PVade Wind-Loading Structural Response Dataset for Solar Photovoltaic Arrays

This dataset contains approximately 3,800 coupled fluid-structure interaction (FSI) simulations of single-axis tracked photovoltaic (PV) arrays, computed using PVade (PV Aerodynamic Design Engineering), an open-source tool for simulating the wind-driven motion and deformation of PV systems (https://www.nlr.gov/computational-science/pvade-photovoltaic-aerodynamic-design-engineering-software, https://pvade.readthedocs.io/en/latest/). Each simulation models the wind loading and structural response of a tracker-mounted PV array under a randomly sampled combination of array geometry and environmental conditions.

For each sample, the following parameters were drawn via random sampling: number of streamwise rows (3-10), number of modules per row (25-121, used to compute panel span), ground cover ratio (0.3-0.5, used to compute row-to-row spacing), wind direction (90-180 deg), and wind speed (4-20 m/s). Tracker angle (-60 to 60 deg) was sampled using one of three schemes across the dataset: a single uniform angle applied to all rows (representing ideal, synchronized tracking), a uniform angle with small row-to-row jitter of +/-5 deg (representing minor real-world tracking misalignment), or fully independent random angles per row (representing larger row-to-row misalignment). These three regimes were included to help machine learning models learn to generalize across realistic variations in tracker positioning. All other array, structural, and solver parameters were held fixed across samples and are provided in baseline_inputs.yaml. The corresponding domain for each sample is computed automatically from the array footprint and wind direction, with buffer distances added upstream, downstream, and laterally.

PVade solves the Arbitrary Lagrangian-Eulerian (ALE) form of the incompressible Navier-Stokes equations using the finite element method with large-eddy simulation (LES) turbulence closure using the Smagorinsky subgrid model. A logarithmic-law inflow velocity profile is applied, with slip boundary conditions on the lateral and upper domain boundaries and no-slip at the ground. The structural response of the PV hardware to the calculated fluid forces is modeled with a nonlinear solver. The fluid and structural solvers are coupled via mesh adaptation and a partitioned FSI scheme with relaxation.
The output of each simulation is provided as an indexed sample_##### directory containing the following items: input_params.yaml (the full realized parameter set for that sample, including the domain bounds and per-row tracker angles); log.txt (solver run log); a mesh/ subdirectory with the fluid and structural meshes and a wall-distance field (in .h5, .xdmf, and .msh formats); and a solution/ subdirectory containing the full time-resolved fluid and structural solution fields (solution_fluid.h5, solution_structure.h5, with accompanying .xdmf files), along with summary time series of lift and drag coefficients (lift_and_drag.csv) and structural acceleration/position (accel_pos.csv). A top-level samples.csv summarizes the key varied parameters (array size, wind conditions, tracker angle, etc.) for every sample, and baseline_inputs.yaml documents all parameters that were held fixed.

The simulations were computed on the Kestrel high-performance computing system at the National Laboratory of the Rockies. This work was funded by a National Laboratory of the Rockies Laboratory Directed Research and Development program. The data was collected, reformatted, and prepared to serve as a training and benchmarking dataset for machine learning and surrogate modeling of wind loads and structural response on tracked PV arrays, though it may also support other aerodynamic or structural analysis studies. A github repo (https://github.com/NatLabRockies/pvade-dataset-explorer) is included that provide scripts and notebooks to support accessing and processing of the data.