Same-day correction of baselines for demand response using long short-term memory

¹ Safety and Occupational Health Engineering, Department of Environmental Engineering, Kyoto University, Yoshida Nihonmatsu, Sakyo, Kyoto, 606-8501, Japan
² Agency for Health, Safety and Environment, Kyoto University, Yoshida Nihonmatsu, Sakyo, Kyoto, 606-8501, Japan

^* Corresponding author: nagaya.taiki.53w@st.kyoto-u.ac.jp

Abstract

In incentive-based the Demand Response, the amount of electricity demand reduction is calculated by subtracting actual electricity demand from the baseline (BL). The BL is the estimated electricity demand of households when no electricity demand suppression is performed. In Japan, the high 4 of 5 method is used to forecast the BL by averaging the actual demand of the day. In this study, we refer to the high 4 of 5 method as BL1. BL2 is the BL to which the value of the same-day adjustment is added based on the actual demand of the day. BL3 is BL1 plus the value of the same-day adjustment predicted using Long Short-Term Memory (LSTM). The average MAE values for BL2 and BL3, calculated using actual electricity demand data from October 15, 2021, to December 24, 2021, were 11.2 kW and 8.1 kW, respectively, with BL3 being 3.1 kW smaller than BL2. To estimate the confidence intervals for BL2 and BL3, we calculated the error by subtracting each BL from the actual value and calculated the ±3σ equivalent for the distribution of the error. The confidence interval calculated for BL3 was found to be ±9.2 kW lower than that for BL2. The F-test for the distribution of the errors for BL2 and BL3 yielded a P-value of 4.05 × 10-50, indicating that the variances of the two distributions were not equally distributed.