Bayesian Compressed Deep Learning for State Estimation of Unobservable Power Systems

Abstract

In recent years, state-of-the-art Deep Learning (DL)-based modeling has been applied to the problem of state estimation of unobservable electrical distribution systems, with promising results. Unfortunately, the definition and training of these flexible models have been largely heuristic, which may result in oversized neural networks, with computationally inefficient layers. In this work, we apply the method of Bayesian Compression for eliminating spurious redundancies of DL-based State Estimation models. Experimental results in four test networks, including two IEEE Test Case Power Networks, corroborate the benefits of the proposed compression approach for obtaining reduced versions of the models without compromising their performance.