Solar Cycle 25 Prediction

Abstract

Solar activity, characterized by sunspot numbers and solar flares, operates on an approximately 11-year cycle. Predicting the amplitude and timing of these cycles is critical for protecting power grids, satellite communications, and orbital infrastructure. This project presents a novel approach using Bidirectional Long Short-Term Memory (LSTM) networks to forecast Solar Cycle 25.

Methodology

Traditional solar cycle predictions rely on dynamo models or simple statistical precursors. Our approach leverages the temporal memory of Recurrent Neural Networks (RNNs). By using a Bidirectional LSTM architecture, the model is able to learn dependencies from both historical past data and "future" context within the training window, providing a more robust understanding of the asymmetric nature of the solar cycle.

Data preprocessing involved normalizing smoothed sunspot numbers from 1749 to 2019. We implemented a sliding window approach with a 24-month horizon to predict the subsequent solar activity levels.

Results & Conclusion

Our findings suggest that Solar Cycle 25 will have a peak amplitude slightly higher than Cycle 24, occurring in early to mid-2025. This prediction aligns with contemporary research suggesting a strengthening of the solar dynamo after several decades of decline. The model achieved a high degree of accuracy on the test set (MAE < 5.0), demonstrating the potential for deep learning in celestial phenomena forecasting.