on the Internet of Things, as claimed in the embodiments of this application; Figure 3 A third flowchart illustrating a flexible photovoltaic intelligent monitoring and management method based on the Internet of Things (IoT) claimed in this application. Figure 4 This is a structural diagram of a flexible photovoltaic intelligent monitoring and management system based on the Internet of Things, which is claimed in the embodiments of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0018] The terms "first," "second," and "third" in this application are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first," "second," or "third" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified. All directional indications (such as up, down, left, right, front, back, etc.) in the embodiments of this application are only used to explain the relative positional relationships and movements between components in a specific orientation (as shown in the figures). If the specific orientation changes, the directional indications also change accordingly. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] According to the first embodiment of the present invention, referring to Figure 1 This invention claims protection for a flexible photovoltaic intelligent monitoring and management method based on the Internet of Things, comprising: S1: Multi-source photovoltaic data is collected in real time by various sensor nodes deployed on the flexible photovoltaic array, and then the multi-source photovoltaic data is cleaned and standardized by edge computing nodes through the Internet of Things gateway before being uploaded to the cloud platform. S2, the cloud platform receives the multi-source photovoltaic data through a distributed message queue, and uses a stream processing engine to perform real-time parsing, correlation and calculation to generate real-time performance indicators at the component level and array level, and compares them with preset thresholds to make instantaneous anomaly judgments. S3, the multi-source photovoltaic data after stream processing is persistently stored in a distributed data lake, and the full historical data in the data lake is periodically batch-processed and analyzed to extract high-order features characterizing component performance degradation, failure modes and environmental correlations; S4, the real-time performance indicators and high-order features are input into the pre-trained ensemble learning model for inference, and the overall health status score and remaining effective lifetime prediction value of each flexible photovoltaic module and array are output. S5, The health status score and predicted value are dynamically displayed through a visualization interface, and a comprehensive view of the running status is provided to the administrator; S6. Based on the health status score, predicted value, and predefined operation and maintenance strategy knowledge base, automatically generate differentiated operation and maintenance instructions and send the operation and maintenance instructions to the corresponding operation and maintenance personnel or control system.

[0021] In this embodiment, S1 continuously collects multi-source monitoring data, including current, voltage, output power, module surface temperature, ambient temperature and humidity, light intensity, module curvature angle, and structural strain values, through various types of IoT sensor nodes deployed on the flexible photovoltaic module body, combiner box, inverter, and support structure. Each sensor node aggregates the data to the IoT gateway through a self-organizing wireless mesh network. The IoT gateway performs protocol parsing and preliminary packaging of the received heterogeneous data and sends it to the edge computing nodes deployed on-site. The edge computing nodes perform integrity verification on the received data packets, use an outlier detection method based on statistical principles to remove obviously erroneous data, and uniformly convert all valid data into standardized data frames with predefined formats and units. After adding a precise timestamp and a unique device identifier, the data is uploaded to the cloud platform through a wide area network transmission module. In S2, the cloud platform's data access layer receives standardized data streams from edge computing nodes through a distributed message queue service. This message queue acts as a buffer and decoupler. The stream processing engine consumes data from the message queue in real time. First, it associates and combines data from different sensors within the same photovoltaic array based on device identifiers, forming data records at the array level. Then, based on pre-set component physical models and logical rules, it calculates current power generation efficiency, power deviation rate, and derived indicators derived from the difference between backsheet temperature and ambient temperature in real time. Finally, it quickly compares the real-time data and derived indicators with the pre-loaded threshold ranges of normal operating parameters for each device in memory, immediately marking any data record exceeding the threshold as an abnormal state. While performing real-time processing, S3's stream processing engine persists data to a data lake built on a distributed file system. This data lake stores all raw and processed data by time partitioning. Batch processing tasks are periodically launched to perform a full scan and analysis of the historical data stored in the data lake. This analysis process includes: cleaning and smoothing the power output data of individual components over long periods to eliminate the impact of short-term environmental fluctuations, thereby extracting long-term performance degradation trends; analyzing the correlation between power data sequences and historical environmental data sequences to identify key environmental factors affecting power generation efficiency and their lag time; calculating the statistical characteristic values ​​of the data sequences and detecting abnormal patterns in historical data to construct a deep feature set characterizing the health status of equipment. In S4, a machine learning model is constructed. The input feature set of this model integrates the real-time derived indicators generated in step S2 and the historical deep features extracted in step S3, and also includes the cumulative operating time of the equipment. This model is used to comprehensively calculate the current health status of each flexible photovoltaic module and output a quantitative health score. At the same time, based on the historical data sequence of the entire life cycle of the module, a time series prediction model is used to predict the future change trajectory of its health score and predict the time required for its health to decline to a preset critical value, i.e., the remaining effective lifespan. S5 provides the real-time status data from step S2, abnormal alarms, health scores generated in step S4, and remaining life prediction results to the front-end visualization system via a web service interface. The visualization system provides the system operator with a dynamic display interface for the overall operating status of the entire photovoltaic power station, detailed status information of each subarray and key component, using a combination of graphics, charts, lists, and maps. S6 is pre-configured with a knowledge base containing various abnormal modes, fault types, equipment information, and their corresponding maintenance operations. The system matches and infers from the knowledge base based on the health score, remaining life prediction, and real-time abnormal status marked in step S2 generated in step S4. It automatically generates differentiated maintenance suggestion instructions with priority, which clearly point to specific equipment or arrays and include the type of maintenance operation to be performed. Finally, the instructions are sent to the mobile terminals of relevant maintenance personnel or the production execution system through the message push interface.

[0022] Furthermore, step S1 also includes: When the edge computing nodes perform data cleaning and standardization, they remove outliers based on the 3σ criterion or interquartile range method, and interpolate sensor data with different sampling frequencies to the same timestamp.

[0023] In this embodiment, step S1, the outlier detection method based on statistical principles, specifically involves: calculating the average value and standard deviation of the data within a sliding time window, identifying data points that significantly exceed the normal range calculated from the average value and standard deviation as outliers and removing them.

[0024] Furthermore, referring to Figure 2 Step S3 also includes: S3.1 Extract the historical time-series dataset of the target flexible photovoltaic module within at least one complete operating cycle from the distributed data lake, and divide it into training set and validation set according to the time series. S3.2, The power output data in the training set and validation set are normalized to standard test conditions using physical formulas or neural network models to eliminate the interference of environmental fluctuations on the analysis; S3.3, use seasonal trend decomposition or variational mode decomposition algorithm to separate the long-term trend term, seasonal periodic term and residual term from the normalized power time series data; S3.4, perform nonlinear fitting on the long-term trend term obtained from the decomposition, and calculate the slope, curvature and the percentage of decay relative to the initial value, as the core feature characterizing the slow decay of performance. S3.5 Analyze the statistical characteristics of the residual terms, calculate the kurtosis, skewness, and root mean square value, and use an isolated forest or a type of SVM algorithm to identify outliers as indicators of potential sudden failures. S3.6 Calculate the time-delay cross-correlation between power data and concurrent environmental data, and extract the maximum correlation coefficient and the corresponding time-delay value as features of environmental correlation.

[0025] In this embodiment, step S3, which extracts the long-term performance degradation trend features, specifically includes: S3.1 Data Selection and Preparation: Select power output data of the target flexible photovoltaic module from the data lake over a period of at least one year, under similar light and temperature conditions (e.g., a fixed time period at noon every day), to form the first time series dataset; S3.2 Data Normalization: Divide the power output value in the first time series dataset by the rated power value of the component at the time of manufacture to obtain the daily relative efficiency value, forming the second time series dataset, in order to eliminate the absolute numerical differences between different components. S3.3 Trend Fitting: Apply a data smoothing algorithm to the second time series dataset to filter out high-frequency fluctuations and noise, and obtain a smoothed efficiency change curve; S3.4 Trend Quantification: Calculate the slope of the smoothing efficiency change curve over a unit of time, and use this slope value as the core trend feature to measure the speed of performance degradation of the component; S3.5, “Identifying key environmental factors affecting power generation efficiency and their lag time” specifically includes: aligning the module power output data with synchronously collected data on various environmental parameters (including light intensity, ambient temperature, module surface temperature, and wind speed) according to time series; calculating the correlation coefficient between the power data and each environmental parameter data at different time offsets; identifying the time offset that maximizes the correlation coefficient, which is the lag time of the environmental factor's impact on power; and recording the maximum correlation coefficient value to measure the significance of the environmental factor's impact. S3.6, “Calculate the statistical characteristic values ​​of the data sequence and detect abnormal patterns in historical data” specifically includes: calculating the statistical distribution characteristics of the daily power output data sequence, including the skewness and steepness of the data distribution; for power data over a longer period of time, using density estimation methods in unsupervised learning, identifying a few data points that are significantly different from the vast majority of data points in terms of statistical characteristics, and marking the operating states represented by these points as potential abnormal patterns.

[0026] Furthermore, referring to Figure 3 Step S4 also includes: S4.1 For the flexible photovoltaic module to be evaluated, construct a multi-dimensional feature vector, including the real-time performance index, higher-order features, and the cumulative runtime and historical fault record statistics of the module; S4.2, The multidimensional feature vector is input into a pre-trained health assessment model, which outputs a continuous health score between 0 (complete failure) and 1 (novel) based on the integration of gradient boosting decision trees or deep neural networks. S4.3, based on the predefined range of the health score, divide the component status into multiple levels such as "healthy", "attention", "abnormal" and "severe"; S4.4 Input the historical feature data sequence of the entire life cycle of the component into the remaining lifetime prediction model based on temporal convolutional network or long short-term memory network to predict the health score decay trajectory at future time points, and extrapolate it to the time point when the health score is lower than the failure threshold. The difference between the time point and the current time is the predicted remaining effective lifetime.

[0027] In this embodiment, step S4, "using the model to comprehensively calculate the current health status of each flexible photovoltaic module," specifically includes: S4.1 Feature Vector Construction: Construct a feature vector for each component to be evaluated. The vector contains the following elements: the real-time power generation efficiency at the current moment, the deviation rate between the current power and the rated power, the performance degradation slope characteristic calculated by step S3, the maximum correlation coefficient and lag time characteristic with each environmental factor, the frequency of occurrence of historical abnormal patterns, and the cumulative number of operating days of the component. S4.2 Model Inference: The constructed feature vector is input into a gradient boosting decision tree model that has been pre-trained using a large amount of historical data; after receiving the input, the model performs weighted calculations through multiple layers of decision trees and outputs a value between 0 and 1 as the current health score of the component. S4.3 Status Grading: The health score is mapped to several predefined continuous intervals, for example, [0.9, 1.0] is "Excellent", [0.8, 0.9) is "Good", [0.6, 0.8) is "Attention", [0.4, 0.6) is "Abnormal", and below 0.4 is "Severe", thereby generating a qualitative status level; S4.4, “Using a time series prediction model to predict the future trajectory of its health score” specifically includes: using the historical health score sequence of the component over a period of time (obtained through daily calculations) as model input; using a recurrent neural network model that can capture time dependencies to predict the health score of the next moment from the previous moment's health status, and so on recursively to predict the health score value for each day in the future; continuing this prediction process until the predicted health score value is lower than a preset failure threshold (e.g., 0.3) for the first time, and the number of days from the current date to the predicted date is the predicted remaining effective lifespan.

[0028] Furthermore, the health assessment model and remaining life prediction model are updated using an online incremental learning approach, continuously fine-tuning the model parameters using newly generated validated data, and a model performance monitoring mechanism is provided to trigger a full retraining of the model when the performance drops to a threshold.

[0029] In this embodiment, the gradient boosting decision tree model adopts an online update mechanism. Whenever new, operation-verified equipment status data is generated, this new data and its label are added to the training dataset to fine-tune and update the existing model parameters with a small learning rate, without having to retrain completely.

[0030] Furthermore, step S6 also includes: The predefined operation and maintenance strategy knowledge base is a knowledge graph based on a graph structure. Nodes include device type, fault mode, symptom characteristics, and operation and maintenance actions. Edges define the relationships and triggering conditions between nodes. The generation of operation and maintenance instructions is accomplished by reasoning and matching on the graph.

[0031] In this embodiment, the knowledge base is constructed using a graph structure. The node types in the graph include: device type, observable symptoms, diagnosed fault modes, and executable maintenance actions. The edges between nodes define the relationships between them. Fault mode nodes are connected to symptom nodes that caused the fault and maintenance action nodes that resolved the fault. The matching and reasoning process involves finding the path in the graph that best matches the currently observed symptoms and states, thereby recommending corresponding maintenance actions.

[0032] Furthermore, step S6 also includes: The generated operation and maintenance instructions include economic optimization suggestions based on predictive maintenance strategies. These suggestions integrate the predicted remaining lifespan, the cost of performing maintenance, electricity price signals, and power generation forecasts to recommend the maintenance time window with the highest return on investment.

[0033] In this embodiment, when generating the operation and maintenance suggestion instruction in step S6, the system also calls the future electricity price information and weather forecast information provided by the external system; the system calculates the expected power generation loss cost, operation and maintenance action cost, and possible benefits due to changes in electricity price within the time window for executing different operation and maintenance actions, and recommends an operation and maintenance time window suggestion with the lowest total cost or the highest benefit after comprehensive comparison.

[0034] Furthermore, step S1 also includes: The sensor nodes are networked using a wireless mesh network and powered by energy harvesting from the attached flexible photovoltaic modules.

[0035] In this embodiment, the wireless mesh network in step S1 adopts an on-demand distance vector routing protocol, and the sensor nodes dynamically update the routing table according to changes in network topology; the energy harvesting technology refers to the sensor nodes having built-in photovoltaic micro batteries that collect electrical energy from the light environment of the flexible photovoltaic modules they are attached to and charge their own batteries.

[0036] Furthermore, the method also includes: S7. Collect the execution results of operation and maintenance instructions and subsequent equipment status data, establish an early warning-feedback closed loop to evaluate the accuracy of early warning and the effectiveness of operation and maintenance, and use the evaluation results to iteratively optimize the preset threshold in step S2, the feature extraction algorithm in step S3, and the evaluation model in step S4. The evaluation of early warning accuracy and operational effectiveness in step S7 is achieved by calculating precision, recall, and F1-score. The optimization process uses an automated machine learning framework to search and tune the model hyperparameters.

[0037] In this embodiment, in step S7, the system records all issued operation and maintenance instructions, equipment status recovery data after instruction execution, and equipment operation data for a subsequent period of time; it periodically analyzes the degree of consistency between the initial anomaly warning and the final equipment status, and calculates the accuracy and false alarm rate of the warning; using these analysis results, it automatically adjusts the threshold parameters used for real-time comparison in step S2, and optimizes the feature weights on which the machine learning model in step S4 depends.

[0038] The specific method for adjusting the threshold parameters is as follows: if the false alarm rate of a certain type of anomaly is consistently too high, then the corresponding threshold range should be appropriately widened; if the false negative rate is too high, then the corresponding threshold range should be tightened.

[0039] According to a second embodiment of the present invention, referring to Figure 4 This invention claims protection for a flexible photovoltaic intelligent monitoring and management system based on the Internet of Things, comprising: One or more processors; A memory that stores one or more programs, which, when executed by one or more processors, enable the one or more processors to implement the IoT-based flexible photovoltaic intelligent monitoring and management method.

[0040] This embodiment provides a complete implementation case of an IoT-based flexible photovoltaic intelligent monitoring and management system. The system is deployed in a flexible photovoltaic power station in a coastal area. This power station uses lightweight flexible modules installed on a curved structure, and the environmental conditions are complex and variable.

[0041] Various types of monitoring equipment were deployed at key locations throughout the photovoltaic power station. Each photovoltaic string is equipped with an embedded electrical parameter monitoring unit to collect DC current and voltage signals in real time. A waterproof temperature sensing patch is attached to the back of each flexible module to monitor its operating temperature. Environmental monitoring stations are erected around the array, equipped with full-radiation illuminance sensors and temperature and humidity sensors with radiation shields. Miniature strain sensor arrays are installed at the edges of the flexible modules and in stress-concentrated areas of the supporting structure to monitor mechanical deformation. All sensing units form a self-healing mesh network via low-power wireless communication modules, with data converging to a regional gateway device. The gateway device adopts an industrial-grade design and has multi-protocol conversion capabilities, supporting the unified conversion of various industrial protocols such as Modbus and CAN into standard IoT communication protocols. Edge computing nodes employ embedded industrial computers to run data preprocessing programs: First, CRC checks and sequence number checks are performed on data packets to ensure the integrity of data transmission; then, a dynamic threshold model is established for each data stream, setting a reasonable range based on physical probability and historical statistical patterns, and marking and filtering data points that exceed this range for multiple consecutive sampling periods; finally, all valid data is reassembled according to a predefined standardized data format, and a precise timestamp synchronized by the BeiDou satellite navigation system and a globally unique device identifier are added, and then uploaded to the cloud platform data center through a redundantly designed wireless communication network.

[0042] The cloud platform employs a highly available distributed message queue service to receive data streams from edge nodes across various locations. This service boasts horizontal scalability, capable of handling peak traffic of tens of thousands of messages per second. The stream processing engine consumes data from the message queue in real time. First, based on the location encoding information in the device identifier, it performs spatiotemporal alignment and correlation fusion of monitoring data of different physical types within the same photovoltaic array within a sliding time window, forming a complete array operating status record. Subsequently, based on a pre-set physical characteristic model and electrical characteristic formula of the photovoltaic modules, it calculates derived indicators in real time, such as the current power generation efficiency, power output deviation rate, and the difference between the module backsheet temperature and the ambient temperature. Finally, it rapidly compares the real-time collected data and the calculated derived indicators with the threshold range of personalized operating parameters dynamically loaded from the configuration database. Any data record exceeding the preset threshold range is immediately marked with an abnormal status, and a corresponding event message is generated and published to the event bus.

[0043] While performing real-time processing, the stream processing engine persists all data to a data lake built on a distributed file system. The data lake uses a columnar storage format and is partitioned by multiple dimensions such as time, geographical location, and device type. Batch processing tasks are periodically launched to perform a full scan and analysis of the historical data stored in the data lake. This analysis process includes: cleaning and denoising the long-term power output data of individual photovoltaic modules; using time series analysis methods to extract the long-term trend characteristics of their performance degradation; analyzing the spatiotemporal correlation between power generation data and historical environmental monitoring data to identify key environmental factors affecting power generation efficiency and their lag effects; calculating the higher-order statistical feature values ​​of the data sequence; and using pattern recognition technology to detect abnormal operating patterns in historical data to construct a comprehensive deep feature index system characterizing the health status of the equipment.

[0044] A multi-input machine learning evaluation model is constructed. The model's input feature set integrates performance metrics generated from real-time processing and historical features extracted from batch analysis, while also considering factors such as the cumulative operating time of the equipment. This model is used to comprehensively calculate the current health status of each flexible photovoltaic module, outputting a quantitative health score. Based on the module's full lifecycle historical data sequence, time-series prediction techniques are used to predict the future trajectory of its health score, forecasting the time required for its health to decline to a preset critical value, i.e., its remaining effective lifespan, thus providing a basis for preventative maintenance decisions.

[0045] Real-time operating status, abnormal alarm information, health assessment results, and predictive information are provided to the front-end visualization system through standardized data service interfaces. The visualization system adopts a component-based design approach, combining various graphics, charts, lists, and geographic information to provide system operators with a comprehensive display interface showing the overall operating status of the entire photovoltaic power station, the operating status of each subarray, and detailed status information of key components. It supports multi-dimensional and multi-granular data drill-down and analysis.

[0046] A knowledge base system is pre-built, containing various anomaly modes, fault types, equipment information, and corresponding maintenance operations. Based on health scores, remaining lifespan predictions, and real-time anomaly states, the system matches and infers from the knowledge base to automatically generate prioritized, differentiated maintenance recommendations. These recommendations explicitly point to specific devices or arrays and include the suggested maintenance operation type. Finally, the recommendations are sent to the mobile terminals of relevant maintenance personnel or the production management system via various message push interfaces, forming a closed-loop management system.

[0047] During data cleaning at edge computing nodes, a statistical process control-based approach is employed. The arithmetic mean and standard deviation of the temperature data stream within a recent time window are calculated in real time, and the normal fluctuation range is dynamically set to the mean plus or minus three times the standard deviation. When multiple abnormally high data points are received consecutively at a given moment, and these values ​​significantly exceed the reasonable upper limit calculated based on historical statistics, these data points are marked as abnormal sensor data and removed during the data preprocessing stage, without being uploaded to the cloud platform, thus ensuring the reliability of the input data.

[0048] Historical power data of the target flexible photovoltaic modules over their complete operating cycle are selected from the distributed data lake. First, the data is sorted and filtered by time series, retaining a subset of data under specific illumination and temperature conditions to form an initial time-series dataset. The power output value is divided by the module's nominal power value to obtain a normalized relative efficiency value sequence, eliminating individual differences between modules. A digital filtering algorithm is applied to the relative efficiency sequence to suppress the effects of high-frequency fluctuations and random noise, extracting a smoothed efficiency change trend curve. The rate of change of this trend curve is calculated and used as the core characteristic indicator for quantifying the module's performance degradation rate.

[0049] The power output time series of the component is precisely time-aligned with the time series of multiple synchronously acquired environmental parameters. The statistical correlation coefficient between the power data and each environmental parameter data is calculated at different time offsets. The time offset that maximizes the correlation is identified; this offset represents the lag time of the environmental factor's impact on power. The maximum correlation coefficient value is recorded as a quantitative indicator of the degree of influence of the environmental factor.

[0050] Calculate the statistical distribution characteristics of daily power output data sequences, including indices of data distribution asymmetry and concentration. For long-term power data sequences, employ density-based clustering analysis to identify discrete points that are significantly separated from the mainstream data points in the feature space. Mark the operating states corresponding to these points as potential anomaly patterns and record their frequency and intensity characteristics.

[0051] For each photovoltaic module to be evaluated, a multi-dimensional feature vector is constructed, which includes the following elements: the real-time calculated power generation efficiency value, the deviation between the current power output and the rated power, the performance degradation rate characteristics obtained from batch analysis, the maximum correlation coefficient with each environmental factor and the corresponding lag time, the frequency characteristics of historical abnormal patterns, and the cumulative operating time of the module. All these features are standardized to form a complete feature vector.

[0052] The constructed feature vector is input into an ensemble learning model that has been pre-trained using a large amount of historical data. This model consists of multiple basic predictors. After receiving the input features, it undergoes internal multi-level nonlinear transformations and weighted ensemble calculations, outputting a continuous value between zero and one, which serves as the current health score for that component.

[0053] The continuous health score is mapped to multiple predefined non-overlapping intervals. Each interval corresponds to a qualitative status level, such as multiple levels from "excellent" to "severe". Each level is assigned a corresponding color code and handling priority, providing an intuitive basis for operation and maintenance decisions.

[0054] Using the historical health score sequence of this component over a considerable period as model input, a recurrent neural network model capable of capturing long-term time dependencies is employed to predict health scores at multiple future time points in an autoregressive manner, forming a complete health decay trajectory prediction. This prediction process continues until the predicted value first falls below a preset failure threshold. The time length from the current moment to this prediction moment is the predicted remaining effective lifespan.

[0055] The machine learning model is updated and maintained using an online incremental learning mechanism. When new, field-verified equipment status data is generated, the system automatically adds this data and its annotation information to the model's training sample set. This allows for gradual adjustment and optimization of the existing model parameters at a relatively low learning rate, enabling the model to continuously adapt to changes in the system's operating status without frequent complete retraining, thus balancing learning efficiency and system stability.

[0056] The knowledge base system is constructed and stored using graph database technology. The graph contains various types of nodes, representing entities such as device type, observable symptoms, fault modes of diagnosis, and executable maintenance actions. Directed edges between nodes define rich semantic relationships between entities, such as causal relationships, indicative relationships, and processing relationships. The system's matching and reasoning process involves multi-hop queries and pathfinding within the graph structure to find the fault diagnosis path that best matches the currently observed symptoms and states, thereby recommending appropriate maintenance solutions.

[0057] When generating maintenance recommendation instructions, the system also calls external services to provide future electricity price curve information and accurate weather forecast data through standard interfaces. The system comprehensively evaluates the expected power generation loss, maintenance costs, and revenue changes caused by electricity price fluctuations within different maintenance time windows, establishes a multi-objective optimization model, and finally recommends a maintenance time window scheme with optimal overall benefits to maximize economic benefits.

[0058] The sensor nodes employ advanced wireless mesh networking technology, with an on-demand distance vector routing protocol at the network layer. Nodes can dynamically update their routing tables based on network topology changes and link quality, enabling self-organizing and self-healing multi-hop data transmission. Power supply utilizes efficient photovoltaic energy harvesting technology. Each sensor node incorporates built-in photovoltaic microcells, collecting sufficient solar energy from the light environment of its attached flexible photovoltaic modules and charging its internal batteries via power management circuitry, achieving energy self-sufficiency.

[0059] The system establishes a complete feedback optimization mechanism, continuously recording all issued maintenance commands, equipment status recovery data after command execution, and equipment operation data over a subsequent period. It periodically performs statistical analysis on the degree of consistency between initial anomaly warnings and final equipment status, calculating performance indicators such as warning accuracy and false alarm rate. Using these analysis results, the system automatically adjusts threshold parameter settings in real-time monitoring and optimizes the feature weights on which the evaluation model relies, forming a closed loop of continuous improvement.

[0060] Threshold parameter adjustments employ a robust method based on statistical process control. The system continuously monitors the statistical performance indicators of various early warning systems. If the false alarm rate for a certain type of anomaly consistently exceeds an acceptable level, its corresponding threshold range is appropriately widened; conversely, if the false negative rate is too high, the threshold range is tightened accordingly. Adjustments are made in small, gradual steps, with safety boundaries set to ensure the system always operates within a reliable parameter range.

[0061] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces, or indirect coupling or communication connection between apparatuses or units, and may be electrical, mechanical, or other forms.

[0062] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated units described above can be implemented in hardware or as software functional units. The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made based on the description and drawings of this application, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

[0063] The specific embodiments of the invention have been described in detail above, but they are only examples, and this application is not limited to the specific embodiments described above. For those skilled in the art, any equivalent modifications or substitutions to the invention are also within the scope of this application. Therefore, all equivalent changes, modifications, and improvements made without departing from the spirit and principles of this application should be covered within the scope of this application.

Claims

1. A flexible photovoltaic intelligent monitoring and management method based on the Internet of Things, characterized in that, include: S1: Multi-source photovoltaic data is collected in real time by various sensor nodes deployed on the flexible photovoltaic array, and then the multi-source photovoltaic data is cleaned and standardized by edge computing nodes through the Internet of Things gateway before being uploaded to the cloud platform. S2, the cloud platform receives the multi-source photovoltaic data through a distributed message queue, and uses a stream processing engine to perform real-time parsing, correlation and calculation to generate real-time performance indicators at the component level and array level, and compares them with preset thresholds to make instantaneous anomaly judgments. S3, the multi-source photovoltaic data after stream processing is persistently stored in a distributed data lake, and the full historical data in the data lake is periodically batch-processed and analyzed to extract high-order features characterizing component performance degradation, failure modes and environmental correlations. S4, the real-time performance indicators and high-order features are input into the pre-trained ensemble learning model for inference, and the overall health status score and remaining effective lifetime prediction value of each flexible photovoltaic module and array are output. S5, The health status score and predicted value are dynamically displayed through a visualization interface, and a comprehensive view of the running status is provided to the administrator; S6. Based on the health status score, predicted value, and predefined operation and maintenance strategy knowledge base, automatically generate differentiated operation and maintenance instructions and send the operation and maintenance instructions to the corresponding operation and maintenance personnel or control system.

2. The flexible photovoltaic intelligent monitoring and management method based on the Internet of Things according to claim 1, characterized in that, Step S1 also includes: When the edge computing nodes perform data cleaning and standardization, they remove outliers based on the 3σ criterion or interquartile range method, and interpolate sensor data with different sampling frequencies to the same timestamp.

3. The flexible photovoltaic intelligent monitoring and management method based on the Internet of Things according to claim 1, characterized in that, Step S3 also includes: S3.1 Extract the historical time-series dataset of the target flexible photovoltaic module within at least one complete operating cycle from the distributed data lake, and divide it into training set and validation set according to the time series. S3.2, The power output data in the training set and validation set are normalized to standard test conditions using physical formulas or neural network models to eliminate the interference of environmental fluctuations on the analysis; S3.3, use seasonal trend decomposition or variational mode decomposition algorithm to separate the long-term trend term, seasonal periodic term and residual term from the normalized power time series data; S3.4, perform nonlinear fitting on the long-term trend term obtained from the decomposition, and calculate the slope, curvature and the percentage of decay relative to the initial value, as the core feature characterizing the slow decay of performance. S3.5 Analyze the statistical characteristics of the residual terms, calculate the kurtosis, skewness, and root mean square value, and use an isolated forest or a type of SVM algorithm to identify outliers as indicators of potential sudden failures. S3.6 Calculate the time-delay cross-correlation between power data and concurrent environmental data, and extract the maximum correlation coefficient and the corresponding time-delay value as features of environmental correlation.

4. The flexible photovoltaic intelligent monitoring and management method based on the Internet of Things according to claim 1, characterized in that, Step S4 also includes: S4.1 For the flexible photovoltaic module to be evaluated, construct a multi-dimensional feature vector, including the real-time performance index, higher-order features, and the cumulative runtime and historical fault record statistics of the module; S4.2, The multidimensional feature vector is input into a pre-trained health assessment model, which outputs a continuous health score between 0 (complete failure) and 1 (novel) based on the integration of gradient boosting decision trees or deep neural networks. S4.3, based on the predefined range of the health score, divide the component status into multiple levels such as "healthy", "attention", "abnormal" and "severe"; S4.4 Input the historical feature data sequence of the entire life cycle of the component into the remaining lifetime prediction model based on temporal convolutional network or long short-term memory network to predict the health score decay trajectory at future time points, and extrapolate it to the time point when the health score is lower than the failure threshold. The difference between the time point and the current time is the predicted remaining effective lifetime.

5. The flexible photovoltaic intelligent monitoring and management method based on the Internet of Things according to claim 4, characterized in that, The health assessment model and remaining life prediction model are updated using an online incremental learning approach. The model parameters are continuously fine-tuned using newly generated validated data, and a model performance monitoring mechanism is provided to trigger a full retraining of the model when the performance drops to a threshold.

6. The flexible photovoltaic intelligent monitoring and management method based on the Internet of Things according to claim 1, characterized in that, Step S6 also includes: The predefined operation and maintenance strategy knowledge base is a knowledge graph based on a graph structure. Nodes include device type, fault mode, symptom characteristics, and operation and maintenance actions. Edges define the relationships and triggering conditions between nodes. The generation of operation and maintenance instructions is accomplished by reasoning and matching on the graph.

7. A flexible photovoltaic intelligent monitoring and management method based on the Internet of Things according to claim 1 or 6, characterized in that, Step S6 also includes: The generated operation and maintenance instructions include economic optimization suggestions based on predictive maintenance strategies. These suggestions integrate the predicted remaining lifespan, the cost of performing maintenance, electricity price signals, and power generation forecasts to recommend the maintenance time window with the highest return on investment.

8. A flexible photovoltaic intelligent monitoring and management method based on the Internet of Things according to claim 1 or 6, characterized in that, Step S1 also includes: The sensor nodes are networked using a wireless mesh network and powered by energy harvesting from the attached flexible photovoltaic modules.

9. The flexible photovoltaic intelligent monitoring and management method based on the Internet of Things according to claim 1, characterized in that, The method further includes: S7. Collect the execution results of operation and maintenance instructions and subsequent equipment status data, establish an early warning-feedback closed loop to evaluate the accuracy of early warning and the effectiveness of operation and maintenance, and use the evaluation results to iteratively optimize the preset threshold in step S2, the feature extraction algorithm in step S3, and the evaluation model in step S4. The evaluation of early warning accuracy and operational effectiveness in step S7 is achieved by calculating precision, recall, and F1-score. The optimization process uses an automated machine learning framework to search and tune the model hyperparameters.

10. A flexible photovoltaic intelligent monitoring and management system based on the Internet of Things, characterized in that, include: One or more processors; A memory having stored one or more programs, which, when executed by one or more processors, cause the one or more processors to implement a flexible photovoltaic intelligent monitoring and management method based on any one of claims 1 to 9.

Images