VMD와 학습 오차 예측을 통한 시계열 예측 모델 성능 개선에 관한 연구

초록

When creating a time series prediction model, research is being conducted to improve the learning performance by extracting time-series characteristics of the data and applying them to the model. However, most studies focus primarily on central seasonality and trend, while neglecting detailed variations. In this study, we propose a time series forecasting technique using Variational Mode Decomposition (VMD) to better capture fine-grained characteristics. As a first step, VMD is used to decompose the time series data into intrinsic modes that reflect different features. These modes are then used to train a Long Short-Term Memory (LSTM) model to produce the initial forecast. In the second step, the error—calculated as the difference between the predicted and actual values—is further decomposed using VMD. An additional LSTM model is trained on these decomposed errors to predict the residuals, which are then added to the initial forecast to obtain the final prediction. The validation indices used were R2, RMSE, and MAPE, and the performance of the predicted values was measured by additionally using detailed time series characteristics derived through variation mode decomposition. As a result of the experiment, the proposed technique demonstrated high and consistent predictive performance, especially when compared to existing models whose accuracy significantly fluctuates depending on the input data size.

키워드