Neural Networks for Fast Predictions of Transients in Shipboard Electric Power Distribution Systems

Shipboard electrical casualty faults can produce destructive voltage transients and outages that impair ship capacity for casualty fight through. Recent work on relay algorithms to detect faults and coordinate multiple circuit protection devices (CPDs) has demonstrated high-speed relay algorithms capable of reliably detecting transient events within several microseconds of their occurrence. These relay algorithms can be used with new switchgear being developed capable of switching power buses and feeders in several microseconds Ultra-high-speed relays need to distinguish accurately between normal and fault transients. The approach advocated is based on learning normal power system load transients and switching patterns as viewed from the perspective of the CPD. Dynamic polynomial neural network (DPNN) predictors will be used to model both normal and fault transients; their coefficient values, future forecasts, state phase-plan trajectories, and model error residuals can be used to distinguish normal from fault transients. The technique is adaptive and does not require substantial pre-training of the neural networks. Traditional prediction methods (e.g., reduced-order modeling) require computational throughput too excessive to be affordable for real-time implementation. Prediction using DPNNs can be done in real time using low-cost, off-the-shelf digital signal processing hardware.