Statistical Topology Identification and Privacy-Aware Model-Free Current Estimation In Smart Distribution Networks

The ubiquitous energy transition requires distribution system operators to perform advanced calculations in near real time. This task is particularly challenging in low-voltage (LV) distribution networks, where topology is often unknown or undocumented. The availability of large smart meter datasets enables the use of advanced statistical analysis for state estimation of the network, without requiring additional knowledge on topology and network parameters. In this paper, a model-free current calculation algorithm based on machine learning is presented, in which neural networks are employed to accurately calculate current flows in the network. The challenge of unknown topological connectivity is addressed through a statistical topology identification algorithm, through estimating a binary adjacency matrix using a statistical heuristic function and measuring cointegration between fully anaonymized and encrypted voltage and current flow data. The heuristic function is based on a linear combination of Jensen-Shannon divergence and Pearson correlation. This way, both similarity between data values and similarity between distributions are quantified. The estimated adjacency matrix then serves as an input to the model-free current calculation. The proposed methodology achieves highly accurate topology identification, correctly identifying 97% of all connections. Model-free current calculation has exhibited near perfect accuracy, with a mean absolute error of 0.05A.