CN-121982275-A - Solar radiation evaluation method and device based on full-sky imager and visual recognition

CN121982275ACN 121982275 ACN121982275 ACN 121982275ACN-121982275-A

Abstract

The invention relates to a solar radiation evaluation method and device based on an all-sky imager and visual recognition, wherein the method comprises the steps of obtaining a sky image of a place to be tested at the current moment and non-image state data corresponding to the place to be tested and the current moment, processing the sky image and the non-image state data by utilizing a pre-trained deep learning model, and the deep learning model is configured to extract cloud field visual features from the sky image through a visual feature extraction network, extract physical priori features from the non-image state data through a structured feature coding network, and fuse the cloud field visual features with the physical priori features to obtain fusion feature vectors.

Inventors

REN SHIQI
MOU XIAOXUAN
JIA SHENGJIE

Assignees

北京中科技达科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260126

Claims (10)

1. A solar radiation assessment method based on an all-sky imager and visual recognition, comprising: acquiring a sky image of a current moment of a place to be detected and non-image state data corresponding to the place to be detected and the current moment; The sky image and the non-image state data are processed by using a pre-trained deep learning model, wherein the deep learning model is configured to extract cloud field visual features from the sky image through a visual feature extraction network, extract physical prior features from the non-image state data through a structured feature coding network, fuse the cloud field visual features with the physical prior features to obtain a fused feature vector, and output a solar radiation parameter estimated value at the current moment through a regression prediction layer based on the fused feature vector.
2. The all-sky imager and visual recognition based solar radiation assessment method of claim 1, wherein the non-image state data comprises: The solar geometric feature is calculated based on longitude and latitude coordinates of the to-be-detected place and the current moment and is used for representing a theoretical position of the sun in the sky; Meteorological environmental characteristics, including one or more of barometric pressure, ambient temperature, or air humidity.
3. The all-sky imager and visual identification based solar radiation assessment method of claim 2, wherein the non-image state data further comprises: And the radiation theory upper limit value is calculated based on a solar radiation physical model of the to-be-detected place under a clear sky condition and is used for restricting the output of the deep learning model.
4. The all-sky imager and vision recognition based solar radiation assessment method of claim 1, further comprising a solar position correction step prior to entering data into the deep learning model: calculating theoretical projection coordinates of the sun based on the non-image state data; the sky image is subjected to visual analysis, and the actual imaging coordinates of the sun in the image are detected; and inputting the corrected solar position coordinates into the structural feature coding network as a part of the non-image state data, or guiding the visual feature extraction network to extract a region of interest of an image as auxiliary information.
5. The method for evaluating solar radiation based on a whole sky imager and visual recognition according to claim 4, wherein the fusing of theoretical projection coordinates and actual imaging coordinates comprises: acquiring the detection confidence coefficient of the actual imaging coordinate; Dynamically adjusting the weights of the theoretical projection coordinates and the actual imaging coordinates according to the detection confidence coefficient; and when the detection confidence coefficient is lower than a preset threshold value, increasing the weight of the theoretical projection coordinate so as to maintain position continuity.
6. The solar radiation assessment method based on full sky imager and visual recognition according to claim 1, wherein the regression prediction layer of the deep learning model comprises a plurality of parallel output branches for outputting total solar radiation, direct solar radiation and scattered solar radiation, respectively, the plurality of parallel output branches sharing parameters of the visual feature extraction network and the structured feature encoding network.
7. A solar radiation assessment device based on an all-sky imager and visual recognition, comprising: The acquisition module is used for acquiring a sky image of a current moment of a place to be detected and non-image state data corresponding to the place to be detected and the current moment; The prediction module is used for processing the sky image and the non-image state data by utilizing a pre-trained deep learning model, wherein the deep learning model is configured to extract cloud field visual features from the sky image through a visual feature extraction network, extract physical prior features from the non-image state data through a structured feature coding network, fuse the cloud field visual features with the physical prior features to obtain a fused feature vector, and output a solar radiation parameter estimated value at the current moment through a regression prediction layer based on the fused feature vector.
8. The all-sky imager and visual identification based solar radiation assessment device of claim 7, further comprising: The correction module is used for calculating theoretical projection coordinates of the sun based on the non-image state data, carrying out visual analysis on the sky image, detecting actual imaging coordinates of the sun in the image, carrying out fusion processing on the theoretical projection coordinates and the actual imaging coordinates to obtain corrected sun position coordinates, and inputting the corrected sun position coordinates serving as a part of the non-image state data into the structural feature coding network or serving as auxiliary information to guide the visual feature extraction network to extract a region of interest of the image.
9. An electronic device comprising one or more processors and storage means for storing one or more programs that, when executed by the one or more processors, cause the one or more processors to implement the all-sky imager and visual identification-based solar radiation assessment method according to any one of claims 1 to 6.
10. A computer readable medium having stored thereon a computer program, wherein the computer program when executed by a processor implements the all-sky imager and visual identification based solar radiation assessment method according to any one of claims 1 to 6.

Description

Solar radiation evaluation method and device based on full-sky imager and visual recognition Technical Field The invention belongs to the technical field of atmospheric environment monitoring and visual identification, and particularly relates to a solar radiation evaluation method and device based on an all-sky imager and visual identification. Background In recent years, photovoltaic power generation, meteorological fine observation and atmospheric environment research have put higher demands on the space-time resolution and continuity of solar radiation parameters (at least including total solar radiation, scattered radiation, direct radiation, etc.). The radiation elements are not only important inputs of photovoltaic power prediction and component performance evaluation, but also key basic data of cloud-radiation interaction research, short-time proximity prediction, energy balance calculation and the like. The existing radiation parameter acquisition mode mainly comprises the following steps: 1) The surface radiation instrument directly measures, such as a total radiometer, a scattered radiometer, a direct radiometer and the like. The scheme has high measurement precision, but has the common problems of limited cost and deployment, that a plurality of sensors and matched tracking devices are usually required for simultaneous observation of multiple elements, the initial assembly cost is high, and the large-scale low-cost layout is difficult to realize. The operation and calibration load is heavy, the instrument needs to be calibrated, cleaned and aligned regularly (especially the direct radiation/scattered radiation relates to shielding or tracking structures), and the performance is easy to deteriorate in environments such as sand dust, salt fog, high humidity and the like, so that data drift or missing measurement is caused. Site scalability is poor-in remote areas, islands, mountains or large-scale photovoltaic power plant scenarios, the risk of data unavailability is further amplified by insufficient operation and maintenance resources. 2) Radiation inversion or interpolation based on satellite/analysis data has a wide coverage range, but usually has the problem that the time resolution and the space resolution are not matched with the rapid change of the local cloud field. For typical scenes such as small-scale cloud clusters, thin cloud evolution, rapid boundary layer change and the like, a satellite or an analysis radiation product can not timely reflect the real irradiance of the ground, so that the power evaluation and prediction accuracy is affected. 3) The conventional algorithm estimation based on the ground-based all-sky imager can provide high-frequency sky images at the site side, but the conventional method relies on fixed threshold values, empirical parameters or simple color ratio/feature engineering. The method is easy to generate obvious errors under the conditions that local information of an image is distorted due to strong glare, halation and overexposure near the sun, radiation related characteristics are difficult to stably extract by a traditional threshold model, influence of clouds with lower optical thickness such as thin clouds and coiled clouds on direct radiation is obvious, but estimation is inaccurate due to unstable color/brightness difference, sky background brightness and color distribution change under different latitudes, different seasons and different aerosol conditions are large, experience parameter generalization capability is insufficient, and most traditional image methods are difficult to simultaneously output multiple factors such as total radiation, scattered radiation and direct radiation or require additional instrument assistance (such as shielding discs and tracking devices) to distinguish between the direct radiation and the scattered component, and system complexity and maintenance cost are increased. Therefore, a technical solution with a stronger generalization capability under different weather bands, different weather and different seasons is needed, so that the all-sky imager can perform stable, real-time and low-cost automatic estimation on solar radiation elements only by relying on visual information (combining available information such as longitude and latitude of a station and time), thereby reducing the dependence on configuration and high-intensity operation and maintenance of the traditional radiation instrument. Disclosure of Invention In order to solve the problems of the background technology, the invention provides a solar radiation evaluation method based on an all-sky imager and visual recognition, which comprises the steps of obtaining a sky image of a place to be tested at the current moment and non-image state data corresponding to the place to be tested and the current moment, processing the sky image and the non-image state data by utilizing a pre-trained deep learning model, wherein the deep learning model is configured t