Short-term and long-term inertia forecasting with low-inertia event prediction in IBR-integrated power systems using a deep learning approach

The integration of renewable energy sources (RES), particularly inverter-based resources (IBRs) such as solar and wind power, has significantly reduced dependence on conventional synchronous generators, thereby decreasing system-wide spinning inertia. This reduction results in rapid changes in the rate of change of frequency (RoCoF), heightening the risk of grid instability. Accurate inertia forecasting is essential for ensuring grid stability, particularly in systems such as the Great Britain (GB) power system, where inertia levels occasionally fall below critical thresholds. However, most traditional and online estimation techniques provide reactive inertia assessments, limiting their effectiveness for proactive grid management. Moreover, existing machine learning (ML)-based models primarily focus on either short-term or long-term forecasting and are often trained on limited datasets, which undermines their robustness and generalisation capabilities. Critically, these models do not prioritise the detection of low-inertia events, which are key moments requiring swift action from grid operators to maintain system stability. To address these limitations, this study proposes a novel hybrid deep learning neural network (DLNN) model that integrates bidirectional long short-term memory (Bi-LSTM) and bidirectional gated recurrent unit (Bi-GRU) architectures to effectively learn complex temporal dependencies in power system dynamics. The model is benchmarked against baseline architectures, including Bi-LSTM, Bi-GRU, and convolutional neural networks (CNNs). The proposed hybrid model achieves superior predictive performance, with a mean absolute percentage error (MAPE) of 2.74%, mean absolute error (MAE) of 4.55 GVAs, root mean square error (RMSE) of 6.65 GVAs, mean squared error (MSE) of 44.22 GVAs${}^2$, and combined accuracy (CA) of 3.70 GVAs. It consistently outperforms the baseline models across seasonal scenarios, achieving MAPE values of 2.09% for Spring, 2.23% for Summer, 2.62% for Autumn, and 2.53% for Winter. For short-term forecasts, the model achieves MAPE values of 1.01% for 12 h and 1.21% for 24 h horizons. In the task of low-inertia event detection, the model demonstrates high precision (0.9538), recall (0.9687), and F1-score (0.9612), highlighting its practical utility in enhancing grid operator decision-making, maintaining frequency stability, and optimising power system operation.