Field validation of a decoder-only Transformer for multi-actuator greenhouse control via imitation learning

Artificial intelligence (AI)-based greenhouse control has been evaluated mainly by simulation or offline prediction, and field evidence remains limited for operating physical actuators with AI-generated signals. This study developed and field-validated an AI-based greenhouse control-signal system combining a decoder-only Transformer actuator-control prediction model trained by imitation learning with a greenhouse data-acquisition and control platform. The platform integrated environmental sensing, model inference, signal transmission, fallback switching, actuator-state logging, and boiler power metering. A rule-based controller served as a deterministic, verifiable surrogate teacher and reference for actuator-level validation. Diagnostic analysis of the initial autoregressive configuration showed that reusing predicted control signals as subsequent inputs reduced operational reliability. Excluding previous control-signal variables removed this feedback path and enabled continuous non-autoregressive operation. In a subsequent snowfall-associated validation conducted after expansion of the training dataset, maximum daily control-rule violation rates were 0.90% or lower. In final validation, maximum daily violation rates remained at or below 1.94%, the absolute 24-h mean internal-temperature difference was 0.42 ± 0.29 °C, and the inter-greenhouse boiler-energy difference was 3.79 ± 0.88 kWh day⁻¹ (1.44 ± 0.46%). Because validation used an uncultivated greenhouse, the findings demonstrate actuator-level field feasibility rather than energy-saving performance or crop-level optimization.