Innovative Time Series Forecasting: Auto Regressive Moving Average vs Deep Networks
PERFORMER – Innovative Time Series Forecasting: Auto Regressive Moving Average vs Deep Networks
Anthony Mouraud, CEA, France—Responsible of Machine Learning and Software En-gineering Projects at CEA Tech Nantes.
Gives an insight in comparative performances of a commonly used model in time series prediction and a state of the art model in deep learning.
“Over “Growing interest in meaningful indicators extraction from the huge amounts of data generated by energy efficient buildings instrumentations has led to focusing on so called smart analysis algorithms. This work pro-poses to focus on statistical and machine learning approaches that make use only of available data to learn relationships, correlations and dependen-cies between signals. In particular, time series forecasting is a key indication to anticipate, prevent and detect anomalies or unexpected behaviors. As part of the PERFORMER European project, this work focuses on real data produced by instrumented buildings recorded time series among environ-mental, energy or occupancy domains. We propose to compare perfor-mances of a classical ARMA approach to a Deep Highway Network on time series forecasting only making use of past values of the series. In recent years, Deep Learning has been extensively used for many classification or detection tasks. The complexity of such models is often an argument to dis-card such approaches for time series prediction w. r. t. more common ap-proaches performances. Here we give a first attempt to evaluate benefits of one of the most up to date models in the literature for time series prediction.