Prototype Aligned Embedding for Time Series Forecasting

Time-series forecasting plays a crucial role in supporting data-drivendecisions across domains like energy systems, finance, and healthcare.Despite their strong predictive power, many deep learning models arestill viewed as "black-boxes", making it difficult to understand howthey arrive at their predictions. This lack of transparency often limits their broader adoption in real-world applications. To address thischallenge, this thesis proposes an interpretable, prototype-based deeplearning framework for multivariate time-series forecasting that integrates both numerical and textual modalities to enhance predictiveaccuracy and transparency. At the core of the proposed framework lies the Prototype Similarity Network (ProSeNet), which learns a set of representative temporal patterns called prototypes that serve as interpretable anchors formodel reasoning. Each forecast is explained in terms of similarity tothese learned prototypes, providing clear insights into the model’sdecision process. To further enhance contextual understanding, theframework is extended into a multimodal architecture where textualembeddings derived from pretrained language models are fused withtemporal features, enabling the model to reason jointly over structured and unstructured data. Comprehensive experiments demonstrate that the proposed approachachieves strong predictive performance while maintaining interpretability. The prototype-based design allows users to trace predictions backto meaningful temporal patterns, and the integration of textual embeddings significantly improves contextual reasoning and accuracy.This work contributes a transparent, context-aware forecasting framework that bridges interpretable representation learning and multimodal reasoning advancing the development of trustworthy and explainable AI systems for time-dependent applications.