Extended Kalman Filter (EKF)

The EKF handles the nonlinear measurement function h(x) by computing a Jacobian matrix—a matrix of first-order partial derivatives—at the current state estimate. This linearizes the power flow equations around the local operating point, allowing the standard Kalman equations to be applied iteratively. This is a first-order Taylor series approximation, making it computationally efficient but potentially inaccurate in highly nonlinear regions.

The EKF operates in a continuous two-step loop:

• Prediction Step (Time Update): The filter projects the state and error covariance forward using a dynamic process model, often a load forecasting function.
• Correction Step (Measurement Update): When new SCADA or PMU measurements arrive, the filter computes the Kalman Gain to optimally weight the prediction against the noisy measurement, minimizing the posterior error covariance.

Unlike static Weighted Least Squares (WLS) estimators, the EKF explicitly incorporates a state transition model. In grid applications, this is typically a short-term load and generation forecast (e.g., Holt-Winters exponential smoothing). This predictive capability allows the EKF to track dynamic system states between measurement scans, providing a crucial forecast-aided state estimation capability for time-varying distribution grids.

A defining feature of the EKF is its propagation of the state error covariance matrix. This matrix provides a real-time, quantitative measure of estimation uncertainty. The covariance grows during the prediction step (increasing uncertainty) and shrinks during the correction step (reducing uncertainty). This allows operators to assess the confidence in voltage and angle estimates at every node.

The EKF's recursive nature means it does not need to batch-process a large history of measurements. It only requires the previous state estimate and the new measurement vector. This low memory footprint and fast computation make it suitable for real-time deployment in substation controllers and Distribution Management Systems (DMS) where latency must be minimized.

The EKF has known vulnerabilities:

• Derivation Complexity: Computing the Jacobian analytically for complex, unbalanced three-phase distribution networks is mathematically intensive.
• Linearization Error: For highly nonlinear conditions (e.g., sudden topology changes), the first-order approximation can introduce large errors, potentially causing filter divergence.
• Gaussian Assumption: The EKF assumes Gaussian noise, which may not hold for pseudo-measurements with non-normal error distributions.

The foundational recursive Bayesian estimator from which the EKF is derived. It provides the optimal state estimate for linear dynamic systems with Gaussian noise by iterating through a predict step (using a state transition model) and an update step (incorporating measurements).

• Assumes linear process and measurement models
• Computes the exact posterior mean and covariance
• Forms the baseline against which nonlinear extensions are measured

The matrix of first-order partial derivatives that lies at the computational heart of the EKF. It linearizes the nonlinear measurement function h(x) around the current state estimate.

• Each element represents the sensitivity of a measurement to a state variable
• For power systems, this includes derivatives of power injections and flows with respect to voltage magnitudes and angles
• A poorly conditioned Jacobian leads to numerical instability in the Gain Matrix calculation

A dynamic estimation framework that provides the EKF with a physically meaningful process model for the predict step. Instead of assuming a static state, it uses time-series forecasting of load and renewable generation.

• Bridges the gap between static Weighted Least Squares snapshots and true dynamic tracking
• Uses Holt-Winters exponential smoothing or ARIMA models to predict state transitions
• Enables predictive observability during communication outages

The statistical engine that weights the EKF's trust between the process model prediction and incoming measurements. The measurement noise covariance R encodes sensor accuracy, while the process noise covariance Q captures model uncertainty.

• Diagonal elements are error variances for individual measurements
• Off-diagonal elements represent correlated errors between sensors
• Tuning Q and R is the primary practical challenge in EKF deployment

The distribution system formulation where the EKF must operate on an unbalanced, multi-phase network model. Unlike transmission systems, distribution grids have single-phase laterals, mutual coupling, and untransposed lines.

• The state vector expands to include per-phase voltage magnitudes and angles
• The Jacobian must model phase-to-phase coupling terms
• EKF linearization must handle the increased dimensionality without divergence