Continuous and Distribution-Free Probabilistic Wind Power Forecasting: A Conditional Normalizing Flow Approach

We present a data-driven approach for probabilistic wind power forecasting based on conditional normalizing flow (CNF). In contrast with the existing, this is distribution-free (as non-parametric and quantile-based approaches) can directly yield continuous probability densities, hence avoiding quantile crossing. It relies base distribution set of bijective mappings. Both shape parameters mappings are approximated neural networks. Spline-based considered owing to its non-affine characteristics. Over training phase, model sequentially maps input examples onto samples distribution, given contexts, where estimated through maximum likelihood. To issue forecasts, one eventually into desired distribution. Case studies open datasets validate effectiveness proposed model, allows us discuss advantages caveats respect state art.