Rapid and Accurate Changepoint Detection of Power System Forced Oscillations

This paper describes a new approach for using changepoint detection (CPD) to estimate the starting and stopping times of a forced oscillation (FO) in measured power system data. As with a previous application of CPD to this problem, the pruned exact linear time (PELT) algorithm is used. However, instead of allowing PELT to automatically tune its penalty parameter, a method of manually providing it is presented that dramatically reduces computation time without sacrificing accuracy. Additionally, the new algorithm requires fewer input parameters and provides a formal, data-driven approach to setting the minimum FO segment length to consider as troublesome for an electromechanical mode meter. A low-order ARMAX representation of the minniWECC model is used to test the approach, where a 98% reduction in computation time is enjoyed with high estimation accuracy.

💡 Research Summary

**
The paper presents a streamlined changepoint‑detection (CPD) framework for locating the start and stop times of forced oscillations (FOs) in power‑system measurement data. Prior work applied the Pruned Exact Linear Time (PELT) algorithm together with the CROPS (Changepoints for a Range of Penalties) procedure to automatically search for the penalty parameter β. While accurate, CROPS repeatedly invokes PELT, leading to high computational cost and an upper bound on the number of detectable changepoints, which limits real‑time applicability.

The authors propose two major improvements. First, they replace the automatic β search with a manually supplied β value. By deriving an upper bound β_max from the total cost of the data set and the cost of a single changepoint (equations 13‑14), they show that a β in the middle of the feasible interval (e.g., 0.5 · β_max or the mean of β(τ₁) from many simulations) yields the same changepoint locations as the exhaustive CROPS search. Consequently, the internal PELT routine is called only once, reducing the algorithmic complexity from O(N·K) to O(N).

Second, they introduce a deterministic post‑processing step that converts the changepoints returned by MATLAB’s ischange function into the actual FO start (ϵ) and stop (η) sample indices. The function provides an estimated segment‑mean vector (\hat Y); taking its backward difference (dY) identifies positive jumps as candidate starts and negative jumps as candidate stops. Consecutive identical‑type indices are merged (keeping the first start and the last stop), and edge cases where the FO is present at the very beginning or end of the record are handled by appending 0 or N‑1 as needed. The resulting ordered set γ yields a clean list of (ϵ_i, η_i) pairs.

A further contribution is a data‑driven rule for the minimum FO segment length (N_{\text{minSL}}). Using a local signal‑to‑noise ratio (SNR) definition (equation 19) and the smallest SNR that biases an ARMA mode meter (SNR_min), together with an upper bound on expected FO amplitude (A_max), the authors compute (N_{\text{minSL}}) via equation 20. Segments shorter than this threshold are discarded, preventing spurious detections caused by noise and ensuring that only physically relevant oscillations are fed to the mode‑meter.

The methodology is validated on a low‑order ARMAX representation of the miniWECC test system. Simulations generate 25‑minute frequency‑deviation records sampled at 3 Hz, with a 10‑minute FO inserted (samples 1535–3334). FO amplitudes span SNR values from –15 dB to +10 dB. For each SNR, 300 Monte‑Carlo trials are performed. The previous approach (PELT + CROPS, with N_maxCP = 10, N_minSL = 2 min, α = 0.7) is compared against the proposed method (β = mean β(τ₁)).

Results show a dramatic reduction in computational time: the average runtime drops from 363 ms (previous) to 8 ms (new), a 98 % speed‑up that brings FO start/stop estimation to the same order of magnitude as FO detection and parameter estimation. Accuracy of ϵ and η estimates remains essentially unchanged; the new method even exhibits slightly lower variance at low SNRs. Subsequent refinement of FO amplitude, frequency, and phase using only the detected FO segment yields estimates indistinguishable from those obtained with the previous method. When these refined FO parameters are supplied as exogenous inputs to an ARMAX mode meter, the resulting mode‑frequency and damping estimates match the baseline, while the biased ARMA‑only meter shows the expected large errors.

Additional experiments explore varying FO durations, FOs that start before the first sample or end after the last sample, and the presence of large ring‑down transients. In all cases the new algorithm maintains accuracy and speed. Tests with alternative CPD search strategies from the Python Ruptures library (Binary Segmentation, Bottom‑Up, Window Sliding) confirm the classic trade‑off: they are faster than PELT but sacrifice detection precision.

In summary, the paper delivers a practical, low‑latency CPD solution for forced‑oscillation localization in power‑system data. By fixing the penalty parameter manually, invoking PELT only once, and providing a clear mapping from changepoints to FO start/stop indices, the authors reduce computational burden by two orders of magnitude while preserving (or slightly improving) estimation fidelity. The data‑driven minimum‑segment‑length rule further enhances robustness against noise‑induced false detections. This work enables near‑real‑time FO handling in advanced mode‑metering tools, facilitating more reliable damping‑estimation and stability monitoring in modern power grids.