CEU Electronic Theses and Dissertations, 2025
| Author | Štros, Sebastian |
|---|---|
| Title | Improving Elastic Gradient Boosting: Catching Concept Drift in Smart Meter Time Series Data |
| Summary | This thesis improves on the elastic Gradient Boosting Decision Tree algorithm (eGBDT) by limiting its memory use and increasing its accuracy. eGBDT is a successful incremental learning algorithm. Because it deletes weak learners which are outdated, it maintains high accuracy on conceptually drifted data. This thesis offers two versions which improve on eGBDT's accuracy and memory use. The first version contribution-based eGBDT (cbeGBDT) improves eGBDT by making weak learner deletion dependent only on performance and not on index as in the original eGBDT. The second version, Accuracy Updated eGBDT (AUeGBDT) adds to cbeGBDT weighting mechanism inspired by the Accuracy Updated Ensemble. Improving eGBDT accuracy and memory use is motivated by the new and important application of incremental learning on smart meter time series data. Smart meter data need to be accurately forecasted to maintain energy grid stability in the times of the massive renewable energy sources roll out and transport electrification. On this regression task, the cbeGBD outperforms the original eGBDT, non-retraining GBDT, but is comparable to a non-icrementally learning GBDT model which retrains on every batch. |
| Supervisor | Karsai Marton |
| Department | Network Science MSc |
| Full text | https://www.etd.ceu.edu/2025/stros_sebastian.pdf |
Visit the CEU Library.
© 2007-2025, Central European University