Single Model, Multiple Climates: Applying Modified Levenberg Marquardt Algorithm to Meteorological Time-Series Prediction
12th International Conference on Control, Decision and Information Technologies, CoDIT 2026, Bari, İtalya, 13 - 16 Temmuz 2026, ss.43-48, (Tam Metin Bildiri)
- Yayın Türü: Bildiri / Tam Metin Bildiri
- Doi Numarası: 10.1109/codit70676.2026.11630966
- Basıldığı Şehir: Bari
- Basıldığı Ülke: İtalya
- Sayfa Sayıları: ss.43-48
- Hacettepe Üniversitesi Adresli: Evet
Özet
Forecasting meteorological time series typically requires training separate models for distinct geographical locations due to the complexity of local climate dynamics. When a single Artificial Neural Network (ANN) is tasked with learning multiple datasets with diverse characteristics, the phenomenon known as "catastrophic forgetting"often degrades performance on previously learned tasks. In this study, the modified Levenberg-Marquardt (LM) algorithm, which utilizes a parameter masking strategy, is applied and validated for multi-site weather forecasting. Real-world hourly data spanning 2022-2024 for three cities representing different climates - Ankara (Continental), Mersin (Mediterranean), and Çanakkale (Transitional) - were utilized. A single shallow ANN architecture, equipped with randomly generated binary mask vectors specific to each city, was trained simultaneously on these datasets. The proposed method organizes the network weights into shared and task-specific subnetworks, allowing distinct climate patterns to be stored within the same model without interference. Experimental results demonstrate that the model accurately predicts temperature variations with different forecast horizons across all three cities, effectively preventing catastrophic forgetting. This study confirms that the modified LM algorithm provides robust generalization capabilities not only for theoretical regression tasks but also for complex, noisy real-world time series.