A hybrid neural network model, comprising of a convolutional neural network and a multilayer perceptron network, has been developed for day-ahead forecasting of regional scale wind power production. This model requires operational weather forecasts as input and also has the capability to ingest data from ensemble forecasts. Even though the training of the model requires significant computational cost, the actual forecasting can be done within a few minutes on any recent personal computer. The proposed model has demonstrated noteworthy performance at a recent international forecasting competition.

A series of probabilistic models were bench-marked during the European Energy Markets forecasting Competition 2020 to assess their relative accuracy in predicting aggregated Swedish wind power generation using as input historic weather forecasts from a numerical weather prediction model. In this paper, we report the results of one of these models which uses a deep learning approach integrating two architectures: (a) Convolutional Neural Network (CNN) LeNet-5 based architectrure; (b) Multi-Layer Perceptron (MLP) architecture -with two hidden layers-. These are concatenated into the Smooth Pinball Neural Network (SPNN) framework for quantile regression. Hyperparameters were optimised to produce the best model for every region. When tuned, the re-forecasts from the model performed favorably compared to other machine learning approaches and showed significant improvement on the original competition results, though failed to fully capture spatial patterns in certain cases when compared to other methods.