Inertia Estimation

Algorithms remain dormant during steady-state operation and activate only upon detecting a significant Rate of Change of Frequency (ROCOF) excursion. A typical trigger threshold is |ROCOF| > 0.05 Hz/s, indicating a generation-load imbalance. This event-driven architecture conserves computational resources and ensures estimation occurs only when the inertial response is physically excited and measurable. The trigger must discriminate between genuine frequency events and measurement noise to prevent false activations.

The core mathematical model is the classical swing equation, which relates the imbalance between mechanical and electrical power to the derivative of frequency. The algorithm solves for the inertia constant H using the relationship:

• 2H * dω/dt = Pm - Pe
• Where dω/dt is the ROCOF measured by PMUs
• Pm - Pe represents the active power mismatch This physics-based approach ensures the estimate is grounded in first principles rather than purely statistical correlation.

To extract a clean ROCOF value from noisy PMU data, algorithms employ curve-fitting techniques on the frequency trajectory immediately following a disturbance. Common approaches include:

• Polynomial fitting: A 2nd or 3rd-order polynomial is fitted to a short window of frequency data
• Moving average filters: Smooth high-frequency noise before differentiation
• Kalman filtering: Recursively estimates the true ROCOF state while rejecting measurement noise The fitted polynomial's derivative at t = 0+ provides the initial ROCOF used in the swing equation calculation.

The estimation window duration is a critical hyperparameter with a fundamental trade-off:

• Short windows (100-300 ms): Capture the true initial inertial response but are highly sensitive to PMU measurement noise and local oscillations
• Long windows (500-1000 ms): Provide noise averaging but risk contamination from primary frequency response (governor action) which artificially inflates the apparent inertia Optimal window selection often uses adaptive techniques that analyze the signal-to-noise ratio in real-time.

For wide-area system inertia, algorithms compute the Center of Inertia (COI) frequency, which is a weighted average of frequency measurements from geographically distributed PMUs. The weighting factors are typically the MVA rating of the generation connected near each PMU bus. This approach:

• Filters out local inter-machine oscillations
• Provides a single, representative frequency for the entire interconnection
• Enables accurate estimation of the total system inertia constant H_sys rather than a localized value

Robust algorithms provide not just a point estimate but a confidence interval for the inertia constant. Sources of uncertainty include:

• PMU measurement errors: Total Vector Error (TVE) in voltage and current phasors
• Unknown power imbalance magnitude: The exact size of the lost generation or load is often estimated
• Ambient oscillations: Background frequency noise that biases ROCOF calculation Outputting an uncertainty range (e.g., H = 4.2 ± 0.3 seconds) allows grid operators to make risk-informed decisions rather than relying on a potentially misleading single value.

The primary input signal for inertia estimation algorithms. ROCOF measures the speed of frequency decline (in Hz/s) immediately following a generation-loss event. A high ROCOF indicates low system inertia, while a gradual decline suggests robust inertial response. PMUs must compute ROCOF with high fidelity, as measurement noise or time-synchronization errors can severely distort the inertia calculation. Advanced algorithms apply low-pass filtering and outlier rejection to clean the raw ROCOF signal before feeding it into the inertia constant solver.

The fundamental physics model linking power imbalance to frequency dynamics. The equation states that the per-unit inertial constant (H) multiplied by the ROCOF equals the mismatch between mechanical input power and electrical output power. Inertia estimation algorithms solve this equation in reverse: by measuring electrical power deviation and frequency response during a disturbance, they calculate the effective H constant. The swing equation assumes a single aggregated machine model, so advanced estimators incorporate spatial variation across multiple PMU locations.

The sensor backbone for inertia estimation. PMUs capture time-synchronized voltage, current, and frequency phasors at 30-120 samples per second, providing the high-resolution data needed to detect the instant of disturbance and track the subsequent frequency trajectory. Without GPS-synchronized PMU data, inertia estimation would rely on slower SCADA scans (2-4 second intervals), missing the critical sub-second dynamics. Key PMU requirements for inertia applications include:

• High-accuracy ROCOF measurement
• Low Total Vector Error (TVE)
• Reliable GPS time-stamping

The minimum frequency point reached during a disturbance before recovery begins. Inertia estimation algorithms analyze the trajectory from the pre-event steady-state frequency down to the nadir. The depth of the nadir is inversely proportional to system inertia: low-inertia grids experience deeper, faster nadirs. Grid operators use inertia estimates to predict the nadir for contingency scenarios and determine whether Under-Frequency Load Shedding (UFLS) thresholds will be breached before primary frequency response can arrest the decline.

A weighted-average frequency representing the aggregate electromechanical behavior of the entire interconnection. The COI frequency is calculated by summing the frequency of each generator weighted by its inertia constant (H) and MVA rating, then dividing by total system inertia. Inertia estimation algorithms often use the COI frequency rather than a single local PMU measurement to avoid local oscillation artifacts. The COI provides a cleaner signal for solving the swing equation, especially during inter-area oscillations where individual bus frequencies diverge.

The emulated inertial response provided by power-electronic-interfaced resources such as wind turbines, battery energy storage systems, and HVDC links. Unlike synchronous machines, these devices do not inherently release kinetic energy during frequency events. However, fast-acting control loops can inject active power proportional to ROCOF, mimicking the behavior of physical inertia. Inertia estimation algorithms must distinguish between true synchronous inertia and synthetic inertia contributions, as the latter depends on control tuning and may exhibit different dynamic characteristics during severe disturbances.