Time Segmented Deep Forecasting of Power Grid Frequency Deviations

Setianingsih, Casi and Hartmann, Bálint (2026) Time Segmented Deep Forecasting of Power Grid Frequency Deviations. In: UNSPECIFIED IEEE. (Submitted)

Preview

Text Setianingsih_Hartmann_33rd_Minisymposium.pdf - Submitted Version Download (615kB) | Preview

Abstract

Frequency deviations reflect short-term power imbalance and contain information about system-wide synchronization and time-delayed coupling. We analyze about one month of synchronized transmission-grid frequency measurements from 17 monitoring points using lagged Pearson cross-correlation for lags 1–10, reported for two aggregations: an all-month baseline and a weekday–weekend split. Weekends show higher coherence than weekdays across all evaluated lags (e.g., at lag 1 the global mean correlation is 0.258 on weekends versus 0.174 on weekdays). From the lag sweep, we extract peak correlation and best lag to summarize an empirical delay structure; most pairs peak at small lags, while some corridors require larger lags (≈8–10 samples) for maximum alignment. To connect these coupling patterns to forecasting design, we run a controlled univariate benchmark on a reference station (24,083,128 samples in Feb 2025, 0.1 s sampling) using LSTM, TCN, and PatchTST under the same time split, train-only preprocessing, and early stopping, with window lengths k∈{1,3,10,20}. Increasing the window from k=1 to k=20 improves accuracy for all models and reduces architecture sensitivity; at k=20 the three models converge to nearly the same nRMSE (≈11.5%).

Actions (login required)

View Item