Increasingly volatile power load consumption patterns due to the energy transition, aswell as possible power saving calls from governments in Europe as a response to the Ukraine war additionally increase importance of this subject.

In the literature we found no definition or metric for adaptability or how we could increase it in forecasting models. Therefore we created our own metrics called Performance Stability Index (PSI) and Learning Efficiency Index (LEI). These metrics rely on measuring performance metrics, like Root-Mean-Square Error (RMSE), iteratively in training and test cycles. The PSI measures how stable a model performs over time, while the LEI measures how well a model learns from new data. Additionally to providing a definition and two measures for adaptability we also want to try to find a driver to increase it. In this thesis, we create an additional boolean feature on the cluster level, that helps highlighting times of high relevancy of a related predictor to a forecasting model. Because the highlighting is done on the cluster level and is filtered through all yearly data, only highly persistent, reliable highlights remain. Therefore we argue that a model using this attribute is going to produce higher PSI and LEI adaptability scores than the same model not using this feature.

In our experiment the results are inconclusive. Although a tiny trend towards better LEI scores for the model using the attribute, and a slight trend toward better PSI scores for the model not using the attribute exist, the values in general are too close together to be conclusive. But we still succeeded in opening the discussion about adaptability and robustness in forecasting. Also by trying to define adaptability and robustness and how to make it measureable.