Generative time series model based on encoder-decoder architecture

Abstract

Encoder-decoder neural network models have found widespread use in recent years for solving various machine learning problems. In this paper, we investigate the variety of such models, including the sparse, denoising and variational autoencoders. To predict non-stationary time series, a generative model is presented and tested, which is based on a variational autoencoder, GRU recurrent networks, and uses elements of neural ordinary differential equations. Based on the constructed model, the system is implemented in the Python3 environment, the TensorFlow2 framework and the Keras library. The developed system can be used for modeling continuous time-dependent processes. The system minimizes a human factor in the process of time series analysis, and presents a high-level modern interface for fast and convenient construction and training of deep models.

Keywords

• prediction
• variational autoencoder
• GRU recurrent neural network
• neural ordinary differential equation
• latent space
• nonstationary time series