Physics-Informed Neural Network (PINN)

The core innovation of PINNs is the integration of partial differential equations (PDEs) or ordinary differential equations (ODEs) into the loss function. Instead of relying solely on data error, the network is penalized for violating known physics, such as the power flow constraints or Kirchhoff's laws. This acts as a powerful regularizer, forcing the model to learn solutions that are physically plausible even in regions with sparse or noisy sensor data.

A PINN's loss function is a composite of multiple terms, balancing empirical data with theoretical law:

• Data Loss: The standard mismatch between network predictions and available measurements (e.g., SCADA voltage readings).
• Physics Loss: The residual of the governing PDE evaluated at collocation points across the domain.
• Boundary/Initial Loss: Penalties for violating known boundary conditions (e.g., voltage at a substation) or initial states. This hybrid approach allows the network to interpolate accurately far beyond the training data distribution.

Unlike traditional numerical methods like Finite Element Analysis (FEA) which require a complex, hand-crafted mesh, PINNs operate on a mesh-free principle. Training points (collocation points) can be randomly sampled throughout the continuous spatio-temporal domain. This is a significant advantage for grid topology optimization where the domain changes dynamically, eliminating the computationally expensive need to re-mesh the network for every switching configuration.

PINNs excel at inverse problems, where the goal is to discover unknown parameters of the governing physics from observed data. In a grid context, a PINN can be trained on voltage and current measurements to infer unobservable parameters like line impedance, inertia constants, or the location of a fault. The unknown parameter is treated as a trainable variable that is optimized alongside the neural network weights.

A critical enabler of the physics loss is automatic differentiation (AD) . The PINN uses the chain rule to compute the exact partial derivatives of the network's output (e.g., voltage) with respect to its inputs (e.g., time and space). These analytically precise derivatives are then plugged into the governing PDE to calculate the physics residual, avoiding the numerical approximation errors inherent in finite difference methods.

Once trained, a PINN acts as a continuously differentiable surrogate model of the physical system. It can instantly predict the grid's state (e.g., voltage profile) for any given input (e.g., load and generation pattern) without running a slow, iterative power flow solver. This millisecond-level inference speed makes PINNs a powerful tool for Model Predictive Control (MPC) and real-time Volt-VAR Optimization (VVO) , where control decisions must be made on a rapid timescale.

PINNs can solve the nonlinear AC power flow equations without iterative Newton-Raphson solvers. By embedding the power balance constraints (P=VI*) as a physics loss term, the network learns to map load injections directly to voltage magnitudes and angles.

• Mechanism: The loss function penalizes violations of the active and reactive power mismatch equations at every bus.
• Advantage: Provides a continuous, differentiable surrogate model of the grid that is trivially parallelizable on GPUs, unlike traditional sequential solvers.

Traditional Weighted Least Squares (WLS) state estimation fails when the grid becomes unobservable due to sensor scarcity. PINNs act as a physics-regularized interpolator.

• Process: The neural network reconstructs the full system state (voltage phasors) from a limited set of SCADA measurements by minimizing both data discrepancy and the residual of the governing power flow equations.
• Outcome: Enables accurate Distribution System State Estimation even when only a fraction of nodes are instrumented, effectively turning physics into a virtual sensor.

PINNs excel at inverse problems, such as identifying unknown inertia constants or damping coefficients in synchronous generators. The network simultaneously approximates the solution to the swing equation while optimizing for the unknown physical parameters.

• Differential Equation: Embeds the second-order swing equation (M * d²δ/dt² + D * dδ/dt = Pm - Pe) into the loss.
• Application: Enables real-time transient stability assessment by discovering degraded mechanical parameters from PMU data without intrusive testing.

Simulating fast electromagnetic transients (EMT) requires solving stiff differential equations with tiny time steps. PINNs offer a mesh-free alternative to traditional EMTP software.

• Architecture: The network takes continuous time as an input and predicts voltage/current waveforms, constrained by the ordinary differential equations of RLC circuits.
• Benefit: Eliminates numerical oscillation artifacts and allows querying the solution at arbitrary temporal resolutions, crucial for analyzing switching surges and lightning strikes.

PINNs can be reformulated to solve constrained optimization problems like Security-Constrained Optimal Power Flow (SCOPF). The objective function (e.g., generation cost minimization) becomes part of the total loss, while inequality constraints (line thermal limits) are enforced via penalty methods.

• Constraint Handling: Violations of voltage limits and N-1 contingency constraints are added as soft penalties, pushing the network toward feasible dispatch solutions.
• Speed: Once trained for a specific topology, the PINN provides near-instantaneous optimal setpoints compared to iterative interior-point solvers.

Pure physics models often miss unmodeled dynamics like inverter lock-up or weather-dependent line sag. Hybrid PINNs combine physics equations with a data-driven neural network that learns the residuals.

• Structure: The total prediction = Physics Solution + Neural Network Correction.
• Use Case: Accurately forecasting renewable generation by coupling the Navier-Stokes-inspired wind model with a data-driven correction for terrain wake effects, outperforming purely statistical or purely physical models.