Application of Recurrent Neural Networks to Model Bias Correction: Idealized Experiments With the Lorenz-96 Model

Systematic biases in numerical weather prediction models cause forecast deviation from reality. While model biases also affect data assimilation and degrade the analysis accuracy, observation information incorporated through data assimilation can provide information for detecting and alleviating such biases. In this study, the application of machine learning to model bias correction is demonstrated, emphasizing the effectiveness of recurrent neural networks. Idealized experiments are performed using the two-scale coupled Lorenz-96 model as the true system and single Lorenz-96 model as the imperfect forecast model, to compare the effectiveness of bias correction methods based on various architectures of neural networks and simple linear regression. The neural networks generally outperformed linear regression, and recurrent neural networks showed the best ability in finding the systematic bias component from the analysis increment data. Bias correction using the recurrent neural networks also gives the most significant improvement in reducing the error growth rate in extended range forecasts. The results suggest that including past time series of the forecast variables improve model bias correction when limited information of the observation is incorporated through data assimilation.

Automated summary

This study demonstrates the application of machine learning, particularly recurrent neural networks, in model bias correction for weather prediction. Idealized experiments using different architectures of neural networks and simple linear regression are performed to compare their effectiveness. The results show that neural networks generally outperform linear regression, and recurrent neural networks perform the best in detecting and reducing systematic biases in weather models.