Abstract

When dimensioning transformer stations, accurate estimation of peak power demand is crucial. However, the data provided by network operators is usually aggregated to 15-minute intervals, which risks overlooking short-duration power peaks that are critical for engineering
dimensioning decisions. This master’s thesis investigates whether machine learning-based super-resolution can be used to reconstruct high-resolution power demand data from low-resolution measurements, with a particular focus on preserving power peaks relevant for infrastructure dimensioning.

In this paper, machine learning-based super-resolution, a method that aims to reconstruct high-resolution data from low-resolution measurements, is applied to industrial power demand data. An SRP-CNN model was developed and trained on power measurements with a temporal resolution of 30 seconds collected at an industrial facility over a period of approximately nine days. Three loss functions were evaluated: Root mean square error (RMSE), Structural Similarity Index Measure (SSIM), and a weighted RMSE. The latter was specifically designed to assign greater emphasis on extreme power values. The hyperparameters of the model were optimized using Bayesian optimization, applied separately for each loss function.

All three models successfully reconstructed the overall structure of the high-resolution signal. At extreme power peaks, however, the reconstruction performance deteriorated, where all models systematically underestimated the peaks magnitudes. The weighted RMSE model demonstrated the strongest ability to reconstruct power peaks, although significant errors remained even for this model. The results indicate that SRP-CNN models can provide valuable support for general load analysis, but that the method should be used with caution when dimensioning as accurate assessment of extreme power peaks is critical.