Deep Learning with Hybrid Noise Reduction Architecture: Time Series Application

Vanessa Haykal (1) , Hubert Cardot (1) , Nicolas Ragot (1)

1 LIFAT - Laboratoire d'Informatique Fondamentale et Appliquée de Tours

Time series forecasting models have fundamental importance to various practical domains. It is desirable that these methods can learn non-linear dependencies and have a high noise resistance. In this paper, we propose a novel architecture using deep learning to address this challenge. We have adapted a new hybrid noise reduction architecture that use recursive error segments for learning and adjusting the predictions. The solution is based on a simultaneous fusion between the outputs of a Convolutional Neural Network (CNN) and a Long Short Term Memory (LSTM) network. This novel model is able to capture di↵erent types of properties which combination can substantially outperform their separate use. Applications involving electricity and financial datasets illustrate the usefulness of the proposed framework.

Mots clés: Deep learning; convolutional neural network; long short-term memory; noise reduction; recursive error segments; time series forecasting

Domaines: Réseau de neurones [cs.NE]; Machine Learning [stat.ML]

Vanessa Haykal, Hubert Cardot, Nicolas Ragot. Deep Learning with Hybrid Noise Reduction Architecture: Time Series Application. [University works] Université de Tours - LIFAT. 2019. ⟨hal-03151227⟩