2B.4 Using Recurrent Neural Networks (RNNs) to Bias Correction of Wind speed forecasting

Monday, 13 January 2020: 2:45 PM
257AB (Boston Convention and Exhibition Center)
Bonyang Ku, KMA, Seoul, Korea, Republic of (South); and M. K. Kim, S. Y. Park, and Y. H. Lee

Numerical weather prediction (NWP) model plays an important role in forecasting the weather, but it still has systematic errors or bias. Various model post-processes using statistics or machine learning are essential for more accurate weather forecast. The Korea Meteorological Administration (KMA) has supported forecasts using the Model Output Statistics (MOS) based on statistics (e.g. multi linear regression) over 10 years. With the rapid development of artificial intelligence and computing technology, deep learning is increasingly being used for accurate weather forecasts as well. Especially, artificial neural network technique that can reflect nonlinear relationship is effective for predicting complex atmosphere. In this study, the artificial neural network technique is applied to the post-processing of the numerical weather prediction (NWP) model and used to correct the bias of the wind speed forecast. We have undertaken it to replace the forecast guidance using the MOS with that based on artificial neural network which can be updated in a relatively short period of time.

We have developed models using Long Short Term Memory networks (LSTMs) explicitly designed to avoid the long-term dependency problem of Recurrent Neural Networks (RNNs) which are commonly used to process sequential data. Initially only surface wind speed is used as a predictor, then several surface variables such as temperature, humidity, stability are added to predictors. The period of training data is from 2016 to 2018 and the overall data are divided into a training period and a validation period to establish the bias correction model. For hyperparameter optimization, we use random search method that is a technique to find best combinations of the hyperparameters for the model. The performance of the developed model is evaluated at several representative sites (land, coastal and mountain) for the test period. Some attempts to improve the performance of models and results will be given at the presentation.

- Indicates paper has been withdrawn from meeting
- Indicates an Award Winner