결정적 및 확률적 성분과 수준변수를 활용한 SMP 예측 성과 비교

Price forecasting with component decomposition enhances accuracy by analyzing various factors, despite its complexity limiting earlier research. Advances in computing have expanded its application. This study compares forecasting performance for the System Marginal Price (SMP) of the Korea Power Exchange using level variables and component decomposition. SMP was divided into deterministic and stochastic components via spline functions and calendar effects. The deterministic component was forecasted with an exponential smoothing model, while ARDL and LSTM models were applied to level variables. Stochastic components were modeled using AR with jump or GARCH family models and integrated with deterministic forecasts. Results showed that component decomposition outperformed level variables. The AR and jump model achieved the highest accuracy, followed by ARDL, which leveraged SMP and energy price cointegration. GARCH-based models performed worse than jump models, and LSTM was slightly less accurate than ARDL. These findings highlight the need for models tailored to data characteristics for effective forecasting.