FALCON-ABC: Fully automated LSTM co-optimization via NAS and ABC for time series forecasting
Applied Soft Computing, cilt.203, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 203
- Basım Tarihi: 2026
- Doi Numarası: 10.1016/j.asoc.2026.116257
- Dergi Adı: Applied Soft Computing
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Applied Science & Technology Source, Compendex, INSPEC
- Anahtar Kelimeler: ABC, GPR, HPO, NAS, Time series forecasting
- Erciyes Üniversitesi Adresli: Evet
Özet
Accurate and computationally efficient forecasting of hydro-meteorological variables remains challenging due to the nonlinear, stochastic nature of climate data and the high computational cost associated with deep learning model design. In this study, a novel surrogate-assisted optimization framework, termed FALCON-ABC, is proposed by integrating Neural Architecture Search (NAS) and Hyperparameter Optimization (HPO) within an Artificial Bee Colony (ABC) optimization scheme enhanced by Gaussian Process Regression (GPR). The proposed framework automatically identifies optimal LSTM architectures and training hyperparameters, reducing the need for manual model design. Experiments were conducted using monthly temperature and rainfall data from the Kocasinan–Yamula meteorological station. The results show that FALCON-ABC substantially improves predictive performance over the baseline LSTM model, increasing the test-set NSE from 0.660 to 0.979 for temperature forecasting and from 0.685 to 0.953 for rainfall forecasting. Additional comparative experiments against Grid Search, Random Search, Particle Swarm Optimization, and Whale Optimization Algorithm further demonstrate that ABC provides the most favorable accuracy–efficiency trade-off within the proposed NAS–HPO framework. A controlled ablation analysis confirms the contribution of the GPR surrogate by reducing the optimization time while preserving high forecasting accuracy. Furthermore, backbone-level comparisons with Transformer Encoder and GAF-based Vision Transformer models empirically support the selection of LSTM as the core forecasting architecture for the present monthly hydro-meteorological dataset. To examine transferability beyond the original station, a supplementary experiment was also conducted on the independent Jena Climate benchmark dataset, where FALCON-ABC achieved reliable performance across several thermodynamic and humidity-related meteorological variables. Overall, the findings indicate that FALCON-ABC provides an effective, automated, and computationally efficient framework for deep learning-based time-series forecasting, with strong potential for broader meteorological forecasting applications.