Probabilistic forecasts with Bayesian Convolutional Long-Short Term Memory Networks: an idealized Lorenz 84 model and real world Arctic application

Abstract:

Deep neural networks are demonstrated to be useful in many fields, including weather forecasting. However, most of these techniques are deterministic and therefore they cannot represent uncertainties that are intrinsic to weather and climate science. To overcome this, we propose a new network structure belonging to Bayesian deep learning (BDL), namely the Bayesian Convolutional Long-Short Term Memory Networks (BayesConvLSTM), to address temporal-spatial sequences and perform ensemble weather forecast. We study the characteristics of BDL in ensemble weather forecasting by using a variant of BayesConvLSTM without spatial structure. We take output from the Lorenz 84 system with seasonal forcing as a model of the ‘truth’ that we intend to forecast. Fundamental concepts of data assimilation are reflected upon in our analysis and we demonstrate that the BDL is able to address uncertainties in the initial conditions and model parameters. We further explore the viable strategy for generating ensemble forecasts with BayesLSTM and we demonstrate that the forecasts with BayesLSTM can stay close to the attractor of the Lorenz system. Based on this positive result, we apply the BayesConvLSTM to a real world problem. We forecast the Arctic sea ice in the Barents sea at weekly to sub-seasonal time scales and evaluate our results against the ensemble forecasts within the sub-seasonal to seasonal prediction project (S2S) and reanalysis. We find that the BayesConvLSTM agree well with the chosen reanalysis dataset (ERA-Interim). Our study shows that BayesConvLSTM are promising to act as fast ensemble forecast tools and they can potentially enhance weather forecasting capabilities of large forecast systems.