This dissertation develops a unified methodological framework for short-term forecasting in energy systems based on Support Vector Regression (SVR). The proposed framework integrates systematic feature engineering, data-driven feature selection procedures, and tailored hyperparameter optimization strategies. In the STLF domain, a seasonality-adjusted model (SSA-SVR, Strategic Seasonality-Adjusted SVR) and a hybrid DNN–SVR model are developed, combining the representational capabilities of deep neural networks with the generalization properties of SVR. To enhance forecasting precision, a structure of separate hour-specific submodels is established. In the ST-NGPF domain, a seasonality-adjusted SVR framework is applied and adapted to the specific dynamics and volatility of market data. Empirical validation is conducted on multiple real-world datasets. The research results demonstrate that the proposed models systematically outperform benchmark statistical and standalone machine learning models.
The doctoral dissertation was supervised by Prof. Kristijan Lenac, PhD, with Prof. Saša Vlahinić, PhD, serving as co-supervisor.
The Doctoral Dissertation Defense Committee consisted of: Prof. Goran Mauša, PhD (Chair), Assoc. Prof. Vedran Kirinčić, PhD (Member) — both from the Faculty of Engineering, University of Rijeka — and Prof. Juraj Havelka, PhD (Member) from the Faculty of Electrical Engineering and Computing, University of Zagreb.
Congratulations to our colleague!
