Physics-Informed Neural Network (PINN)

Unlike standard neural networks that only minimize data error, PINNs add a physics residual term to the loss function. This term penalizes solutions that violate governing differential equations, such as the power flow equations or Kirchhoff's laws. The network learns to satisfy both observed data and fundamental physical constraints simultaneously, acting as a continuous regularizer that prevents overfitting to noisy sensor measurements.

Traditional finite element solvers require computationally expensive mesh generation for complex grid topologies. PINNs operate as mesh-free function approximators, learning continuous spatiotemporal solutions directly from sampled collocation points. This allows seamless handling of irregular geometries like substation layouts and eliminates the curse of dimensionality when modeling high-dimensional parameter spaces in transient stability assessment.

PINNs excel in data-scarce environments common in distribution grids where sensor density is low. By leveraging the embedded physics as a strong inductive bias, the network can accurately infer voltage magnitudes and phase angles at unmonitored buses from only a handful of phasor measurement unit readings. This bridges observability gaps without requiring full network instrumentation.

Beyond forward simulation, PINNs naturally solve inverse problems to discover unknown grid parameters. The network can simultaneously learn the system state while identifying hidden variables such as:

• Line impedance degradation from aging conductors
• Inertia constants of unmonitored rotating machinery
• Reactive power injection from unmodeled distributed energy resources This dual capability makes PINNs powerful tools for model calibration and anomaly detection.

PINNs model grid behavior as a continuous function of time, not discrete snapshots. This enables interpolation between SCADA polling intervals and prediction of inter-sample dynamics that traditional state estimators miss. The continuous representation captures fast electromechanical oscillations and sub-cycle transients critical for wide-area monitoring systems.

A single PINN architecture can encode coupled physical domains simultaneously. For transformer monitoring, the network jointly solves:

• Electromagnetic field equations for winding currents
• Thermal diffusion equations for hotspot temperature prediction
• Fluid dynamics for oil circulation modeling This unified approach captures cross-domain interactions that sequential solvers miss, enabling accurate predictive maintenance forecasts.

The core innovation of PINNs is the composite loss function that penalizes violations of governing physical laws. In grid applications, this includes:

• Data loss: Mismatch between predicted and measured voltage magnitudes at sensor buses.
• Physics loss: Residuals of the power flow equations (Kirchhoff's laws) evaluated at collocation points throughout the network.
• Boundary loss: Constraints at slack buses and generator terminals. This multi-objective optimization allows the network to learn accurate state representations from sparse, noisy SCADA data by leveraging known algebraic constraints.

PINNs rely on automatic differentiation (AD) to compute exact partial derivatives of the neural network's outputs with respect to its spatio-temporal inputs. Unlike numerical differentiation, AD propagates derivative information through the computational graph at machine precision. This is critical for evaluating physics residuals: the network must compute terms like ∂V/∂t and ∂²θ/∂x² to enforce the differential form of the telegrapher's equations or swing equation directly in the loss function during training.

A set of points in the input domain where the physics loss is evaluated, distinct from the points where sensor data exists. In a grid PINN:

• Points are sampled throughout the network topology, including unobserved buses and transmission lines.
• Strategies include uniform random sampling, Latin hypercube sampling, or adaptive sampling that concentrates points in high-gradient regions.
• The density and distribution of collocation points directly control how strongly the physical equations constrain the solution in unmeasured regions, enabling super-resolution of the grid state beyond sensor coverage.

PINNs function as a continuous-time data assimilation engine, merging observational data with a dynamic model. This contrasts with traditional Kalman filtering approaches:

• Kalman filters require explicit state-space models and linearized updates.
• PINNs learn the underlying state implicitly and handle strong nonlinearities natively. In grid digital twins, PINNs can assimilate asynchronous PMU and SCADA streams to reconstruct a time-continuous estimate of voltage phasors and line flows, effectively performing super-resolution in both space and time.

A trained PINN acts as a reduced order model (ROM) — a computationally lightweight surrogate that approximates the behavior of a high-fidelity simulator. Key properties:

• Inference speed: Once trained, a PINN evaluates in milliseconds, enabling real-time contingency analysis.
• Parametric PINNs: Extensions that condition the network on parameters like load levels or topology switches, creating a family of solutions from a single model.
• This replaces the need to run iterative Newton-Raphson power flow solvers for every scenario, dramatically accelerating what-if analysis in the digital twin.

PINNs can be extended to output predictive uncertainty alongside state estimates, critical for operational decision support. Techniques include:

• Bayesian PINNs: Placing probability distributions over network weights to capture epistemic uncertainty from limited data.
• Ensemble PINNs: Training multiple models with different initializations and physics-loss weights to estimate variance.
• Heteroscedastic loss: Training the network to output a mean and variance, learning to identify regions of high aleatoric uncertainty from sensor noise. This provides grid operators with confidence bounds on predicted voltage violations.