DOI: 10.62520/fujece.1915506 ISSN: 2822-2881

Comparison of Deep Learning, Machine Learning, and Statistical Models for Precipitation and Temperature Forecasting Performance: A Case Study of Hakkari Province

Ertuğrul Gül
Show PDF Cite
This study compares three forecasting approaches for monthly mean temperature and total precipitation over Hakkari, southeastern Turkey: LSTM (deep learning), Random Forest (machine learning), and Prophet (statistical). ERA5 reanalysis data spanning 1940–2025 served as input, and model performance was evaluated using five metrics (RMSE, MAE, R², NSE, and KGE).Analysis of the 85-year ERA5 record reveals statistically significant warming across all seasons (0.221°C/10 years, p < 0.001), with spring warming being the strongest (1.578°C/10 years, p < 0.001). Long-term precipitation change remains negligible (+0.60 mm/10 years). Because snow cover and evapotranspiration were not directly modeled, interpretations related to snow-rain phase shifts and drought risk are treated as physically plausible implications rather than direct empirical conclusions.Under the leakage-safe evaluation, Prophet produced the best temperature performance (RMSE = 1.51°C, MAE = 1.15°C, R2 = 0.976, NSE = 0.976, KGE = 0.952), followed by Random Forest (RMSE = 1.96°C, MAE = 1.59°C, R2 = 0.960, NSE = 0.960, KGE = 0.839) and LSTM (RMSE = 2.52°C, MAE = 2.08°C, R² = 0.933, NSE = 0.933, KGE = 0.791).Precipitation forecasting was consistently more demanding than temperature forecasting. Prophet achieved the best precipitation performance (RMSE = 50.15 mm, MAE = 33.88 mm, R² = 0.601, NSE = 0.601, KGE = 0.681), while Random Forest (RMSE = 52.26 mm, MAE = 35.56 mm, R² = 0.567, NSE = 0.567, KGE = 0.672) and LSTM (RMSE = 53.80 mm, MAE = 36.80 mm, R² = 0.542, NSE = 0.542, KGE = 0.680) showed similar limited skill. These findings suggest that strongly seasonal variables such as temperature remain tractable across modeling frameworks, whereas monthly precipitation exhibits only moderate predictability under leakage-safe conditions when same-month information is excluded and lagged predictors are used. Future studies may benefit from incorporating additional spatial predictors, snow-related variables, and hydrometeorological indicators, as well as testing hybrid deep learning and ensemble learning architectures.

More from our Archive