227. COMPARISON BETWEEN CONTRASTIVE LEARNING AND RECURRENT NEURAL NETWORKS FOR POWER SYSTEM INERTIA ESTIMATION (Споредба помеѓу контрастно учење и рекурентни невронски мрежи за за проценка на инерција кај електроенергетски системи)
Abstract
Abstract: The lack of power system inertia is becoming a potential issue as penetration of renewable energy sources in the power system increases. This is a result of an agenda set at worldwide level, to maximize integration of renewables and turn away from fossil fuels. Along with the potential problem of lack of power system inertia comes the difficulty of estimating equivalent power system inertia in a system that is becoming increasingly influenced by power electronics. While model-based analyses are possible, they do become increasingly difficult to solve. As a way to circumvent the inconvenience of estimating equivalent power system inertia, Machine Learning has proven to be a viable option. Recurrent, Convolutional, Physics Informed Neural Networks, including other types of regression focused approaches have been previously analyzed on this topic, and proven to be potentially useful. This paper makes a comparison between two approaches to estimation of equivalent power system inertia. The first approach is proposed by the authors, and it involves combination of Contrastive Learning and Ridge Regression. The second approach is Recurrent Neural Networks, which have been previously implemented on this kind of problem. Both methods are tested on simulated data from the IEEE 24-bus system. Different performance metrics are compared, on different dataset sizes. The results obtained from the study show that the method proposed by the authors produces better results in cases when there is deficiency of training data, leading to the conclusion that the proposed methodology may be potentially useful for such cases.