Modeling Long- and Short-Term Temporal Patterns with Deep Neural Networks

TLDR

Multivariate time‑series forecasting is crucial in domains such as solar energy, electricity demand, and traffic prediction, yet real‑world data often contain intertwined long‑term and short‑term patterns that traditional autoregressive and Gaussian process models struggle to capture. This work introduces LSTNet, a deep learning framework designed to jointly model long‑ and short‑term temporal dependencies. LSTNet combines convolutional layers to capture short‑term local dependencies, recurrent layers to learn long‑term trends, and a traditional autoregressive component to address scale sensitivity, with all data and code publicly available. On real‑world datasets with complex repetitive patterns, LSTNet outperformed several state‑of‑the‑art baselines, achieving significant performance gains.

Abstract

Multivariate time series forecasting is an important machine learning problem across many domains, including predictions of solar plant energy output, electricity consumption, and traffic jam situation. Temporal data arise in these real-world applications often involves a mixture of long-term and short-term patterns, for which traditional approaches such as Autoregressive models and Gaussian Process may fail. In this paper, we proposed a novel deep learning framework, namely Long- and Short-term Time-series network (LSTNet), to address this open challenge. LSTNet uses the Convolution Neural Network (CNN) and the Recurrent Neural Network (RNN) to extract short-term local dependency patterns among variables and to discover long-term patterns for time series trends. Furthermore, we leverage traditional autoregressive model to tackle the scale insensitive problem of the neural network model. In our evaluation on real-world data with complex mixtures of repetitive patterns, LSTNet achieved significant performance improvements over that of several state-of-the-art baseline methods. All the data and experiment codes are available online.