Uncertainty Quantification of energy demand forecast for use in decentralized energy systems

Abstract

This thesis implements two methods namely Quantile Regression (QR) and Uncertainty Binning Method (UBM). Both the methods have different approaches to generate probabilistic load forecasting (PLF). QR is considered as a benchmark model and compared with the newly developed UBM model which generates PLF by leveraging the development of deterministic models. During literature review, lack of proper evaluation metric was found to be one of the reasons for less adoption of probabilistic forecasting. Current popular metrics do not provide the full picture of probabilistic forecasting evaluation. Moreover, not many literatures include graphical evaluation of probabilistic forecasting. Therefore, three new metrics developed by Saber [1] namely Graphical Calibration Measures (GCM), Quantile Calibration Score (QCS) and Percentage Quantile Calibration Score (PQCS) were implemented with modification. Two more percentage metrics which determine the percentage of measured values above the upper quantile limit and below the lower quantile limit were calculated to assist in understanding more about the prediction uncertainty. These metrics could fill some of the gaps found in the current performance evaluation metrics. In this thesis, the quantification of uncertainty in load demand predictions was conducted for a commercial building located in Munich, Germany. For the given dataset, the probabilistic forecasting generated by UBM was approx. 5 % more reliable but less sharp than QR. Overall, UBM was superior than QR based on Winkler Score. Prediction Interval Coverage Probability (PICP), Mean Prediction Interval Width (MPIW) and Winkler Score for UBM model was found to be 84.65 %, 9.65 and 12.88 respectively. It also revealed that UBM was much faster and easy to implement than QR. The requirement of an existing deterministic model with availability of several months of historical point forecasts as training datasets was the only limitation for UBM compared to QR. That means, UBM could be implemented immediately only if there are already existing point forecast models in application. The results indicate that UBM provides a simplistic approach to quantify uncertainty in the load demand prediction with high efficiency. The consistency of UBM in obtaining good probabilistic forecasting was found to be better compared to QR. Moreover, UBM also performed efficiently for another test dataset with highly variable load demand. PLF generated by UBM could be used as a tool to improve decisionmaking in energy management of the commercial building and other applications