Speakers
Description
Atmospheric electric field forecasting represents a significant challenge in space weather monitoring applications due to the complex interaction among geomagnetic disturbances, atmospheric dynamics, and local meteorological variability. This study proposes a hybrid machine learning model based on Stacking Ensemble Learning for atmospheric electric field forecasting using geomagnetic and meteorological variables acquired through high-temporal-resolution environmental sensors.
The research integrates an electric field mill, a magnetometer, and a meteorological station for environmental data acquisition. Independent variables were selected according to their atmospheric-physical relevance, geomagnetic interactions, statistical dependence, and predictive contribution to atmospheric electric field variability. Meteorological variables considered include temperature, humidity, dew point, atmospheric pressure, wind speed, precipitation, solar radiation, and ultraviolet index, owing to their influence on atmospheric conductivity, ionization processes, cloud electrification, and charge transport mechanisms. In addition, geomagnetic field measurements were incorporated because of their relationship with ionospheric disturbances and space weather variability. The final feature set was further supported through correlation analysis, multicollinearity assessment, and feature importance evaluation within the machine learning framework.
Prior to predictive model development, an Exploratory Data Analysis (EDA) will be conducted to identify statistical distributions, temporal patterns, outliers, missing values, and nonlinear relationships among geomagnetic and meteorological variables associated with atmospheric electric field variability. The study is conceptually grounded in Faraday’s law of electromagnetic induction, which establishes the relationship between temporal variations in magnetic fields and the generation of induced electric fields.
The proposed methodology employs a stacking ensemble architecture composed of both regularized linear regression and nonlinear learning models, including Ridge Regression, Elastic Net, Random Forest, XGBoost, and LightGBM. These models are integrated through a meta-model designed to optimize predictive performance and reduce generalization error. Furthermore, temporal feature engineering techniques, including lag variables and rolling statistics, are incorporated to capture dynamic patterns associated with atmospheric and geomagnetic disturbances.
Predictive performance will be evaluated using Time Series Split cross-validation and regression metrics such as the coefficient of determination (R²), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Additionally, model interpretability will be assessed using feature importance analysis and SHAP values to identify the relative influence of geomagnetic and meteorological variables on atmospheric electric field behavior.
The expected result is to demonstrate that the integration of geomagnetic and meteorological variables through co-learning models significantly improves predictive accuracy compared to independent traditional approaches, contributing to the development of intelligent atmospheric monitoring and forecasting systems for space weather applications.