Drones, combinatorial Optimization and deep Learning for inspecting solar thermal power Plants

• Autores: Miguel Ángel Pérez Cutiño
• Directores de la Tesis: José Miguel Díaz-Báñez (dir. tes.) , J. Valverde (dir. tes.)
• Lectura: En la Universidad de Sevilla ( España ) en 2025
• Idioma: inglés
• Número de páginas: 161
• Enlaces
  • Tesis en acceso abierto en: Idus
• Resumen

  • Concentrated Solar Power plants are a promising system for sustainable energy, offering large-scale clean electricity generation with thermal energy storage capabilities. Parabolic Trough technology, the most mature and deployed CSP configuration, requires efficient operation and fault detection systems to maximize energy output and reduce maintenance costs. Inspired by these challenges, this thesis addresses a wide range of computational problems, integrating techniques from discrete algorithms, optimization, computer vision, and machine learning.

    We first address fundamental path planning challenges for drone-assisted inspections, leveraging the grid-like topology inherent to PT solar fields. We develop optimal algorithms to cover rectangular grid graphs with a set of limited-length tours that depart from and return to a base station. We aim to minimize either (1) the total number of tours required to cover the grid, and (2) their total length. We provide a closed-form expression to obtain the minimum number of tours required to cover the grid, and design an efficient algorithm that obtains a set of tours that is an optimal solution to both problems. In addition, we analyze the Euclidean k-Matching problem, a generalization of the Euclidean 3-Matching problem, which models the covering of a point set with tours constrained by a maximum number of waypoints. We demonstrate that for any fixed k bigger than 3, the problem is NP-hard, which provides a theoretical justification for existing heuristic methods and solves a long-standing open problem. We then propose novel solar tracking optimization models that account for mechanical degradation in PT solar plants. We introduce the Min-Tracking-Motion and Maximal Energy Collection problems and solve them under realistic constraints using Greedy and Dynamic Programming techniques. Through detailed simulations, we demonstrate the efficiency and usability of the proposed system by estimating energy yields exceeding 95% of the current production while reducing the tracking motion by approximately 10%.

    On the perception side, we contribute several new resources and tools for automated inspection. RTSet is introduced as the first dataset for broken-glass detection in PT receiver tubes. Through feature selection, class balancing, and dual optimization with advanced training techniques, we significantly boost classification recall. We further enhance inspection through drone-based aerial analysis using RGB and thermal imagery. Our luminescence-based filter assess HCE visibility and reveal biases in our custom dataset, while our ensemble of CNNs, combined using hypernetwork as meta-learners, consistently outperforms traditional voting systems in challenging conditions. We also introduce a novel pipeline for inspecting Ball Joint Assemblies, integrating synthetic data augmentation and geometric analysis. Our framework automatically detects leaks and misalignments using segmentation neural networks and angular deviation metrics, achieving reliable and robust results. Finally, we build and release AerialCSP, a synthetic dataset for CSP inspection, fostering research on automatic visual inspection of PT solar fields. Moreover, AerialCSP demonstrates its value in pretraining artifical neural networks for the task of object detection, reducing the needs for labeled data and being especially beneficial for rare defects. Our results show substantial performance gains in real-world tasks and validate the effectiveness of combining 3D modeling, background inpainting, and domain adaptation strategies.

    Through this interdisciplinary approach, this thesis advances the state-of-the-art in intelligent operation and inspection of CSP plants, with practical contributions for scalable deployment and sustainability.