The global transition to renewable energy forces electricity systems to integrate large shares of variable generation, making investment planning increasingly complex. Energy System Optimisation Models (ESOMs) jointly optimise capacity investment and operational dispatch to support these decisions, but their computational cost grows rapidly with increasing temporal resolution. Temporal aggregation reduces this cost by replacing the full time series with a smaller set of representative partitions. A known weakness of standard aggregation methods is the smoothing effect: extreme demand peaks and near-zero renewable generation periods are averaged with their neighbours, causing the optimisation model to underestimate capacity requirements and incur avoidable costs when its investment plan is evaluated against the full time series.

This thesis proposes two extensions to chronology-preserving hierarchical clustering that actively counteract this smoothing effect. Post-hoc Extreme Correction (PEC) runs standard Ward hierarchical clustering and then replaces the centroid of any partition containing an extreme time step with the corresponding maximum or minimum value. Extreme-Aware Clustering (EAC) integrates extreme-event information directly into the merge priority, delaying the mixing of extreme and non-extreme time steps and thereby forming more coherent partitions before any representative correction is applied. Both strategies are embedded within the Fully Flexible Temporal Resolution (FFTR) framework, enabling automated, data-driven, variable-specific temporal partitions.

We evaluate all three strategies --- baseline Hierarchical Clustering (HC), PEC, and EAC, together with Uniform Temporal Resolution (UTR) as an additional baseline --- using load duration curves, Sum of Squared Error analysis, and comparison against the full-resolution benchmark. We find that EAC's advantage over HC is conditional rather than universal: EAC outperforms HC when temporal variability is high and cluster counts are low, but HC achieves comparable or lower regret when variability is moderate and the resolution assignment for non-profiled assets is consistent. We additionally find that the resolution assigned to non-profiled assets and inter-location flows has a first-order effect on investment quality: aligning these to the common highest resolution of their connected profiled endpoints dramatically reduces HC regret, substantially reordering the method rankings at low to moderate resolutions. These results suggest that the resolution assignment strategy matters as much as the choice of clustering algorithm, and that EAC is a robust improvement over HC specifically when extreme events dominate the cost structure. ... Read more

Machine learning can be used to classify patients in a hospital. Here, the classifier has to minimize the cost of misclassifying the patient and minimize the costs of the tests. Unfortunately, obtaining features may be costly, e.g., taking blood tests or doing an x-ray scan. Furthermore, it is possible that acquiring those test results may take a few days. To train such a classifier, several machine learning algorithms exist. Decision trees appear as favourites since, unlike other classifiers, decision trees do not need all feature values to classify an instance. Current approaches, however, only use heuristics to find a local optimum. Although heuristics are relatively fast, they are not optimal and therefore may not capture well the underlying characteristics of the given dataset. We propose an optimal approach to train cost-sensitive decision trees while also considering these delayed tests. Here we show, smaller trees with higher accuracy and a lower cost can be constructed, compared to a heuristic approach. We use dynamic programming to allow us to skip many calculations, which speeds up the programming time to seconds. ... Read more