TFT-Flood for Nationwide Flood Prediction in Turkey: A Three-Branch Hybrid Deep Learning Architecture
34th Signal Processing and Communications Applications Conference, SIU 2026, İstanbul, Türkiye, 7 - 10 Temmuz 2026, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/siu71813.2026.11636290
- Basıldığı Şehir: İstanbul
- Basıldığı Ülke: Türkiye
- Anahtar Kelimeler: deep learning, flood prediction, LSTM, natural disaster early warning, temporal fusion transformer, time series classification
- İstanbul Medipol Üniversitesi Adresli: Evet
Özet
This paper proposes TFT-Flood, a novel hybrid deep learning model that provides flood forecasting at the district level for Turkey. TFT-Flood uses the Temporal Fusion Transformer and consists of three separate modules: Static Enrichment, Temporal Processing and Risk Quantification that serve individual responsibilities. TFT-Flood is evaluated under a strict no leakage protocol for 6372 eight-day sequences for all 973 districts of the TUCBS open-data flood database with NASA POWER backfill, where the date of prediction is masked and cumulative engineered features are calculated on the basis of the observation period only. Although the marginal distribution of each meteorological feature is very similar for flood and non-flood data, TFT-Flood can successfully exploit temporal knowledge embedded in the dataset multivariate structure. The TFT-Flood ensemble (8 random seeds) achieves 0.909 AUC. In a chronology-based time split, the model exhibits performance degradation but is still better than a random predictor at a statistical level, implying a limited potential for cross-period generalization.