Lag-Enhanced LSTM for Power Grid Frequency Prediction

This paper investigates the impact of lagged features on short-term frequency prediction using three univariate Long Short-Term Memory (LSTM) models: without lag, with lag steps 1 to 3, and with lag steps 1 to 10. The baseline model yielded a MAPE of 8.08%, which was reduced by 89.4% to 0.86% when using lag steps 1 to 3. However, extending the window to 10 lag steps (equivalent to 200 ms) slightly worsened performance, raising the MAPE to 1.08%. Correlation and delay analyses revealed that frequency synchronization across the grid is highly time-sensitive and spatially structured, with most predictive information concentrated in the immediate past. These directional differences highlight that correlation degradation is not symmetric, and the direction of interaction influences how frequency deviations propagate. Identifying such asymmetric and time-sensitive relationships is essential for refining model input design, particularly for lag-aware LSTM architectures that rely on the physical structure and timing of signal transmission. These findings suggest that compact lag windows offer the best balance between model accuracy and efficiency, aligning with the physical dynamics of the grid.