Applying Neural Networks to Large-Scale Distribution System Analysis: an Empirical Computational Perspective

Abstract

Research has shown that Neural Networks (NNs) are capable of accurate and quick low voltage (LV) grid analysis. Therefore, NNs could be a viable method for the middle long-term scenario tool (MLT), a tool created and used by Enexis, one of the major distribution system operator (DSO) in the Netherlands, to analyze future scenarios of LV grids. However, the tool analyzes a substantial amount of LV grids, and each would require a NN. This amount of NNs necessitates a single network architecture and training method for all NNs, which can be achieved by knowing hyperparameters beforehand, since determining hyperparameters is computationally costly. This paper estimates how long it would take to train a substantial amount of NNs, determines if hyperparameters are shareable between NNs of similar-sized LV grids and if hyperparameters are predictable based on LV grid sizes. The results of hyper-parameter sharing show comparable performance between NNs, however, differences start to occur for larger LV grids. Predicting hyperparameters based on LV grid size gives an unsatisfactory performance.