Data-driven power system stability assessment

Abstract

This report aims to construct a data driven DSA system to provide system stability assessment. Three algorithms ELM, RVFL and SCNs as well as method decision making were used in DSA system. Owing to the fact that these three algorithms are all randomization-oriented, decision making allows as much as matrix of input weights to be tried, thus improving accuracy in assessment. Firstly, three individual algorithms were tried. In this phase, parameters tuning for three algorithms was an essential step. After relatively better parameters setting was acquired, individual algorithm was tested in two ways. The first strategy is to running algorithm for one time and record the corresponding accuracy of assessment. The second strategy is based on the philosophy of decision making. Algorithm was allowed to repeat for multiple times, then the averaged results would indicate the final accuracy. Secondly, based on the parameter settings developed in the last section, three algorithms were combined in a hybrid DSA system. Each algorithm was repeated for multiple times, then the final results was analyzed based on three algorithms via decision making method. From the observations in this report, conclusions could be made from two perspectives. The first conclusion was acquired via the process of parameters tuning. For three algorithms, the setting of hidden neurons, setting of input features and activation function were tuned to a relatively better range. The second conclusion was made concerning with the implementation of decision making. With the help of decision making, accuracy in ELM and RVFL could be improved. However, for SCNs, there was no distinguished benefits in accuracy due to the nature of incremental learning. For hybrid DSA system, decision making still provided significance to mitigate the negative impact of uncertainty lied in the performance of randomized algorithms.