Hierarchical Electrical Load Forecasting of Industrial Production Systems in the Manufacturing Industry based on Deep Learning

Climate change is one of the most significant challenges of the 21st century. As one of the countermeasures, the German government plans to convert the power supply entirely to renewable energies by 2045. However, with an increased share of renewable energies, the flexibility of electricity generation is decreasing, while the need for flexibility is increasing due to the weather dependence of renewable electricity generation. A potential measure to counter this problem is the creation of new flexibilities on the demand side, especially in the industry. For a factory to adapt its future energy demand to changing power generation requirements at short notice, it is necessary to know the future power consumption as precisely as possible. Therefore, in the context of this dissertation, a very short-term load forecasting model for production systems was developed based on deep learning methods. The model development phase is based on the Cross Industry Standard Process for Machine Learning process. The three Deep Learning algorithms, Long Short-Term Memory, Gated Recurrent Unit, and Convolutional Neural Network, were selected for model development because they are particularly suitable for forecasting complex time series. Furthermore, an architecture was designed to integrate historical information and information on the forecast horizon into the model development process without compromising the generalisation capability due to increasing influencing factors (curse of dimensionality). In a preliminary study, three concepts were investigated. Additionally, Neural Architecture Search methods were used for modelling. Finally, based on the concept that emerged from the preliminary study, a hierarchical load forecast of the production system was carried out based on the use case of the ETA research factory. Therefore, a forecasting model was created for the five primary consumers first, and their results were integrated into the forecasting model of the production system. The results show that the selected Deep Learning techniques are well suited for industrial load forecasting. Furthermore, the hierarchical structure for forecasting the production system leads to improved load forecasting. For the first time in such a work, a novel architecture for integrating historical information and information on the forecasting horizon into the modelling process has been implemented. Additionally, the developed approach for integrating information on the forecasting horizon improved the load forecast in more than 83 % of the models