Previous Abstract Return to Session B3 Next Abstract

Session B3: Atmospheric Effects

Application of Machine Learning to Ionospheric Scintillation Forecast
Yunxiang Liu, Zhe Yang, Y. Jade Morton, University of Colorado Boulder
Location: Pavilion Ballroom West
Alternate Number 3

Objective

The objective of this paper is to propose a method to forecast the scintillation by using a machine learning algorithm.

Introduction

Ionospheric scintillation refers to rapid and random amplitude and phase fluctuations of the signals propagating through ionosphere plasma irregularities. Strong scintillation could severely impact the acquisition and tracking of the GNSS receivers, resulting in performance degradation in accuracy, continuity, and integrity. It will be desired to prepare for impact in advance, especially for safety critical applications like aviation. Therefore, it is very important to have the capability of scintillation forecast. In [1], a machine learning algorithm named neural network was proposed to forecast the level of scintillation with lead time of hours. Furthermore, the authors in [2] proposed to apply another machine learning algorithm named support vector machine (SVM) to predict the scintillation in a 1-hr and 3-hr lead time. However, both methods mentioned above neglected the fact that scintillation forecast is essentially a time series prediction problem and failed to incorporate the temporal information. In forecast tasks, the temporal information is very important. Therefore, we propose to use a recurrent neural network (RNN) to forecast the scintillation, where the temporal information can be captured.

Methodology

In this paper, we propose to use an RNN to forecast the scintillation. In this method, the output label is the scintillation level (defined based on average rate of TEC index (ROTI) [3]) in the near future and the input features are a time series of measurements from the past and the present. Here, the measurements include scintillation level, satellite-wise ROTI, time, geomagnetic indices, etc. In short, given the time series of observed measurements from the past a few hours as the input features, the proposed method aims to predict what the scintillation level will be several hours later.

To implement this, an RNN architecture named long short-term memory (LSTM) is applied [4]. This architecture is capable of capturing the temporal dynamic behavior of the scintillation by feeding a time series of measurements. As a result, a better scintillation forecast performance can be expected. Different choices of measurements as input features will be tested, and the importance of each measurement will be investigated and discussed.

Anticipated Results

To evaluate the performance, a dataset containing two years of measurements from a high latitude international GNSS service (IGS) station will be obtained and preprocessed. The proposed method will be compared with other forecast models, including conventional time series models and convectional machine learning algorithms. The performance comparison will be discussed. A list of compared models is shown below:
• Autoregression: A time series model that uses observations from previous time steps as input to a regression equation to predict the value at the next time step.
• Vector-Autoregression: A stochastic process model used to capture the linear interdependencies among multiple time series.
• SVM [2]: One of the best shallow machine learning algorithms that deals with prediction task.
• Full-connected Neural Network [1]: A class of artificial neural networks that deals with prediction tasks.
• Recurrent neural network: A class of artificial neural networks that deals with time series data.

Reference

[1] de Lima G R T, Stephany S, de Paula E R, et al. Prediction of the level of ionospheric scintillation at equatorial latitudes in Brazil using a neural network[J]. Space Weather, 2015, 13(8): 446-457.
[2] McGranaghan R M, Mannucci A J, Wilson B, et al. New capabilities for prediction of high-latitude ionospheric scintillation: A novel approach with machine learning[J]. Space Weather, 2018, 16(11): 1817-1846.
[3] Pi X, Mannucci A J, Lindqwister U J, et al. Monitoring of global ionospheric irregularities using the worldwide GPS network[J]. Geophysical Research Letters, 1997, 24(18): 2283-2286.
[4] Hochreiter S, Schmidhuber J. Long short-term memory[J]. Neural computation, 1997, 9(8): 1735-1780.



Previous Abstract Return to Session B3 Next Abstract