Kp forecasting with a recurrent neural network

Ernest Scott Sexton,Xuanye Ma,Katariina Nykyri

doi:10.1051/swsc/2019020

Abstract

In an effort to forecast the planetary Kp-index beyond the current 1-hour and 4-hour predictions, a recurrent neural network is trained on three decades of historical data from NASA’s Omni virtual observatory and forecasts Kp with a prediction horizon of up to 24 h. Using Matlab’s neural network toolbox, the multilayer perceptron model is trained on inputs comprised of Kp for a given time step as well as from different sets of the following six solar wind parameters, Bz, n, V, |B|, σB and $ {\sigma }_{{B}_z}$. The purpose of this study was to test which combination of the solar wind and Interplanetary Magnetic Field (IMF) parameters used for training gives the best performance as defined by correlation coefficient, C, between the predicted and actually measured Kp values and Root Mean Square Error (RMSE). The model consists of an input layer, a single nonlinear hidden layer with 28 neurons, and a linear output layer that predicts Kp up to 24 h in advance. For 24 h prediction, the network trained on Bz, n, V, |B|, σB performs the best giving C in the range from 0.8189 (for 31 predictions) to 0.8211 (for 9 months of predictions), with the smallest RMSE.

Highlights

1.1 Artificial Neural Networks (ANNs)The concept of machine learning was first introduced by Samuel (1959), it was not until the early 2000s that the private and scientific communities were able to leverage this powerful idea to solve increasingly difficult problems, largely due in part to the explosion of both computing power and our ability to record, store, and process large amounts of data
Nonlinear in nature, forecasting is a prime candidate for the application of neural networks, which have been shown to be effective in modeling the behavior of highly complex systems
Data from 12/31/16 were initialized and the model predicted 24 h in advance. These were replaced with data from 1/1/16 and the process repeated for 272 days

Summary

Introduction

The concept of machine learning was first introduced by Samuel (1959), it was not until the early 2000s that the private and scientific communities were able to leverage this powerful idea to solve increasingly difficult problems, largely due in part to the explosion of both computing power and our ability to record, store, and process large amounts of data. One subset of the overarching field of machine learning is time series prediction, in which a computer’s ability to forecast future events based on prior data is tested. Chaotic time series are notoriously difficult to predict and are seen in fields ranging from signal processing (Lapedes & Farber, 1987) to economics (Kaastra & Boyd, 1996). The Artificial Neural Network (ANN) achieved an error

Objectives

Methods

Results

Conclusion