Artificial Intelligence (AI) Advanced Techniques for Real-Time Inertia Estimation in Renewable-Based Power Systems

Abstract

In modern power systems with increasing renewable energy integration, accurate estimation of time-varying inertia is critical for ensuring grid stability and resilience. This paper proposes an innovative ensemble of dual-stage hybrid deep learning methods that leverage the strengths of different hybrid networks. The goal is to provide computationally efficient, accurate, and robust inertia estimation in real-time under ambient, dynamic, and challenging operating conditions. The ensemble model comprises a fusion of hybrids of long short-term memory and multilayer perceptron (LSTM-MLP), convolutional-LSTM (ConvLSTM), and temporal convolutional network-LSTM (TCN-LSTM). First, the proposed model is trained and evaluated against cutting-edge methods, using a modified IEEE 39-bus system's data comprising electric power (PE), photovoltaic (PV) power, mechanical power (PM), and rate-of-change-of-frequency (RoCoF). The results showcase the root mean squared error (RMSE) of the proposed method achieves 0.027411, outperforming the above individual methods in inertia prediction by 21.86%, 61.81%, 46.03% respectively. Second, we evaluate the model resilience to unavailability of individual input variables and noise in input variables, achieving low average RMSE values of 0.068782 and 0.044712, respectively, for PM and PE inputs. Additionally, the proposed method demonstrates effective generalization in estimating inertia on new data with entirely different distributions from training data. The results also highlight high sensitivity of RoCoF and PV input to noise, thus providing valuable insights for optimizing model performance in real-world applications. Finally, the approach is further validated on a modified IEEE 68-bus system, achieving significant accuracy and robustness, highlighting its capability to handle larger systems and enhancing its practical applicability.