Automated IoT-enabled greenhouse microclimate prediction using multivariate sequence-to-sequence LSTM-based encoder-decoder

Precision agriculture implemented in controlled environment agriculture (CEA) can facilitate sustainable farming by providing accurately managed environments for crop cultivation. However, effectiveness relies on the timely prediction and management of internal microclimates. This research presents a novel framework for multi-step ahead greenhouse microclimate prediction that integrates the Internet of Things (IoT) sensing and machine learning for smart agriculture applications using sequence-to-sequence (Seq2Seq) multivariate encoder-decoder long short-term memory (MED-LSTM) networks, forecasting the dynamic and interdependent environmental factors including temperature, humidity, CO₂ concentration, and light condition. The proposed approach employs both internal and external temporal information to capture complex dynamics and variable interdependencies, generating accurate 24 hour-ahead forecasts at a 1-hour scale. The proposed MED-LSTM model achieved MAEs of 0.89, 8.31, 9.33, and 131.45 for room temperature, humidity, CO₂, and PAR, respectively. Prediction error distribution on a scale of 24 hours was also discussed. The results demonstrate that the proposed method significantly enhances the precision of microclimate forecasting against baseline models, enabling more proactive and efficient greenhouse management. Practical implications include improved energy efficiency, resource conservation, and crop quality optimization, contributing to sustainable agricultural practice. This research advances the application of IoT-enabled sensing and machine learning in smart farming and highlights the potential for innovative solutions to global food production challenges.