Computational intelligence-based time series forecasting

Time series forecasting is a crucial area of data science that is essential for decision-making across multiple domains such as transportation, finance, meteorology, and energy management. Computational Intelligence (CI) provides a flexible approach to solve forecasting tasks that is not limited to traditional statistical, machine learning, or deep learning-based methods. This thesis proposes three novel CI-based forecasting paradigms to deal with various forecasting scenarios. The first paradigm proposes a decomposition-based forecasting framework to address complex sequential data with multi-resolution information. The proposed framework addresses the data leakage issue and the inconsistency between training and testing phases caused by decomposition. Experimental results demonstrate that the proposed framework outperforms existing methods significantly. The second paradigm presents a dynamic performance-based ensemble forecasting framework that addresses time-varying characteristics in sequential data. The proposed Bayesian optimization-based dynamic ensemble (BODE) overcomes the limitations of single model-based techniques by incorporating ten distinct model candidates, including statistical methodologies, machine learning-based models, and deep neural network models. The proposed framework outperforms current statistical, machine learning, and deep learning methods. The third paradigm proposes a graph ensemble deep random vector functional link network (GEdRVFL) for node-wise traffic forecasting. The proposed model combines deep neural networks with graph machine learning to capture the spatiotemporal dynamics of the transportation network. Experimental results demonstrate that the proposed model outperforms the most advanced models in four out of five scenarios. Overall, the proposed CI-based forecasting paradigms have shown promising results in various forecasting scenarios and can aid decision-making across multiple domains.Doctor of Philosoph