This article provides a comprehensive survey of deep learning architectures used for time series forecasting, covering both one-step-ahead and multi-horizon predictions. It highlights the integration of temporal information through encoder and decoder designs, and discusses recent advancements in hybrid models that combine statistical and deep learning components. The survey also explores how deep learning can facilitate decision support through interpretability and counterfactual prediction methods. The authors emphasize the importance of domain knowledge in improving model performance and outline future research directions, including continuous-time and hierarchical models.This article provides a comprehensive survey of deep learning architectures used for time series forecasting, covering both one-step-ahead and multi-horizon predictions. It highlights the integration of temporal information through encoder and decoder designs, and discusses recent advancements in hybrid models that combine statistical and deep learning components. The survey also explores how deep learning can facilitate decision support through interpretability and counterfactual prediction methods. The authors emphasize the importance of domain knowledge in improving model performance and outline future research directions, including continuous-time and hierarchical models.