Temporal Fusion Transformers

TFT employs instance-wise feature selection at each time step, automatically identifying which exogenous inputs (e.g., voltage, frequency, load) are most relevant for predicting rotor angle stability. This gating mechanism suppresses noisy or irrelevant variables, improving generalization on sparse PMU data streams.

• Uses Gated Residual Networks (GRNs) for non-linear processing
• Applies softmax-based attention over input features
• Eliminates manual feature engineering for transient stability assessment

Unlike standard transformers, TFT modifies the multi-head attention mechanism to produce explainable attention weights that reveal which past time steps the model focuses on when predicting future rotor angle trajectories. This is critical for control room operators who need to understand why a stability margin is degrading.

• Shares values across heads, aggregating with learnable importance weights
• Exposes temporal attention patterns for forensic analysis
• Enables compliance with NERC reliability auditing requirements

TFT integrates time-invariant metadata—such as generator inertia constants, transformer impedance, and network topology—through dedicated encoder networks. These static features condition the temporal dynamics, allowing a single model to generalize across different fault locations and grid configurations.

• Encodes generator parameters (H-constant, damping coefficient)
• Conditions recurrent layers on network topology embeddings
• Enables zero-shot transfer to unseen contingency scenarios

TFT produces full probabilistic forecasts by predicting multiple quantiles (e.g., 10th, 50th, 90th percentiles) of the post-fault rotor angle trajectory simultaneously. This quantifies the uncertainty in transient stability predictions, enabling risk-based decision-making for remedial action schemes.

• Outputs prediction intervals for critical clearing time estimates
• Uses pinball loss for asymmetric error penalization
• Supports conservative vs. aggressive RAS arming strategies

TFT combines LSTM encoders and decoders with self-attention, processing raw time-series data locally before applying global attention. This hybrid design captures both short-term electromechanical oscillations and long-range dependencies spanning multiple swing cycles.

• LSTM encoder compresses historical PMU windows into context vectors
• LSTM decoder generates multi-step rotor angle predictions
• Attention layers operate on flattened recurrent outputs for efficiency

TFT is architected for low-latency deployment in online stability monitoring systems. Once trained, the model processes streaming synchrophasor data and outputs stability assessments within milliseconds, meeting the sub-cycle response requirements of wide-area protection schemes.

• Inference latency typically < 10 ms on GPU hardware
• Supports batch processing of multiple contingency scenarios
• Compatible with edge deployment in substation IEDs via ONNX export