A Review of Time Series Prediction Models Based on Deep Learning

Time series data, capturing the evolution of specific variables over time, underpins critical forecasting tasks across diverse domains. While traditional models struggle with the inherent complexity (nonlinearity, non-stationarity, seasonality) of real-world time series, deep learning (DL) has emerged as a transformative solution, leveraging its prowess in nonlinear modeling and automated feature extraction. This paper provides a comprehensive review of recent advancements in DL-based time series prediction models. We systematically analyze the characteristics, strengths, and limitations of prominent model families, including Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs) (e.g., LSTM, GRU), Transformer variants (e.g., Informer, Autoformer, iTransformer), Graph Neural Networks (GNNs), and hybrid models. On this basis, we establish a structured, task-oriented framework for model selection and cross-comparison, evaluating these models across essential dimensions such as sequence length handling, multivariate support, interpretability, computational efficiency, and real-time performance. Finally, addressing the challenges and issues faced by deep learning in time series prediction, this paper prospects future research directions, including enhancing model interpretability, optimizing model structures, and integrating multi-source data and domain knowledge.