Decentralized prediction of electrical time series in smart grids using long short-term memory neural networks

In the modern power grid framework, Renewable Energy Sources must be integrated into the existing energy systems to optimally deal with load, power and electromagnetic imbalance issues. In this context, smart grids have a pivotal role in transforming the aggregation of decentralized power sources. In order to implement these complex systems and to enable such an integration, machine learning techniques must be investigated and adopted where necessary. The realization of smart systems to solve dispatchability problems must rely on various learning schemes, in particular forecasting of time series regarding load, price, and power. In this paper, an emerging machine learning paradigm is proposed, which makes use of a distributed architecture based on the Long Short-Term Memory Network model, a type of recurrent neural network that is tailored to the aforementioned forecasting problems