Hybrid Deep Neural Network Model for Multi-Step Energy Prediction of Prosumers

Abstract

The transition toward decentralized renewable energy transforms the energy grid, with the prosumers playing an active role in the local energy management. Accurate day-ahead prediction of their energy demand is a prerequisite to ensure the stability and efficiency of the power grid by balancing the energy demand with the production, while incorporating renewable energy as much as possible. In this paper, we propose a 24-steps-ahead energy prediction model that integrates clustering and multilayer perceptron classification models used to detect the classes of energy profiles and multilayer perceptron regression models used to fine-tune the energy prediction, considering the energy data streams. We introduce new features derived from the raw energy data collected from prosumers, such as the profile peaks and valleys, concerning the energy baseline and describe a software infrastructure for integrating the real-time energy data streams with the hybrid deep learning models training and prediction. The evaluation tests consider energy datasets that are closer to the real-time energy data streams from prosumers. The results show that, even on energy data streams, the model offers a good prediction accuracy for small- and medium-scale prosumers.