Computational study demonstrates improved multivariate time series forecasting across complex temporal datasets, highlighting the benefits of combining convolutional and recurrent neural...
Key Points
To develop a deep learning architecture capable of simultaneously modeling short-term local dependencies and long-term recurring trends in multivariate time series data.
Designed the Long- and Short-term Time-series network (LSTNet) architecture, integrating Convolutional Neural Networks (CNNs) for local inter-variable patterns with Recurrent Neural Networks (RNNs) for long-term trends.
Coupled the deep neural components with a linear autoregressive path to overcome neural network scale insensitivity.
Evaluated the framework on complex real-world benchmark datasets, including solar power generation, electricity consumption, and traffic flow.
Achieved statistically significant forecasting accuracy gains over state-of-the-art baselines and traditional statistical models across multiple real-world benchmarks.
Demonstrated qualitative improvements in modeling both high-frequency fluctuations and prolonged seasonal or cyclical shifts without losing scale sensitivity.