Unified Attention Recurrent Neural Network for Bias Correction of MJO Prediction
Yiyi Guo
This project addresses the critical issue of systematic biases, specifically rapid amplitude damping and phase distortion, in global subseasonal Madden-Julian Oscillation (MJO) predictions from dynamical models. Recognizing that deep learning offers a promising solution for bias correction but that existing direct mapping strategies are suboptimal across different forecast lead times, we developed a unified attention recurrent neural network (UAR) framework. Our methodology involves processing full sequences of dynamical forecasts using a masked input tensor to extract temporally contextualized features and introduces a novel phase-aware loss function that explicitly penalizes errors in both amplitude and phase. Comprehensive evaluations on operational S2S forecast datasets (JMA, CNRM, BoM) demonstrate that our method significantly improves MJO forecast accuracy and lead-time consistency. This work is primarily of interest to researchers and practitioners in subseasonal-to-seasonal (S2S) forecasting, climate modeling, and deep learning applications in atmospheric science. Keywords include: MJO, Subseasonal Forecasting, Bias Correction, Deep Learning, Recurrent Neural Network, Attention Mechanism
Please use this thread for discussion about this talk.