Intelligent operation and maintenance and energy efficiency optimization control method for photovoltaic power station

Abstract

This invention discloses an intelligent operation and maintenance and energy efficiency optimization control method for photovoltaic power plants, belonging to the cross-technical field of new energy power generation and smart grids. This method integrates cloud-edge-device collaborative sensing, physical information neural network digital twins, multimodal deep learning fault diagnosis, hierarchical multi-objective optimization, constrained multi-agent reinforcement learning, and federated learning technologies to construct a fully intelligent closed-loop system. It addresses the pain points of traditional technologies such as data silos, poor model generalization, delayed fault warnings, and the disconnect between operation and maintenance and energy efficiency. This invention can accurately sense equipment status, provide early warnings of faults, and achieve global energy efficiency optimization, significantly improving power generation efficiency, reducing operation and maintenance costs, and extending equipment lifespan. It is adaptable to intelligent operation and maintenance of photovoltaic power plants throughout their entire lifecycle and conforms to industry standards.

Description

Technical Field

[0001] This invention belongs to the field of cross-technology of new energy power generation and smart grid, specifically involving a method for intelligent operation and maintenance and energy efficiency optimization control of photovoltaic power plants. It is particularly applicable to the full life cycle operation and maintenance management, fault prediction and handling, and maximization of power generation efficiency of centralized ground photovoltaic power plants, distributed industrial and commercial rooftop photovoltaics, residential distributed photovoltaics, photovoltaic-storage integrated power plants, and photovoltaic clusters in complex terrains and extreme climates such as plateaus, mountains, deserts, and coastlines. This method deeply integrates core technologies such as IoT cloud-edge-device collaborative sensing, Physical Information Neural Network (PINN) digital twins, multimodal deep learning fault diagnosis, hierarchical multi-objective optimization, constrained multi-agent reinforcement learning, and federated learning incremental iteration. It constructs a fully intelligent closed-loop system encompassing data perception, physical twins, health assessment, collaborative optimization, operation and maintenance execution, and self-evolutionary iteration. This completely breaks down the industry barriers that separate photovoltaic power plant operation and maintenance from energy efficiency control. It fundamentally solves the core pain points of existing technologies, such as poor generalization of pure data-driven models, delayed fault warnings, insufficient multi-system collaboration, imbalance between short-term power generation revenue and long-term equipment lifespan, and weak adaptability to extreme operating conditions. It is fully compatible with the industry standards of "artificial intelligence + new energy" and the intelligent operation needs of large-scale, clustered, and refined photovoltaic power plants. It can be directly applied to the design and deployment of new photovoltaic power plants and the intelligent upgrading and transformation of existing photovoltaic power plants. Background Technology

[0002] Photovoltaic power generation, with its core advantages of being clean and pollution-free, renewable, and having continuously decreasing costs throughout its entire life cycle, has achieved rapid development in terms of scale, clustering, and large capacity. Photovoltaic power plants are rapidly expanding towards centralized systems of 100MW or more, distributed full coverage, deployment in complex terrains, and synergistic integration of photovoltaics and energy storage. However, while the industry is developing rapidly, existing photovoltaic power plant operation and maintenance and energy efficiency control technologies still suffer from six fatal bottlenecks that are incompatible with industry development, severely restricting power plant revenue, operation and maintenance cost control, and equipment lifespan. These specific shortcomings are as follows:

[0003] 1. Severe fragmentation of multi-source data, making it impossible to break down data silos: Photovoltaic power plants involve multiple brands and models of equipment, including photovoltaic modules, string inverters, centralized inverters, combiner boxes, transformer substations, energy storage systems, weather stations, tracking brackets, and inspection equipment. Inconsistent interface protocols, heterogeneous data formats, and asynchronous data collection timestamps among different manufacturers lead to high data acquisition costs and extreme difficulty in cross-device data fusion, resulting in severe "data silos." Existing technologies can only achieve simple data collection and storage for single devices, failing to establish a comprehensive mapping between environmental meteorological parameters, equipment physical characteristics, real-time health status, power generation efficiency, and grid dispatch requirements. Massive amounts of operational data can only be used for basic real-time monitoring, unable to provide accurate data support for status perception, fault diagnosis, and energy efficiency optimization, thus failing to fully realize the value of the data.

[0004] 2. Digital twin models have extremely low accuracy and insufficient generalization ability: Most existing photovoltaic digital twin solutions are static geometric models, which can only restore the physical location and appearance of power station equipment. They cannot reproduce the real physical processes of photovoltaic module light-to-electricity conversion, inverter energy conversion, equipment aging and degradation, environmental coupling effects, and fault evolution. Some pure data-driven twin models do not embed physical mechanism constraints. Under complex operating conditions such as extreme high temperature, strong wind and sand, sensor failure, and large-scale data loss, the model prediction deviation increases by more than 50%, and the generalization ability is extremely poor. They cannot provide reliable virtual mapping support for optimization decisions and are difficult to adapt to the long-term operation of power stations in complex scenarios such as plateaus and mountains.

[0005] 3. Weak fault diagnosis and early warning capabilities, resulting in huge losses from passive operation and maintenance: Traditional photovoltaic operation and maintenance is based on the core model of "post-event repair + regular manual inspection", which relies entirely on the experience judgment of operation and maintenance personnel. The failure rate is as high as 30% and the misjudgment rate exceeds 25%. It is unable to achieve early identification and accurate location of hidden "sub-health" faults such as module microcracks, PID potential-induced degradation, inverter IGBT aging, poor junction box contact, and bracket jamming. Existing AI diagnostic solutions are based only on electrical timing or single-modal data of images, without integrating multi-dimensional features such as thermal infrared, acoustics, and vibration. The fault identification accuracy is low and the early warning lead time is less than 7 days, resulting in the equipment operating with defects for a long time, with an average annual power generation loss of more than 15% and a shortened equipment lifespan of 3-5 years.

[0006] 4. Limited and singular energy efficiency optimization methods, lacking multi-objective balancing capabilities: Existing energy efficiency optimization technologies only focus on the independent MPPT control of a single string or a single inverter. In scenarios with local shading, inconsistent component degradation, and uneven dust accumulation, they are prone to getting stuck in local optima and cannot achieve power plant-level global power maximization. At the same time, the optimization objective only pursues the maximization of instantaneous power generation, without incorporating the accumulation of thermal / electrical / mechanical stress, lifespan loss, operation and maintenance costs, grid dispatch instructions, and time-of-use pricing into a unified optimization framework. This creates a vicious cycle of "short-term gains and long-term losses," failing to balance the three core operational objectives of power generation revenue, operation and maintenance costs, and equipment lifespan, resulting in a persistently high levelized cost of electricity (LCOE) throughout the entire life cycle.

[0007] 5. Lack of multi-timescale response capability and poor adaptability to extreme conditions: Existing optimization schemes are mostly static day-ahead scheduling modes with a time granularity of only 1 hour, which cannot adapt to second / millisecond-level changes in operating conditions such as rapid cloud movement, minute-level changes in irradiance, and grid frequency fluctuations; lacking a hierarchical architecture of "global day-ahead optimization - real-time rolling optimization - event-triggered rapid response", it is impossible to balance the economy of global optimization with the real-time performance of extreme conditions, resulting in a long-term curtailment rate of more than 15% for power plants, large fluctuations in grid-connected power, frequent overload operation of equipment, and extremely poor grid adaptability and operational stability.

[0008] 6. Operation and maintenance are completely separated from energy efficiency optimization, lacking a closed-loop self-evolution mechanism: In existing technologies, operation and maintenance management and energy efficiency optimization are two independent system modules. Fault diagnosis results cannot be automatically converted into energy efficiency optimization control strategies, and energy efficiency optimization schemes are not dynamically adjusted in conjunction with the real-time health status of equipment, resulting in the common industry problem of "operation and maintenance and optimization being two separate entities." At the same time, the model parameters are fixed, making it impossible to achieve continuous iterative optimization based on real-time power plant operation data, operation and maintenance feedback, and equipment aging trends. As the equipment's operating years increase and the operating environment changes, the system's diagnostic accuracy and optimization performance continuously decline, failing to adapt to the full life cycle operation needs of the power plant.

[0009] In recent years, although there have been attempts at technological improvements within the industry, existing photovoltaic power plant operation and maintenance solutions based on digital twins only focus on equipment status monitoring and basic operation and maintenance, failing to achieve deep collaboration between operation and maintenance and energy efficiency optimization. Furthermore, digital twin models lack physical mechanism constraints and have insufficient generalization ability under extreme conditions. Some technologies attempt to introduce reinforcement learning and multi-objective optimization algorithms to improve energy efficiency control, but they fail to incorporate equipment health status, operation and maintenance costs, and safety constraints into the optimization loop, resulting in extremely poor feasibility for implementation and inability to be practically applied to engineering scenarios. Existing technologies are completely unable to meet the industry standards and the refined and intelligent operation requirements of large-scale photovoltaic power plants. Therefore, developing an intelligent control method for photovoltaic power plants that combines high status perception accuracy, strong fault early warning capabilities, global multi-objective collaborative optimization, deep integration of operation and maintenance and energy efficiency, continuous self-evolution, and adaptation to extreme conditions has become a core technical problem that the photovoltaic industry urgently needs to solve. Summary of the Invention

[0010] Addressing the six core shortcomings of existing technologies, this invention aims to provide a method for intelligent operation and maintenance and energy efficiency optimization control of photovoltaic power plants. Through the organic integration of six core modules—cloud-edge-device data collaboration, physical twins, multimodal fault diagnosis, hierarchical optimization, intelligent operation and maintenance closed-loop, and reinforcement learning self-evolution—three major breakthrough innovations are achieved: First, overcoming the generalization bottleneck of purely data-driven models by constructing a high-precision digital twin model with physical mechanism constraints, enabling accurate real-time status perception at the component level; second, overcoming the industry pain point of the separation between operation and maintenance and energy efficiency by constructing a full-process linkage mechanism of fault diagnosis, energy efficiency optimization, operation and maintenance execution, and effect feedback, achieving deep collaboration between predictive operation and maintenance and global energy efficiency optimization; and third, overcoming the limitations of single-objective optimization by constructing a hierarchical multi-objective optimization framework to achieve global Pareto optimality by maximizing power generation revenue, minimizing operation and maintenance costs, maximizing equipment lifespan, and optimizing grid connection stability. This invention can significantly improve the power generation efficiency of photovoltaic power plants, reduce operation and maintenance costs, extend equipment life, and improve grid connection stability. It has outstanding novelty, inventiveness and practicality, and is fully adaptable to the operation needs of photovoltaic power plants in all scenarios and throughout their entire life cycle.

[0011] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: a method for intelligent operation and maintenance and energy efficiency optimization control of photovoltaic power plants, comprising the following steps:

[0012] Step S1: Multi-source heterogeneous data collection and preprocessing through cloud-edge-device collaboration to build a standardized data system for photovoltaic power plants across all dimensions;

[0013] Step S2: Construct a physical information-driven digital twin model of a photovoltaic power station to achieve real-time synchronization and multi-physics collaborative simulation between the physical power station and the virtual model;

[0014] Step S3: Based on multimodal fusion deep learning and digital twins, achieve accurate prediction of equipment failures and full life cycle health status assessment;

[0015] Step S4: Construct a hierarchical multi-objective collaborative optimization decision-making framework, and realize global energy efficiency dynamic optimization control based on constrained multi-agent reinforcement learning;

[0016] Step S5: Based on fault warning and energy efficiency optimization results, realize the automatic generation, closed-loop management and collaborative execution of intelligent operation and maintenance tasks;

[0017] Step S6: Based on deep reinforcement learning, model self-evolution and cloud-based full lifecycle management are implemented to achieve continuous iterative optimization throughout the entire process.

[0018] Step S1: Multi-source heterogeneous data acquisition and preprocessing through cloud-edge-device collaboration

[0019] This step employs a three-tiered distributed architecture of edge perception, edge preprocessing, and cloud fusion, completely resolving the problems of data fragmentation, time asynchrony, high noise interference, and numerous missing values ​​in existing technologies. This provides high-quality data input for subsequent model training and decision-making. The specific implementation is as follows:

[0020] 1. Multi-source heterogeneous data acquisition across all dimensions: Through intelligent sensing devices deployed at the edge layer, six categories of core operational data are collected, covering the full-dimensional operational status of the power plant: Electrical data of equipment operation, acquired through component-level voltage / current / temperature sensors, inverter communication interfaces, and combiner box acquisition units, with key power parameters acquired at a frequency of 100Hz and conventional parameters at a frequency of 1Hz; Environmental meteorological data, acquired through distributed weather stations at a frequency of once per minute, simultaneously accessing satellite cloud images and numerical weather forecast data for the next 72 hours; Multimodal sensing data, acquired through drones, quadrupedal inspection robots, fixed infrared thermal imagers, and acoustic sensors, including visible light images, thermal infrared images, equipment operating noise, and mechanical vibration data; Mechanical posture data, real-time acquisition of tracking support tilt angle, drive motor current, and wind load vibration data; Grid dispatch data, connecting to the grid dispatch system to obtain active / reactive power dispatch instructions, time-of-use pricing, and grid connection point parameters; Equipment ledger data, synchronizing equipment factory parameters, historical maintenance records, and fault handling work orders.

[0021] 2. Real-time preprocessing at the edge: The PTP high-precision time protocol is used to achieve microsecond-level time synchronization of all sensing devices, solving the problem of clock drift among multiple devices; outlier cleaning is completed through an improved 3σ principle and isolated forest algorithm, and the dynamic sliding window width is adaptively adjusted according to the data volatility to avoid misjudging normal data fluctuations such as irradiation mutations as anomalies; for long-term missing data with more than 10 consecutive sampling points, the Spatiotemporal Graph Neural Network (STGNN) is used for topological interpolation, with an average relative error of ≤3% for completion; wavelet transform algorithm is used to denoise time series data, Gaussian filtering algorithm is used to denoise image data, and Min-Max normalization method is used to map all data to the [0,1] interval to eliminate the influence of dimensions; mutual information method is used to screen core features, and PCA principal component analysis is combined to compress the original 100-dimensional features to 20 core principal components, greatly reducing data redundancy.

[0022] 3. Cloud-based data fusion and storage: The federated learning fusion algorithm is used to achieve collaborative fusion of multi-source data, construct a unified and standardized feature vector, completely break down data silos, and achieve a data fusion accuracy of ≥99%; the edge database stores nearly one month of real-time data for rapid on-site decision-making and control command execution; the cloud adopts a hybrid architecture of TDengine time-series database + MongoDB unstructured database to achieve encrypted storage and rapid retrieval of data throughout the entire lifecycle, which complies with the power industry's data security standards.

[0023] II. Step S2: Constructing a Physical Information-Driven Digital Twin Model of a Photovoltaic Power Plant

[0024] This step constructs a four-layer digital twin architecture: physical entity layer, virtual simulation layer, data interaction layer, and fusion inference layer. It deeply integrates physical mechanisms and data-driven approaches, completely solving the pain points of existing digital twin models, such as lack of physical constraints, poor generalization, and low simulation accuracy. The specific implementation is as follows:

[0025] 1. Basic Model Construction: Based on power plant design drawings, equipment rated parameters, and geographical topology information, a 1:1 high-fidelity virtual model is constructed using the Unity3D engine to accurately map all physical equipment and geographical environment; it integrates multi-physics coupling mechanism models such as the photovoltaic module single diode five-parameter photo-to-electricity conversion model, one-dimensional heat conduction thermodynamic model, tracking mechanism kinematic model, inverter power conversion efficiency model, and energy storage charging and discharging loss model, to completely reproduce the entire physical process of photovoltaic energy conversion, equipment aging, and fault evolution; it realizes bidirectional real-time data interaction between the virtual model and the physical power plant through a standardized API interface, with a data interaction latency of ≤200ms, a synchronization frequency of 1 time / minute under normal operating conditions, and an increased synchronization frequency of 100ms / time under extreme operating conditions.

[0026] 2. Construction of PINN Physical Information Neural Network: The photovoltaic multiphysics mechanism equations are transformed into residual loss terms and embedded in the neural network training process. The total loss function is a weighted sum of data loss, mechanism residual loss, boundary condition loss, and initial condition loss. An 8-layer fully connected residual network structure is adopted, and the Tanh activation function is used to ensure smooth output. The model is trained using the AdamW optimizer and cosine annealing learning rate strategy. An alternating optimization method of mechanism constraint initialization and measured data fine-tuning is adopted to ensure the physical consistency of the model. After training, the model can output real-time state parameters at the component level. The average absolute error of junction temperature prediction is ≤1.8℃, the average absolute error of current prediction is ≤2.5%, and the occlusion recognition accuracy is ≥95%. Even under extreme conditions with 30% missing data, the prediction deviation is still ≤5%.

[0027] 3. Collaborative Simulation Applications: Based on digital twin models, four categories of collaborative simulations are conducted to provide accurate data support for subsequent fault diagnosis, energy efficiency optimization, and operation and maintenance scheduling: fault evolution simulation, simulating the occurrence and development process of different types and levels of faults; energy efficiency simulation, simulating power generation efficiency under different environments, parameters, and strategies; operation and maintenance simulation, simulating the implementation effects of different inspection, cleaning, and disposal schemes; and extreme operating condition simulation, simulating the power plant operation status under strong winds, high temperatures, sandstorms, and sudden changes in cloud cover, and predicting risks in advance.

[0028] III. Step S3: Multimodal fusion fault prediction and health status assessment

[0029] This step integrates multi-dimensional features from sequential electrical data, thermal infrared images, and acoustic vibration to achieve early warning of latent faults and full lifecycle health management of equipment. The specific implementation is as follows:

[0030] 1. Multimodal diagnostic model construction: A dual-stream cross-attention architecture is adopted. The timing branch uses a TCN+LSTM hybrid network to extract electrical timing features, the image branch uses a lightweight MobileNetV3 network to extract thermal infrared image features, and the acoustic branch uses CNN+STFT to extract vibration spectrum features. The cross-attention mechanism realizes multimodal feature alignment and weighted fusion, outputting fault type, fault level, fault location, and fault cause, while predicting the equipment operation trend and potential fault warning information for the next 7-45 days.

[0031] 2. Model Training and Optimization: A dataset is constructed using historical fault data, normal operation data, and simulated fault data generated by digital twin simulation of the power plant. It covers 12 typical faults, including component microcracks, hot spots, PID attenuation, and inverter IGBT aging. The training set accounts for 70%, the validation set for 20%, and the test set for 10%. The model is trained using the Adam optimization algorithm with an initial learning rate of 0.001. The model achieves a fault identification accuracy of ≥98%, a false alarm rate of ≤3%, a warning lead time of ≥7 days for serious faults, and a warning lead time of ≥30 days for latent faults.

[0032] 3. Health Status Assessment: Based on the output of digital twin theory and the actual measured data, the health factor HI is calculated, and four equipment states are divided into healthy, sub-healthy, performance degradation, and fault. The degradation source is separated and accurately located through PSO particle swarm parameter identification, distinguishing the degradation causes such as dust accumulation, aging, PID, poor contact, etc. The Wiener process degradation model + particle filter algorithm is used to evaluate the remaining service life (RUL) of the equipment, and output quantitative results with a 95% confidence interval, providing an accurate basis for operation and maintenance decisions.

[0033] 4. Tiered early warning and handling: Serious faults immediately lock the equipment and trigger triple alarms of sound and light + SMS + cloud; general faults push maintenance reminders to the terminal; and sub-healthy status is included in the predictive maintenance plan to achieve proactive fault prediction and tiered handling.

[0034] IV. Step S4: Hierarchical Multi-Objective Cooperative Optimization Control

[0035] This step constructs four core optimization objectives, combines three-layer time-scale optimization with constrained multi-agent reinforcement learning, and achieves global Pareto optimality. The specific implementation is as follows:

[0036] 1. Optimization Objectives and Constraints: Optimization objectives include maximizing power generation revenue over the entire life cycle, minimizing operation and maintenance costs, maximizing the average lifespan of equipment, and minimizing curtailment rate; constraints cover four main categories: safe operation of equipment (module temperature ≤ 65℃, energy storage SOC 10%-90%), grid connection standards, environmental safety restrictions (outdoor operations are prohibited when wind speed ≥ 15m / s), and operation and maintenance resource restrictions, ensuring that the optimization strategy can be implemented.

[0037] 2. Three-layer hierarchical optimization architecture: The day-ahead global optimization layer uses the NSGA-II algorithm to solve the 24-hour optimal control plan every morning, and combines the equipment health status to limit the operating load of low-HI equipment; the real-time rolling optimization layer uses the MPC model predictive control algorithm to correct the control plan every 10 minutes, and dynamically adjusts the weights to adapt to different operating conditions such as sunny weather, high temperature, and peak grid conditions; the event-triggered fast optimization layer responds to emergency events such as sudden changes in irradiance and grid disturbances in milliseconds, and uses the multi-peak MPPT fast scanning algorithm to lock the global maximum power point, and ensures the safety of equipment and grid through emergency control.

[0038] 3. Multi-agent collaborative optimization: Four intelligent agents are constructed: tracking bracket, cleaning scheduling, inverter MPPT, and energy storage charging and discharging. The PPO near-end strategy optimization + SafetyShield safety layer algorithm is adopted to embed hard constraints into the optimization process. The reward function integrates four indicators: power generation gain, operation and maintenance cost, equipment stress, and life extension, and the dynamic weights are adapted to different scenarios. The model is first trained in a digital twin environment with tens of thousands of rounds of simulation. After verification, it is deployed to the edge node to achieve global collaborative optimization of multiple subsystems.

[0039] V. Step S5: Closed-loop management of intelligent operation and maintenance tasks

[0040] This step achieves seamless integration of fault diagnosis, energy efficiency optimization, and operation and maintenance execution, completely resolving the industry pain point of the disconnect between operation and maintenance and optimization. The specific implementation is as follows:

[0041] 1. Automatic task generation: Based on fault warning results, energy efficiency optimization requirements, and digital twin simulation results, the system automatically generates standardized operation and maintenance tasks such as fault handling, component cleaning, regular inspection, and equipment maintenance, clearly defining the task type, priority, location, completion deadline, required resources, operation guidelines, and safety specifications.

[0042] 2. Intelligent scheduling: An improved genetic algorithm is used to prioritize tasks, with serious fault handling taking precedence over energy efficiency optimization tasks. The allocation of operation and maintenance resources and travel routes are optimized. Automated inspection equipment is prioritized in complex scenarios such as plateaus and mountains, and the system is dynamically adapted to extreme weather and new fault warnings.

[0043] 3. Task execution monitoring: Automated tasks are executed autonomously by drones and robots, and data is transmitted back in real time; manual tasks push intelligent work orders to the terminal, complete safety verification before execution, and are monitored throughout the process to ensure operational safety.

[0044] 4. Closed-loop evaluation: Quantitatively evaluate the operation and maintenance effect, check the equipment status recovery of fault handling, check the dust removal rate and energy efficiency improvement during cleaning, update the equipment ledger if the standard is met, and regenerate the task if the standard is not met, forming a closed-loop process of early warning-generation-scheduling-execution-evaluation-optimization.

[0045] VI. Step S6: Model Self-Evolution and Cloud-based Full Lifecycle Management

[0046] This step establishes a continuous optimization mechanism to completely resolve the pain points of model solidification and performance degradation. The specific implementation is as follows:

[0047] 1. Cloud Management Platform: Build a B / S architecture cloud platform that integrates seven major functions: data management, digital twin simulation, fault management, energy efficiency management, operation and maintenance scheduling, model management, and report statistics. It enables remote visual control and supports "minor staffing" or even "unmanned operation".

[0048] 2. Data Value Mining: Statistical analysis of full lifecycle operation data to uncover fault patterns, energy efficiency influencing factors, and equipment aging trends, providing data support for model iteration.

[0049] 3. Self-evolution mechanism: The central learning engine is driven by the PPO algorithm. Incremental learning is triggered when conditions such as ≥50 new fault samples and prediction error >3% are met. The EWC elastic weight integration algorithm avoids catastrophic forgetting. Federated learning is used to achieve model sharing optimization under the data privacy and security of multiple power plants. Through model version management and canary release, the system can be guaranteed to operate stably and continuously adapt to equipment aging and changes in operating conditions.

[0050] Compared with existing technologies, this invention possesses outstanding substantive features and significant beneficial effects, and its inventiveness, novelty, and practicality far surpass those of traditional solutions.

[0051] 1. High-precision state perception, breaking through the generalization bottleneck: The PINN physical information digital twin model deeply integrates physical mechanisms and data-driven approaches, maintaining high prediction accuracy even under extreme conditions. The data fusion accuracy rate is ≥99%, providing a reliable virtual mapping for optimized decision-making and completely solving the problem of poor generalization of pure data-driven models.

[0052] 2. Comprehensive fault management to significantly reduce operational losses: The multimodal diagnostic model has a fault identification accuracy of ≥98%, provides early warnings 7-45 days in advance, reduces unplanned downtime by more than 60%, reduces power generation losses by 40%-50%, and can identify latent faults 1-3 months in advance, avoiding equipment from operating with defects.

[0053] 3. Global multi-objective optimization to achieve optimal benefits throughout the entire life cycle: Three-level hierarchical optimization + multi-agent collaboration improves power generation efficiency by 8%-15%, reduces operation and maintenance costs by 30%-40%, extends equipment life by 8-10 years, reduces curtailment rate by 20%-30%, and reduces levelized cost of electricity (LCOE) by 15%-25% throughout the entire life cycle.

[0054] 4. Deep collaboration in operation and maintenance efficiency, completely solving the problem of disconnect between the two systems: Closed-loop linkage throughout the entire process improves operation and maintenance efficiency by more than 40%, reduces labor costs by more than 50%, enables "unmanned operation" in complex scenarios, and significantly reduces safety production risks.

[0055] 5. Self-evolution throughout the entire lifecycle, ensuring continuous and stable system performance: Incremental + federated learning ensures that the system adapts to equipment aging and changes in operating conditions, with performance degradation of ≤5% after 5 years of operation, far superior to traditional fixed models.

[0056] 6. High adaptability and feasibility, outstanding engineering value: It is compatible with existing power station equipment, requires no large-scale hardware modification, is suitable for photovoltaic power stations in all scenarios, meets the latest industry standards, has low modification costs and high promotion value. Attached Figure Description

[0057] Figure 1 This is a flowchart illustrating the overall process framework of the intelligent operation and maintenance and energy efficiency optimization control method of the present invention.

[0058] Figure 2 This is a diagram illustrating the architecture of the multi-source data acquisition and physical information digital twin model of the present invention.

[0059] Figure 3 This is a schematic diagram of the multimodal fault diagnosis and hierarchical optimization control principle of the present invention;

[0060] Figure 4 This is a flowchart of the intelligent operation and maintenance closed-loop management and model self-evolution of the present invention. Detailed Implementation

[0061] The invention will be further described in detail below with reference to specific embodiments. This embodiment takes a 100MW centralized photovoltaic power station integrating photovoltaic and energy storage in the Qinghai Plateau as an example. The power station has an average altitude of 3300 meters, with winter temperatures as low as -40℃ and summer temperatures as high as 40℃. It experiences more than 60 days of sandstorms per year. The system is equipped with 250,000 photovoltaic modules, 500 string inverters, 1250 combiner boxes, 20 transformer substations, and a 10MW / 20MWh energy storage system. 80% of the system consists of tracking photovoltaic brackets. The original operation and maintenance mode suffered from high failure detection rate, low energy efficiency, and high operation and maintenance costs. By adopting the method of this invention, the core pain points are completely solved. The specific implementation process is as follows:

[0062] S1 Cloud-Edge Collaborative Multi-Source Data Acquisition and Preprocessing

[0063] The edge layer deploys component-level voltage / current / temperature sensors, 12 distributed weather stations, 6 visible light + infrared dual-camera drones, 4 quadruped inspection robots, and 30 fixed infrared thermal imagers to achieve full-dimensional data acquisition. On-site deployment of 8 industrial-grade edge servers completes data synchronization, anomaly cleaning, missing data imputation, and noise reduction preprocessing, with long-term missing data imputation error ≤2.8%. The cloud uses TDengine time-series database + MongoDB unstructured database to achieve multi-source data fusion storage, with a data fusion accuracy of ≥99.2%, providing high-quality data input for subsequent models.

[0064] S2 Physical Information Digital Twin Model Construction

[0065] A 1:1 virtual model was built using Unity3D to accurately map all equipment and the plateau terrain environment. A multi-physics coupling mechanism model was integrated to construct a PINN physical information neural network. After three years of training with historical operating data and mechanism constraints, the component junction temperature prediction error was ≤1.6℃, the current prediction error was ≤2.2%, and the prediction deviation was still ≤5% even under the condition of 30% missing data. Based on the model, four simulations were carried out: fault evolution, energy efficiency, operation and maintenance, and extreme operating conditions. These simulations simulated scenarios such as plateau sandstorms, high temperatures, and sudden changes in cloud cover, providing accurate data support for subsequent diagnosis, optimization, and scheduling.

[0066] S3 Multimodal Fault Diagnosis and Health Assessment

[0067] A multimodal fusion deep learning model was constructed and trained using over 50,000 historical and simulation data points. The model achieved a fault identification accuracy of 98.7%, a latent fault warning lead time of ≥30 days, and a false alarm rate of ≤2.6%. During implementation, the oxidation fault of the B3 group string terminal of inverter #4 array #8 was successfully diagnosed, with a warning issued 32 days in advance, preventing equipment burnout and power generation loss. Through the assessment of health factor HI and remaining service life RUL, the health status of the equipment was accurately classified, the causes of degradation were quantified, and precise basis was provided for operation and maintenance decisions.

[0068] S4 hierarchical multi-objective collaborative optimization control

[0069] Four optimization objectives were identified: power generation revenue, operation and maintenance costs, equipment lifespan, and curtailment rate. Hard constraints were set, such as component temperature ≤65℃ and energy storage SOC 10%-90%. A three-level hierarchical optimization was implemented, with the NSGA-II algorithm used to complete the day-ahead global optimization every early morning, and the MPC algorithm used for rolling correction every 10 minutes. In the event of sudden irradiance changes, the global MPP was locked within 100ms. Four intelligent agents, namely tracking brackets, cleaning and scheduling, inverters, and energy storage, were constructed and deployed after tens of thousands of rounds of digital twin simulation training. As a result, the power plant's power generation efficiency increased by 9.2%, the curtailment rate decreased from 18.5% to 3.8%, and the annual ancillary service revenue increased by 1.2 million yuan.

[0070] S5 Intelligent Operation and Maintenance Closed-Loop Management

[0071] Based on fault warning and energy efficiency optimization results, standardized operation and maintenance tasks are automatically generated. An improved genetic algorithm is used to intelligently schedule resources, prioritizing the use of drones and robots to replace manual inspections and reducing the risk of manual operations in complex areas. The entire task execution process is monitored, and the effect is quantitatively evaluated after completion. The component cleaning and dust removal rate reaches 96.5%, and the equipment HI value recovers to above 0.92 after fault handling. After implementation, operation and maintenance efficiency is improved by 45%, labor costs are reduced by 52%, and the average fault repair time is shortened by 75%, completely solving the problem of the disconnect between operation and maintenance and optimization.

[0072] S6 Model Self-Evolution and Cloud Management

[0073] A cloud-based full lifecycle management platform was built to achieve remote and visual control. Maintenance personnel can view the status and issue instructions in real time through computers and apps. Data mining analysis was conducted to analyze the impact of high-altitude sandstorms and high temperatures on faults and energy efficiency, and to optimize early warning thresholds and cleaning strategies. When incremental learning conditions are met, the EWC algorithm is used to iterate the model, and federated learning is used to achieve shared optimization among multiple power plants. After 6 months, the fault diagnosis accuracy rate increased to 99.1%, and the system's adaptability continued to improve.

[0074] Summary of Implementation Results

[0075] In this embodiment, after adopting the method of the present invention, the annual power generation of the 100MW plateau photovoltaic power station increased from 150 million kWh to 166.5 million kWh, an increase of 11%; the annual operation and maintenance cost decreased from 8 million yuan to 5.2 million yuan, a decrease of 35%; unplanned downtime decreased by 65%; the average service life of the equipment increased from 18 years to 26 years, an increase of 8 years; the curtailment rate decreased from 18.5% to 3.8%, a decrease of 14.7 percentage points; "less manned operation" was achieved in the plateau environment, the workload of manual operation and maintenance was reduced by 60%, the safety production risk was greatly reduced, and the operational efficiency was significantly improved.

[0076] The above description is only a preferred embodiment of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent operation and maintenance and energy efficiency optimization control of a photovoltaic power station, comprising the steps of photovoltaic power station operation data collection, equipment state diagnosis, power generation energy efficiency optimization and operation and maintenance task management, characterized in that, Includes the following steps: S1 adopts a three-level distributed architecture of end-layer perception, edge preprocessing, and cloud fusion to complete the collection and standardized preprocessing of multi-source heterogeneous data from photovoltaic power plants, and build a full-dimensional standardized data system covering environmental meteorology, equipment electrical, mechanical pose, multimodal perception, power grid dispatch and equipment ledger. S2 constructs a four-layer digital twin architecture consisting of a physical entity layer, a virtual simulation layer, a data interaction layer, and a fusion inference layer. It embeds a photovoltaic multi-physics coupling mechanism model in the virtual simulation layer and constructs a photovoltaic power station digital twin model with physical mechanism constraints based on the physical information neural network PINN, thereby realizing bidirectional real-time synchronization and multi-physics collaborative simulation between the physical power station and the virtual model. S3, based on the digital twin model of step S2 and the standardized data of step S1, uses a multimodal fusion deep learning model to complete the accurate prediction of photovoltaic equipment faults and the full life cycle health status assessment, and outputs graded early warning information. S4 takes maximizing power generation revenue throughout the entire life cycle, minimizing operation and maintenance costs, maximizing equipment lifespan, and minimizing curtailment rate as optimization objectives. Combining the equipment health status assessment results output from step S3, it constructs a hierarchical multi-objective collaborative optimization decision framework of day-ahead global optimization, real-time rolling optimization, and event-triggered rapid optimization. Based on constrained multi-agent reinforcement learning, it realizes dynamic optimization control of global energy efficiency of photovoltaic power plants. Based on the fault classification and early warning information in step S3 and the energy efficiency optimization and control results in step S4, S5 automatically generates standardized operation and maintenance tasks, completes the intelligent scheduling, collaborative execution and quantitative evaluation of the operation and maintenance tasks, and forms a closed-loop management of the entire operation and maintenance process. S6 builds a model self-evolution mechanism based on deep reinforcement learning. It combines the full life cycle operation data and maintenance feedback data of photovoltaic power plants, and completes the continuous iterative optimization of the model through incremental learning and federated learning to achieve cloud-based full life cycle management. 2.The intelligent operation and maintenance and energy efficiency optimization control method of the photovoltaic power station of claim 1, characterized in that: In step S2, the total loss function of the physical information neural network PINN is a weighted sum of data loss, mechanism residual loss, boundary condition loss and initial condition loss; the mechanism residual loss is obtained by transforming the photovoltaic multiphysics coupling mechanism equation, and the multiphysics coupling mechanism model includes the photovoltaic module photo-to-electric conversion model, thermodynamic model, inverter power conversion model, energy storage charging and discharging loss model and tracking mechanism kinematic model. 3.The intelligent operation and maintenance and energy efficiency optimization control method of the photovoltaic power station of claim 1, characterized in that: In step S2, the physical information neural network PINN adopts an 8-layer fully connected residual network structure, with the Tanh function as the activation function, and is trained using the AdamW optimizer and cosine annealing learning rate strategy. After training, the digital twin model has a prediction deviation of ≤5% under extreme conditions with 30% missing data, an average absolute error of ≤1.8℃ for component junction temperature prediction, and an average absolute error of ≤2.5% for current prediction. 4.The intelligent operation and maintenance and energy efficiency optimization control method of a photovoltaic power station according to claim 1, characterized in that: In step S1, the edge preprocessing includes: using the PTP high-precision time protocol to achieve microsecond-level time synchronization of all sensing devices; cleaning outliers using the improved 3σ principle and the isolated forest algorithm; using the spatiotemporal graph neural network STGNN to complete topological interpolation of long-term missing data, with an average relative error of ≤3%; denoising the data using wavelet transform and Gaussian filtering; and reducing the dimensionality of features using Min-Max standardization and principal component analysis to construct a standardized feature vector. 5.The intelligent operation and maintenance and energy efficiency optimization control method of the photovoltaic power station of claim 1, characterized in that: In step S3, the multimodal fusion deep learning model adopts a dual-stream cross-attention architecture, including a temporal branch, an image branch, and an acoustic branch; the temporal branch uses a TCN+LSTM hybrid network to extract electrical temporal features, the image branch uses a MobileNetV3 network to extract thermal infrared image features, and the acoustic branch uses CNN+Short Time Fourier Transform to extract vibration spectrum features. The cross-attention mechanism is used to align and weightedly fuse multimodal features, and output fault information and fault warning information for the next 7-45 days.

6. The intelligent operation and maintenance and energy efficiency optimization control method for photovoltaic power plants according to claim 5, characterized in that: In step S3, the fault identification accuracy of the multimodal fusion deep learning model is ≥98%, the false alarm rate is ≤3%, the early warning lead time for serious faults is ≥7 days, and the early warning lead time for latent faults is ≥30 days. The full life cycle health status assessment includes calculating the equipment health factor HI to classify the equipment health status, separating the attenuation source through particle swarm parameter identification, and using the Wiener process degradation model and particle filtering algorithm to assess the remaining service life of the equipment.

7. The intelligent operation and maintenance and energy efficiency optimization control method for photovoltaic power plants according to claim 1, characterized in that: In step S4, the hierarchical multi-objective collaborative optimization decision-making framework is set with four types of constraints: equipment safe operation, power grid connection specifications, environmental safety restrictions, and operation and maintenance resource restrictions. The day-ahead global optimization layer uses the NSGA-II algorithm to solve the 24-hour optimal control plan every morning. The real-time rolling optimization layer uses the MPC model predictive control algorithm to correct the control plan every 10 minutes. The event-triggered fast optimization layer provides millisecond-level response to irradiance mutations and power grid disturbance events.

8. The intelligent operation and maintenance and energy efficiency optimization control method for photovoltaic power plants according to claim 7, characterized in that: In step S4, the constrained multi-agent reinforcement learning constructs four agents: tracking support, cleaning scheduling, inverter MPPT, and energy storage charging and discharging. It adopts the PPO near-end policy optimization algorithm combined with the SafetyShield safety layer algorithm. The reward function integrates power generation gain, operation and maintenance cost, equipment stress and life extension indicators. The model is first trained by digital twin environment simulation and then deployed to edge nodes. 9.The intelligent operation and maintenance and energy efficiency optimization control method of the photovoltaic power station of claim 1, characterized in that: In step S5, an improved genetic algorithm is used to prioritize and schedule resources for maintenance tasks. After the task is completed, the maintenance effect is quantitatively evaluated. If the target is not met, the maintenance task is regenerated, forming a closed loop of the entire process of early warning, generation, scheduling, execution, evaluation and optimization. In step S6, the incremental learning adopts the EWC elastic weight integration algorithm. Incremental learning is triggered when there are ≥50 new fault samples or the model prediction error is >3%. The federated learning realizes model sharing optimization under the privacy and security of multiple power plants. 10.The intelligent operation and maintenance and energy efficiency optimization control method of the photovoltaic power station of claim 1, characterized in that: The method is suitable for centralized ground photovoltaic power station, distributed industrial and commercial roof photovoltaic, household distributed photovoltaic, photovoltaic storage integrated power station, and photovoltaic cluster under complex terrain and extreme climate of plateau, mountain, desert and coastal area.

Images