Answers are grounded in this paper and cite their sources.
Sparse greenhouse surrogates are often selected by prediction accuracy, although their purpose is feedback control. We compare three sparse identification of nonlinear dynamics (SINDy) libraries and 15 controllers in a GreenLight tomato greenhouse using reanalysis weather. The identification ladder comprises 54 distinct recipes at 20 seeds. Physical features reduce mean one-step temperature root mean square error from 1.86 to 1.68 °C but increase median 24 h error from 2.67 to 23.79 °C; CO2 and humidity show the same reversal. Labelled numerical-completion and planner-cap variants yield a mixed cohort of 1048 full trajectories across four 60-day test seasons. The raw ensemble has the highest reference-setting mean margin (+4.32 EUR m−2), with 4140 climate-corridor variable-steps; the low-threshold intermediate controller achieves +3.07 with 3699, while proximal policy optimisation achieves +0.45 with 1330. Training-selected horizons and price-dependent selection retain the raw ensemble’s largest mean among four principal sparse configurations at all nine prices, without adjusted separation from the nearest comparator at nominal prices. Weather interventions, crop-state continuations and paired coefficient edits assess mechanisms and robustness. The inner model predictive controller predicts climate and resource use, not crop revenue; seasonal margin is evaluated separately. Results support complementary multi-step, interpretability and closed-loop checks within the tested sparse-identification setting.
No takes yet. Share an insight, caveat, or question.
Naumov et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: