Machine learning-based greenhouse temperature forecasting model with structural and geospatial features

Accurate forecasting of greenhouse internal temperature is important for proactive climate management, energy-efficient operation, and stable crop production. This study evaluates a GRU-based forecasting model that augments conventional environmental measurements with deterministic structural and solar-geometry descriptors, including latitude, longitude, roof slope, roof azimuth, solar altitude, solar azimuth, and roof-specific incidence angles. Using winter data collected from an experimental greenhouse with spatially distributed sensors and an on-site weather station, we compared a baseline GRU with an improved GRU incorporating the added descriptor set. The improved model increased the average R² from 0.622 to 0.662 and reduced the average RMSE from 1.935 to 1.801°C across sensor locations. A repeated rolling evaluation using 7-day training and 1-day testing windows showed consistent improvement over the available dataset, supported by paired Wilcoxon tests. These results indicate that structural and solar-geometry descriptors can improve short- to medium-term greenhouse temperature forecasting under the present single-site winter conditions. Broader validation across seasons and sites remains necessary.