Research on cross-building energy storage management system based on reinforcement learning

Abstract

This study considers a cross-building energy storage system in which the objective function of each step is a piecewise linear function of decision variables and state variables. Therefore, the objective function can be modeled as piecewise linear programming and then transformed into a mixed integer linear programming (MILP) problem. However, as a multi-stage stochastic programming problem in which we utilize approximate dynamic programming (ADP) to tackle the computational issues, we need to solve the objective function multiple times. To further decrease computational cost, we propose several approximate algorithms to determine variable splitting, which degrades the problem to a linear programming problem. We use approximate techniques to solve the problem and design experiments to verify our conclusion. Numerical experiments show that our algorithm greatly reduces the time needed to solve the problem under the condition of minimal loss of accuracy. The simulation experiment in a Python environment further proves that the cross-building energy storage management system based on energy routers and control centers is better than the energy system of each building working alone to maximize the benefits.