Model Predictive Control with Pattern Learning of Prediction and Control Trajectory and its Application to a Battery EMS Problem

While the model predictive control method can be widely applied to energy systems and industrial fields, it is difficult to evaluate the reproducibility in advance, and there is a limit to the quality assurance of the control system. In this paper, we propose the prediction/control trajectory of the model predictive control as a control parameter and tried to simplify the model predictive control algorithm by limiting it to a finite number of scenario patterns using the k-means clustering method. This approach does not require optimization calculations and can reduce the computational load and ensure the reproducibility of control operations. The proposed method was applied to the energy management system (EMS) problem with battery storage. The target is an energy system consisting of fluctuating demand load and photovoltaic power generation as a variable renewable energy resource, and storage battery and commercial power reception, and the proposed method is applied to the optimization of storage battery operation. Focusing on the net load pattern as the predicted trajectory and the storage battery's state of charge (SOC) pattern as the control trajectory, the proposed pattern learning model predictive control EMS method was applied to each pattern. The former corresponds to output feedback control and the latter corresponds to state feedback control. Also, state observer-based control is considered where the state is estimated by neural network. As a result of evaluation by simulation of EMS operation for one year based on actual data of PV power generation and demand, control loss increase of only about 8% compared to the optimal control case, with only 3 representative patterns of prediction and control trajectory parameters. So, the effectiveness of the proposed method was confirmed. We also confirmed the robustness against demand forecast error and representative pattern selection error.