[Paper Review] Detection of Clouds in Multiple Wind Velocity Fields using Ground-based Infrared Sky Images

Guillermo Terrén-Serrano, Manel Martínez‐Ramón|arXiv (Cornell University)|May 7, 2021

Solar Radiation and Photovoltaics4 citations

TL;DR

This paper proposes an unsupervised learning method to detect multiple cloud layers in ground-based infrared sky images by modeling temperature and motion vector distributions using mixture models and a sequential hidden Markov model (HMM). The approach uses weighted Lucas-Kanade optical flow on posterior probabilities from Gaussian Mixture Models (GMMs) to estimate wind velocity fields, with the HMM outperforming Bayesian metrics in identifying cloud layer count, achieving higher detection accuracy than traditional methods.

ABSTRACT

Horizontal atmospheric wind shear causes wind velocity fields to have different directions and speeds. In images of clouds acquired using ground-based sky imagers, clouds may be moving in different wind layers. To increase the performance of an intra-hour global solar irradiance forecasting algorithm, it is important to detect multiple layers of clouds. The information provided by a solar forecasting algorithm is necessary to optimize and schedule the solar generation resources and storage devices in a smart grid. This investigation studies the performance of unsupervised learning techniques when detecting the number of cloud layers in infrared sky images. The images are acquired using an innovative infrared sky imager mounted on a solar tracker. Different mixture models are used to infer the distribution of the cloud features. The optimal decision criterion to find the number of clusters in the mixture models is analyzed and compared between different Bayesian metrics and a sequential hidden Markov model. The motion vectors are computed using a weighted implementation of the Lucas-Kanade algorithm. The correlations between the cloud velocity vectors and temperatures are analyzed to find the method that leads to the most accurate results. We have found that the sequential hidden Markov model outperformed the detection accuracy of the Bayesian metrics.

Motivation & Objective

• To improve intra-hour global solar irradiance (GSI) forecasting by detecting multiple cloud layers in ground-based infrared sky images.
• To address the challenge of varying wind velocity fields across different atmospheric cloud layers, which affect solar irradiance fluctuations.
• To develop an unsupervised learning framework that avoids manual labeling and enables real-time cloud layer detection.
• To compare the performance of Bayesian model selection metrics and sequential HMMs in determining the optimal number of cloud layers.
• To enhance solar forecasting accuracy by modeling cloud dynamics independently per layer using temperature and motion vector features.

Proposed method

• Uses a ground-based infrared sky imager mounted on a solar tracker to capture sequential thermal images of the sky.
• Applies a weighted Lucas-Kanade optical flow algorithm to compute motion vectors, with weights derived from posterior probabilities of a Gaussian Mixture Model (GMM) of pixel temperatures.
• Employs mixture models (GMM, BGaMM) to infer the underlying distribution of cloud features, including temperature and velocity vectors.
• Uses Bayesian model selection metrics (e.g., minimum ICL) and a sequential hidden Markov model (HMM) to determine the optimal number of cloud layers.
• Applies the MAP criterion to select the most likely scenario (one or two cloud layers) based on the joint distribution of temperature and velocity vector features.
• Implements a temporal HMM to model the evolution of cloud layer states across image sequences, improving detection robustness.

Experimental results

Research questions

• RQ1 Can unsupervised learning techniques effectively detect multiple cloud layers in infrared sky images using only temperature and motion vector features?
• RQ2 How does the performance of a sequential hidden Markov model compare to Bayesian model selection metrics in identifying the number of cloud layers?
• RQ3 To what extent do posterior probabilities from mixture models improve motion vector estimation and cloud layer detection accuracy?
• RQ4 Can the proposed method detect cloud layers moving in different wind velocity fields without labeled training data?
• RQ5 How does the integration of temperature and velocity vector distributions enhance cloud layer detection over single-feature approaches?

Key findings

• The sequential hidden Markov model outperformed all Bayesian metrics in cloud layer detection accuracy, achieving superior performance in identifying the correct number of cloud layers.
• The highest detection accuracy among Bayesian metrics was 81.11%, achieved using a GMM on temperature features with the minimum ICL criterion.
• The BGaMM model performed poorly due to overfitting, even with regularization, and was deemed impractical for cloud layer detection.
• Posterior probabilities from the GMM enabled accurate, independent estimation of motion vectors for each cloud layer, improving feature extraction for forecasting.
• The proposed method successfully modeled cloud dynamics in multiple wind velocity fields by leveraging both temperature and velocity vector distributions.
• The algorithm's unsupervised nature eliminates the need for labeled data, enabling automatic and scalable deployment in real-time solar forecasting systems.