Adapting a deep convolutional RNN model with imbalanced regression loss for improved spatio-temporal forecasting of extreme wind speed events in the short to medium range

Abstract. The number of wind farms and amount power production in Europe, both on- offshore, have increased rapidly the past years. To ensure grid stability on-time (re)scheduling maintenance tasks to mitigate fees energy trading, accurate predictions speed are needed. Particularly, extreme events high importance farm operators as timely knowledge these can prevent damages offer economic preparedness. This work explores possibility adapting a deep convolutional recurrent neural network (RNN)-based regression model spatio-temporal prediction short medium range (12 h lead time 1 intervals) through manipulation loss function. this end, multi-layered long short-term memory (ConvLSTM) is adapted with variety imbalanced functions that been proposed literature: inversely weighted, linearly weighted squared error-relevance area (SERA) loss. Forecast performance investigated for various intensity thresholds events, comparison made commonly used mean error (MSE) absolute (MAE) results indicate inverse weighting method most effectively shift forecast distribution towards tail, thereby increasing forecasted ranges, considerably boosting hit rate reducing root-mean-squared (RMSE) those ranges. also show, however, such improvements invariably accompanied by pay-off terms overcasting false alarm ratio, which increase threshold. balances trade-off, MAE scoring slightly better than MSE It concluded provides an effective way adapt learning task its application forecasting range.