Time series prediction and forecasting using Deep learning Architectures

Nature brings time series data everyday and everywhere, for example, weather data, physiological signals and biomedical signals, financial and business recordings. Predicting the future observations of a collected sequence of historical observations is called time series forecasting. Forecasts are essential, considering the fact that they guide decisions in many areas of scientific, industrial and economic activity such as in meteorology, telecommunication, finance, sales and stock exchange rates. A massive amount of research has already been carried out by researchers over many years for the development of models to improve the time series forecasting accuracy. The major aim of time series modelling is to scrupulously examine the past observation of time series and to develop an appropriate model which elucidate the inherent behaviour and pattern existing in time series. The behaviour and pattern related to various time series may possess different conventions and infact requires specific countermeasures for modelling. Consequently, retaining the neural networks to predict a set of time series of mysterious domain remains particularly challenging. Time series forecasting remains an arduous problem despite the fact that there is substantial improvement in machine learning approaches. This usually happens due to some factors like, different time series may have different flattering behaviour. In real world time series data, the discriminative patterns residing in the time series are often distorted by random noise and affected by high-frequency perturbations. The major aim of this thesis is to contribute to the study and expansion of time series prediction and multistep ahead forecasting method based on deep learning algorithms. Time series forecasting using deep learning models is still in infancy as compared to other research areas for time series forecasting.Variety of time series data has been considered in this research. We explored several deep learning architectures on the sequential data, such as Deep Belief Networks (DBNs), Stacked AutoEncoders (SAEs), Recurrent Neural Networks (RNNs) and Convolutional Neural Networks (CNNs). Moreover, we also proposed two different new methods based on muli-step ahead forecasting for time series data. The comparison with state of the art methods is also exhibited. The research work conducted in this thesis makes theoretical, methodological and empirical contributions to time series prediction and multi-step ahead forecasting by using Deep Learning Architectures.