The deliverable reports the implementation of the 1st version of the Bright energy prediction tool. Machine learning models are used to analyze the energy data streams and to predict the energy demand, generation, and flexibility. In the prediction process, two types of features have been used, ones that are derived from the monitored energy data like for example profile peaks, valleys, baseline, etc. and one related to the context in which the energy data was collected like for example weather data or time (i.e., season, day of the week, time of day, etc.). We have used hybrid models, based on an ensemble learner that aggregate partial prediction results from classifiers trained to predict the energy classes of the peaks, valleys, and baseline and regressors implemented using neural networks trained to fine-tune the prediction results. For the next steps, we plan to test the forecasting tool in the context of the Bright pilots to facilitate the implementation of Demand Response progress and management of local energy communities. Finally, we plan to consider new social-related features into the forecasting process which may potentially improve the prediction results.
