Howind, S. (2024). Load profile forecasting of manufacturing processes with a data-driven model [Diploma Thesis, Technische Universität Wien]. reposiTUm. https://doi.org/10.34726/hss.2024.120307
energy awareness; production planning; machine learning; energy monitoring
en
Abstract:
In order to mitigate climate change, countries all over the world invest in replacing fossil fuel power plants with electricity generation from renewable energy sources, whose availability is subject to changing weather conditions. Besides storing excess electricity generation in electrical storage systems, energy demand has to be adapted to the availability of renewable energy tomaintain the balance between electricity generation and consumption in the electricity grid at all times. An incentive to do so are real-time energy tariffs. Manufacturing scheduling can take energy costs into account as an optimization criterion but depends on forecasts of the power profile of the individual manufacturing processes at the time of scheduling.This thesis proposes a data-driven energy model architecture that generates a power profile from the process parameters. The architecture is aimed at discrete production, especially production types like batch production or job production, with processes of smaller quantities that run on adaptable machines with adaptable process parameters. Unlike model-driven approaches, the proposed data-driven energy model is easily adaptable to various use cases by training it with historical data, and unlike a simple lookup table, it can interpolate betweenparameter combinations from the historical data.Different architectures were tested on synthetic energy data based on a real use case of batterypack assembly and on energy consumption data recorded in an experiment series conducted onan industrial robot. Of the tested model architectures, an Ensemble Long Short-Term Memoryarchitecture and a Long Short-Term Memory-Sequence-to-Sequence architecture generally showedthe best prediction accuracy while the Neural Network architecture proved to be unsuitable forthe task. Altogether, an absolute prediction error of 5% with the most suitable architectures inthe respective cases can be expected.