Machine Learning for Time Series Prediction of Energy Data

Abstract - Zusammenfassung

The rise of renewable energy sources threatens the power grid’s stability. This can be remedied by discharging batteries in electric vehicles back into the grid at the right moment to compensate for peak loads. The ReNuBiL project (Reallabor Nutzerorientiertes Bidirektionales Laden, English: real-world laboratory for user-oriented bidirectional charging) explores bidirectional charging of electric vehicles and the scheduling of bookings based on time series predictions of energy data and the need for grid interaction. This work presents three artificial neural network-based machine learning models such as a basic dense neural network model, a convolution model and an autoregressive Long Short-Term Memory (LSTM) model that predict three energy time series one day into the future. The predicted time series comprise the electricity price, the power consumption of the Audimax building of the University of Lübeck and the renewable power generation. The models are trained with TensorFlow and Keras using hardware acceleration, such as the NVIDIA DGX2 machine at the AI Lab Lübeck. An improvement in performance of at least 54% compared to statistical baselines was achieved for all three time series using the models in optimal configurations. Preprocessing steps such as adding temperature data and seasonal adjustment using median windowing are tested and the latter is found to improve performance. A qualitative prediction approach is also evaluated. Finally, this work offers suggestions on improving these models in future research.