Probabilistic Weather Prediction with Bayesian Neural Networks

Abstract:

Ensemble forecasts provide valuable probabilistic ranges of likely weather scenarios. However, dynamic model ensembles are computationally expensive to produce and are typically under-dispersive and therefore generate unreliable forecast spread statistics. Dynamic ensembles thus require statistical post-processing to more accurately capture the correct probability distribution of the future state of the atmosphere. This study proposes a flexible alternative to ensemble generation via Bayesian Neural Networks (BNN) for probabilistic distributions developed from post-processing deterministic model forecasts. Using North American West Coast integrated vapor transport (IVT) as a case study variable and a newly released 34-year deterministic reforecast (West-WRF) for model training, the BNN produces sharp and reliable probabilistic distributions for IVT between 0-72 hours. The BNN is shown to outperform and be more computationally efficient than 1) more traditional pattern matching statistical ensemble methods (The Analog Ensemble) and 2) a dynamic model ensemble (the Global Ensemble Forecasting System). Further, this study informs future reforecast development by testing the length of training data required to develop an accurate statistical ensemble. This study demonstrates the value of deep learning for increased computational efficiency by questioning the need for dynamic ensembles on short time scales.