Forecasting Occurrence and Intensity of Geomagnetic Activity With Pattern‐Matching Approaches. Issue 6 (24th June 2021)

• Record Type: Journal Article
• Authors: Haines, C.; Owens, M. J.; Barnard, L.; Lockwood, M.; Ruffenach, A.; Boykin, K.; McGranaghan, R.
• Is Part Of: Space weather. Volume 19:Issue 6(2021)
• Page Start: n/a
• Page End: n/a
• Publication Date: 2021-06-24
• Subjects: forecast -- geomagnetic activity -- machine learning -- support vector machine -- analogue ensemble; Space environment -- Periodicals; 551.509992
• DOI: 10.1029/2020SW002624
• Languages: English
• ISSNs: 1542-7390

Abstract

Abstract: Variability in near‐Earth solar wind conditions gives rise to space weather, which can have adverse effects on space‐ and ground‐based technologies. Enhanced and sustained solar wind coupling with the Earth's magnetosphere can lead to a geomagnetic storm. The resulting effects can interfere with power transmission grids, potentially affecting today's technology‐centered society to great cost. It is therefore important to forecast the intensity and duration of geomagnetic storms to improve decision making capabilities of infrastructure operators. The 150 years aa H geomagnetic index gives a substantial history of observations from which empirical predictive schemes can be built. Here we investigate the forecasting of geomagnetic activity with two pattern‐matching forecast techniques, using the long aa H record. The techniques we investigate are an Analogue Ensemble (AnEn) Forecast, and a Support Vector Machine (SVM). AnEn produces a probabilistic forecast by explicitly identifying analogs for recent conditions in the historical data. The SVM produces a deterministic forecast through dependencies identified by an interpretable machine learning approach. As a third comparative forecast, we use the 27 days recurrence model, based on the synodic solar rotation period. The methods are analyzed using several forecast metrics and compared. All forecasts outperform climatology on the considered metrics and AnEn and SVM outperform 27 days recurrence. A Cost/Loss analysis reveals the potential economic value is maximized using the AnEn, but the SVM is shown as superior by the true skill score. It is likely that the best method for a user will depend on their need for probabilistic information and tolerance of false alarms. Plain Language Summary: Space weather has the potential to disrupt society and the economy on a large scale. One such major impact is on power grids, which can be damaged by disturbances in Earth's magnetic field caused by space weather events. As a result, it would be useful to have an accurate forecast of space weather that can help power grid operators make decisions about taking mitigating action. In this work, we test three forecasting techniques which utilize long historical records to exploit patterns in the data and hence predict future disturbances in Earth's magnetic field. We find that all three of the techniques provide valuable information and the best method depends on the individual needs of the forecast user. Key Points: Pattern‐matching techniques are an effective way to forecast geomagnetic activity The analogue ensemble and support vector machine outperform 27 days recurrence and climatology The best forecast approach for the end user will depend on their need for probabilistic forecast information