Performance Evaluation of LSTM, GRU, and Transformer Models with Incremental Learning for Real-Time Streaming Load Forecasting in the West Kalimantan Power System
Accurate short-term load forecasting is essential for reliable and efficient power system operation, yet its accuracy is influenced by prediction horizon and load dynamics. This study evaluates Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Transformer models under baseline and incremental learning frameworks using Supervisory Control and Data Acquisition (SCADA) load data from the West Kalimantan power system. The models are trained on 30 -minute resolution data from January 2018 to July 2025 and evaluated in a real-time streaming setting from August 2025 to March 2026, where predictions are generated sequentially and model updates are performed only after new observations become available, without access to future data. The evaluation focuses on a 24 -hour forecasting horizon, with incremental learning implemented using update intervals of 6h, 12h, 24h, and 48h. Results show that forecasting errors increase with longer prediction horizons across all models. The Transformer follows this trend but exhibits lower performance than recurrent-based models, particularly at shorter horizons. Among the evaluated models, GRU achieves the best overall forecasting performance, reducing MAE from 15.67 to 15.30 (2.3%) at the 24-hour update interval, with RMSE and MAPE of 20.90 and 3.45%, respectively. The Diebold-Mariano (DM) test confirms statistically significant differences between baseline and incremental forecasting performance $(\boldsymbol{p}<\mathbf{0. 0 0 1})$. Overall, the performance gain from incremental learning is modest under relatively stable load conditions, indicating that its effectiveness depends on data variability.