Deep Learning Approaches for the Detection of Ramping Artifacts in Schedule Deviation Data in Belgium

In order to contribute to the goals set by the Paris Agreement to combat climate change, economies are increasingly incorporating renewable energy sources into their energy mix. As the share of renewables in the energy mix increases, so does the level of volatility and flexibility. This effect requires stakeholders to implement advanced methods to maintain grid stability and, by this, to balance the supply and demand of electricity in real time. Decisions on different balancing measures are supported by near real-time forecasts published by the responsible transmission system operator (TSO). If such control data is corrupted by ramping artifacts, this could induce unnecessary costs or threaten stability. Using operational data from the Belgian TSO Elia, we present a solution to detect these artifacts. Specifically, we use the knowledge of the internal structure of the artifacts to generate a synthetic dataset as a foundation for training a supervised deep learning model. The proposed model architecture contains an 1d-CNN feature extractor combined with transformer encoder blocks. This architecture is analyzed and compared to other supervised deep learning models and a classical unsupervised model, the Isolation Forest, in two training scenarios, a classification and a regression task. Testing on an unseen synthetic test set, we manage to outperform all other evaluated models, achieving an Fβ-Score of 0.99 in the classification task and 0.75 in the masking task.