Deep Learning-based Seasonal Forecasting Over K-means-derived Climate Zones in Türkiye

The primary aim of this study is to develop an integrated approach that redefines climate zones and enhances seasonal forecasts through multivariate clustering, climate extremes, and deep learning in Türkiye. Daily observations of minimum (TN) and maximum (TX) temperatures and total precipitation (RR) from 82 stations covering 1993–2022 were analyzed. K-means clustering (k = 2–15) was tested, with evaluation metrics indicating k = 5, 8, and 10 as the most meaningful solutions; k = 10 offered the best balance between regional gradients and local heterogeneity. Climate extremes were characterized using indices recommended by the Expert Team on Climate Change Detection and Indices (ETCCDI), and PCA/biplot analyses were employed to interpret cluster sensitivities. Results showed that precipitation extremes dominate the first principal component (PC1), while temperature extremes load on both PC1 and PC2, highlighting distinct climatic drivers. Seasonal forecasts for TN, TX, RR were generated using a Long Short-Term Memory (LSTM) deep learning model trained within each redefined clusters. The model revealed clear spatial and seasonal asymmetries: RMSE for TN and TX was lowest in summer (≈ 1–2 °C), moderate in spring and autumn (≈ 2–3 °C), and highest in winter (up to 4–5 °C), while RR errors were smallest in summer (≈ 5–40 mm) and largest in winter, exceeding 100 mm in the wettest clusters. Overall, the integrated k-means–PCA–LSTM framework demonstrates that cluster-based forecasts reduce errors relative to aggregated approaches, providing a more robust basis for climate risk assessment and adaptation strategies in Türkiye.