Particle Filter

The foundational mechanism that recursively updates particle weights as new measurements arrive. At each time step, particles are propagated through the state transition model and then reweighted according to the likelihood function—how probable the new observation is given each particle's predicted state. Without resampling, this leads to weight degeneracy, where after a few iterations, one particle holds nearly all the weight while others become negligible, wasting computational resources on irrelevant state hypotheses.

A corrective mechanism that eliminates particles with negligible weights and replicates those with high weights, focusing computational effort on regions of high posterior probability. Common schemes include:

• Multinomial resampling: Draws N independent samples from the categorical weight distribution
• Systematic resampling: Uses a single random offset with evenly spaced intervals for lower variance
• Residual resampling: Handles the deterministic replication of particles with large weights separately from stochastic sampling
• Stratified resampling: Divides the cumulative weight interval into equal strata for more uniform coverage

Resampling prevents weight collapse but introduces sample impoverishment, where particle diversity is lost among replicated survivors.

Unlike the Kalman filter, which assumes linear dynamics and Gaussian noise, particle filters impose no parametric constraints on the state distribution. They naturally represent:

• Multimodal posteriors: When a grid state could be in one of several distinct configurations due to switching events or fault scenarios
• Heavy-tailed noise: Impulsive measurement errors from sensor malfunctions or communication glitches
• Hard constraints: Physical limits like generator ramp rates or voltage bounds enforced by rejecting invalid particles

This flexibility is critical for probabilistic power flow where renewable generation forecasts often exhibit bimodal behavior—sunny vs. cloudy regimes—that Gaussian approximations cannot capture.

The choice of importance density from which particles are drawn dramatically affects filter efficiency. The simplest choice—using the transitional prior as the proposal—is computationally cheap but ignores the latest measurement, causing particles to be placed blindly before weighting. The optimal proposal incorporates the new observation, guiding particles toward high-likelihood regions. In grid applications, this might mean using a fast surrogate model or linearized power flow approximation as the proposal, then correcting with the full nonlinear AC power flow during weighting.

After repeated resampling cycles, the particle set collapses to a few distinct values, losing the diversity needed to adapt to sudden state changes. Countermeasures include:

• Regularization: Convolving resampled particles with a kernel density to jitter them into a continuous distribution
• Markov Chain Monte Carlo moves: Applying a reversible MCMC kernel after resampling that leaves the target distribution invariant while restoring diversity
• Auxiliary particle filters: Pre-weighting particles before propagation using a point estimate of the future observation, ensuring high-quality particles survive
• Adaptive resampling: Triggering resampling only when the effective sample size drops below a threshold, reducing unnecessary diversity loss

Particle filters trade computational burden for representational power. Key scaling considerations:

• Curse of dimensionality: The number of particles required grows exponentially with state dimension, limiting practical use to moderate-dimensional systems without Rao-Blackwellization—analytically marginalizing out linear-Gaussian substructures
• Real-time constraints: For phasor measurement unit data streaming at 30–60 samples per second, particle count must be balanced against execution time
• Parallelization: Weight computation and resampling are embarrassingly parallel across particles, enabling GPU acceleration for grid-scale deployment

Typical grid state estimation applications use 1,000–10,000 particles depending on the network size and required update rate.

A variance reduction technique that draws particles from a proposal distribution rather than the true posterior, then assigns importance weights to correct the bias.

• Concentrates computational effort on high-probability regions of the state space
• Critical for rare event simulation in grid contingency analysis
• The sequential importance sampling (SIS) algorithm forms the backbone of every particle filter
• Without resampling, leads to weight degeneracy where one particle dominates all others

A class of algorithms that construct a Markov chain whose stationary distribution equals the target posterior. Unlike particle filters, MCMC operates on a single chain rather than a population.

• Metropolis-Hastings and Gibbs sampling are common variants
• Used in Bayesian grid state estimation when the system is static
• Particle filters with MCMC move steps rejuvenate the particle set after resampling
• Computationally more expensive per sample but handles high-dimensional spaces well

The statistical paradigm that updates the posterior distribution of grid states by combining a prior distribution with a likelihood function derived from new SCADA measurements.

• Particle filters are a computational method for performing recursive Bayesian estimation
• The prediction step propagates particles through the state transition model
• The update step reweights particles based on how well they explain observed measurements
• Enables principled handling of non-Gaussian sensor noise and nonlinear power flow equations

A probabilistic model representing complex distributions as a weighted sum of Gaussian components. Often used as an alternative to particle filters when a parametric representation is preferred.

• Each component captures a distinct mode of the posterior (e.g., multiple possible grid states)
• Expectation-Maximization (EM) algorithm fits GMM parameters to data
• Particle filters can approximate any distribution shape, while GMMs assume Gaussian components
• GMMs provide a compact representation but may miss fine-grained non-Gaussian structure

A deterministic sampling technique that propagates a minimal set of sigma points through nonlinear functions to estimate output mean and covariance.

• Forms the core of the Unscented Kalman Filter (UKF) — a competitor to particle filters
• Requires only 2n+1 sigma points for an n-dimensional state, versus thousands of particles
• Assumes the posterior remains approximately Gaussian after nonlinear transformation
• Particle filters outperform UKF when distributions become strongly multimodal or heavy-tailed

A non-symmetric measure quantifying how one probability distribution diverges from a reference distribution. Used to evaluate particle filter performance.

• Measures the information loss when approximating the true posterior with a particle set
• KL divergence = 0 only when the two distributions are identical
• Guides adaptive resampling — resample only when the effective sample size drops below a threshold
• Also used in variational inference to fit simpler distributions to complex posteriors