Closed-loop simulation testing of a probabilistic DR framework for Day Ahead Market participation applied to Battery Energy Storage Systems

In this manuscript, we test the operational performance decrease of a probabilistic framework for Demand Response (DR). We use Day Ahead Market (DAM) price scenarios generated by a Combined Quantile Regression Deep Neural Network (CQR-DNN) and a Non-parametric Bayesian Network (NPBN) to maximise profit of a Battery Energy Storage System (BESS) participating on the DAM for energy arbitrage. We apply the generated forecast time series to a stochastic Model Predictive Control (MPC), and compare the performance using a point and perfect forecast. For the probabilistic forecasts, we test two control strategies; 1) minimising the Conditional Value at Risk (CVaR) for making costs, and 2) minimising the expected value of the cost. We apply the MPC in a closed-loop simulation setting and perform a sensitivity analysis of the profit by changing the ratio between battery capacity and the max power, the cluster reduction method, and the number of scenarios used by the MPC. We show that the proposed framework works, but the approach does not increase profit compared to a deterministic point forecast. This can possibly be explained by the deterministic forecast capturing the shape of the price curve with less noise than a probabilistic forecast without enough scenarios. We show that the value of a good forecast becomes smaller as the charging time of the battery becomes larger, due to the battery being unable to exploit small price differences optimally.