Forecasting the energy demand of individual actors based on time series is characterized by a huge amount of data due to the large number of energy consumers. Currently, every commercial customer with a consumption > 100 MWh per year is subjected to a recording power metering (RLM) and an individual consumption forecast is created for each customer. In order to be able to map this complexity with mathematical models, flexible data-driven solutions are necessary. On the one hand, this extensive data situation poses challenges to the handling of the analysis, but on the other hand, it also enables data-driven modeling of complex behavior patterns. However, a core problem then lies in the transfer of this model to the individual forecast for each actor.
The goal of AGENS is to develop flexible neural network (NN) based models that are capable of modeling overall complexity using large amounts of data. To enable robust prediction per actor, an improvement of the data quality for each individual consumer is necessary. For this purpose, so-called Generative Adversarial Neural Networks (GAN) are developed as the core of the project, which enable data augmentation. For a successful training of the GAN, statistical pre-analyses of the data have to be performed in order to determine their characteristic patterns and to feed these into the GAN. Based on the extended data set, the goal is to subsequently be able to train robust models with calibrated uncertainties and to ensure their applicability in an industrial setting. The subproject "Dynamic Neural Networks" of the University of Bremen investigates on the theoretical side the mathematical-theoretical relationships between time-recurrent neural networks (RNNs), which are based on discrete data (time series), and the continuous concepts of Neural Ordinary Differential Equations (NODE). The aim is to estimate the approximation quality taking into account model uncertainties.
These models will be tested in cooperation with industrial partners for the prediction of electricity demand. The main subject of this subproject is the development and analysis of dynamic neural networks with a focus on energy forecasting.
