[1909.11865v1] Probabilistic Forecasting using Deep Generative Models
The results show that despite the better performance of the AnEn method, the Conditional Variational Autoencoder used in this work produces reliable probabilistic forecasts using a fraction of the time and memory needed by the AnEn

Abstract: The Analog Ensemble (AnEn) method tries to estimate the probability
distribution of the future state of the atmosphere with a set of past
observations that correspond to the best analogs of a deterministic Numerical
Weather Prediction (NWP). This model post-processing method has been
successfully used to improve the forecast accuracy for several weather-related
applications including air quality, and short-term wind and solar power
forecasting, to name a few. In order to provide a meaningful probabilistic
forecast, the AnEn method requires storing a historical set of past predictions
and observations in memory for a period of at least several months and spanning
the seasons relevant for the prediction of interest. Although the memory and
computing costs of the AnEn method are less expensive than using a brute-force
dynamical ensemble approach, for a large number of stations and large datasets,
the amount of memory required for AnEn can easily become prohibitive.
Furthermore, in order to find the best analogs associated with a certain
prediction produced by a NWP model, the current approach requires searching
over the entire dataset by applying a certain metric. This approach requires
applying the metric over the entire historical dataset, which may take a
substantial amount of time. In this work, we investigate an alternative way to
implement the AnEn method using deep generative models. By doing so, a
generative model can entirely or partially replace the dataset of pairs of
predictions and observations, reducing the amount of memory required to produce
the probabilistic forecast by several orders of magnitude. Furthermore, the
generative model can generate a meaningful set of analogs associated with a
certain forecast in constant time without performing any search, saving a
considerable amount of time even in the presence of huge historical datasets.

Figure 1: CVAE general architecture training (KL-Vanishing Problem and Cyclic Annealing Scheduling)Figure 2: Cyclic Annealing with Beta > 1 (KL-Vanishing Problem and Cyclic Annealing Scheduling)Figure 3: CVAE decoder architecture inference (CVAE Inference) (Dispersion and Rank Histogram)Figure 4: Ensemble members Mean Squared Error (MSE) (solid line) with bootstraping (red range) and mean ensemble variance (dashed line) for a) CVAE, b) AnEn (Dispersion and Rank Histogram) (Dispersion and Rank Histogram)Figure 5: 21 ensemble members Rank Histogram for a) CVAE, b) AnEn (Dispersion and Rank Histogram)Figure 6: CRPS results for CVAE 21 ensemble members (Probability Score)Figure 7: Memory usage comparisons between CVAE and AnEn models (Discussion)Figure 8: RH for wind direction (Wd), surface pressure (P), and 2-m temperature (T) produced by AnEn (left column) and CVAE (right column) models (Discussion)Figure 9: Runtime comparisons between CVAE and AnEn models (Discussion)›