Abstract
Reconstructing missing values in meteorological time series is essential for reliable renewable-energy assessment, but on-device imputation must respect the power, memory and latency budget of edge hardware. We benchmark 46 PyPOTS imputation models—four naïve baselines, TRMF and 41 neural networks—on two tropical coastal stations (UCC and CUC, Colombian Caribbean) resampled to 15 min and subjected to a PyGrinder missingness scenario, running each model on an NVIDIA Jetson Orin Nano (25 W) and measuring MAE, RMSE, power, memory and time. A hybrid Entropy–CRITIC weighting scheme with TOPSIS and MARCOS ranked the alternatives. The best neural networks cut the mean-imputer error by 40–45% (MAE) and 28–40% (RMSE), yet resource usage spanned two orders of magnitude, uncorrelated with accuracy (|ρ|≤0.15). Because both methods placed 64–76% of the weight on resource criteria, naïve baselines and TRMF topped the station-averaged and UCC rankings, TiDE was the most rank-stable neural network, and SCINet led at CUC under the untuned configuration. TOPSIS and SPOTIS agreed closely (ρ≥0.98); MARCOS agreed moderately with both (ρ≈0.68–0.93). Station-to-station agreement was low for TOPSIS (ρ=0.52)—a confound rather than a station effect: the same masking configuration produced an asymmetric gap structure at the shorter, previously incomplete UCC record and a uniform one at the longer, complete CUC record, so station-level conclusions are not directly comparable. An automated per-model hyperparameter-search protocol, which reorders the CUC top, is provided in an appendix. These results offer a reproducible, hardware-validated framework for edge-constrained imputation model selection.