Time series forecasting is pivotal in domains ranging from energy management to financial analytics, where accurate predictions hinge on effectively capturing inherent temporal correla- tions. Traditional approaches often address these correlations in...
Time series forecasting is pivotal in domains ranging from energy management to financial analytics, where accurate predictions hinge on effectively capturing inherent temporal correla- tions. Traditional approaches often address these correlations in isolation, leading to suboptimal performance across varying data complexities. This dissertation introduces a correlation-driven progressive framework that systematically models three fundamental temporal relations: inter- nal patterns within univariate sequences, interactions among variables in multivariate data, and fusions between numerical time series and exogenous textual modalities. By grounding the framework in these correlations, it establishes a logical progression where each model is ex- plicitly designed to capture a specific correlation type, enhancing predictive robustness and interpretability in deep learning-based TSF. The framework commences with DelayNet, a lightweight convolutional architecture de- signed to capture intra-series correlations in univariate datasets through optimized temporal convolutions and delay mechanisms. This model achieves 5-20% reductions in mean squared error, alongside substantial decreases in computational runtime and parameters compared to established Deep Learning (DL) baselines, while maintaining or surpassing predictive accu- racy. Extending this base, LSPatch-T extends the approach to inter-variate correlations in multi- variate settings. Through a self-supervised pre-training strategy incorporating semantic patch- ing and hybrid tokenization, LSPatch-T captures cross-variable dependencies, such as covari- ations, yielding 10–20% average MSE improvements across benchmarks like ETTh1, ETTh2, and Weather datasets. This enhances generalization by enabling seamless transitions between temporal, variate, and patch views. Culminating the progression, TextTimeFuse integrates cross-modal relationships by fusing textual contexts with time series data using cross-attention and pre-trained language model embeddings. Evaluations on the Time-MMD dataset yield 3.3% mean squared error and 9.7% mean absolute error gains over unimodal counterparts across nine domains. Through rigorous experimental validations, this work contributes: (1) advanced sparse con- volutional techniques for intra-series correlation capture; (2) transfer learning paradigms for inter-variate modeling; and (3) multimodal fusion mechanisms for cross-modal enhancements. The correlation-driven paradigm offers a cohesive pathway for TSF advancements, promoting innovative AI applications in dynamic environments. Keywords: Deep Learning, Time Series Forecasting, Temporal Convolutional Network, Trans- former, Multi-modal Time Series