CN-121999391-A - Method for treating light efficiency loss of photovoltaic power station based on reinforcement learning and multispectral analysis

CN121999391ACN 121999391 ACN121999391 ACN 121999391ACN-121999391-A

Abstract

The invention belongs to the technical field of photovoltaic power generation, and discloses a method for treating light efficiency loss of a photovoltaic power station based on reinforcement learning and multispectral analysis. The method comprises the steps of synchronously obtaining electrical parameters of a photovoltaic module and surface multispectral image data through a multisource data acquisition module, carrying out edge calculation filtering denoising, YOLO model defect detection and multispectral feature extraction to achieve accurate identification of light effect loss causes (hot spots, dirt, shielding and component aging), constructing a multidimensional light effect loss assessment model to quantify loss degree, combining reinforcement learning to dynamically generate an optimal treatment strategy, executing targeted measures through a hierarchical response mechanism, and utilizing digital twin and feedback data to iteratively optimize model parameters. The invention realizes the closed-loop management of 'accurate identification-dynamic management-continuous optimization' of light efficiency loss, and obviously improves the power generation efficiency of the photovoltaic power station and the life cycle economy of the equipment.

Inventors

XIE ZHIQI
ZUO TIANCAI
FENG HUAN
Zhou jinjiang
HU JINGWEI
SU QIAN
LUO YU
LI LIN
TANG XIAOBO

Assignees

贵州乌江水电开发有限责任公司

Dates

Publication Date: 20260508
Application Date: 20251215

Claims (7)

1. The method for treating the light efficiency loss of the photovoltaic power station based on reinforcement learning and multispectral analysis is characterized by comprising the following steps of: Step 1, synchronously acquiring multi-source data, namely acquiring electric parameters of a photovoltaic module in real time through a sensor network, carrying out aerial photography on a photovoltaic array by using an unmanned plane carrying a multi-spectrum camera to acquire surface multi-spectrum image data, endowing uniform time stamps on the two types of data to realize time-space synchronization, and transmitting the two types of data to a central processing unit; Step 2, data preprocessing and cause identification, namely filtering, denoising and abnormal quantification are carried out on the electrical parameters based on edge calculation to obtain binary state parameters, after denoising and enhancing treatment are carried out on the multispectral image, detecting surface defects of the component through a YOLO model, and identifying the cause type of light efficiency loss and extracting characteristic parameters by combining the multispectral band reflection characteristics and temperature distribution; Step 3, multi-dimensional light efficiency loss assessment, namely constructing a light efficiency loss assessment model, fusing the electrical parameter abnormality degree, the loss cause characteristic parameter and the environmental factor, calculating a real-time light efficiency loss index, and quantifying the influence of loss on the power generation efficiency; Step 4, reinforcement learning dynamic optimization, namely constructing a reinforcement learning model by taking 'light efficiency loss index minimization and treatment cost minimization' as double targets, taking loss index, cause type, environmental condition and equipment state as state input, taking a treatment strategy as action output, and generating an optimal treatment strategy through historical data training; Step 5, grading treatment and linkage response, namely grading treatment according to the light efficiency loss index, executing corresponding strategies including component cleaning, shielding removal, angle optimization, fault maintenance or replacement, and linkage Mesh ad hoc network pushing tasks to an operation and maintenance terminal; And 6, model self-evolution optimization, namely collecting light effect recovery data and equipment operation data after treatment, feeding back to the reinforcement learning model by combining a digital twin simulation result, adjusting state transition probability and rewarding coefficient, and updating a treatment strategy library.
2. The method of claim 1, wherein the electrical parameters in step 1 comprise operating current, open circuit voltage, short circuit current, maximum output power, and module back plane temperature of the photovoltaic module, and the multispectral image data comprises 400-760nm visible light band, 760-1100nm near infrared band, and 8-14 μm infrared thermal imaging band.
3. The method according to claim 1, wherein the abnormal quantification in the step 2 is specifically that comparing the electrical parameter with a preset normal threshold, quantifying a value u=0 when the parameter is within the threshold, and u=1 when the parameter exceeds the threshold; the loss-causing characteristic parameters comprise hot spot area, hot spot-environment temperature difference, dirt coverage, shelter projection area and aging crack length.
4. The method of claim 1, wherein the light efficiency loss index in step 3 has a value ranging from 0 to 1, a larger value representing a more serious light efficiency loss; The evaluation model fuses five core factors, namely a power attenuation rate, a hot spot relative temperature difference, a dirt coverage, a shielding area ratio and a crack length ratio, wherein the weights of the five core factors are calibrated through a hierarchical analysis method to ensure that the total weight is 1, and the total weight is more than 60% when the power attenuation rate and the hot spot relative temperature difference weight are the highest.
5. The method of claim 1, wherein the reinforcement learning model in step 4 employs a DQN (Deep Q-Network) architecture, and the state space includes a light efficiency loss index, a loss cause type, an environmental parameter, and a device state; The rewarding function comprehensively considers the difference value of loss indexes before and after treatment, the treatment cost per unit area and the light efficiency recovery time, and balances the influence of three indexes on rewarding results through coefficients; the environment parameters comprise illumination intensity, ambient temperature and wind speed, and the equipment state comprises service life of the component and historical failure times.
6. The method of claim 1, wherein the abatement level in step 5 comprises: Level I (light loss, light efficiency loss index 0-0.2) performing conventional high pressure water washing (pressure 0.8-1.2 MPa), fine tuning component tilt angle (+ -1-2 DEG) based on real-time illumination data; class II (moderate loss, light efficiency loss index 0.2-0.5) adopting chemical cleaning (neutral cleaner concentration 2-5%), removing shielding material, optimizing group string current matching; Grade III (severe loss, light efficiency loss index 0.5-0.8) is to stop for maintenance, replace hot spot components (forced replacement when temperature difference exceeds 20 ℃) or repair aged cracks; and IV (extremely severe loss, light efficiency loss index is more than or equal to 0.8), emergency stop, replacement of the whole group of fault components and troubleshooting of the group string circuit.
7. The method according to claim 1, wherein the self-evolution optimization in step 6 is specifically implemented by constructing a multi-physical field model of the photovoltaic module by digital twinning, and simulating long-term effects of different governance strategies; The method comprises the steps of combining the actual light effect recovery rate after treatment with equipment health data, adjusting experience playback Chi Quan weight and Q value update rate of a reinforcement learning model, and updating an optimal parameter threshold value in a treatment strategy library in each quarter; The multi-physical field model is a composite model of a light field, a thermal field and an electric field.

Description

Method for treating light efficiency loss of photovoltaic power station based on reinforcement learning and multispectral analysis Technical Field The invention belongs to the technical field of photovoltaic power generation, and particularly relates to a method for controlling the light efficiency loss of a photovoltaic power station based on reinforcement learning and multispectral analysis, which is suitable for optimizing the light efficiency and managing the operation and maintenance of various photovoltaic arrays of a large-scale ground photovoltaic power station, a roof distributed photovoltaic power station and the like. Background As global energy structures are transformed to clean energy, photovoltaic power generation, one of the core forms of renewable energy, continues to grow in installed capacity. The power generation efficiency of a photovoltaic power station directly determines the economic benefit and the energy supply capacity of the photovoltaic power station, and the light efficiency loss is a key problem for limiting the full life cycle benefit of the photovoltaic power station. According to industry data statistics, annual energy generation capacity attenuation of the photovoltaic module caused by light effect loss in the operation process can reach 2-5%, and when serious, the module is burnt out due to hot spot effect, so that direct economic loss is caused. The light efficiency loss causes have diversity and complexity, mainly comprise hot spot effect, surface dirt coverage, external shielding, component aging and the like, so a light efficiency loss treatment method capable of realizing accurate identification of causes, dynamic generation strategy and continuous optimization iteration is needed to solve the defects of the prior art and improve the operation and maintenance efficiency and economic benefit of a photovoltaic power station. Disclosure of Invention The invention aims to provide a method for treating light efficiency loss of a photovoltaic power station based on reinforcement learning and multispectral analysis, which realizes accurate diagnosis and efficient treatment of light efficiency loss through multisource data fusion, intelligent algorithm optimization and hierarchical response and solves the problems of fuzzy cause identification, strategy staticization and low operation and maintenance efficiency in the prior art. In order to achieve the above object, the embodiment of the present invention provides the following technical solutions: A method for treating light efficiency loss of a photovoltaic power station based on reinforcement learning and multispectral analysis comprises the following steps: Step 1, synchronously acquiring multi-source data, namely acquiring electric parameters of a photovoltaic module in real time through a sensor network, carrying out aerial photography on a photovoltaic array by using an unmanned plane carrying a multi-spectrum camera to acquire surface multi-spectrum image data, endowing uniform time stamps on the two types of data to realize time-space synchronization, and transmitting the two types of data to a central processing unit; Step 2, data preprocessing and cause identification, namely filtering, denoising and abnormal quantification are carried out on the electrical parameters based on edge calculation to obtain binary state parameters, after denoising and enhancing treatment are carried out on the multispectral image, detecting surface defects of the component through a YOLO model, and identifying the cause type of light efficiency loss and extracting characteristic parameters by combining the multispectral band reflection characteristics and temperature distribution; Step 3, multi-dimensional light efficiency loss assessment, namely constructing a light efficiency loss assessment model, fusing the electrical parameter abnormality degree, the loss cause characteristic parameter and the environmental factor, calculating a real-time light efficiency loss index, and quantifying the influence of loss on the power generation efficiency; Step 4, reinforcement learning dynamic optimization, namely constructing a reinforcement learning model by taking 'light efficiency loss index minimization and treatment cost minimization' as double targets, taking loss index, cause type, environmental condition and equipment state as state input, taking a treatment strategy as action output, and generating an optimal treatment strategy through historical data training; Step 5, grading treatment and linkage response, namely grading treatment according to the light efficiency loss index, executing corresponding strategies including component cleaning, shielding removal, angle optimization, fault maintenance or replacement, and linkage Mesh ad hoc network pushing tasks to an operation and maintenance terminal; And 6, model self-evolution optimization, namely collecting light effect recovery data and equipment operation data after treatment, feedi