Speaker
Description
Abstract
Prediction of total solar irradiance has experienced significant growth and has become an important field of study in solar physics, space climate, and space weather that employs machine learning (ML) as a tool to overcome the challenges of the nonlinear and dynamic nature of solar activities and intricate spatiotemporal relationships in solar irradiance forecasting.
Important factors influencing the prediction of total solar irradiance include continuum image, magnetogram image, humidity, temperature, time of day and date, cloudiness index, latitude, longitude, and topography. Its accurate prediction is critical for understanding solar variability and long-term climate trends, and contributes to the monitoring of solar activity pertinent to both space- and ground-based technologies.
Due to the complexity of the physical modeling approach, the use of machine learning for modeling and forecasting solar irradiance has increased. These improvements, combined with the availability of ground—and space-based data from solar photosphere image databases, have significantly enhanced space weather applications.
This study proposes a novel multistage model that integrates Self-Organizing Map, AutoEncoder, and Long Short-Term Memory (SOM-AE-LSTM) to enhance total solar irradiance prediction. This article focuses specifically on forecasting total solar irradiance from hours to days based on space weather observations from the Solar Dynamics Observatory (SDO). The space-based data considered encompasses the data set from the phase before and after solar cycle 24 maximum, spanning a year for the training set and 3 months for the test to understand the physical mechanism of these two distinct phases.
This article employs a self-organizing map for the classification of features of solar photosphere images, specifically continuum and magnetogram images that contain sunspot and active region information. This model generates the feature map that serves as input to the autoencoder model. The encoder part is used to extract the compressed features, and the decoding part is used for the reconstruction of the images classified by SOM. Then, the output of the encoder is employed as input for the long short-term model to perform the irradiance prediction. The model considered for this article is trained and evaluated using the mean squared error (MSE), the correlation coefficient (R), the coefficient of determination ($R^2$), and the mean absolute percentage error (MAPE) as performance metrics.
The study demonstrates good results in capturing both spatial and temporal dependencies associated with solar activity. The results, including forecast accuracy and comparative analysis with baseline models, will be presented at the conference.
Keywords: Solar irradiance / TSI / SSI / Som-Autoencoder-LSTM
