Machine‐learning‐based estimation of electrical activity in the weather research and forecasting model

Abstract

Machine-learning-based parameterizations of cloud electrical activity were developed using the Weather Research and Forecasting model-electric (WRF-ELEC) variable LIGHT as the training target. Four predictive models were built using output variables from WRF simulations as predictors: Linear Regression (LR), Lightning Potential Index (LPI), Random Forest (RF), and Neural Networks (NN). Among them, the RF model achieved the best performance, significantly outperforming both the LR and the LPI-based approaches. Although the NN model showed comparable accuracy to RF during training and validation, its performance declined substantially when applied to an unbalanced dataset, highlighting its sensitivity to data distribution and normalization. In contrast, the RF model demonstrated robust generalization under such conditions, making it a promising alternative to the full WRF-ELEC implementation. Notably, all models – particularly RF – tended to predict contiguous regions of lightning activity rather than isolated columns. This behavior suggests that around a column with LIGHT > 0, the values of the input variables in the neighbor columns could be similar. The lightning-related variable LIGHTDENS, which better represents the spatial extent of electrical activity, could be a good alternative of a more suitable target for future model development.