Method and system for improving power supply quality of photovoltaic transformer area

Abstract

The invention discloses a method and system for improving the power supply quality of a photovoltaic transformer area. The method comprises the following steps: packaging a functional module into an independent micro-service by utilizing a Docker container technology, and configuring an automatic capacity expansion and contraction strategy; collecting and preprocessing multi-source data; performing feature analysis and decomposition on the processing task to realize optimal scheduling of computing resources; identifying and classifying the power supply quality scene by adopting a machine learning method; and establishing a multi-dimensional power supply quality evaluation index system according to a classification result, and performing evaluation and attribution analysis. According to the invention, the problems of low data processing efficiency and lack of refined analysis capability in the existing photovoltaic area power supply quality management are solved, and efficient power supply quality evaluation and attribution are realized.

Description

Technical Field

[0001] This application relates to the field of power distribution network management technology, and in particular to a method and system for improving the power supply quality of photovoltaic power distribution areas. Background Technology

[0002] Photovoltaic distribution areas in the power distribution network are areas supplied by distribution transformers that integrate photovoltaic power generation equipment, and are an important component of energy transition and green development. With the continuous increase in photovoltaic power generation penetration, the power supply quality management of photovoltaic distribution areas is becoming increasingly complex and important.

[0003] Currently, common photovoltaic power supply quality management mainly relies on the monitoring and analysis of basic electrical quantities such as voltage and current. For example, by deploying smart meters and power quality monitoring equipment at key nodes, indicators such as voltage deviation and three-phase imbalance are collected to achieve power supply quality assessment and early warning.

[0004] More advanced technical solutions employ multi-source data fusion to comprehensively analyze electrical quantity data, meteorological data, and electricity load data. This approach establishes data analysis models to analyze the impact of factors such as photovoltaic output and load changes on power supply quality, thereby providing decision support for power supply quality management.

[0005] However, existing technologies suffer from the following problems: Firstly, data processing efficiency is low, especially when dealing with massive amounts of heterogeneous data, traditional centralized processing methods struggle to meet real-time requirements; secondly, they lack the ability to perform refined analysis for different scenarios, making it impossible to accurately identify and analyze power quality issues under different operating conditions. These problems limit the effective management and improvement of power quality in photovoltaic power distribution areas. Summary of the Invention

[0006] In view of this, this application provides a method and system for improving the power supply quality of photovoltaic power distribution areas, which solves the problems of low data processing efficiency and lack of refined analysis capabilities in the prior art.

[0007] This application provides a method for improving the power supply quality of a photovoltaic power distribution area, including: By using Docker container technology, the data acquisition, preprocessing, and analysis modules are encapsulated as independent microservices, and an automatic scaling strategy is configured to obtain a decoupled management architecture for the data processing pipeline. Collect electrical quantity data, meteorological data, and equipment status data; clean, standardize, and assess the quality of the data to obtain preprocessed data. The preprocessed data processing tasks are characterized and decomposed, a task mapping optimization model is established, and tasks are allocated to heterogeneous computing resources to achieve optimal scheduling of computing resources. Using the preprocessed data and the task mapping optimization model, machine learning methods are employed to identify and classify power quality scenarios, thereby obtaining scenario classification results. Based on the scenario classification results, a multi-dimensional power supply quality assessment index system is established to conduct power supply quality assessment and problem attribution analysis, and output a power supply quality analysis report.

[0008] According to one embodiment of this application, the decoupled management architecture for obtaining the data processing pipeline includes: Build a message queue system, set up raw data topics and preprocessed data topics, and obtain the data stream processing architecture; Using the aforementioned data stream processing architecture, a topic subscription model is adopted to realize inter-service communication, forming a service communication mechanism; Based on the service communication mechanism, a unified service gateway layer is established to complete the decoupled management of the data processing pipeline.

[0009] According to one embodiment of this application, the identification and classification of power supply quality scenarios to obtain scenario classification results includes: Using the preprocessed data, feature extraction calculations are performed to construct statistical features and temporal features, thereby obtaining feature vectors; Using the aforementioned feature vectors, an unsupervised learning method is applied to perform coarse-grained classification of the running states, resulting in preliminary classification results. Based on the preliminary classification results, a classifier is trained to output scene recognition results for photovoltaic fluctuations and load abrupt changes.

[0010] According to one embodiment of this application, the step of using machine learning methods to identify and classify power quality scenarios to obtain scenario classification results includes: Using the preprocessed data, an ensemble learning framework is built to fuse the results of multiple base classifiers and generate a basic classification model. Import the basic classification model, use the sliding window method to segment the real-time data, and construct a feature calculation result set; Based on the feature calculation result set, multiple model instances are deployed for load balancing and output real-time scene recognition results.

[0011] According to one embodiment of this application, the establishment of a multi-dimensional power supply quality evaluation index system includes: An evaluation index system is constructed using voltage quality, power quality, and power supply reliability indicators to form an evaluation framework; By combining the scenario classification results and the evaluation framework, the evaluation strategy and indicator weights are dynamically adjusted to obtain the evaluation parameters. Based on the evaluation parameters, the fuzzy comprehensive evaluation method is applied to transform the quantitative indicators and generate a power supply quality evaluation report.

[0012] According to one embodiment of this application, the step of performing power quality assessment and problem attribution analysis, and outputting a power quality analysis report, includes: Collect data on factors affecting power supply quality, establish a causal relationship analysis model, and obtain key impact paths; Using the aforementioned key influence paths, Bayesian networks are applied to analyze the probabilistic dependencies between variables and obtain the contribution of influencing factors. Based on the contribution of the influencing factors, perform a problem attribution analysis and output diagnostic results and optimization suggestions.

[0013] According to one embodiment of this application, the feature analysis and decomposition of the execution processing task includes: By using static code analysis techniques, a program dependency graph is constructed to identify data dependencies between tasks and obtain task units that can be executed in parallel. Dynamic performance analysis tools are used to monitor task execution status and generate task computation characteristics. Based on the task computation characteristics, perform classification labeling for computationally intensive and I / O intensive tasks, and output the task classification results.

[0014] According to one embodiment of this application, the method further includes: By utilizing the data dependencies between the tasks, a directed acyclic graph is constructed, and the critical path analysis method is used to obtain the task constraints. Based on the task constraints, a hierarchical partitioning algorithm is applied to process the set of tasks that can be executed in parallel, forming a task decomposition scheme. Import the task decomposition scheme, perform task granularity threshold control, and output the final task set.

[0015] According to one embodiment of this application, establishing a task mapping optimization model and generating an initial mapping scheme includes: Deploy a distributed monitoring agent to collect information on computing resource usage, build a resource profile database, and form a resource utilization assessment model; Using the resource utilization evaluation model, the task mapping optimization model is designed using a genetic algorithm to generate an initial mapping scheme; The initial mapping scheme is optimized by performing crossover and mutation operations, and a dynamic scheduling scheme for computing resources is output.

[0016] This application also provides a system for improving the power supply quality of photovoltaic power distribution areas, including: A microservice architecture based on Docker container technology is used to encapsulate data collection, preprocessing, and analysis modules into independent microservices, configure automatic scaling strategies, and achieve decoupled management of the data processing pipeline. The data acquisition and preprocessing module is used to acquire electrical quantity data, meteorological data, and equipment status data, and to clean, standardize, and assess the quality of the data to obtain preprocessed data. The task mapping module is used to perform feature analysis and decomposition on the processing tasks of the preprocessed data, establish a task mapping optimization model, allocate tasks to heterogeneous computing resources, and achieve optimal scheduling of computing resources. The scene recognition and classification module is used to identify and classify power supply quality scenes based on the preprocessed data and the task mapping optimization model, and to obtain scene classification results. The power supply quality assessment module is used to establish a multi-dimensional power supply quality assessment index system based on the scenario classification results, conduct power supply quality assessment and problem attribution analysis, and generate a power supply quality analysis report.

[0017] This application embodiment also provides a computer device, the computer device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform the above-described method for improving the power supply quality of the photovoltaic substation.

[0018] This application also provides a computer-readable storage medium that stores computer instructions for causing a computer to execute the above-described method for improving the power supply quality of a photovoltaic power distribution area.

[0019] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the method for improving the power supply quality of the photovoltaic power distribution area described above.

[0020] This application has the following technical effects: 1) By adopting a microservice architecture based on Docker container technology and an automatic scaling strategy, the decoupled management of the data processing pipeline was achieved, improving system processing efficiency and resource utilization. 2) By using static task mapping technology for heterogeneous systems based on serial-parallel decomposition, optimal scheduling of computing resources was achieved, improving the real-time performance of data processing; 3) By using machine learning methods to identify and classify power quality scenarios, refined analysis for different scenarios is achieved, improving the accuracy of power quality assessment; 4) Through multi-dimensional power supply quality assessment and attribution analysis, the accurate location and cause diagnosis of power supply quality problems were achieved, providing effective decision support for improving power supply quality. Attached Figure Description

[0021] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0022] Figure 1 A flowchart of a method for improving power supply quality in photovoltaic power distribution areas provided in this application embodiment; Figure 2 This application provides a schematic diagram of a microservice process based on Docker container technology in an embodiment of the present application. Figure 3 A schematic diagram of the decoupled management process of the data processing pipeline provided in the embodiments of this application; Figure 4 A flowchart illustrating the implementation of the task mapping optimization model provided in this application embodiment; Figure 5 A flowchart illustrating the method for power quality scene identification and classification provided in this application embodiment. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure 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 disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0024] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0025] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0026] like Figure 1 As shown in the embodiment of this application, a method for improving the power supply quality of a photovoltaic power distribution area is provided, including: S1: Utilize Docker container technology to encapsulate data acquisition, preprocessing, and analysis modules into independent microservices, configure automatic scaling strategies, and obtain a decoupled management architecture for the data processing pipeline; Regarding the specific implementation of the microservice architecture built using Docker container technology in S1, one embodiment of this application proposes a pipelined, high-efficiency serverless inference system based on decoupled management. This system uses Docker container technology to build basic service units, dividing system functions into multiple independent microservice modules such as data acquisition services, data preprocessing services, and analysis and modeling services. Each microservice module is encapsulated in an independent Docker container, containing the runtime environment, dependency libraries, and application code. For example, the data acquisition service container integrates SDKs and drivers for various data acquisition interfaces, while the preprocessing service container contains tool libraries required for data cleaning and transformation. This containerized deployment method ensures service independence and environmental consistency.

[0027] At the container orchestration level, the system uses Kubernetes as its container orchestration platform to achieve unified management of microservices. Kubernetes defines the deployment requirements for each service through configuration files, including container images, resource limits, and the number of replicas. The system is configured with an automatic scaling strategy based on CPU utilization and memory usage. When the service load exceeds a preset threshold (e.g., CPU utilization exceeds 80%), the system automatically adds service instances; when the load decreases, it automatically reduces the number of instances, thereby achieving elastic scaling of resources.

[0028] To achieve efficient communication between services, a message queue system based on Apache Kafka was built. The message queue adopts a topic-based publish-subscribe pattern, setting independent topics for different types of data streams, such as raw data topics and preprocessed data topics. Producer services send data to the corresponding topics, while consumer services subscribe to relevant topics as needed. The message queue supports data persistence and message replay to ensure the reliability of data processing, and also enables parallel processing capabilities through message partitioning.

[0029] The system establishes a unified service gateway layer, using Spring Cloud Gateway to implement unified API management. The gateway layer implements functions such as request routing, load balancing, and authentication / authorization. By configuring routing rules, external requests are correctly forwarded to the appropriate microservices; it also integrates a JWT (JSON Web Token) authentication mechanism to ensure access security. The gateway also implements request rate limiting and circuit breaking mechanisms, automatically initiating service degradation when service anomalies are detected to ensure system stability.

[0030] S2: Collect electrical quantity data, meteorological data, and equipment status data; clean, standardize, and assess the quality of the data to obtain pre-processed data. This application presents an embodiment of a multi-source data quality assessment method for data preprocessing and quality evaluation. At the data acquisition interface level, a unified data access framework is designed. For electrical quantity data, real-time communication is established with smart meters and power quality monitoring equipment via the MQTT protocol; for meteorological data, it interfaces with meteorological stations and third-party meteorological services via a REST API; and for equipment status data, operating parameters of devices such as photovoltaic inverters and distribution transformers are collected via the ModBus protocol. The system configures a dedicated data acquisition adapter for each type of data source to achieve preliminary data format conversion and verification.

[0031] The data cleaning process employs a pipelined processing architecture. First, the system identifies outliers by combining static and dynamic thresholds; for example, voltage data exceeding ±7% of the rated voltage is considered abnormal. For detected outliers, different processing strategies are applied based on data characteristics: short-term outliers are corrected using linear or spline interpolation methods; long-term outliers are marked as invalid data. For missing values, the system combines time-series characteristics and similar daily patterns, using machine learning methods (such as LSTM networks) to complete the data.

[0032] Data standardization mainly includes two aspects: format unification and unit conversion. Regarding format unification, the system defines a unified data model, including fields such as timestamps, measured values, and quality labels. All source data must be converted to this unified format. Regarding unit conversion, the system establishes a complete unit conversion system. For example, power data from different sources is uniformly converted to kilowatts (kW) to ensure data consistency.

[0033] Data quality evaluation employs a multi-dimensional indicator system. The completeness dimension assesses the missing data rate and the proportion of valid data; the accuracy dimension assesses the anomaly rate and fluctuation range; and the timeliness dimension assesses data latency and update frequency. The system sets scoring criteria for each dimension and calculates a comprehensive score through weighted averages. Simultaneously, a data quality monitoring panel is established to display the real-time quality status of various data types and automatically triggers an alarm mechanism when quality indicators fall below thresholds. S3: Perform feature analysis and decomposition on the processing tasks of the preprocessed data, establish a task mapping optimization model, allocate tasks to heterogeneous computing resources, and achieve optimal scheduling of computing resources. Based on a specific implementation of static task mapping for heterogeneous systems using serial-parallel decomposition, this application proposes an efficient task analysis and resource scheduling method in one embodiment. The construction of a task feature analysis model is fundamental to the entire task mapping system. This model first identifies the computational characteristics of tasks through a combination of static code analysis and dynamic performance analysis. At the static analysis level, the system uses a program dependency graph (PDG) to analyze data dependencies between tasks and identify parallelizable task units. At the dynamic analysis level, performance profiling tools (such as perf) are used to collect features such as CPU utilization, memory access patterns, and I / O intensity of tasks. For example, computationally intensive tasks such as FFT transformations are marked as suitable for GPU processing, while I / O intensive tasks such as data queries are marked as suitable for CPU processing.

[0034] The task decomposition algorithm employs a top-down recursive decomposition strategy. First, a directed acyclic graph (DAG) is constructed based on the data dependencies of the tasks. Then, critical path analysis is used to identify the sequential constraints of the tasks. For a set of tasks that can be executed in parallel, the system further decomposes them into appropriately granular subtasks based on their computational complexity and data communication overhead using a hierarchical partitioning algorithm. To avoid scheduling overhead caused by over-partitioning, the system sets a minimum task granularity threshold; partitioning stops when the computational complexity of a subtask falls below this threshold.

[0035] The resource utilization assessment model monitors the usage of heterogeneous computing resources in real time. The system deploys a distributed monitoring agent to collect metrics such as CPU core utilization, GPU computing unit utilization, and memory bandwidth utilization. Through time series analysis, the system can predict short-term resource usage trends. Simultaneously, a resource profile database has been established to record the execution efficiency metrics of different types of tasks on various computing resources, providing a basis for task mapping decisions.

[0036] The task mapping optimization algorithm employs a heuristic search method to minimize the overall execution time while satisfying task dependency constraints. The algorithm considers multiple factors, including task computational characteristics, resource utilization, and communication overhead, and constructs a mapping optimization model based on a genetic algorithm. During optimization, new mapping schemes are generated through crossover and mutation operations, and task completion time and resource utilization are used as fitness functions for evaluation. To handle dynamically changing workloads, the system also implements an online adjustment mechanism, which can dynamically optimize the mapping scheme based on actual execution conditions.

[0037] S4: Using the preprocessed data and the task mapping optimization model, machine learning methods are used to identify and classify power supply quality scenarios to obtain scenario classification results; This application presents an embodiment of a machine learning-based scene recognition method for power quality scene identification and classification. The feature engineering pipeline first preprocesses and extracts features from the original time-series data. In the time domain, the system calculates statistical characteristics (mean, standard deviation, skewness, kurtosis, etc.) and time-series characteristics (trend, periodicity, autocorrelation, etc.) of signals such as voltage, current, and power. In the frequency domain, the system extracts spectral features of the signals using Fast Fourier Transform (FFT) and wavelet transform to identify the energy distribution in specific frequency bands. Furthermore, the system constructs features based on domain knowledge, such as photovoltaic power output fluctuation rate and load change rate. To improve the expressive power of the features, the system employs methods such as Principal Component Analysis (PCA) and autoencoders for feature dimensionality reduction and combination.

[0038] The scene classification model employs a hierarchical classification architecture. In the first layer, unsupervised learning methods (such as DBSCAN clustering) are used for coarse-grained classification of operating states. In the second layer, a specialized classifier is trained for each scene category to achieve refined recognition. The system uses an ensemble learning framework, fusing the results of multiple base classifiers such as Random Forest and XGBoost to improve classification accuracy and robustness. For typical scenarios, such as photovoltaic fluctuation scenarios, the model identifies them by analyzing the correlation between power curve features and meteorological conditions; for load change scenarios, it makes judgments by detecting the rate of power change and its duration.

[0039] The model training and validation process employs an online learning strategy. The system maintains a dynamically updated training dataset containing labeled historical scene data. During model training, cross-validation is used to evaluate model performance, and hyperparameters are optimized using grid search. To address the data imbalance problem, the system employs oversampling techniques such as SMOTE to augment minority class samples. Simultaneously, a model performance monitoring mechanism is implemented, triggering a model retraining process when the classification accuracy falls below a threshold.

[0040] The online scene recognition service employs a streaming processing architecture. The system segments real-time data using a sliding window method, with the window size dynamically adjusted based on the time scale of scene features. To improve real-time performance, the system implements parallel processing and incremental update mechanisms for feature computation. Simultaneously, load balancing is achieved by deploying multiple model instances to ensure responsiveness in large-scale data processing scenarios. Recognition results are pushed to downstream analysis modules via a message queue, and scene changes are displayed in real-time on the monitoring interface.

[0041] S5: Based on the scenario classification results, establish a multi-dimensional power supply quality assessment index system, conduct power supply quality assessment and problem attribution analysis, and output a power supply quality analysis report.

[0042] To address the specific implementation of multi-dimensional power quality assessment and attribution analysis, one embodiment of this application proposes a scenario-adaptive assessment and attribution method. The power quality assessment index system employs a multi-layered architecture. At the top level, three main dimensions are established: voltage quality, power quality, and power supply reliability. The voltage quality dimension includes indicators such as voltage deviation rate and three-phase imbalance, with the voltage deviation rate calculation considering time-weighted factors and penalizing the duration exceeding the limit. The power quality dimension covers indicators such as harmonic content and flicker, calculating each harmonic component in real time using a harmonic analysis algorithm. The power supply reliability dimension includes indicators such as power supply reliability rate and fault recovery time, establishing a reliability assessment model based on historical fault data. The system sets grading standards for each indicator and adjusts the weighting coefficients of each indicator according to the characteristics of the photovoltaic power distribution area.

[0043] The multi-dimensional evaluation model is implemented based on a scenario-adaptive evaluation framework. The system first dynamically adjusts the evaluation strategy according to the scenario identification results. For example, in a photovoltaic fluctuation scenario, the weights of voltage and power fluctuation indicators are increased; in a load change scenario, the focus is on voltage stability and harmonic characteristics. The evaluation model employs a fuzzy comprehensive evaluation method, transforming quantitative indicators into qualitative evaluation results by establishing membership functions. Simultaneously, the system implements spatiotemporal correlation analysis, enabling the assessment of the spatial distribution characteristics and temporal evolution patterns of power supply quality problems.

[0044] The attribution analysis model employs a multi-model fusion approach to pinpoint the root cause of problems. First, a causal graph of power quality influencing factors is constructed using a causal inference model to identify key impact paths. The system uses Bayesian networks to capture probabilistic dependencies between variables and automatically discovers potential causal links through structural learning algorithms. Building upon this, a feature importance analysis model based on gradient boosting trees is deployed to quantify the contribution of each influencing factor. For complex fault scenarios, the system also integrates an expert system rule base, fusing domain knowledge for cause diagnosis. To improve the accuracy of the analysis, the system implements time-series alignment and correlation analysis of multi-source data, enabling the identification of causal relationships at different time scales.

[0045] The analysis report generation module employs an intelligent reporting engine. Based on the evaluation results and attribution analysis output, the system automatically generates structured analysis reports. The report content includes a comprehensive power quality score, key indicator anomaly statistics, problem attribution analysis results, and improvement suggestions. For visualization, the system integrates various chart types, such as radar charts to display multi-dimensional evaluation results, heatmaps to show the spatiotemporal distribution characteristics of problems, and Sankey diagrams to display causal chains. The system provides differentiated report templates for different user levels; management users can view high-level statistical analysis, while technical personnel can delve into detailed technical indicators. Simultaneously, the system features intelligent recommendation functionality, automatically generating optimization suggestions based on historical processing experience for identified problems and evaluating the cost-effectiveness of implementing these suggestions. These analysis results are ultimately displayed through a web interface and a mobile app, supporting real-time querying and interactive analysis.

[0046] To ensure the practicality of the analysis results, the system has established a continuous optimization mechanism. By recording user feedback and actual improvement effects, the parameters and weights of the evaluation model are continuously optimized. Simultaneously, the system regularly performs offline analysis to uncover deeper patterns in power quality issues, providing decision support for preventative maintenance and long-term planning. Regarding data security, the system implements multi-level access control and data anonymization to ensure the security of sensitive information. Furthermore, the analysis results can be exported to standard formats (such as PDF and Excel) for easy data sharing and integration with other systems.

[0047] Among them, such as Figure 2 As shown, S1 specifically includes: S1.1: Based on Docker container technology, basic service units are built, and functional modules such as data collection, preprocessing, and analysis are encapsulated as independent microservices; The system employs Docker container technology to build its basic service units. Docker, as a lightweight virtualization technology, can package applications and their dependencies into a portable container, ensuring consistent application operation across different environments. In the implementation scheme of this application, the system divides the various functional modules required for improving the power supply quality of the photovoltaic power distribution area into microservices with clearly defined boundaries. Specifically, the data acquisition microservice is responsible for communicating with various sensors, smart meters, and weather stations to achieve real-time acquisition of multi-source heterogeneous data; the data preprocessing microservice focuses on data cleaning, standardization, and quality assessment; and the analysis microservice is responsible for executing core algorithms such as scene recognition, power supply quality assessment, and problem attribution. Each microservice is encapsulated in an independent Docker container, containing its dedicated runtime environment, dependency libraries, and configuration files. This encapsulation ensures isolation between microservices, avoids dependency conflicts, and simplifies service deployment and upgrade processes.

[0048] S1.2: Use Kubernetes to implement container orchestration and cluster management, and configure automatic scaling strategies; The system utilizes Kubernetes as its container orchestration platform to achieve unified management of microservice clusters. Kubernetes provides the ability to automatically deploy, scale, and manage containerized applications through declarative configuration. In practice, the system creates a Deployment configuration for each microservice, defining key parameters such as container images, resource requirements, and health check policies. Notably, the system is configured with a resource utilization-based automatic scaling rule (HorizontalPod Autoscaler). When the CPU utilization of a microservice consistently exceeds a preset threshold (e.g., 75%), Kubernetes automatically increases the number of instances for that service; conversely, when the load decreases, it reduces the number of instances to conserve computing resources. This dynamic adjustment mechanism ensures that the system can flexibly allocate resources according to the actual workload, meeting performance requirements while optimizing resource utilization. Furthermore, Kubernetes' service discovery and load balancing features simplify communication between microservices, improving the overall reliability and resilience of the system.

[0049] S1.3: Build a message queue system to realize asynchronous processing of data streams and communication between services; The system constructs a data stream processing architecture based on message queues, realizing loosely coupled communication between services and asynchronous processing of data streams. In traditional synchronous call patterns, strong dependencies between services can easily lead to increased system vulnerability and performance bottlenecks. To address this issue, this application adopts Apache Kafka as a distributed message queue system, establishing a publish-subscribe communication mechanism. Kafka's high throughput and persistence capabilities make it particularly suitable for processing large-scale real-time data streams in photovoltaic power stations. System administrators can create and configure topics, monitor message traffic, and adjust the number of partitions and replication factor as needed through Kafka's management interface to balance data processing performance and reliability. In actual operation, the data acquisition microservice acts as a producer, publishing raw data to the corresponding topics; the preprocessing microservice consumes this raw data and produces preprocessed results; and the analysis microservice subscribes to the preprocessed data topics to obtain the input required for analysis. This data stream-driven architecture not only improves system throughput but also enhances system fault tolerance; even if a microservice is temporarily unavailable, data can be safely cached in the message queue awaiting processing.

[0050] S1.4: Establish a service gateway to achieve unified API management and security authentication.

[0051] The system establishes a unified service gateway layer, serving as the sole entry point for external access to internal microservices. Implemented using Spring Cloud Gateway, it is built on Spring WebFlux and provides non-blocking reactive processing capabilities, efficiently handling concurrent requests. The gateway layer primarily implements three key functions: request routing, security authentication, and traffic control. For request routing, the gateway forwards requests along different paths to the corresponding microservices by configuring routing rules. For security authentication, the gateway integrates OAuth2.0 and JWT authentication mechanisms, achieving centralized user authentication and authorization. For traffic control, the gateway implements request rate limiting, circuit breaking, and degradation mechanisms. When an abnormal response or timeout is detected in a microservice, it can quickly trigger circuit breaking protection, returning a default response or cached data to prevent cascading failures. Furthermore, the gateway provides auxiliary functions such as request logging and monitoring metric collection, providing strong support for system operation and maintenance and problem diagnosis.

[0052] Specifically, such as Figure 3 As shown, S1.3 can be further refined as follows: S1.3.1: Build a message queue system, set the raw data topic and the preprocessed data topic, and obtain the data stream processing architecture; The system features a detailed design for the message queue's topic structure and data flow. First, a hierarchical topic naming convention is established, such as "raw-data.electrical", "raw-data.meteorological", and "processed-data.electrical", to distinguish data streams from different sources and processing stages. For each topic, the system configures an appropriate number of partitions based on data volume and real-time requirements to support parallel processing. For example, electrical quantity data, due to its high sampling frequency and high real-time requirements, has more partitions configured for its raw data topic; while meteorological data, being relatively low-frequency, has fewer topic partitions. Furthermore, the system sets up data retention strategies for key topics, retaining important data for a longer period to support historical data analysis. Through this carefully designed topic structure, the system constructs a clear and efficient data flow processing architecture, providing a reliable data channel for subsequent data processing and analysis.

[0053] S1.3.2: Using the aforementioned data stream processing architecture, a topic subscription model is adopted to realize inter-service communication, forming a service communication mechanism; Based on the aforementioned data stream processing architecture, the system implements a topic-based publish-subscribe pattern, forming a complete service communication mechanism. Each microservice can be either a producer or a consumer, or both, depending on its functional role. To ensure message processing reliability, the system implements a message acknowledgment mechanism on the consumer side, committing offsets only after a message has been successfully processed. For messages that fail to process, the system employs retry strategies and dead-letter queues to ensure that abnormal data is not lost and does not affect the normal data flow. Furthermore, the system supports two consumption modes for different types of consumption scenarios: a high-throughput batch consumption mode and a low-latency single-message consumption mode. The data preprocessing service uses the batch consumption mode to improve processing efficiency, while the real-time alarm service uses the single-message consumption mode to reduce latency. Through flexibly configurable consumer groups, the system also achieves message load balancing, ensuring that parallel processing capacity increases linearly with the increase of consumer instances.

[0054] SS1.4 can be further refined as follows: S1.4.1: Based on the service communication mechanism, establish a unified service gateway layer to complete the decoupled management of the data processing pipeline.

[0055] The system utilizes the aforementioned service communication mechanisms to establish a unified service gateway layer, achieving decoupled management of the data processing pipeline. The service gateway is not merely the entry point for API calls, but also the core hub of the entire microservice architecture. Tightly integrated with the message queue system, the gateway layer can convert external synchronous API requests into internal asynchronous message processing flows, further enhancing the system's resilience and responsiveness. To support different types of clients, the gateway provides multiple communication protocol interfaces such as REST API, WebSocket, and MQTT. Regarding authentication and authorization, the gateway implements fine-grained access control, granting different API access permissions to different user roles (such as system administrators, operations personnel, and analysts). Furthermore, the gateway layer implements an API version management mechanism, supporting the coexistence of multiple API versions and ensuring backward compatibility during system upgrades. Through these designs, the service gateway successfully decouples external requests from internal implementation, forming a flexible and scalable data processing pipeline architecture, providing a solid technical foundation for real-time monitoring and analysis of power supply quality in photovoltaic power distribution areas.

[0056] S2 specifically includes: S2.1: Establish a multi-source data acquisition interface, including electrical quantity data, meteorological data, equipment status data, etc. The system establishes a comprehensive multi-source data acquisition interface framework to obtain a complete data profile of the photovoltaic power station operation. This framework adopts a layered design, implementing adapters for various communication protocols at the bottom layer, including the DL / T645 protocol for smart meters, the IEC61850 protocol for power quality analyzers, the Modbus protocol for weather stations, and the SunSpec protocol for photovoltaic inverters. The middle layer constructs a unified data access abstraction layer, shielding the characteristics of different protocols at the bottom layer and providing standardized data acquisition interfaces upwards. Regarding interface configuration, the system uses declarative configuration files; administrators only need to define information such as data source type, communication parameters, and acquisition cycle, completing the access of new data sources without writing code. For electrical quantity data, the system collects multiple indicators including three-phase voltage, current, active power, reactive power, power factor, and harmonic content, with sampling frequencies ranging from milliseconds to seconds; meteorological data includes parameters such as irradiance, temperature, humidity, wind speed, and wind direction, with sampling frequencies ranging from minutes; equipment status data covers information such as transformer oil temperature, load rate, and photovoltaic inverter operating status, with sampling frequencies ranging from minutes to hours. To ensure the reliability of data acquisition, the system also implements a monitoring and retry mechanism for data acquisition tasks. When an acquisition anomaly is detected, the retry process is automatically triggered or an alarm notification is issued.

[0057] S2.2: Implement the data cleaning process, including outlier handling, missing value filling, and duplicate value removal; The system implements a complete data cleaning workflow to ensure the quality of data used in subsequent analyses. Data cleaning is a multi-stage pipeline process, including outlier detection and handling, missing value imputation, duplicate value removal, and data smoothing. For outlier detection, the system combines statistical methods and domain knowledge rules. Statistical methods mainly include Z-score-based detection, box plot methods, and sliding window-based time-series anomaly detection; domain knowledge rules are based on reasonable ranges set according to the physical characteristics of the power system, such as voltage deviation not exceeding ±7% of the rated value and three-phase current imbalance not exceeding 15%. For detected outliers, the system adopts different processing strategies based on the degree and duration of the anomaly: minor and transient anomalies are corrected using median filtering or spline interpolation; significant or long-term anomalies are marked as invalid data and recorded in the anomaly log for subsequent analysis. In the missing value imputation stage, the system selects an appropriate imputation method based on the data type and missing value pattern. For data with strong time series characteristics, such as electrical quantity data, time series prediction models such as ARIMA or LSTM are used for filling. For data with strong correlation to external factors, such as the relationship between photovoltaic output and solar irradiance, multivariate regression models are used for estimation. In addition, the system also implements duplicate value detection and removal functions. By comparing timestamps and data content, duplicate records caused by communication retransmissions and other reasons are identified and removed to ensure the uniqueness and consistency of the data.

[0058] S2.3: Perform data standardization processing to unify data formats and units of measurement; Data standardization is a crucial step in achieving multi-source heterogeneous data fusion and analysis. The system first constructs a unified data model, defining a standardized representation structure for all data types. This model includes basic metadata fields (such as data source ID, measurement point ID, timestamp, data quality flags, etc.) and specific types of measurement value fields. Regarding format unification, the system standardizes the time representations from different sources, uniformly adopting a UTC timestamp plus time zone information format, resolving potential time representation differences between different devices. For numerical data, the system establishes a complete unit conversion engine, supporting automatic conversion between various physical quantity units. For example, power data from equipment from different manufacturers is uniformly converted to kilowatts (kW), and voltage data is uniformly converted to volts (V). The system maintains an extensible unit conversion rule library, and administrators can add new conversion rules through the configuration interface. Furthermore, for enumerated data types such as device status, the system defines a standard status code mapping table, mapping proprietary status descriptions from different manufacturers to a unified status definition. This comprehensive standardization ensures that data from different times and sources can be compared and analyzed in the same spatiotemporal dimension, providing a reliable data foundation for subsequent scene identification and quality assessment.

[0059] S2.4: Construct a data quality evaluation index system, including dimensions such as completeness, accuracy, and timeliness.

[0060] The system constructs a comprehensive data quality evaluation index system to achieve quantitative assessment and monitoring of data quality. This index system adopts a multi-level design structure, with four main dimensions at the top level: completeness, accuracy, timeliness, and consistency. Under the completeness dimension, the system calculates indicators such as data missing rate and effective record ratio, assessing the completeness of data by monitoring data coverage in long-term series. Under the accuracy dimension, the system sets indicators such as outlier ratio and data fluctuation range to assess the degree of conformity between data and actual physical quantities. Under the timeliness dimension, the system monitors data acquisition latency, processing latency, and update frequency to assess the real-time level of data. Under the consistency dimension, the system examines the logical relationships between different data sources, such as the deviation between calculated and measured power values, and the balance relationship of upstream and downstream measurement points, to assess the internal consistency of the data. For each specific index, the system sets multi-level scoring standards and obtains a comprehensive score for each dimension and an overall quality score through weighted calculation. These scoring results are displayed in real time through a visual dashboard and support historical trend analysis. When a quality index falls below a preset threshold, the system automatically triggers an alarm mechanism to notify relevant personnel for inspection and handling. In addition, the system has established a classification and diagnosis function for data quality problems, which can provide possible cause analysis and improvement suggestions for different types of quality problems. For example, when a certain measuring point is found to have continuous missing data, the system will automatically analyze and prompt that the problem may be caused by communication interruption or equipment failure.

[0061] like Figure 4 As shown, S3 specifically includes: S3.1: Establish a task feature analysis model to identify the task's computational intensity and data dependencies; The system implements a sophisticated task feature analysis model, laying the foundation for efficient scheduling of heterogeneous computing resources.

[0062] S3.1.1: Using static code analysis techniques, construct a program dependency graph, identify data dependencies between tasks, and obtain task units that can be executed in parallel; The system employs static code analysis techniques to deeply analyze the program structure. Specifically, the system first modularizes various data processing and analysis algorithms, with each module having clearly defined input and output interfaces. By analyzing the source code or configuration files of these modules, the system constructs a detailed program dependency graph (PDG), which represents the data flow relationships between modules in the form of a directed graph. For example, the voltage quality analysis module depends on the output of the voltage data preprocessing module, while the harmonic analysis module depends on the preprocessing results of both voltage and current data. Based on this dependency analysis, the system can identify modules without direct or indirect dependencies, which can be marked as parallelizable task units. Furthermore, the system analyzes the internal loop structures and conditional branches of the modules, identifying potential data-level parallelism opportunities, such as vector operations and matrix operations. This static analysis not only provides a theoretical upper limit for task parallelism but also offers important references for subsequent task decomposition and scheduling optimization.

[0063] S3.1.2: Employ dynamic performance analysis tools to monitor task execution status and generate task computation characteristics; The system further enriches the execution characteristic profile of tasks through dynamic performance analysis. It deploys lightweight performance monitoring probes to run representative tasks on different types of computing resources, collecting various performance metrics during execution. These metrics include low-level performance data such as CPU utilization, memory access patterns, cache hit rate, and I / O throughput, as well as high-level metrics such as execution time and resource consumption. Through statistical analysis of this data, the system can generate performance characteristic descriptions for each task module, such as the average CPU time, peak memory usage, and disk I / O counts for a module processing 10MB of data. To obtain a more comprehensive performance profile, the system designed a series of test cases covering scenarios with different data sizes and parameter configurations, thereby understanding the trend of task performance changing with input size. This dynamically collected performance data is stored in a performance knowledge base and continuously updated and improved through online learning algorithms to ensure it reflects the latest execution environment characteristics.

[0064] S3.1.3: Based on the task computation characteristics, perform classification labeling for computationally intensive and I / O intensive tasks, and output the task classification results; Based on the task feature data collected in the first two steps, a fine-grained classification of task types was performed. First, the system categorized tasks into two main types based on the ratio of CPU computation to I / O operations: computationally intensive and I / O intensive. Computationally intensive tasks, such as FFT transformations and machine learning model training, have execution times primarily limited by processor performance; I / O intensive tasks, such as database queries and file read / write operations, have execution times primarily limited by storage system performance. Furthermore, the system also identified the parallel computing characteristics of the tasks, such as whether they are suitable for GPU acceleration and whether they involve a large number of vector operations. Based on these classification results, the system assigned a type label and a resource preference label to each task. For example, computationally intensive tasks suitable for GPU acceleration were labeled "COMPUTE_INTENSIVE:GPU_PREFERRED"; database operations involving a large number of random accesses were labeled "IO_INTENSIVE:MEMORY_PREFERRED". This labeling information is used not only for subsequent resource allocation decisions but also for optimizing task scheduling strategies. For example, the system will try to avoid assigning multiple I / O-intensive tasks to the same storage node to reduce resource contention; while for compute-intensive tasks, it can consider mixing them with I / O-intensive tasks for scheduling to achieve a balanced utilization of compute and I / O resources.

[0065] S3.2: Implement a task decomposition algorithm to break down complex tasks into subtasks that can be executed in parallel; The system implements an efficient task decomposition algorithm, which transforms complex data processing flows into a parallel task network with appropriate granularity.

[0066] S3.2.1: Using the data dependencies between the tasks, construct a directed acyclic graph and use the critical path analysis method to obtain the task constraints; Based on the previously identified task dependencies, the system constructs a complete directed acyclic graph (DAG) for task execution. In this graph, nodes represent specific task modules, and edges represent data dependencies. The system employs critical path analysis to identify the longest path from the start node to the end node. The sequence of tasks along this path determines the minimum completion time of the entire process. Tasks on the critical path are marked as high priority, and the system prioritizes ensuring they receive sufficient computing resources. Simultaneously, the system also identifies tasks not on the critical path; these tasks have some time leeway and can be more flexibly scheduled in resource allocation. Through this critical path analysis, the system obtains the temporal constraints for task execution, including the earliest start time and the latest finish time. These constraints provide crucial references for subsequent task decomposition and scheduling.

[0067] S3.2.2: Based on the task constraints, apply a hierarchical partitioning algorithm to process the set of tasks that can be executed in parallel, forming a task decomposition scheme; The system employs a hierarchical partitioning algorithm for fine-grained decomposition of the identified set of parallelizable tasks. This algorithm uses a top-down recursive approach, breaking down large tasks into multiple subtasks based on their data processing characteristics and computational complexity. The decomposition process considers both data parallelism and task parallelism. For tasks with data parallelism, such as large-scale matrix operations and batch data processing, the system decomposes them into multiple subtasks processing different data partitions based on the data size and available resources. For complex operations with task parallelism, the system breaks them down into multiple functionally independent subtasks based on their internal logical structure. During the decomposition process, the system also considers the principle of data locality, grouping subtasks accessing the same data together to reduce data transmission overhead. Furthermore, the system implements a task merging mechanism, merging related tasks with excessively fine granularity into more suitable task units to balance parallelism and scheduling overhead. Through this multi-level, multi-dimensional task decomposition, the system forms a task decomposition scheme that fully utilizes parallel computing resources while minimizing communication overhead.

[0068] S3.2.3: Import the task decomposition scheme, execute task granularity threshold control, and output the final task set; The initial task decomposition scheme was optimized in granularity to ensure that the task size was neither too large, leading to insufficient parallelism, nor too small, resulting in excessive scheduling overhead. An adaptive task granularity threshold control mechanism was introduced, dynamically adjusting the minimum task granularity threshold based on the quantity and performance characteristics of available computing resources. When the system detects that the estimated execution time or computational load of a subtask is below the threshold, it will stop further decomposition or consider merging it with other related subtasks. This threshold is not fixed but dynamically adjusted based on system load and resource utilization. For example, when the system load is low, smaller granular tasks can be accepted to fully utilize idle resources; while when the load is high, the granularity threshold is increased to reduce scheduling overhead. Furthermore, the system considers task priority and criticality. For tasks on the critical path, a more aggressive decomposition strategy may be adopted to maximize parallelism; while for non-critical tasks, a more conservative strategy may be adopted to reduce resource consumption. Through this fine-grained control, the system ultimately outputs a final task set that balances parallel efficiency and management overhead.

[0069] S3.3: Construct a resource utilization assessment model to dynamically monitor the system resource status; A comprehensive resource utilization assessment model was constructed to achieve dynamic monitoring and performance evaluation of heterogeneous computing resources. This model collects multi-level resource status indicators, including basic indicators at the physical resource layer such as CPU core utilization, memory usage, network bandwidth utilization, and storage I / O queue length, as well as application-level indicators at the container layer such as resource quota usage and task queue length. The system deploys a distributed monitoring agent to collect these indicators in real time on each computing node and stores and indexes them using a time-series database. Based on this historical monitoring data, the system establishes a time-series analysis model of resource usage patterns, capable of predicting short-term resource usage trends. For example, by analyzing periodic patterns in historical data, the system can predict potential load peaks in certain periods and plan resources in advance. Furthermore, the system implements anomaly detection functionality, which can promptly issue warnings and initiate corresponding processing mechanisms when abnormal fluctuations in resource utilization occur. This comprehensive resource monitoring and assessment capability provides real-time and accurate decision-making basis for subsequent task mapping optimization.

[0070] S3.4: Design a task mapping optimization algorithm to achieve optimal allocation of tasks to heterogeneous computing resources; The system is designed with an advanced task mapping optimization algorithm to achieve optimal allocation of tasks to heterogeneous computing resources.

[0071] S3.4.1: Deploy a distributed monitoring agent to collect computing resource usage information, build a resource profile database, and form a resource utilization evaluation model; A fine-grained distributed monitoring agent was deployed to collect detailed usage information for various computing resources. The monitoring agent employs a lightweight design to minimize its impact on system performance while ensuring comprehensive and real-time data collection. Collected metrics include CPU core utilization, memory allocation and access patterns, network latency and throughput, and storage device I / O performance. This raw monitoring data, after preprocessing and aggregation, is stored in a resource profile database. This database not only records real-time resource status but also maintains historical performance data and resource configuration information, forming a complete performance profile for each computing node. Based on this data, the system constructs a resource utilization evaluation model capable of assessing and predicting the execution efficiency of different types of tasks on various computing resources. For example, the model can predict the execution time of the FFT algorithm on a specific GPU or the completion time of data query operations on different configured storage nodes. This performance prediction capability based on historical data provides a quantitative evaluation basis for task mapping decisions.

[0072] S3.4.2: Using the resource utilization evaluation model, the task mapping optimization model is designed using a genetic algorithm to generate an initial mapping scheme; Based on the previously established resource assessment model, a task mapping optimization framework based on a genetic algorithm was designed. This framework abstracts the task mapping problem into a combinatorial optimization problem, aiming to minimize the overall execution time and balance resource utilization while satisfying task dependency constraints. The system first defines a gene encoding scheme, using integer sequences to represent the mapping relationship between tasks and resources. For example, the gene "2,4,1,3" indicates that task 1 is assigned to resource 2, task 2 is assigned to resource 4, and so on. Then, a fitness function is designed, comprehensively considering multiple objectives such as execution time, resource utilization balance, and energy consumption. Based on this encoding and evaluation mechanism, the system generates an initial population, with each individual representing a possible task mapping scheme. By simulating the evolutionary process, including selection, crossover, and mutation operations, the system iteratively optimizes these schemes, gradually approaching the global optimum. In practical applications, the system dynamically adjusts the population size and number of iterations according to the problem size and time budget, achieving a balance between optimization quality and computational overhead. Through this heuristic search method, the system can efficiently find near-optimal task mapping schemes in complex solution spaces, maintaining good performance even with large-scale task sets and heterogeneous computing environments.

[0073] S3.4.3: Optimize the initial mapping scheme by performing crossover and mutation operations, and output a dynamic scheduling scheme for computing resources.

[0074] The initial mapping scheme underwent in-depth optimization to further improve resource allocation efficiency. The system first applies a crossover operation, selecting two high-performing schemes from the current pool of schemes and exchanging some of their mapping relationships to generate new candidate schemes. The selection of the crossover point considers task dependencies and resource affinity, ensuring that the generated new schemes still satisfy the constraints. Then, the system applies a mutation operation, randomly selecting and changing the mapping relationships of certain tasks to increase scheme diversity and avoid getting trapped in local optima. The probability and magnitude of the mutation operation are dynamically adjusted according to the progress of the optimization process; a larger mutation rate is used initially to explore the solution space, while the mutation rate is reduced later to promote convergence. For each newly generated scheme, the system uses a resource utilization evaluation model to predict its execution performance and updates the population based on the prediction results. After multiple iterations, the system selects the best-performing scheme as the final resource allocation decision. To adapt to dynamic changes in workload and resource status, the system also implements an online adjustment mechanism, periodically re-evaluating the current scheme and triggering a remapping process when necessary. This dynamic optimization mechanism ensures that the system can always maintain near-optimal resource utilization efficiency, achieving efficient computing resource scheduling even in complex and changing operating environments.

[0075] like Figure 5 As shown, S4 specifically includes: S4.1: Construct a feature engineering pipeline to extract time-series data features and statistical features; A complete feature engineering pipeline was constructed to extract rich feature representations from multi-source heterogeneous data from photovoltaic power stations. This pipeline adopts a modular design, encompassing multiple stages such as data segmentation, feature extraction, feature selection, and feature transformation, adaptable to the feature representation needs of different data types. First, in the data segmentation stage, the system divides continuous time-series data into fixed-length or variable-length time windows according to the time scale requirements of the analysis task. Shorter time windows (e.g., seconds or minutes) are used for rapidly changing phenomena such as voltage fluctuations, while longer time windows (e.g., hours or days) are used for relatively slow processes such as load changes. The system supports multiple window types, including fixed windows, sliding windows, and event-triggered windows, to adapt to different analysis scenarios. Second, in the feature extraction stage, the system implements a rich feature calculation method library, covering multiple categories such as statistical features, time-series features, frequency domain features, and domain-specific features. This feature library supports configurable feature combination strategies, allowing analysts to select appropriate feature sets according to specific application requirements. Finally, the system also implements automatic feature selection and dimensionality reduction functions. Through methods such as analysis of variance, correlation analysis, and principal component analysis, it selects the most distinctive feature subset, reduces data dimensionality, and improves model performance. This comprehensive feature engineering pipeline provides high-quality data representation for subsequent scene classification and is the foundation for achieving accurate scene recognition.

[0076] S4.1.1: Using the preprocessed data, perform feature extraction calculations to construct statistical features and time-series features, and obtain feature vectors; In-depth feature extraction calculations were performed on the preprocessed data, constructing multi-dimensional feature vectors. Regarding statistical features, the system calculated basic statistics such as mean, standard deviation, median, maximum, and minimum to characterize the central tendency and dispersion of the data; it also calculated higher-order statistics such as skewness and kurtosis to depict the shape characteristics of the data distribution; and, for anomaly detection, the system extracted robust statistical features such as interquartile range and outlier ratio. Regarding time-series features, the system adopted a multi-level feature extraction strategy, including trend features, seasonal features, periodic features, and autocorrelation features. Trend features were obtained by fitting linear or polynomial models to characterize the long-term changing trend of time-series data; seasonal and periodic features were extracted using Fourier analysis or seasonal decomposition methods to characterize the periodic changing patterns of the data; and autocorrelation features were obtained by calculating time-delay correlation coefficients to reflect the time-dependent structure of the time-series data. For electrical quantity data, the system also specifically extracted power system-specific features, such as the duration distribution of voltage deviation rate, power factor change rate, and statistical characteristics of three-phase imbalance. To characterize complex events, the system designed multivariate cross-features, such as the correlation coefficient between irradiance and photovoltaic output, and the time-series matching degree between load changes and voltage fluctuations. After standardization and combination, these multi-dimensional features form a high-dimensional feature vector, comprehensively representing all aspects of the photovoltaic power station's operating status.

[0077] S4.2: Establish a scenario classification model, including typical scenarios such as normal operation, photovoltaic fluctuations, and sudden load changes; A complete scenario classification model framework was established to identify typical scenarios in the operation of photovoltaic (PV) power grids. Based on in-depth analysis of the operating characteristics of PV power grids, the system defined a series of key scenario categories, including normal operation, PV fluctuations, load surges, harmonic interference, and voltage anomalies. Each scenario category has its specific electrical characteristics and formation mechanisms. For example, PV fluctuation scenarios are usually directly related to weather conditions (such as cloud cover changes and showers), manifested as rapid fluctuations in PV output and corresponding changes in system voltage; load surge scenarios are mostly caused by the sudden connection or disconnection of large loads, manifested as step changes in power and current, and possible voltage dips or surges. To accurately characterize these scenarios, the system established a scenario feature knowledge base, recording the typical characteristic patterns and discrimination criteria of various scenarios. Simultaneously, the system also adopted a data-driven adaptive classification framework, which can continuously optimize scenario classification criteria and identify newly emerging scenario types as data accumulates. This hybrid approach, combining domain knowledge and data-driven methods, makes the scenario classification model both professionally interpretable and possesses adaptive learning capabilities, enabling it to efficiently cope with the complex and ever-changing operating states of PV power grids.

[0078] S4.2.1: Using the aforementioned feature vectors, an unsupervised learning method is applied to perform coarse-grained classification of the running states, forming preliminary classification results; Unsupervised learning methods were employed to coarsely classify the operating status of photovoltaic power stations. This step aims to discover naturally occurring state clusters from massive amounts of data, providing a preliminary framework for subsequent fine-grained classification. The system primarily uses two types of unsupervised learning algorithms: distance-based clustering algorithms and density-based clustering algorithms. Distance-based clustering algorithms, such as K-means and hierarchical clustering, are suitable for discovering clusters with regular shapes and similar sizes; while density-based clustering algorithms, such as DBSCAN and OPTICS, are better suited for discovering clusters with irregular shapes and varying densities, and are more resistant to outliers. In practical applications, the system automatically selects appropriate algorithms or combines multiple algorithms based on data characteristics. To address the clustering challenges in high-dimensional feature spaces, the system first applies dimensionality reduction techniques such as Principal Component Analysis (PCA) or t-SNE to map feature vectors to a lower-dimensional space, reducing the impact of the "curse of dimensionality." During the clustering process, the system employs an adaptive parameter adjustment mechanism, such as dynamically determining the optimal number of clusters using evaluation metrics like the silhouette coefficient or the Davies-Bouldin index. The clustering algorithm outputs are post-processed, including small cluster merging and outlier cluster labeling, to form a preliminary classification result of the running state. The advantage of this unsupervised learning method is that it does not rely on predefined labels and can discover potential, possibly unrecognized patterns in the data, providing valuable knowledge guidance for subsequent supervised learning.

[0079] S4.2.2: Based on the preliminary classification results, train the classifier and output the scene recognition results of photovoltaic fluctuations and load changes; Based on the preliminary classification results, a refined scene classifier was further trained. This stage employed supervised learning methods, mapping the clusters obtained from unsupervised clustering to expert-defined scene categories to establish a labeling system. During the label mapping process, the system combined automatic matching and expert verification: first, preliminary mapping relationships were automatically generated by comparing the feature centers of the clusters with feature patterns in the scene knowledge base; then, domain experts reviewed and corrected these mappings to ensure label accuracy. Based on this labeled training data, the system trained a series of specialized classifiers, each responsible for identifying specific types of scenes. For example, for photovoltaic fluctuation scenarios, a specialized classifier combining meteorological data and electrical features was constructed, capable of distinguishing photovoltaic fluctuations caused by different factors such as cloud cover changes and temperature fluctuations; for load mutation scenarios, a classifier based on power change features and time-series patterns was constructed, capable of identifying various subtypes such as load initiation, shedding, and periodic changes. These specialized classifiers adopted algorithmic models suitable for their respective task characteristics, such as Support Vector Machines (SVM), Random Forests, and Deep Neural Networks, and optimized model parameters through cross-validation. Through this hierarchical classification strategy, the system can output refined scene recognition results, providing accurate scene context for subsequent power quality assessment and problem attribution.

[0080] S4.3: Implement the model training and validation process and optimize model parameters; A complete model training and validation process was implemented to ensure the accuracy and stability of the scene classification model. This process employs an automated machine learning (AutoML) framework, covering multiple stages including data partitioning, feature selection, model selection, hyperparameter optimization, and performance evaluation. First, the system uses a stratified sampling strategy to divide the labeled dataset into training, validation, and test sets, ensuring consistent distribution of scene categories across subsets. Then, based on the characteristics of different scene types, the system automatically performs feature importance evaluation, selecting the most discriminative feature subset. In the model selection stage, the system chooses multiple candidate models from a predefined model library, including traditional machine learning algorithms and deep learning models, for initial training and comparison. For each model, the system applies methods such as grid search or Bayesian optimization for hyperparameter tuning to find the optimal parameter combination. In the model evaluation stage, the system not only focuses on common metrics such as overall accuracy, precision, and recall, but also pays special attention to the confusion between different scene categories and the model's performance on boundary and noisy samples. To improve the model's generalization ability, the system also applies techniques such as data augmentation and regularization to reduce the risk of overfitting. The entire training and validation process adopts a pipeline design, supporting automated execution and result recording, enabling model developers to quickly iterate and optimize while ensuring the repeatability and traceability of experiments.

[0081] S4.3.1: Using the preprocessed data, build an ensemble learning framework, fuse the results of multiple base classifiers, and generate a basic classification model; A robust ensemble learning framework was built based on preprocessed data. By fusing the results of multiple base classifiers, a base classification model with high accuracy and robustness was generated. This framework employs a multi-level ensemble strategy, including methods such as Bagging, Boosting, and Stacking. In selecting base classifiers, the system considers model diversity and complementarity, incorporating different types of algorithms such as decision trees, support vector machines, k-nearest neighbors, and neural networks. For each base classifier, the system determines its optimal configuration parameters through cross-validation. During the ensemble process, a weighted voting method was used, with weights dynamically allocated based on the performance of each base classifier on the validation set. To address class imbalance in scene classification, the system employs oversampling and special weight adjustment strategies for minority classes (such as specific types of voltage anomalies) to improve their recognition accuracy. Furthermore, the system implements model interpretability analysis, revealing the internal mechanisms of model decision-making through methods such as feature importance analysis and partial dependency graphs, enhancing the credibility and interpretability of the results. Through this multi-model fusion and optimization, the final generated basic classification model not only performs well in conventional scenarios, but also copes well with newly emerging or marginal scene types, significantly improving the overall reliability of scene recognition.

[0082] S4.4: Deploy online scene recognition services to achieve real-time scene classification; A high-performance online scene recognition service architecture was designed, deploying the trained model to the production environment to achieve real-time monitoring and classification of the photovoltaic power station's operating status. This architecture adopts a microservice design concept, modularizing the scene recognition function into independent service units, supporting elastic scaling and fault tolerance. The service interface uses a RESTful API design, providing standardized data input and output formats for easy integration with other systems. To handle high-concurrency data streams, the system implements a message queue-based streaming processing mechanism. The data acquisition module pushes real-time data to the queue, while the scene recognition service consumes data from the queue at a specified processing rate. This decoupled design not only improves system throughput but also enhances service stability, enabling it to handle data spikes and temporary service unavailability. In terms of deployment architecture, the system uses containerization technology and a service orchestration platform to achieve automated deployment and version management of the model service. To ensure service quality, the system also implements a comprehensive monitoring and alarm mechanism, tracking key indicators such as service response time, processing latency, and resource utilization in real time. When indicators exceed thresholds, alarms and self-healing processes are automatically triggered. Through this highly available and scalable service architecture, the system can continuously provide stable and reliable scene recognition services to meet the needs of real-time monitoring of photovoltaic power stations.

[0083] S4.4.1: Import the basic classification model, use the sliding window method to segment the real-time data, and construct a feature calculation result set; A highly efficient real-time data processing workflow is implemented, using a sliding window method to segment and extract features from continuous data streams. This sliding window mechanism employs an overlapping design, where new windows overlap with previous windows to a certain extent. This design captures cross-window events and transition states, avoiding information loss that might occur with hard segmentation. The window size is dynamically set; the system adaptively adjusts window parameters based on the temporal characteristics and real-time processing requirements of different scenarios. For example, rapid voltage fluctuations typically require smaller windows (e.g., seconds) for accurate capture, while load changes may require larger windows (e.g., minutes) to fully reflect their characteristics. For data within each window, the system invokes the aforementioned feature engineering pipeline to calculate a series of time-series and statistical features. To improve computational efficiency, the system implements an incremental feature calculation method, processing only new data within the window and updating existing feature values, significantly reducing redundant calculations. Furthermore, a feature caching mechanism is designed to cache frequently used intermediate calculation results, further improving computational speed. Through these optimization techniques, the system can process large amounts of high-dimensional data at near real-time speeds, generating stable and continuous feature calculation result sets, providing timely and accurate data support for scene recognition.

[0084] S4.4.2: Based on the feature calculation result set, deploy multiple model instances for load balancing and output real-time scene recognition results.

[0085] Based on the feature calculation result set, multiple model instances were deployed for load balancing, achieving high-performance real-time scene recognition. This step adopted a distributed inference architecture, deploying the scene recognition model on multiple computing nodes to form a model service pool. The front-end load balancer dynamically allocated requests based on the load and health status of each model instance, ensuring balanced utilization of system resources and rapid response. To adapt to recognition tasks of varying complexity, the system implemented a multi-layered model scheduling strategy: simple scenarios were processed quickly by lightweight models, while complex scenarios were routed to more powerful models for in-depth analysis. This hierarchical processing mechanism significantly improved the system's average response speed. Regarding the integration of processing results, the system designed a temporal consistency check mechanism to ensure reasonable continuity of recognition results across consecutive windows, avoiding frequent jumps in scene types. When a potential misclassification is detected, the system triggers a secondary analysis process, integrating information from multiple windows for verification and correction. The final scene recognition results are output in multiple ways: firstly, they are pushed to a monitoring panel in real time for intuitive viewing by operations and maintenance personnel; secondly, they are provided to downstream analysis systems, such as the power quality assessment module, through standardized interfaces. The system also maintains a historical database of recognition results, supporting retrospective queries and trend analysis. Through this efficient and reliable real-time processing mechanism, the system can continuously output accurate scene recognition results, providing important scene context information for the refined management of power supply quality.

[0086] S5 specifically includes: S5.1: Establish a power supply quality assessment index system, including dimensions such as voltage quality and reliability; A comprehensive and systematic power supply quality assessment index system has been established, providing a scientific basis for the accurate evaluation of power supply quality in photovoltaic power distribution areas. This index system adopts a multi-level structural design. The primary indexes include three core dimensions: voltage quality, power quality, and power supply reliability, comprehensively covering all key aspects of power supply quality. The voltage quality dimension mainly focuses on basic electrical characteristics such as voltage amplitude, frequency, and three-phase balance, which directly affect the normal operation of electrical equipment. The power quality dimension focuses on waveform distortion and interference factors, such as harmonic content, flicker, and transient phenomena, which have a particularly significant impact on sensitive equipment. The power supply reliability dimension emphasizes the evaluation of the continuity and stability of power supply, including indicators such as outage frequency, outage time, and power supply reliability rate. Each primary index is further divided into multiple secondary indexes. For example, voltage quality includes voltage deviation rate, voltage qualification rate, and voltage fluctuation rate; power quality includes total harmonic distortion rate, flicker severity, and voltage imbalance; and power supply reliability includes the System Average Outage Frequency Index (SAIFI) and the System Average Outage Duration Index (SAIDI). The system sets evaluation standards for each indicator based on national standards and industry norms, and makes appropriate adjustments to suit the characteristics of the photovoltaic power distribution area, ensuring the professionalism and applicability of the evaluation standards. This comprehensive indicator system design enables the system to objectively and comprehensively evaluate the power supply quality of the photovoltaic power distribution area from multiple dimensions, laying a solid foundation for subsequent differentiated assessments and problem attribution.

[0087] S5.1.1: Construct an evaluation index system and form an evaluation framework using voltage quality, power quality, and power supply reliability indicators; Based on the aforementioned indicator system, a structured evaluation framework was constructed, enabling quantitative evaluation of power supply quality. This framework first standardizes various indicators, converting indicators with different units of measurement and numerical ranges into a unified scoring scale for comprehensive evaluation. During standardization, the system employs different conversion methods based on the characteristics of the indicators: for indicators with clear standard limits, such as voltage deviation rate, a piecewise linear function is used to map actual measured values ​​to the scoring interval; for statistical indicators, such as the 95th percentile of harmonic content, a probability distribution model is used for standardization; and for composite indicators, such as the combined impact of voltage fluctuations and flicker, a multivariate function model is used for calculation. The standardized indicators are organized into a hierarchical scoring system, forming a complete evaluation framework. This framework supports flexible combination and weight adjustment of indicators, adapting to the needs of different evaluation scenarios. The system also designs a preprocessing workflow for evaluation data, including outlier detection, data smoothing, and missing value handling, ensuring the quality and reliability of the evaluation data. Furthermore, to support the traceability and transparency of the evaluation process, the system implements a detailed evaluation log recording function, recording raw data, calculation processes, and evaluation results for easy post-event review and verification. Through this rigorous and systematic evaluation framework, the system can transform complex and diverse power quality indicators into structured and quantitative evaluation results, providing a scientific basis for subsequent analysis and decision-making.

[0088] S5.1.2: Combining the scenario classification results and the evaluation framework, perform dynamic adjustments to the evaluation strategy and indicator weights to obtain evaluation parameters; By combining scenario classification results, the system dynamically adjusts the evaluation strategy and indicator weights, making the evaluation process more accurate and targeted. This dynamic adjustment mechanism is based on the concept that "the key factors for power supply quality differ in different scenarios," and achieves adaptive optimization of the evaluation model through scenario awareness. Specifically, the system first establishes a scenario-indicator importance mapping table, recording the relative importance of each indicator in different scenarios. For example, in a photovoltaic output fluctuation scenario, the weights of voltage deviation rate and voltage fluctuation rate indicators will be increased accordingly, because these indicators better reflect power supply quality problems in such scenarios; while in a load change scenario, indicators such as transient response and voltage recovery time are more important. The system constructs this mapping relationship by combining machine learning methods and expert knowledge: on the one hand, it uses scenario-problem correspondence patterns in historical data to automatically discover key indicators in different scenarios through correlation analysis and feature importance assessment; on the other hand, it combines the domain knowledge of power system experts to verify and optimize the mapping relationship discovered by the model. When the system identifies the current operating scenario, it automatically adjusts the evaluation strategy according to the scenario characteristics, including selectively activating specific indicators, adjusting indicator weights, and optimizing scoring criteria. Furthermore, the system considers the impact of contextual factors such as time, space, and external environment on the assessment. For example, it assigns higher weight to power supply reliability indicators during peak summer electricity consumption periods and pays special attention to insulation performance indicators during periods of severe air pollution. Through this intelligent dynamic adjustment mechanism, the system obtains the optimal assessment parameter settings for the current specific scenario, making power supply quality assessments more accurate and targeted.

[0089] S5.1.3: Based on the evaluation parameters, the fuzzy comprehensive evaluation method is applied to transform the quantitative indicators and generate a power supply quality evaluation report; Based on the aforementioned evaluation parameters, the fuzzy comprehensive evaluation method is applied to transform various quantitative indicators into an overall evaluation result, generating a detailed power supply quality evaluation report. The fuzzy comprehensive evaluation method is particularly suitable for handling the complexity and fuzziness in power supply quality evaluation, effectively integrating the influence of multiple indicators to form a scientifically sound overall judgment. The system first constructs a set of evaluation factors (i.e., various power supply quality indicators) and a set of evaluation comments (such as "Excellent," "Good," "Qualified," and "Unqualified" levels). Then, based on fuzzy membership theory, a membership function is established for each indicator, transforming the rigid quantitative indicators into membership degrees for each evaluation level. For example, a voltage deviation rate of 4.5% might have a membership degree of 0.7 for the "Good" level and 0.3 for the "Qualified" level. The system employs various types of membership functions, including trapezoidal functions, Gaussian functions, and sigmoid functions, to adapt to the characteristics of different indicators. After constructing the fuzzy relation matrix, the system combines the previously determined dynamic weight vector and obtains the final comprehensive evaluation result through fuzzy synthesis operations. To enhance the discriminative power and practicality of the evaluation results, the system also employs fuzzy clustering and fuzzy optimization theories for result refinement. The final power quality assessment report not only includes the overall evaluation results but also details the scores for each dimension's indicators, key issues, and historical comparative analysis. The report utilizes various visualization methods to present the evaluation results, including radar charts displaying multi-dimensional scores, trend charts showing indicator changes, and heatmaps identifying problem areas, making the evaluation results intuitive and easy to understand. Furthermore, the system provides a drill-down function for the evaluation results, allowing users to delve deeper from the overall evaluation to the raw data of specific indicators, gaining a comprehensive understanding of the evaluation's basis and process. Through this scientific and comprehensive evaluation method, the system can generate power quality assessment reports that are both theoretically sound and practically valuable, providing reliable decision support for the operation and management of photovoltaic power stations.

[0090] S5.2: Implement a multi-dimensional evaluation model to conduct differentiated evaluations for different scenarios; A multi-dimensional evaluation model has been implemented, enabling differentiated assessments for different scenarios and further improving the accuracy and practicality of power supply quality evaluation. This multi-dimensional evaluation model surpasses the traditional "one-size-fits-all" evaluation method, establishing a more flexible and refined evaluation system. First, the system implements multi-scale evaluation from a time dimension, including evaluation mechanisms for different time windows such as real-time evaluation, daily evaluation, monthly evaluation, and annual evaluation. The evaluation focus differs at different time scales: real-time evaluation mainly focuses on anomalies and fluctuations in instantaneous electrical parameters; daily evaluation focuses on intraday variation patterns and specific time-period issues in power supply quality; monthly and annual evaluations focus more on long-term trends and seasonal issues. Second, a multi-level evaluation framework has been constructed from a spatial dimension, covering three levels: transformer level, line level, and node level. Transformer-level evaluation mainly focuses on the overall power supply quality performance of distribution transformers; line-level evaluation focuses on voltage distribution and power flow characteristics on feeders; and node-level evaluation provides refined evaluation for specific needs of important user access points or sensitive areas. In addition, the system also implements targeted evaluation from a scenario dimension, establishing specialized evaluation models for typical scenarios such as photovoltaic output fluctuations, load abrupt changes, and low-voltage ride-through. Each scenario model has its specific combination of indicators and weight settings, enabling accurate evaluation of power supply quality performance under that scenario. For example, the photovoltaic output fluctuation scenario model focuses on evaluating voltage regulation capability and power balance performance; while the load change scenario model emphasizes transient response characteristics and voltage recovery capability. The system automatically selects and switches evaluation models based on scenario identification results, achieving scenario-adaptive evaluation. Finally, the system also supports personalized evaluation based on user needs. Managers can customize evaluation indicators and weights according to specific concerns to obtain evaluation results that better meet actual needs. This multi-dimensional and differentiated evaluation model greatly improves the pertinence and practical value of power supply quality evaluation, providing a powerful tool for the refined management of photovoltaic power distribution areas.

[0091] S5.3: Construct an attribution analysis model to identify the main influencing factors of power supply quality problems; An advanced attribution analysis model was constructed, enabling accurate identification of the main influencing factors and root causes of power supply quality problems. Unlike traditional methods relying on experience-based judgment, this model employs a data-driven and knowledge-integrated approach, systematically analyzing the causes of complex problems and quantifying the impact of each factor. First, the system established a comprehensive database of power supply quality influencing factors, covering grid-side factors (such as network structure and line impedance), photovoltaic-side factors (such as power generation fluctuations and inverter characteristics), load-side factors (such as load characteristics and distribution), and external environmental factors (such as weather conditions and temperature changes). Each category of factors is further subdivided into multiple specific indicators, forming a multi-level influencing factor system. Second, the system designed an attribution analysis framework based on multi-model fusion, integrating the advantages of various analytical methods. At the correlation analysis level, the system uses Pearson correlation coefficient and mutual information to identify the statistical association between power supply quality indicators and influencing factors. At the causal inference level, the system uses Granger causality test and structural equation modeling to explore the direction and strength of causal relationships between variables. At the path analysis level, the system uses Bayesian networks and decision trees to construct the propagation paths and hierarchical structures of influencing factors. These analytical results are comprehensively evaluated and cross-validated, forming a multi-faceted understanding of the causes of problems. In addition, the system also implements anomaly pattern recognition, quickly identifying known types of problems by matching patterns with typical cases in the historical database. For newly emerging complex problems, the system uses anomaly detection and cluster analysis methods to discover potential problem patterns. Finally, the system combines attribution analysis with the previous scenario classification results to achieve scenario-aware targeted attribution. In different scenarios, the system activates different attribution logics and focuses on different influencing factors, further improving the accuracy and efficiency of the analysis. Through this comprehensive and refined attribution analysis model, the system can deeply analyze the root causes of power supply quality problems, providing a clear direction for subsequent optimization and improvement.

[0092] S5.3.1: Collect data on factors affecting power supply quality, establish a causal relationship analysis model, and obtain key impact paths; The system has achieved comprehensive data collection on factors affecting power quality and the construction of a causal relationship analysis model. The scope of data collection extends beyond traditional electrical quantities to include various potential influencing factors. On the photovoltaic (PV) power generation side, the system collects operating parameters such as PV panel temperature, irradiance, tilt angle, and power generation efficiency, as well as equipment information such as inverter operating status and power control strategies. On the load side, the system acquires user load curves, electricity consumption characteristics, and large equipment start-up and shutdown records, among other electricity consumption behavior data. On the grid side, the system collects infrastructure information such as network topology, line parameters, and reactive power compensation equipment status. On the environmental side, the system integrates meteorological condition data such as temperature, humidity, wind speed, and cloud cover. This multi-source heterogeneous data, after preprocessing and feature extraction, serves as the basis for causal relationship analysis. Regarding the construction of the causal relationship model, the system first adopts a domain knowledge-guided approach, establishing a preliminary causal hypothesis graph based on power system theory and expert experience. This graph clarifies the direction and nature of possible causal relationships between variables. Then, the system applies causal discovery algorithms to verify these hypotheses with data. The system implements various causal discovery methods, including constraint-based algorithms (such as the PC algorithm) and scoring-based algorithms (such as greedy equivalence search). These algorithms infer causal relationship structures from data by analyzing conditional independence and statistical association patterns. To handle the specific characteristics of time-series data, the system also employs time-delay causal analysis methods, such as time-series causal networks, to capture causal effects at different time scales. Finally, by combining theoretical knowledge and data-driven results, the system constructs a probabilistic causal graph of power supply quality influencing factors, clearly demonstrating the dependencies and influence paths among various factors. The system pays particular attention to the identification of key influence paths, calculating important propagation paths from root causes to observed indicators through path analysis algorithms, providing a structured framework for subsequent quantitative analysis and intervention strategy design.

[0093] S5.3.2: Using the aforementioned key influence paths, apply Bayesian network analysis to determine the probabilistic dependencies between variables and obtain the contribution of influencing factors; Based on the previously identified key influencing paths, the Bayesian network method was applied to deeply analyze the probabilistic dependencies between variables and quantify the contribution of each influencing factor. As a probabilistic graphical model, Bayesian networks are particularly suitable for expressing uncertainties and conditional dependencies in complex systems, making them an ideal tool for analyzing the multi-factor influence mechanisms of power supply quality. The system first determines the topology of the Bayesian network through a structure learning algorithm, which reflects the direct dependencies between variables. During structure learning, the system combines a hybrid approach based on scoring and constraints, while also considering prior constraints from domain knowledge to avoid generating physically unreasonable connections. After the network structure is determined, the system estimates the conditional probability table (CPT) through a parameter learning algorithm. These probability parameters quantitatively describe the degree of influence of parent nodes on child nodes. For continuous variables, the system uses a conditional Gaussian network or discretization; for discrete variables, the CPT parameters are directly estimated statistically. After the complete Bayesian network is constructed, the system uses probabilistic inference algorithms to analyze influencing factors. Specifically, the system uses sensitivity analysis methods, such as variable intervention analysis and evidence influence analysis, to assess the degree of influence of each factor on power supply quality indicators. Variable intervention analysis simulates the probability distribution of indicator changes after intervention in specific factors, quantifying the strength of the intervention effect; evidence influence analysis assesses the degree of impact of observed specific factor values ​​on indicator prediction. The system also implements multivariate joint influence analysis to identify interactions and synergistic effects among factors. These analytical results are integrated into a comprehensive contribution score, reflecting the relative importance of each factor to power quality issues. For example, the analysis might show that in a region with low voltage, the reactive power control strategy of the photovoltaic inverter contributes 40%, excessive line impedance contributes 35%, and the mismatch between peak load and photovoltaic output contributes 25%. This quantitative contribution analysis provides a scientific basis for accurately identifying key influencing factors and formulating targeted optimization strategies.

[0094] S5.3.3: Based on the contribution of the influencing factors, perform a problem attribution analysis and output diagnostic results and optimization suggestions; Based on the analysis of the contribution of influencing factors, an in-depth problem attribution analysis was performed, outputting professional diagnostic results and optimization suggestions. The problem attribution analysis adopted a hierarchical structure: First, the system identified the main influencing factors according to their contribution ranking; these factors typically explain the variation in the main parts of the problem. Then, these main factors were analyzed in depth to explore the specific reasons for their abnormal states. Finally, the system analyzed the interactions between multiple factors to identify possible compound effects and systemic problems. In terms of the diagnostic mechanism, the system integrated rule-based and model-based methods: rule-based diagnosis relies on a predefined expert knowledge base, containing diagnostic rules and processing procedures for common problems; model-based diagnosis utilizes a data-driven approach, discovering potential problems through anomaly pattern detection and deviation analysis. These two methods complement each other, ensuring both rapid and accurate diagnosis of known problems and the ability to discover new problems. For the diagnosed problems, the system generated hierarchical and categorized optimization suggestions. When generating suggestions, the system considered multiple factors: technical feasibility assessment including required technical conditions and implementation difficulty; economic analysis including implementation costs and expected benefits; and risk assessment including implementation risks and uncertainty analysis. The system also supports multi-solution comparison, generating multiple alternatives simultaneously for comprehensive comparison to help decision-makers select the optimal solution. To enhance the practicality of the recommendations, the system incorporates a historical case library, providing successful solutions and lessons learned for similar problems. For example, for the identified issue of "low voltage due to insufficient reactive power control in the photovoltaic inverter," the system not only provides direct suggestions to "adjust the inverter's reactive power control parameters," but also offers specific parameter setting ranges, implementation steps, expected results, verification methods, and historical successful cases of similar adjustments. This comprehensive, specific, and targeted diagnosis and recommendations provide a direct and feasible action guide for the rapid resolution of power quality issues.

[0095] S5.4: Generate analysis reports and optimization suggestions to support decision-making.

[0096] The system generates comprehensive and practical analysis reports and optimization suggestions, providing strong support for decision-making. The analysis reports employ a multi-layered structure to meet the information needs of different user roles: the high-level summary is for management decision-makers, concisely presenting core findings and key recommendations; the technical details section is for technical professionals, providing in-depth data analysis and technical solutions; and the operational guidance section is for implementers, detailing specific implementation steps and precautions. The report content covers multiple dimensions: the power supply quality status assessment displays the scores and trends of various indicators; the problem analysis section describes in detail the identified problems and their impact scope and severity; the cause diagnosis section presents the results of attribution analysis, including causal relationship diagrams and contribution quantification analysis; and the improvement suggestions section proposes system-generated optimization solutions, including short-term emergency measures and long-term improvement strategies. In terms of content presentation, the system cleverly combines text descriptions, data tables, and visualization charts, making complex analysis results intuitive and easy to understand. Particularly in visualization, the system implements various advanced charts: radar charts display multi-dimensional assessment results; time-series charts show the trends of indicators over time; heat maps identify the spatiotemporal distribution characteristics of problems; Sankey diagrams show the flow relationships between factors; and decision tree diagrams intuitively present problem classification and diagnostic logic. Furthermore, the system supports interactive reporting, allowing users to drill down into charts for more detailed data and explanations. Regarding optimization recommendations, the system not only provides technical solutions but also incorporates economic and feasibility analyses, offering implementation cost estimates, expected benefit assessments, and return on investment analyses for each recommendation to help decision-makers make comprehensive trade-offs. The system also features a recommendation tracking mechanism to record the adoption status and implementation effects of recommendations, forming a closed-loop management system. These analysis reports and optimization recommendations are distributed through multiple channels, including a web portal, mobile applications, email push notifications, and regular briefings, ensuring relevant personnel can access information promptly. Through this comprehensive, professional, and easy-to-understand decision support information, the system effectively facilitates the rapid identification and efficient resolution of power supply quality issues in photovoltaic power distribution areas, achieving continuous improvement in power supply quality.

[0097] This application embodiment also provides a computer device, the computer device comprising: At least one processor; and, A memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform the above-described method for improving the power supply quality of the photovoltaic substation.

[0098] This application also provides a computer-readable storage medium that stores computer instructions for causing a computer to execute the above-described method for improving the power supply quality of a photovoltaic power distribution area.

[0099] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the method for improving the power supply quality of the photovoltaic power distribution area described above.

[0100] This application has the following technical effects: 1) By adopting a microservice architecture based on Docker container technology and an automatic scaling strategy, the decoupled management of the data processing pipeline was achieved, improving system processing efficiency and resource utilization. 2) By using static task mapping technology for heterogeneous systems based on serial-parallel decomposition, optimal scheduling of computing resources was achieved, improving the real-time performance of data processing; 3) By using machine learning methods to identify and classify power quality scenarios, refined analysis for different scenarios is achieved, improving the accuracy of power quality assessment; 4) Through multi-dimensional power supply quality assessment and attribution analysis, the accurate location and cause diagnosis of power supply quality problems were achieved, providing effective decision support for improving power supply quality.

[0101] This application also provides a system for improving the power supply quality of photovoltaic power distribution areas, including: a microservice architecture based on Docker container technology, a data acquisition and preprocessing module, a task mapping module, a scene recognition and classification module, and a power supply quality assessment module. Through the collaborative work of these modules, the system achieves comprehensive monitoring, evaluation, and optimization of the power supply quality of photovoltaic power distribution areas, effectively solving the problems of low data processing efficiency and lack of refined analysis capabilities in existing technologies.

[0102] It should be noted that the above embodiments are merely preferred technical solutions of this application and are not intended to limit the scope of protection of this application. Those skilled in the art can make various modifications or improvements to this application under the guidance of this application without departing from the spirit of this application, and all such modifications and improvements fall within the scope of protection of this application.

[0103] This disclosure also provides a computer-readable storage medium storing a computer program. When executed by a processor, the computer program performs the steps of the method and system for improving power supply quality in photovoltaic power distribution areas as described in the above-described method embodiments. The storage medium can be a volatile or non-volatile computer-readable storage medium.

[0104] In addition, this disclosure also provides a computer program product, which stores a computer program. When the computer program is run by a processor, it executes the steps of the method and system for improving the power supply quality of photovoltaic power distribution areas provided in any of the above embodiments of this disclosure. For details, please refer to the above method embodiments, which will not be repeated here.

[0105] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium, which can be a volatile or non-volatile computer-readable storage medium. In another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0106] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the devices and apparatuses described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed devices, apparatuses, and methods can be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0107] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0108] In addition, the functional units in the various embodiments of this disclosure 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.

[0109] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0110] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method for improving the quality of power supply in a photovoltaic district, characterized by, The application relates to a power supply quality analysis method and device. The application comprises the following steps: Data acquisition, preprocessing and analysis function modules are encapsulated into independent microservices by using a Docker container technology, an automatic expansion and contraction strategy is configured, and a decoupled management architecture of a data processing pipeline is obtained; Electrical quantity data, meteorological data and equipment state data are collected, the data is cleaned, standardized and quality evaluated, and preprocessed data is obtained; Feature analysis and decomposition are performed on a processing task of the preprocessed data, a task mapping optimization model is established, the task is allocated to heterogeneous computing resources, and optimal scheduling of the computing resources is realized; The preprocessed data and the task mapping optimization model are used to identify and classify a power supply quality scene by using a machine learning method, and a scene classification result is obtained; 2. The method of claim 1, wherein, According to the scene classification result, a multi-dimensional power supply quality evaluation index system is established, power supply quality evaluation and problem attribution analysis are performed, and a power supply quality analysis report is output. The decoupled management architecture of the data processing pipeline comprises the following steps: A message queue system is built, original data topics and preprocessed data topics are set, and a data stream processing architecture is obtained; By using the data stream processing architecture, a topic subscription mode is used to realize inter-service communication, and a service communication mechanism is formed; 3. The method of claim 1, wherein, According to the service communication mechanism, a unified service gateway layer is established, and decoupled management of the data processing pipeline is completed. The identification and classification of the power supply quality scene to obtain the scene classification result comprises the following steps: Feature extraction calculation is performed on the preprocessed data, statistical features and time sequence features are constructed, and a feature vector is obtained; The feature vector is used to perform coarse-grained division on the running state by using an unsupervised learning method, and a preliminary classification result is formed; 4. The method of claim 1, wherein, According to the preliminary classification result, a classifier is trained, and a scene identification result of photovoltaic fluctuation and load mutation is output. The identification and classification of the power supply quality scene by using the machine learning method to obtain the scene classification result comprises the following steps: The preprocessed data is used to build an ensemble learning framework, the results of multiple base classifiers are fused, and a basic classification model is generated; The basic classification model is imported, a sliding window method is used to segment and process real-time data, and a feature calculation result set is constructed; 5. The method of claim 1, wherein, According to the feature calculation result set, multiple model instances are deployed for load balancing processing, and a real-time scene identification result is output. The multi-dimensional power supply quality evaluation index system comprises the following steps: Voltage quality, power quality and power supply reliability indexes are used to construct an evaluation index system, and an evaluation framework is formed; The scene classification result and the evaluation framework are combined, an evaluation strategy and dynamic adjustment of index weights are performed, and evaluation parameters are obtained; 6. The method of claim 1, wherein, According to the evaluation parameters, a fuzzy comprehensive evaluation method is applied to convert various quantitative indexes, and a power supply quality evaluation report is generated. The power supply quality evaluation and problem attribution analysis comprise the following steps: Power supply quality influence factor data is collected, a causal relationship analysis model is established, and a key influence path is obtained; The key influence path is used to analyze the probability dependence relationship between variables by using a Bayesian network, and the contribution degree of an influence factor is obtained; According to the contribution degree of the influencing factors, a problem attribution analysis is performed, and a diagnosis result and optimization suggestion are output.

7. The method of claim 1, wherein, The feature analysis and decomposition of the execution processing task include: Using static code analysis technology, a program dependency graph is constructed, data dependency relationships between tasks are identified, and task units that can be executed in parallel are obtained; Using dynamic performance analysis tools, the task execution state is monitored, and task computing features are generated; According to the task computing features, the classification and labeling of computing-intensive and I / O-intensive tasks are performed, and the task classification result is output.

8. The method of claim 7, wherein, The method further includes: Using the data dependency relationships between tasks, a directed acyclic graph is constructed, and using the critical path analysis method, task constraint conditions are obtained; According to the task constraint conditions, a hierarchical partitioning algorithm is applied to process the set of tasks that can be executed in parallel, and a task decomposition scheme is formed; The task decomposition scheme is imported, the task granularity threshold control is performed, and the final task set is output.

9. The method according to claim 7 or 8, characterized in that, The establishment of the task mapping optimization model and the generation of the initial mapping scheme include: Deploying a distributed monitoring agent, collecting computing resource usage, constructing a resource portrait database, and forming a resource utilization evaluation model; Using the resource utilization evaluation model, using a genetic algorithm to design the task mapping optimization model, and generating an initial mapping scheme; Performing cross and mutation operations on the initial mapping scheme to optimize, and outputting a dynamic scheduling scheme of computing resources.

10. A system for improving the quality of power supply in a photovoltaic district, characterized by, It includes: A microservice architecture based on Docker container technology is used to encapsulate data collection, preprocessing, and analysis function modules into independent microservices, configure automatic scaling strategies, and realize decoupled management of data processing pipelines; The data collection and preprocessing module is used to obtain electrical quantity data, meteorological data, and equipment state data, clean, standardize, and quality assess the data, and obtain preprocessed data; The task mapping module is used to perform feature analysis and decomposition on the processing tasks of the preprocessed data, establish a task mapping optimization model, assign tasks to heterogeneous computing resources, and realize optimal scheduling of computing resources; The scene recognition and classification module is used to identify and classify power supply quality scenes based on the preprocessed data and the task mapping optimization model using machine learning methods, and obtain a scene classification result; The power supply quality evaluation module is used to establish a multi-dimensional power supply quality evaluation index system based on the scene classification result, perform power supply quality evaluation and problem attribution analysis, and generate a power supply quality analysis report.

Images