Fabio Chiappina, a.i. solutions; Katie Kosak, a.i. solutions; John Smigo III, a.i. solutions
Keywords: space weather, atmospherics, solar cycle, solar activity, solar, cycle, solar proxy, proxy, proxies, machine learning, online learning, ML, AI, AI/ML, LSTM, RNN, MLP, neural, network, neural network, Schatten, F10.7, Ap, Kp
Abstract:
Solar activity strongly influences the Earth’s atmosphere. Solar storms and associated phenomena, including unexpected solar flares and coronal mass ejections, can interact with the Earth’s magnetic field and atmosphere, increasing the atmospheric density and therefore the drag experienced by spacecraft for several days. If space mission operators are using density models that are not updated frequently to account for these short-term, high-intensity, nearly unpredictable events, the resulting modeling errors can ultimately cause higher fuel expenditure and increase collision risk. Accurately forecasting future solar activity is thus critical for safe and efficient space operations.
Several observable metrics, known collectively as solar proxies, vary according to the 11- year solar cycle and grant insight into the transient and long-term behavior of the Sun. Among the most valuable of these solar proxies are the 10.7cm solar radio flux, or F10.7, Ap geomagnetic index, and sunspot number, all of which are commonly forecasted by existing solar proxy models, and some of which are subsequently used as inputs to atmospheric density models for orbit propagation. But existing solar proxy forecasting models face limitations. The Schatten model, considered industry standard for forecasting F10.7, adequately captures the long-term behavior of the 11-year solar cycle, but is not designed to model day-to-day variations, which are critical for applications that require a precise picture of atmospheric conditions within the next few days. The Schatten model is also not programmatically data-driven; instead, it requires manual tuning to fit historical data, limiting the learnable patterns to those identifiable by the subject matter experts tuning the model. Recent advancements in machine learning have facilitated the development of purely data-driven models trained on historical solar proxy data; yet these models fail to capitalize on any a-priori understanding of the underlying physical principles governing solar activity, as captured by models like Schatten. Furthermore, most existing solar proxy models consider only a single solar proxy at a time, failing to leverage relationships between proxies to improve their predictive accuracy. By synthesizing data across multiple solar proxies and combining modern machine learning techniques for timeseries forecasting with a physical understanding of solar activity patterns, the accuracy of solar proxy forecasts – and, therefore, downstream applications like atmospheric density models – can be improved.
In this investigation, a wide variety of novel neural network architectures for solar proxy forecasting are evaluated. The core of each model architecture is either a multi-layer perceptron (MLP) or a long short-term memory (LSTM) network, with the optional inclusion of dropout and/or regularization layers. The impacts of model depth, layer sizes, and other configuration parameters on the predictive accuracy of the models are explored. The accuracy of a set of models each forecasting on a single solar proxy is compared against that of a single model forecasting on multiple proxies simultaneously. Hybrid forecasting models, which make predictions by synthesizing past solar proxy data with Schatten F10.7 forecasts, are also analyzed, and their forecasting accuracy is compared to that of their purely data-driven counterparts. Finally, the best performing models are compared against the forecasting accuracy of the Schatten F10.7 model, and performance improvements, particularly in short-term forecasts, are highlighted.
Preliminary analysis suggests that our approach can significantly improve the accuracy of solar proxy forecasts in the near-term ( < 7 day forecast horizon). Using a March 2016 forecast that was produced using a Schatten model immediately after manual tuning by a subject matter expert, the next-day F10.7 predictions generated with our approach demonstrated an accuracy improvement of 10% (from 88% to 98% accuracy) over the classical Schatten predictions. We propose an autonomous daily pipeline that collects new daily solar proxy data points, updates the model, and publishes online highly accurate short-term predictions for F10.7 that can be fed directly into high-precision orbit propagation tools such as FreeFlyer and STK. This innovation will allow satellite operators to incorporate the latest solar activity into their mission planning, leading to more efficient and safer operations in space. Date of Conference: September 16-19, 2025
Track: Atmospherics/Space Weather