BoxCox multi-output linear regression for 10.7 cm solar radio flux prediction

Abstract

Abstract We consider the problem of predicting the mid-term daily 10.7cm solar radio flux (F10.7), a widely-used solar activity index. A novel approach is proposed for this task, in which BoxCox transformation with a proper parameter is first applied to make the data satisfy the property of homoscedasticity that is a basic assumption of regression models, and then a multi-output linear regression model is used to predict future F10.7 values. The experiment shows that the BoxCox transformation significantly improves the predictive performance and our new approach works substantially better than the prediction from the US Airforce and other alternative methods like Auto-regressiveModel, Multi-layer Perceptron, and Support Vector Regression.