Forecasting Multivariate Time Series Data Using Neural Networks

Abstract

Over the last few years, neural networks have become extremely popular, and their usage is increasing rapidly. This project has investigated the use of neural networks for one-step time series forecasting on highly random data. Multi-layer perceptron (MLP), convolutional neural networks (CNN), recurrent neural networks (RNN), and long short-term memory (LSTM) cells are tested to see if they can give a binary classification accuracy above 50% using this data. The assignment focuses on designing a small embedded neural network with low latency.

The different neural network architectures are built using a deep learning library in Python, called Keras. This is a high-level software framework, built on top of either Tensorflow or Theano, for fast and easy prototyping of neural networks.

The conclusion of the study is that only the CNN satisfied the requirements of the assignment during the work of this thesis. None of the other architectures showed sign of learning generalized patterns and structures from the dataset in question. The CNN showed the most promising results, being able to extract information about the training set that increased the classification accuracy of the test. This leads the way for further development and an eventual hardware implementation of the inference phase reducing the run-time latency.