Exploratory Analysis of Deep Learning Models for Forecasting Meteorological Parameters in the Agricultural Sector

This study compares GRU, LSTM, CNN-GRU, and CNN-LSTM architectures for multivariate agricultural weather forecasting using 134,376 hourly ERA5 observations from Ioannina, Greece, covering January 2011 through April 2026. Targets include reference evapotranspiration, vapour pressure deficit, wind speed, and sine/cosine wind-direction components at 24-hour and 168-hour horizons. According to the abstract, CNN-GRU achieves the highest composite WQS, while CNN-LSTM produces nearly identical results with substantially fewer parameters. Gains over recurrent baselines are modest: 1.22–1.63% at 24 hours and 0.44–0.45% at 168 hours.