Improving probabilistic wind speed forecasts with deep learning

This project aims at developing post-processing models for deriving probabilistic weather forecasts from NWP forecast data using deep learning techniques.

The first part of the project involves improving probabilistic wind speed forecasts for consecutive lead times using an autoregressive model. Post-processing multiple lead times simultaneously is challenging because of the inherent temporal dependencies. Classical approaches often involve processing lead times individually and subsequently employing empirical copula methods to handle such dependencies. Building on previous work, we instead consider the ARMOS model which incorporates temporal dependencies through the autoregressive property of forecast errors and can be used to obtain an explicit multivariate probability distribution for the weather variable in question. As such, it is a generalization of the widely used Ensemble Model Output Statistics (EMOS) used for estimating marginal distributions.

For the purpose of this project, the model is applied to deterministic forecasts from the Harmonie-Arome model of KNMI, yielding a multivariate parametric forecast distribution for hourly wind speeds up to 48 hours ahead. We model the marginal conditional distributions as truncated normal distributions, and the model parameters are estimated both linearly and as the output of a neural network with convolutional and optional LSTM layers which can detect spatial patterns and temporal dependencies, respectively. We compare the resulting models to a variant of EMOS adapted to deterministic forecasts that is paired with a copula method.

The ARMOS model has so far shown good performance in modeling temporal dependencies explicitly without the need to use a copula method. Moreover, the network models outperform the classical approach of estimating the distribution parameters linearly. We will provide an update on these results as well as an outlook on planned future work.