Improving fault tolerance in diagnosing power system failures with optimal hierarchical extreme learning machine

Accurate diagnosis of failures has a pivotal role to play in the stable operation of power systems. Neural networks have shown promising fault tolerance in solving this problem. However, the widely used BP and RBF networks have a tedious training process and are difficult to provide approving generalization performance. In this work, an optimal hierarchical extreme learning machine (HELM) via adaptive quadratic interpolation learning differential evolution (AQILDE) is designed to address this issue. HELM has good generalization performance but its optimal structure is hard to achieve. Thus, we present AQILDE to automatically search the structure parameters of HELM, including the number of hidden layers, the number of neurons per hidden layer, and the regularization factor. In addition, individual coding method and improved training target function are proposed to ensure the generalization performance and structural compactness. The size of decision variables can be adjusted during the training process. Both regression loss and classification loss are integrated into the target function. The feasibility of AQILDE-based HELM is evaluated in a 14-bus power system and a practical fault in the Siping power grid, China. Simulation results show that it has better generalization performance and diagnoses varied fault scenarios correctly with higher fault credibility.