Data-driven parameterization of fault detection algorithms in high voltage direct current systems using classification trees

Abstract

Reliable fault detection in High Voltage Direct Current (HVDC) systems is critical, but its effectiveness depends on the optimal parameterization of protection algorithms. Traditionally, this parameter selection is a heuristic process, reliant on subjective, experience-based tuning that lacks objectivity and reproducibility. This paper addresses this gap by introducing a novel data-driven framework to automate the parameterization process, deliberately distinguishing its scope from the development of new detection algorithms. Within this framework, Classification and Regression Trees (CART) are applied to four prominent non-unit protection techniques: the Rate of Change of Current (ROCOC), the Rate of Change of Voltage (ROCOV), the reactor voltage-based method (L), and Mathematical Morphology (MM). The models are trained on extensive fault data from both Line-Commutated Converter (LCC) and Modular Multilevel Converter (MMC) systems, simulated in PSCAD/EMTDC (Power Systems Computer Aided Design/Electromagnetic Transients including DC). This process yields optimized, transparent decision trees that provide directly implementable if-else rules for protection relays. The efficacy of the CART-derived parameters was rigorously validated using a Finite State Machine (FSM) implementation against a comprehensive suite of unseen fault scenarios. The results confirm the framework’s effectiveness, identifying ROCOV as the superior algorithm for the LCC system and ROCOC for the MMC system. This outcome highlights the approach’s ability to produce technology-specific solutions. By replacing a subjective art with a systematic and objective science, the proposed framework offers a reproducible and interpretable pathway to enhancing the reliability and performance of protection schemes in modern HVDC grids.