Systematic evaluation of exogenous weather-conditioned generative augmentation for greenhouse microclimate time series

Greenhouse microclimate prediction and control require sufficient, diverse indoor air temperature (TA) and relative humidity (RH) time series, but stable acquisition is often constrained by sensor noise, long-term degradation, and site-specific differences. Generative paradigms have rarely been compared for exogenous weather-conditioned greenhouse microclimate augmentation. This study evaluated variational autoencoder-, generative adversarial network-, raw diffusion-, and latent diffusion-based models within a common framework using EnergyPlus-based pseudo-measurements to control data-related confounding across paradigms. Marginal, temporal, and cross-variable fidelity was assessed using Wasserstein distance (WD), autocorrelation-function mean squared error (ACF-MSE), and TA-RH correlation difference (Δρ). Across three training runs, the exogenous-conditioned raw-space denoising diffusion probabilistic model (Raw-DDPM) achieved the lowest mean WD and ACF-MSE: 0.0593 and 1.31 × 10 −5 for TA, and 0.2573 and 3.71 × 10 −5 for RH, with over tenfold WD reductions versus unconditional generation. Coupling rankings varied across runs, with the Transformer-based time-series generative adversarial network (TTS-GAN) achieving the lowest mean Δρ. Time-only benefits were model-dependent. Three-run evaluation of 320 input-subset configurations identified outdoor temperature and humidity as the largest main-effect inputs for their targets; solar radiation supported marginal and temporal fidelity but not coupling on average. Adding the counter indoor variable improved WD in a mean 70.8% of TA and 82.3% of RH paired subsets, supporting temperature–humidity coupling as conditioning information, although exact minimizing subsets varied across runs. The framework supports model and conditioning-input selection for greenhouse and other sensor-based systems according to the statistical properties required for analysis.