Over here, we report our experience on dealing various deep learning algorithms for time series prediction and forecasting tasks. Additionally, we propose some advance hybrid models with improved performance and accuracy for temporal sequences.
In our lab, we work with different types of time series data for example, biomedical signals, meteorological / weather data and sequential traffic patterns. Recently, Deep learning approaches are invading rapidly and providing remarkable performance in almost all domains. To model the complex time series data, one possibility is to develop more robust and meaningful features that are capable to detain the appropriate information and can accurately represent the behavior of underlying data. However, it is obvious that computing domain specific features for each task is expensive, time-consuming and needs expertise of the data. Alternative to this, is to use unsupervised feature learning. It is not something new, this choice of representation has an enormous effect on the performance of machine learning algorithms. Deep learning aims to learn good or in other sense highly meaningful representations from low level raw features.
1. Meteonowcasting using Deep Learning architecture