Risk-Based Thresholding for Reliable Anomaly Detection in Concentrated Solar Power Plants

Yorick Estievenart, Sukanya Patra, Souhaib Ben Taieb (✉) · 2025

Read Risk-Based Thresholding for Reliable Anomaly Detection in Concentrated Solar Power Plants by Yorick Estievenart, Sukanya Patra, Souhaib Ben Taieb (✉) (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2503-19146.txt.

Introduction

Yorick Estievenart, Sukanya Patra, and Souhaib Ben Taieb build a safer alarm system for solar power plants that focus the sun's heat. Receivers at these plants can freeze, warp, or rust, and cameras watch them in infrared. Picking the line between normal and odd images is the weak point of past alarms. Their 2025 frame sets that line with risk control, backed by math promises on error rates.

When a call looks risky, the system can pass the image to a human expert instead of guessing. A forecasting method scores how likely each new image is, given those before it. Trials ran for months at two real plants. Since the plant data are secret, the team shares a made-up data set that mimics many plants.

Catalog record

CREATORYorick Estievenart, Sukanya Patra, Souhaib Ben Taieb (✉)
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE47 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2503-19146

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Risk-Based Thresholding for Reliable Anomaly Detection in Concentrated Solar Power Plants Yorick Estievenart, Sukanya Patra, Souhaib Ben Taieb (✉) Abstract Efficient and reliable operation of Concentrated Solar Power (CSP) plants is essential for meeting the growing demand for sustainable energy. However, high-temperature solar receivers face severe operational risks, such as freezing, deformation, and corrosion, resulting in costly downtime and maintenance. To monitor CSP plants, cameras mounted on solar receivers record infrared images at irregular intervals ranging from one to five minutes throughout the day. Anomalous images can be detected by thresholding an anomaly score, where the threshold is chosen to optimize metrics such as the F1-score on a validation set. This work proposes a framework, using risk control, for generating more reliable decision thresholds with finite-sample coverage guarantees on any chosen risk function. Our framework also incorporates an abstention mechanism, allowing high-risk predictions to be deferred to domain experts. Second, we propose a density forecasting method to estimate the likelihood of an observed image given a sequence of previously observed images, using this likelihood as its anomaly score. Third, we analyze the deployment results of our framework across multiple training scenarios over several months for two CSP plants. This analysis provides valuable insights to our industry partner for optimizing maintenance operations. Finally, given the confidential nature of our dataset, we provide an extended simulated dataset 1 1 1 https://tinyurl.com/macmnjyt , leveraging recent advancements in generative modeling to create diverse thermal images that simulate multiple CSP plants. Our code is publicly available 2 2 2 https://github.com/yoest/reliable-ad-csp . 1 Introduction The global transition toward greener and more sustainable renewable energy sources is hindered by two critical challenges: (i) on-demand generation and (ii) dispatchability. Concentrated Solar Power (CSP) plants offer a promising solution, leveraging thermal energy storage to provide electricity even when sunlight is unavailable. Among the various CSP configurations, central tower-based plants are the most prevalent, using an array of mirrors to concentrate sunlight onto a receiver, where a heat transfer medium absorbs and stores the energy. However, the extreme operating temperatures make these systems highly susceptible to failures such as metal fatigue and tube blockages, directly impacting their efficiency, reliability, and operational lifespan. To mitigate these risks, thermal imaging from infrared cameras is used to monitor CSP plants. Nonetheless, the sheer volume and complexity of thermal image data render manual monitoring impractical, necessitating the development of an automated, data-driven Predictive Maintenance (PdM) pipeline. This problem naturally aligns with anomaly detection (AD), where the goal is to identify abnormal behaviours. Despite significant progress in both deep and shallow AD research [ 20 , 32 ] , existing image- and video-based approaches fall short in addressing the problem of detecting anomalous behaviours of operational CSP plants due to three key challenges. First, the lack of interpretability of the anomaly scores hinders decision-making in high-stakes applications without an appropriate thresholding strategy [ 28 ] . Traditional approaches rely on performance metrics such as F1-score or GMean to determine thresholds depending on the available labelled samples. These methods do not guarantee that the results will remain consistent in a deployment setting. Moreover, they assume that all CSP plants define risk similarly and follow the same operational strategies. In reality, this often differs (e.g., deploying a maintenance team may be preferable to replacing a tower component). Second, deep learning-based AD models are often perceived as unreliable [ 28 ] due to the uncertainty in predictions stemming from their inability to properly estimate the decision boundary, particularly when training data is limited. Thus, practitioners are hesitant to use the predictions even when the associated uncertainty is minimal, severely limiting their adoption in real-world applications. Also, unlike classical image- and video-based AD data, CSP plant monitoring involves thermal images without semantic content, lacks a fixed frame rate, and exhibits significant non-stationarity and temporal dependencies due to pronounced daily seasonal patterns. As a result, conventional image- and video-based anomaly detection methods are inappropriate. A recent forecasting-based AD method, ForecastAD [ 1 ] , attempts to address these challenges by measuring per-pixel errors between predicted and observed thermal images. However, reconstruction-based AD methods suffer a critical flaw: models trained on normal data can inadvertently reconstruct and misclassify anomalous images as normal [ 10 , 22 ] , leading to unreliable detection. To overcome these limitations, we propose a principled, robust AD framework tailored for CSP plant monitoring. First, we introduce a risk-controlling thresholding strategy for anomaly scores that satisfies finite-sample performance guarantees on any chosen risk function (e.g., false positive rate or F1-score)—a critical requirement for reliable predictive maintenance (PdM) in industrial settings. To enhance trust and adoption, we integrate a machine-learning-with-abstention framework [ 27 ] with adaptive thresholds that account for the overlap between normal and anomalous score distributions. This approach defers high-risk predictions to domain experts, ensuring human intervention when uncertainty is high. Furthermore, we propose an AD method based on density forecasting, DensityAD , which leverages conditional normalizing flows to model the likelihood of an observed sample being normal, given past thermal images and timestamps. This approach mitigates the limitations of reconstruction-based methods and enables likelihood-based thresholding for more effective anomaly detection. Our key contributions are: • We propose a framework for computing reliable anomaly detection thresholds with finite-sample performance guarantees for any chosen risk function. The framework includes an abstention mechanism that defers decisions to domain experts under high uncertainty. • We develop an unsupervised AD method that computes anomaly scores using density forecasting by estimating the conditional likelihood of an observed infrared image given a sequence of previously observed images. • We conduct an extensive deployment analysis of our framework across multiple real-world scenarios over several months, using data from two CSP plants. This analysis provides valuable insights to our industry partner for maintenance operations. • We release a simulated dataset by leveraging recent advancements in generative modelling to create diverse infrared images that emulate real-world data from CSP plants. Our work not only advances the state of anomaly detection in renewable energy systems but also serves as an important milestone for future research in robust, data-driven PdM strategies for critical infrastructure monitoring. 2 Anomaly detection in thermal images from CSP plants In the following, we describe our AD use case and the associated dataset. Use-case . Concentrated Solar Power (CSP) plants are designed to harness solar energy for large-scale electricity generation while addressing two major challenges commonly associated with renewable energy sources – on-demand generation and dispatchability. Among the four primary CSP technologies currently in use, namely, Solar Tower, Parabolic Trough, Linear Fresnel, and Dish-Stirling systems, this study specifically focuses on the operational aspects of Solar Tower-based CSP plants. These plants comprise two critical components: the Thermal Solar Receiver and the Steam Generator. Positioned atop a central tower, the Solar Receiver functions as a solar furnace, absorbing concentrated sunlight reflected by an array of heliostats—movable mirrors strategically arranged on the ground around the tower. A high-capacity heat transfer medium, such as molten salts, circulates through vertical heat exchanger tubes configured as panels within the receiver, absorbing the thermal energy from the concentrated sunlight. This heated transfer medium is subsequently stored in a Thermal Energy Storage (TES) system. It is later used to generate superheated

Read the complete text