Load Classification and Forecasting for Temporary Power Installations

摘要

Temporary Power Installations (TPIs) serve energy at events (e.g., festivals), typically from on-site generation. As they become more prominent, there is a greater need for efficient configuration and optimal usage. Predictive modeling can help in this regard, however, this is particularly challenging due to limited data and high configuration diversity. We present approaches for: (1) offline load classification, prior to the TPI to improve system efficiency, and (2) online load forecasting, during TPI operation to improve system reliability. First, TPI attributes and load data are used as features for clustering, and TPI attributes are mapped to the obtained clusters using a classifier. Second, forecasting real-time load data is framed as a regression problem to predict load at least two hours ahead. A case-study using real-world data measured at festivals shows that: (1) load patterns cluster in practice and can be predicted from TPI attributes beforehand, and (2) by modeling residuals, load forecasting accuracy can be improved online. Our improved forecasts thereby enable more efficient TPI configuration.