Unified Attention Recurrent Neural Network for Bias Correction of MJO Prediction

Abstract

In global subseasonal forecasting using dynamical models, correcting the systematic biases of Madden–Julian Oscillation (MJO) predictions has proven critical, particularly due to issues of rapid amplitude damping and phase distortion. To address these biases, recent studies have demonstrated that deep learning offers a promising solution by learning mappings that directly translate biased dynamical model outputs to ground truth MJO indices, effectively serving as post‐processing bias correctors. These approaches implicitly assume that a single neural network can extract all relevant features necessary for bias correction across different forecast time steps. However, we observe that the relationships between forecast lead times and their corresponding corrections are more intricate: features relevant to short range forecasting error differ significantly from those that govern long range forecasting errors. Consequently, direct mapping strategies often result in suboptimal performance, excelling in either short or long range forecasts, but rarely both. To address this challenge, we propose a unified attention recurrent neural network framework that processes the full sequence of dynamical forecasts using a masked input tensor, enabling the model to extract temporally contextualized features across all lead times. Additionally, we introduce a phase regulated loss function that explicitly penalizes errors in both amplitude and phase, capturing the cyclical structure of MJO more effectively than traditional mean squared error based objectives. Comprehensive evaluations on operational subseasonal to seasonal reforecast data sets demonstrate that our method significantly improves forecast accuracy and lead‐time consistency.