Scenario Based Probabilistic Energy Demand Forecasting using Autoencoders and Gaussian Mixture Models

Short-term energy demand scenarios have a significant impact on the operation cost and power system reliability due to the stochastic nature of energy consumers' behaviour. In order to obtain the scenarios containing the information of forecast errors and fluctuations of short-term energy demand forecasting, a scenario generation method based on autoencoders and Gaussian mixture models is proposed. Within this paper, the use of autoencoders in order to achieve the projection of daily demand profiles in a latent characteristics space is described. The first step is the definition of the latent space using an autoencoder and the second step is a Gaussian mixture model to be fitted on the latent space. The proposed scenario generation method is based on sampling from the estimated mixture model on the latent space. After the evaluation of the proposed methodology, using energy demand data of Greece on a regional level, compared to Quantile Random Forests and Quantile Regression Neural Networks, the results conclude that the proposed methodology outperforms both these methodologies.