An Approach to Time Series Forecasting with Spiking Neural Networks (Preprint)

Abstract

Analyzing time series data is critical in applications as di-verse as demand forecasting and speech recognition. Timely and energy-efficient predictions can play a key role on edge devices, where power requirements can be stringent, but rapid reactions are required. In re-cent years, several Deep Learning (DL) algorithms have been successfully applied to solve time series data problems; however, they inherently lack the ability to process time as a dimension and often have high SWaP (Size, Weight, and Power) demands. Spiking Neural Networks (SNNs) are regarded as a new avenue in which to solve time series problems but with lower-SWaP needs due to their more closely biologically inspired algorithmic architecture. In this work, we propose an SNN pipeline to process and forecast time series. We develop a novel data spike-encoding mechanism, and two loss functions that optimise the prediction of the time series in their original domain. Our data encoding system is inspired by NM event sensors as a means to increase the interoperability of our algorithm with NM technology as well as to prepare data to be processed by the SNN; our loss function takes into account the number of events transmitted throughout the network to keep a lower power consumption while ensuring convergence to top-level solutions. We develop an SNN in such a way that is compatible for hardware deployment on NM chips. We utilize the Panama short-term electricity load forecasting dataset and the electricity transformer temperature dataset to evaluate our solution on a univariate forecasting task. Collectively, our results show that SNNs can outperform the baseline model on the forecasting task, while also ensuring low power consumption and fast information processing. This underlines how SNNs can be a perfect fit for other time series forecasting tasks, including real-time signal processing and tasking on edge devices.

Open PDF

Document Details

Document Type
Technical Report
Publication Date
Jul 01, 2023
Accession Number
AD1206875

Entities

People

  • Alex Vicente-sola
  • Davide L. Manna
  • Gaetano Di Caterina
  • Paul Kirkland
  • Trevor J Bihl

Organizations

  • Air Force Research Laboratory
  • University of Strathclyde

Tags

Communities of Interest

  • Energy and Power Technologies
  • Materials and Manufacturing Processes

DTIC Thesaurus Topics

  • Air Force
  • Air Force Facilities
  • Air Force Research Laboratories
  • Algorithms
  • Artificial Intelligence
  • Artificial Intelligence Software
  • Automated Speech Recognition
  • Coding
  • Computer Science
  • Deep Learning
  • Electrical Engineering
  • Energy Consumption
  • Engineering
  • Information Processing
  • Information Systems
  • Machine Learning
  • Neural Networks
  • Recognition
  • Signal Processing
  • Standards

Fields of Study

  • Computer science

Readers

  • Neural Network Machine Learning.
  • Systems Analysis and Design

Technology Areas

  • AI & ML
  • AI & ML - Machine Learning Algorithms
  • AI & ML - Neural Networks