Abstract
Daily maximum and minimum temperature forecasting is critical across multiple sectors, yet most studies are confined to single stations or homogeneous climates.This study proposes a CNN-LSTM for one-day-ahead Tmax and Tmin forecasting across nine stations spanning Algeria's three climate zones.The model is evaluated alongside eight approaches including Persistence, Climatology, ARIMA, SARIMA, ANN, SVR, LSTM, and the Global Forecast System using a two-year test period.Results reveal pronounced asymmetry between Tmax and Tmin.For Tmax, GFS consistently outperforms CNN-LSTM across all stations.Conversely, CNN-LSTM outperforms GFS for Tmin at eight of nine stations, with GFS showing particularly poor skill at Dar El Beida (R² = 0.22).CNN-LSTM consistently beats baselines and classical models, yet performs indistinguishably from SVR and LSTM at most stations, indicating limited gains from architectural complexity.Seasonal analysis shows a consistent spring-difficult pattern for Tmax versus station-specific structure for Tmin.Findings support a targeted role for data-driven forecasting as a complement to NWP for Tmin in data-sparse regions, not as a substitute for Tmax.