Optimizing the Training Strategy of Bi-LSTM Models for Real-Time Electricity Demand Forecasting

This paper studies the optimization of training strategies for a Bi-LSTM (Bidirectional Long Short-Term Memory) model to improve real-time electricity demand forecasting accuracy. Unlike most existing research that emphasizes hyperparameter tuning, this work examines three key training strategy variables essential in practice: training period, window size, and retraining frequency. Using four years of South Korea’s power demand and weather data, a sequential search identifies the optimal configuration. Results show that a 24-month training period, a 7-day window, and a 3-day retraining cycle achieve the most accurate and stable forecasts, with an average MAPE of 0.84%. The study also assesses the trade-off between accuracy and computational cost, confirming the practicality of the proposed strategy and underscoring that training strategy optimization is as critical as model architecture tuning."