Projects per year
Abstract
Stream interaction region (SIR) is one of the space weather phenomena that accelerates the upstream particles of the interface region in interplanetary space and causes geomagnetic storms. SIRs are largescale structures that vary temporally and spatially, both in latitudinal and radial directions. Predicting the arrival times of interface regions (IRs) is crucial to protect our navigation and communication systems. In this work, a 1D ensemble system comprised of a Longshortterm memory (LSTM) model and a Convolution Neural Network (CNN) model—LCNN is introduced to classify the observed IR time series and give the prediction interval nowcast of its transit time to the observer. The outcomes of the two models are combined in a way to boost the accuracy of the predictor and prevent error propagation between them. The implemented technique is time series classification on datasets from STEREO A and B spacecrafts. The LCNN prediction system of IRs provides advanced Notice Time (NT) interval between [20, 160] minutes with sensitivity around 93% and geometric mean score gmean of 91.7%, and the skills decrease with increasing the prediction time. The LCNN demonstrates an enhancement in the prediction with respect to using only either the CNN or LSTM models. The predicted probabilities are recalibrated so that the predicted frequency of IRs becomes on average consistent with the observed frequency. Application of the method is useful to provide a classification of IRs by inputting a time series and estimating the likelihood of occurrence of an IR and its arrival time on the observer.
Key Points
Deep neural networks are used for time series classification and prediction
The adapted machine learning ensemble model is tailored for the Space Weather community, but applicable to many other communities
The adopted recalibration technique seems reliable for robust classification
Plain Language Summary
Stream interaction regions (SIRs) accelerate the upstream particles of the interaction/interface regions (IRs) in interplanetary space and cause geomagnetic storms. That raises the hazards on satellites' components and Earth's communication and navigation systems. Predicting the arrival time of IRs using machine learning regression techniques is challenging, particularly for such a complex problem of multidimensions time series of different time scales. In this work, a 1D ensemble system comprised of a Longshortterm memory (LSTM) model and a Convolution Neural Network (CNN) model—LCNN is introduced to classify the observed IR time series and give the prediction interval nowcast of its transit time to the observer. We used a nowcasting data set from STEREOA and B spacecrafts. The LCNN provides advance imminent IR at least 20 min and not longer than 160 min with a sensitivity of around 93% and the latter decreases with increasing the prediction time. The predicted probabilities are recalibrated so that the predicted frequency of IRs becomes on average consistent with the observed frequency. Application of the method is useful to provide a classification of IRs by inputting a time series and estimating the likelihood of occurrence of an IR and its arrival time on the observer.
Key Points
Deep neural networks are used for time series classification and prediction
The adapted machine learning ensemble model is tailored for the Space Weather community, but applicable to many other communities
The adopted recalibration technique seems reliable for robust classification
Plain Language Summary
Stream interaction regions (SIRs) accelerate the upstream particles of the interaction/interface regions (IRs) in interplanetary space and cause geomagnetic storms. That raises the hazards on satellites' components and Earth's communication and navigation systems. Predicting the arrival time of IRs using machine learning regression techniques is challenging, particularly for such a complex problem of multidimensions time series of different time scales. In this work, a 1D ensemble system comprised of a Longshortterm memory (LSTM) model and a Convolution Neural Network (CNN) model—LCNN is introduced to classify the observed IR time series and give the prediction interval nowcast of its transit time to the observer. We used a nowcasting data set from STEREOA and B spacecrafts. The LCNN provides advance imminent IR at least 20 min and not longer than 160 min with a sensitivity of around 93% and the latter decreases with increasing the prediction time. The predicted probabilities are recalibrated so that the predicted frequency of IRs becomes on average consistent with the observed frequency. Application of the method is useful to provide a classification of IRs by inputting a time series and estimating the likelihood of occurrence of an IR and its arrival time on the observer.
Original language  English 

Article number  e2022SW003326 
Number of pages  19 
Journal  Space Weather 
Volume  21 
Issue number  3 
DOIs  
Publication status  Published  16 Mar 2023 
Keywords
 Machine Learning in Heliophysics
 COMPUTATIONAL GEOPHYSICS
 Neural networks, fuzzy logic, machine learning
 EXPLORATION GEOPHYSICS
 Gravity methods
 GEODESY AND GRAVITY
 Transient deformation
 Tectonic deformation
 Time variable gravity
 Gravity anomalies and Earth structure
 Satellite geodesy: results
 Seismic cycle related deformations
 HYDROLOGY
 Estimation and forecasting
 Extreme events
 Time series analysis
 INFORMATICS
 Forecasting
 Machine learning
 Temporal analysis and representation
 INTERPLANETARY PHYSICS
 Solar wind plasma
 IONOSPHERE
 MAGNETOSPHERIC PHYSICS
 MATHEMATICAL GEOPHYSICS
 Prediction
 Probabilistic forecasting
 Persistence, memory, correlations, clustering
 Stochastic processes
 OCEANOGRAPHY: GENERAL
 Ocean predictability and prediction
 Time series experiments
 NATURAL HAZARDS
 Monitoring, forecasting, prediction
 NONLINEAR GEOPHYSICS
 Probability distributions, heavy and fat‐tailed
 Scaling: spatial and temporal
 POLICY SCIENCES
 RADIO SCIENCE
 Interferometry
 Ionospheric physics
 SEISMOLOGY
 Continental crust
 Earthquake dynamics
 Earthquake source observations
 Earthquake interaction, forecasting, and prediction
 Seismicity and tectonics
 Subduction zones
 SPACE PLASMA PHYSICS
 Stochastic phenomena
 SPACE WEATHER
 Policy
 Research Article
Fingerprint
Dive into the research topics of 'Prediction Interval of Interface Regions: Machine Learning Nowcasting Approach'. Together they form a unique fingerprint.Projects
 1 Finished

Solar System Physics at Aberystwyth University
Morgan, H., Cook, T., Gorman, M., Li, X., Pinter, B. & Taroyan, Y.
Science and Technology Facilities Council
01 Apr 2019 → 31 Dec 2022
Project: Externally funded research