Paper Info
Title
Efficient Time Series Forecasting Using Time Delay Neural Networks with Domain Pre-Transforms
Abstract
Many applications that involve prediction or forecasting of the state of a dynamic system can be formulated as a time series forecasting problem. In these cases, the state of the system is represented as a time series. Among countless examples are weather forecasting, predicting energy consumption of a system, predicting future position of a vehicle, or prediction of economic metrics. Time Delay Neural Networks (TDNNs) are one of the main tools for this purpose that have recently received attention. In this paper, we examine how a transform layer applied to the beginning layer of the TDNN (called pre-transform layer) can significantly improve learning time, while also moderately improving accuracy. We take the example of vehicle speed time series and show that TDNNs can significantly improve the prediction accuracy; applying a Discrete Cosine Transform (DCT) or a mix of DCT and Haar transforms improves the accuracy even more, while significantly reducing the training time. We also observe that these shallow networks (e.g., 3 layers) are able to learn the inverse transform automatically. This means that the network output does not have to go through an inverse transform. The results are verified using two different data sets.
Year
DOI
Venue
2019
10.1109/MWSCAS.2019.8884826
Midwest Symposium on Circuits and Systems Conference Proceedings
Keywords
Field
DocType
time series forecasting,time delay neural networks,DCT,Haar,transforms,mixed transforms
Time series,Data set,Control theory,Haar,Computer science,Discrete cosine transform,Algorithm,Time delay neural network,Artificial neural network,Energy consumption,Weather forecasting
Conference
ISSN
Citations 
PageRank 
1548-3746
0
0.34
References 
Authors
0
2
Name
Order
Citations
PageRank
Masoumeh Kalantari Khandani100.34
Wasfy B. Mikhael27676.27