Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Enhancing Solar Driver Forecasting with Multivariate Transformers

Sergio Sanchez-Hurtado, Victor Rodriguez-Fernandez, Julia Briden, Peng Mun Siew, Richard Linares

TL;DR

The paper addresses forecasting solar driver indices $F_{10.7}$, $S_{10.7}$, $M_{10.7}$, and $Y_{10.7}$ to improve space weather predictions. It proposes a multivariate time-series Transformer (PatchTST) that uses an $L=18$ day lookback to predict a $T=6$ day horizon, employing patching, channel independence, and a distribution-distance weighted loss to balance high- and low-activity periods. The key contributions include showing improved forecast accuracy over the SET benchmark, particularly during high solar activity, and providing open-source code for the community. The approach has practical significance for space weather forecasting and JB2008-based thermospheric models, enabling more reliable risk mitigation for satellites and communications.

Abstract

In this work, we develop a comprehensive framework for F10.7, S10.7, M10.7, and Y10.7 solar driver forecasting with a time series Transformer (PatchTST). To ensure an equal representation of high and low levels of solar activity, we construct a custom loss function to weight samples based on the distance between the solar driver's historical distribution and the training set. The solar driver forecasting framework includes an 18-day lookback window and forecasts 6 days into the future. When benchmarked against the Space Environment Technologies (SET) dataset, our model consistently produces forecasts with a lower standard mean error in nearly all cases, with improved prediction accuracy during periods of high solar activity. All the code is available on Github https://github.com/ARCLab-MIT/sw-driver-forecaster.

Enhancing Solar Driver Forecasting with Multivariate Transformers

TL;DR

The paper addresses forecasting solar driver indices , , , and to improve space weather predictions. It proposes a multivariate time-series Transformer (PatchTST) that uses an day lookback to predict a day horizon, employing patching, channel independence, and a distribution-distance weighted loss to balance high- and low-activity periods. The key contributions include showing improved forecast accuracy over the SET benchmark, particularly during high solar activity, and providing open-source code for the community. The approach has practical significance for space weather forecasting and JB2008-based thermospheric models, enabling more reliable risk mitigation for satellites and communications.

Abstract

In this work, we develop a comprehensive framework for F10.7, S10.7, M10.7, and Y10.7 solar driver forecasting with a time series Transformer (PatchTST). To ensure an equal representation of high and low levels of solar activity, we construct a custom loss function to weight samples based on the distance between the solar driver's historical distribution and the training set. The solar driver forecasting framework includes an 18-day lookback window and forecasts 6 days into the future. When benchmarked against the Space Environment Technologies (SET) dataset, our model consistently produces forecasts with a lower standard mean error in nearly all cases, with improved prediction accuracy during periods of high solar activity. All the code is available on Github https://github.com/ARCLab-MIT/sw-driver-forecaster.
Paper Structure (7 sections, 1 equation, 3 figures)

This paper contains 7 sections, 1 equation, 3 figures.

Table of Contents

  1. Introduction
  2. Current literature comparison
  3. Multivariate models
  4. Results
  5. Data Splitting and Loss Function
  6. Comparison with SET benchmark
  7. Discussion

Figures (3)

  • Figure 1: Distribution of F10.7, S10.7, M10.7, and Y10.7 data splits for training, validation and test sets, with solar activity thresholds shown in the background for each index.
  • Figure 2: Comparison of solar activity for F10.7. Note: The weights ($w_s$ for $s \in S = \{\text{low, moderate, elevated, high}\}$) represent the differences between historical ($H_s$) and training ($T_s$) distributions, normalized to sum 1, defined as $w_s = |H_s - T_s| \cdot \frac{1}{\sum_{s \in S} |H_s - T_s|}$.
  • Figure 3: Comparison of SET benchmark against PatchTST ensemble with wMSE and wMAE losses. Note: The data is categorized with the SET dataset for better accuracy, specifically using different values for S10.7. However, we could not use the SET dataset for training as it only includes values between 2012 and 2018, which is insufficient for our training process.