Implementation of Electricity Cost-Optimized BEMS Utilizing Deep Learning-Based Power Demand Forecasting

Abstract

In this study, an energy management system (EMS) that integrates demand forecasting and energy storage system (ESS) operation optimization is proposed and its economic feasibility is analyzed using actual power demand data from industrial sites. Based on the actual data of a specific company in Korea, seasonal demand patterns were identified and LSTM, BiLSTM and CNN-LSTM models were implemented using sliding window methods. Each model was tuned with hyperparameters by Bayesian optimization and the optimal model was selected by comparing performance based on Mean Absolute Percentage Error (MAPE) for the last 12 months of data. Based on the prediction results, mixed integer linear programming (MILP) based ESS charge-discharge optimization model was constructed. The optimization model, which is performed on a daily basis, reflects actual operating conditions such as charge/discharge efficiency, SOC constraints, rated power and peak limits and the domestic industrial electricity tariff structure was applied to the model to calculate the cost savings. The economic analysis utilized NREL's battery cost forecasts to derive break-even points for various combinations of ESS capacity and operating hours and to suggest optimal combination. This study provides an empirical methodology to evaluate the technical and economic feasibility of introducing ESS through analysis and simulation based on actual electricity demand data.