Optimal Operation of Community Energy Storage using Stochastic Gradient Boosting Trees

This paper proposes an algorithm for the optimal operation of community energy storage systems (ESSs) using a machine learning (ML) model by solving a nonlinear programming (NLP) problem iteratively to obtain synthetic data. The NLP model minimizes the network's total energy losses by setting the community ESS's operation points. The optimization model is solved recursively by Monte Carlo simulations in a distribution system with high PV penetration, considering uncertainty in exogenous parameters. Obtained optimal solutions provide the training dataset for a stochastic gradient boosting trees (SGBT) ML algorithm following an imitation learning approach. The predictions obtained from the ML model have been compared to the optimal ESS operation to assess the model's accuracy. Furthermore, the ML model's sensitivity has been tested considering the sampling size and the number of predictors. Results showed a 98% of accuracy for the SGBT model compared to optimal solutions. This accuracy was obtained even after a reduction of 83% in the number of predictors.