Forecast-Aided State Estimation

FASE is fundamentally a recursive Bayesian estimator, most commonly implemented via the Extended Kalman Filter (EKF) or Unscented Kalman Filter (UKF). The algorithm operates in two phases:

• Prediction Step: The state transition model propagates the previous state estimate forward using forecasted load changes, producing an a priori estimate and its covariance.
• Update Step: Real-time measurements (SCADA, PMU, AMI) are assimilated to correct the prediction, yielding the a posteriori state estimate. This recursion provides a natural mechanism for bad data detection by flagging measurements with high innovation residuals against the forecast.

The foundational recursive Bayesian algorithm that forms the mathematical backbone of forecast-aided state estimation. It operates in a two-step cycle: a time update (prediction) where a dynamic model projects the state forward, and a measurement update (correction) where real-time sensor data refines the estimate.

• Minimizes the error covariance by optimally weighting the forecast against new measurements
• The process model captures the temporal evolution of bus voltages and branch currents
• Requires a linear system model; nonlinear power flow equations necessitate extensions like the EKF or UKF

Synthetic data points that supplement sparse real-time sensor data to achieve numerical observability in under-instrumented distribution grids. In forecast-aided state estimation, these are not static historical averages but dynamically updated forecasts of load and generation.

• Derived from short-term load forecasting models, AMI data aggregation, or renewable generation predictions
• Assigned higher variance (lower weight) in the covariance matrix compared to physical measurements
• The quality of the forecast directly dictates the accuracy of the prior state estimate

A nonlinear extension of the Kalman Filter that handles the non-convex power flow equations by computing a first-order Taylor series approximation at each time step. The EKF linearizes the measurement function around the current operating point using the Jacobian matrix.

• Propagates the state estimate through the full nonlinear dynamic model before applying the linearized correction
• Susceptible to divergence if the system is highly nonlinear or the initial estimate is poor
• Widely implemented in transmission-level dynamic state estimators and increasingly adapted for distribution systems

A derivative-free alternative to the EKF that avoids linearization errors by propagating a minimal set of sigma points through the true nonlinear power flow equations. The UKF captures the posterior mean and covariance accurately to the third order for Gaussian distributions.

• Selects 2n+1 sigma points around the current state estimate, where n is the state dimension
• Computationally more intensive than the EKF but significantly more accurate for highly nonlinear distribution networks
• Particularly effective when combined with forecast-aided approaches that provide a strong prior

The critical pre-processing step that determines whether a unique state estimation solution can be computed from the available measurement set and network topology. Forecast-aided state estimation relaxes observability requirements by treating forecasts as virtual measurements.

• Identifies observable islands where the state is solvable and unobservable branches where it is not
• Numerical observability is assessed via the rank and condition number of the gain matrix
• Forecast injection points must be strategically placed to restore observability in the most critical unmonitored regions

A square matrix that encodes the uncertainty and correlation of all measurement and forecast errors in the state estimation problem. The diagonal elements represent the variance of each individual input, while off-diagonal elements capture statistical dependencies.

• Forecasts are assigned higher variance values than physical sensor measurements, reflecting their lower certainty
• The process noise covariance matrix models the uncertainty in the dynamic model's prediction step
• Proper tuning of these matrices is essential for the filter to optimally balance the forecast prior against incoming measurements