What is the Optimal Path for a Drone Exploring a PV Plant?

What is the Optimal Path for a Drone Exploring a PV Plant? M. Moriconi1*, G. Rocca1*, L. Grassi1, G. Mottola1, 2, M. Paolucci1, C. T. Recchiuto1, and A. Sgorbissa1 Abstract— Inspecting photovoltaic (PV) plants is essential to ensure optimal performance. Drones can be employed to acquire both optical and thermal data for anomaly detection. However, while visual servoing can accurately guide a drone along individual PV panel rows, the inter-panel transitions between subsequent rows rely on the Global Navigation Satellite System (GNSS) and are therefore subject to larger errors in positioning accuracy. To address this problem, this paper presents a path-planning solution that minimizes the portion of the route dependent on GNSS navigation by formulating three variants of the Traveling Salesman Problem (TSP) and analyzing their impact on path length and inter-panel transitions in simulated models of actual PV plants. I. INTRODUCTION Solar photovoltaic (PV) plants play a crucial role in renewable energy generation but require periodic inspections to detect panel malfunctions caused by discoloration, cracks, and hot spots [1]. Automated drone-based inspection has emerged as a promising solution for collecting optical and thermal data, streamlining fault detection and maintenance. Inspection systems can efficiently employ visual servoing [2] to control UAV’s movement along each PV panel row. However, transitions between rows typically rely on Global Navigation Satellite System (GNSS) signals, which are susceptible to various sources of positioning error, with errors that can reach several meters [3]. More critically for our application, when the drone moves from the end of a PV panel row to the start of the next one, GNSS errors vary over time with low-frequency components. This variation becomes more significant when the traveled distance is greater, and thus, the flight time between panels is longer. Figure 1 shows the two inspection phases. The segments where visual servoing is used are parallel to the PV panel rows. They do not overlap with the panels because the drone is positioned slightly offset rather than directly above them, to achieve an optimal viewing angle with the onboard camera, given the panels’ inclination. The GNSS transitions between PV panel rows are also shown. An effective inspection strategy should not only minimize the total flight distance but also reduce the maximum transition length between PV panel rows. Longer GNSS-dependent transitions increase the likelihood of positioning errors varying dynamically, which can cause the UAV to reach a location with no visible panel rows, making it impossible to resume inspection. 1Department of Informatics, Bioengineering, Robotics and Systems Engineering, University of Genoa, Via All’Opera Pia 13, 16145 Genoa, Italy 2Department of Engineering, University of Messina, Contrada di Dio, 98166 Sant’Agata, Messina, Italy *Equal contribution. Corresponding author: [email protected] This research was supported by the European Union under the SOLARIS project (GA no. 101146377). Fig. 1: Coverage of the Riso Hybrid PV plant in Denmark (red dot: UAV; green segments: inspection path). To address these challenges, we propose a novel optimization strategy based on a modified Traveling Salesman Problem (TSP) formulation, which explicitly penalizes long inter-panel transitions while ensuring complete coverage. II. METHODOLOGY We model the PV inspection task as a TSP, where the waypoints include the takeoff/landing point as well as the two endpoints of each PV panel row, and edges connecting waypoints within the same row are mandatorily included in the solution [4]. A solution to our inspection path planning problem must: (i) connect all waypoints, including the takeoff/landing point; (ii) traverse all intra-panel edges, ensuring full coverage of each panel. In searching for a solution, unlike the standard TSP, we explicitly distinguish between three types of edges: •Start/end transitions: The initial and final segments connecting the takeoff/landing point to the first and last PV panel rows. These legs are typically executed under manual or semi-autonomous control and are not critical in terms of GNSS accuracy. •Intra-panel transitions: Segments that traverse PV panel rows between their two endpoints, ensuring visual coverage. These are executed using visual servoing for precise navigation, without relying on GNSS. •Inter-panel transitions: Transitions between different PV panel rows that rely on GNSS-based navigation. These segments are prone to positional inaccuracies, particularly over longer distances. Unfortunately, standard path-planning approaches primarily aim to minimize the total flight distance, but they do not focus on minimizing the length of inter-panel transitions. To address this issue, we compare: 1) TSP: A conventional formulation in which the UAV must visit all panel waypoints while minimizing the 2025 I-RIM Conference October 17-19, Rome, Italy ISBN: 9788894580570 10.5281/zenodo.17629908 264 TABLE I: Comparison of total flight distance (TFD), total start/end distance (TSED), max inter-panel distance (MIPD) in Riso (R) and Montalto di Castro (M) for different strategies. All distances are in meters (lower values in bold). PV Plant Metric TSP C-TSP C-H-TSP R TFD 1175.90 ±43.45 1205.48 ±50.97 1223.91 ±52.11 TSED 176.82 ±68.77 151.19 ±50.79 280.44 ±52.11 MIPD 71.05 ±2.59 31.20 ±0.19 15.49 ±0.00 M(-20%) TFD 4086.32 ±27.11 4089.14 ±27.19 4250.22 ±39.53 TSED 44.09 ±29.00 43.15 ±28.27 290.45 ±39.53 MIPD 51.81 ±2.24 49.84 ±0.47 24.90 ±0.00 M(-10%) TFD 4486.91 ±17.84 4493.41 ±16.62 4685.65 ±102.13 TSED 44.34 ±34.76 40.53 ±29.04 296.86 ±102.13 MIPD 47.62 ±18.02 43.71 ±12.19 9.60 ±0.00 M(Full) TFD 5011.28 ±13.93 5025.40 ±12.85 5255.10 ±95.38 TSED 45.15 ±31.27 41.81 ±31.12 326.14 ±95.38 MIPD 56.02 ±1.84 49.55 ±0.23 9.06 ±0.00 total Euclidean path length. 2) C-TSP (Cautious TSP): A modified version of the TSP that squares the inter-panel distances, making longer GNSS-based transitions significantly more costly. 3) C-H-TSP (Cautious Homeless TSP): A variant of CTSP that assigns zero cost to the start and end transitions. This enables the solver to focus solely on optimizing inter-panel transitions, which are the most critical segments in terms of positional accuracy. III. EXPERIMENTS We performed simulation experiments to validate our pathplanning strategies, based on two real-world PV plants. The first is the Riso Hybrid Power Plant (R) located at DTU in Denmark (Figure 1), while the second is a segment of the large-scale PV plant in Montalto di Castro (M), Italy. We considered four configurations. In R, all 17 PV panel rows are included. In M, which includes 88 PV panel rows, we considered: 70 panel rows, with 20% randomly removed to simulate unexpected maintenance or temporary inaccessibility; 79 panel rows, with 10% removed; all 88 panel rows. Each configuration was evaluated using the TSP, C-TSP, and C-H-TSP strategies. For each strategy, we simulated 12 different takeoff/landing locations. For each path computed over the PV plant in a given configuration, we recorded: (i) total flight distance: sum of all segments (start, intra-panel, inter-panel, and end); (ii) total start/end distance: sum of the first and last segments from/to takeoff/landing; (iii) maximum inter-panel distance: longest GNSS-reliant transition. The results in Table I highlight that C-H-TSP consistently yields the shortest maximum inter-panel distance, often reducing it by more than 70% compared to the TSP baseline. This confirms its effectiveness in minimizing the longest GNSS-dependent transitions, which are most error-prone. Please note that the standard deviation of the maximum interpanel distance in C-H-TSP is always zero, as the solution is invariant to the takeoff/landing point, since the start/end transitions are excluded from optimization and therefore have no impact on the remaining path. The start and end 0 50 100 150 200 X (m) 100 125 150 Y (m) (a) TSP: a longer inter-panel jump is visible. 0 50 100 150 200 X (m) 100 125 150 Y (m) (b) C-H-TSP: longer start/end transitions. Fig. 2: Riso Hybrid configuration. Blue: intra-panel traversal. Red: inter-panel transitions. Black: start/end transitions. legs are longer than those in the other strategies, but this is not an issue, as the drone can be manually controlled during takeoff and landing. C-TSP does not offer a clear advantage: it slightly increases the total path length over TSP and fails to reduce inter-panel distances as effectively as CH-TSP. To illustrate the differences between the proposed strategies, Figure 2 shows that the TSP solution on the R layout introduces long diagonal jumps between distant panel rows to minimize overall distance. The C-H-TSP strategy, on the other hand, produces a more compact path. IV. CONCLUSIONS This work addressed the problem of planning UAV inspection paths in PV plants by accounting for the dual nature of drone navigation: reliable visual servoing along panel rows versus less accurate GNSS-based transitions between PV panel rows. We proposed three solutions with different cost-modeling assumptions. The results confirm that the CH-TSP method is effective in minimizing both the average and maximum inter-panel distances. A significant advantage of the C-H-TSP formulation is that it completely decouples the optimized inspection path from the UAV’s takeoff/landing location. This means that the takeoff and landing points can be selected a posteriori, giving operators maximum flexibility in real-world deployments. Ongoing work is focused on testing the system in a mockup PV plant outside our facility. REFERENCES [1] C. Prasshanth, S. B. Narayanan, N. V. Sridharan et al., “Fault detection in photovoltaic systems using unmanned aerial vehicle-captured images and rough set theory,” Sol. Energy, vol. 290, 2025. [2] E. Velasco-Sánchez, L. F. Recalde, B. S. Guevara et al., “Visual servoing NMPC applied to UAVs for photovoltaic array inspection,” IEEE robot. autom. lett., vol. 9, no. 3, p. 2766–2773, Mar. 2024. [3] K.-D. Park, W.-J. Yoon, and J.-S. Lee, “Reduction of multipath effect in GNSS positioning by applying pseudorange acceleration as weight,” Sensors (Basel), vol. 24, no. 21, October 26 2024. [4] D. L. Applegate, R. E. Bixby, V. Chvatál et al.,The Traveling Salesman Problem: A Computational Study. Princeton University Press, 2006. 265