High-Fidelity Wind Speed Network Reconstruction from Sparse Monitoring Stations Using QR Pivoting, Shallow Decoder Networks, and Shallow Recurrent Decoders

Accurate reconstruction of spatially distributed wind speed is important for wind-resource assessment and monitoring when station coverage is sparse. This study compares three data-driven approaches for spatiotemporal wind-field reconstruction: a QR-pivoting linear model, the Shallow Decoder Network (SDN), and the Shallow Recurrent Decoder (SHRED). Monthly mean wind-speed observations from 26 stations in Türkiye’s Marmara region were evaluated across input configurations ranging from one station to full-network coverage. Under sparse conditions, SHRED showed the strongest and most stable performance, achieving a mean relative error (MRE) of 0.120 and a mean coefficient of efficiency (MCE) of 0.878 with a single input station. Its recurrent structure enabled effective use of temporal information and reduced sensitivity to the selected station across the tested realizations. As sensor density increased, SDN and QR improved substantially. QR outperformed SHRED on both MRE and MCE from 20 stations onward and slightly surpassed SDN at 25 stations. However, SDN achieved the highest full-network accuracy with all 26 stations (MRE = 0.034; MCE = 0.990). These findings demonstrate that reconstruction performance depends on monitoring-network density, with SHRED being particularly effective under sparse conditions and SDN and QR becoming increasingly competitive as spatial coverage increases.