COMPARATIVE ASSESSMENT OF RANDOM FOREST, XGBOOST AND MLP MODELS FOR PREDICTIVE GREENHOUSE CONTROL USING WIRELESS SENSOR NETWORK DATA
EVALUAREA COMPARATIVĂ A MODELELOR RANDOM FOREST, XGBOOST ȘI MLP PENTRU CONTROLUL PREDICTIV AL UNEI SERE, FOLOSIND DATE DINTR-O REȚEA DE SENZORI WIRELESS
DOI : https://doi.org/10.35633/inmateh-79-51
Authors
Abstract
Control of greenhouse microclimate requires the interpretation of several environmental and substrate-related parameters, including air temperature, relative humidity, light intensity, carbon dioxide concentration and soil moisture. This paper compares three machine learning models, Random Forest, XGBoost and multilayer perceptron, for 30-minute-ahead prediction of irrigation, ventilation and shading states in a greenhouse. The models were trained using data collected from a wireless sensor network with four sensor nodes installed in the protected cultivation area. After preprocessing, scaling, error removal, temporal synchronization and resampling at 10-minute intervals, a structured dataset of 7063 records was obtained. Individual sensor readings, spatially aggregated variables, time-related features, rolling averages and short-term differences were used to describe both the current and recent evolution of the greenhouse environment. The dataset was split chronologically into training, validation and test subsets, preserving the temporal order of the measurements. The best average performance was obtained by XGBoost, with a mean F1-score of 0.9237, followed by Random Forest with 0.9045 and MLP with 0.5862. Random Forest achieved the best results for irrigation and ventilation control, with F1-scores of 0.9630 and 0.9611, respectively, while XGBoost achieved the best result for shading control, with an F1-score of 0.8571. The results indicate that tree-based ensemble models are more suitable than the tested MLP architecture for predictive greenhouse control based on multi-sensor WSN data. The proposed framework supports the selection of a final control model according to prediction performance, stability and interpretability.
Abstract in Romanian



