Abstract
This paper investigates how existing forecasting models can be enhanced to accurately forecast the electric load at factory level, enabling industrial companies to shift consumption to times of low energy costs. The model architecture must outperform state-of-the-art models and be sufficiently robust for use in multiple factories with low effort for specific applications. Moreover, this work focuses on the processing of high-resolution input data available almost in real time from multiple submeters after the main meters. The theory of load forecasting and related works are summarized in a first step including the requirements of forecasting models applied at factory level. Based on existing models, a new hybrid machine-learning model is proposed, combining a decision tree-based typical load profiler with a convolutional neural network that extracts features from multidimensional endogenous inputs with measurements of the preceding two weeks for multi-step-ahead load forecasts updated almost in real time. Furthermore, a multi-model approach is presented for calculating bottom-up forecasts with submeter data aggregated to a main-meter forecast. In a case study, the forecasting accuracy of the hybrid model is compared to both base models and a seasonal naïve model calculating the load forecasts for three factories. The results indicate that the proposed typical-load-profile-supported convolutional neural network for all three factories achieves the lowest forecasting error. Furthermore, it is validated that a reduction in data transfer delay leads to better forecasts, as the forecasting accuracy is higher with near real time data than with a data transfer delay of one day. Thus, a model architecture is proposed for robust forecasting in digitalized factories.