Triad Conditioned Forecasting for Risk-Aware Autonomous Irrigation

Autonomous irrigation operates under incomplete sensing, measurement staleness, uncertain forecasts, and imperfect actuation. We study a triad-conditioned framework that separates environmental belief, executed-control history, and candidate-plan information. The environmental state is represented by adaptively routing among a direct temporal expert, a reliability-weighted random-feature kernel-mean expert, and a conservative fallback. The routed belief conditions conditional variational autoencoders (CVAEs) and recurrent Gaussian predictors. Because tail-risk control is meaningful only for calibrated scenarios, forecasts pass through chronological conformal calibration and state- and loss-space readiness checks before entering executed-action-aware mean–CVaR model predictive control (MPC). Experiments show that complexity is not uniformly beneficial: a unimodal GRU–Gaussian attains the best neural RMSE (0.1939) and energy score (0.3214), Random Forest gives the best overall point RMSE (0.1854), and the No-KME CVAE outperforms the full KME–RFF–attention CVAE. Principal CVAEs cover only 32.21–57.54% of nominal 95% intervals, and simulated controllers violate moisture constraints in 42.38–50.71% of decision steps. We therefore present the system as an auditable diagnostic architecture, not a deployed safety certificate. The results motivate adaptive representation routing, calibration-gated control, loss-tail backtesting, measured actuator models, and validated fallback operation.