Building Energy and ESS Charging Discharging Prediction: Performance Comparison of LSTM and Transformer Based on Structurally Different Input Data

초록

This study fairly compares Long Short-Term Memory (LSTM) and Transformer models by enforcing a single shared pipeline with identical physical constraints-state-of-charge (SOC) bounds, charge/discharge efficiencies, and rated power limits. We evaluate using mean absolute error (MAE) and root mean square error (RMSE) under two cases. In Case 1, models train on a mixed dataset combining a 2020 photovoltaic (PV) plant with January-July 2023 building data and are tested on August 2023. In Case 2, training uses only January-July 2023 building data and testing is again August 2023. All preprocessing (normalization, imputation, and Korea Meteorological Administration (KMA) weather matching) is fit exclusively on the training split, and group-aware chronological splits by time and site prevent leakage. We further assess regime robustness by decomposing results into clear-day and overcast-day subsets. For visualization, we focus on August 2023 calendar days that were pre-flagged as adverse by KMA forecasts using temperature, humidity, wind speed, and precipitation; these days are ranked into severity tiers, and only this prioritized subset is plotted. To analyze operational implications, we derive a proxy ESS charge/discharge signal from the power-balance equation and examine the resulting charge/discharge and grid-import profiles, which provide qualitative evidence of peak-spike mitigation and PV-aligned charging. Value-based metrics (cost, peak reduction, self-consumption) are deferred to future work.

키워드