Multi-energy systems, as emerging power system architectures integrated by multiple converter-controlled wind power plants and electrolyzer facilities, face new stability challenges, among which subsynchronous oscillations (SSOs) have emerged as a critical concern. Despite extensive research on SSO analysis, achieving reliable real-time identification remains challenging due to the nonlinear, time-varying nature, and the limitations of conventional signal-processing techniques. To address this challenge, this paper proposes a data-driven framework based on a one-dimensional convolutional and long shortterm memory neural network (1D-ConvLSTM) for the early identification of low-damping SSO in multi-energy systems. A stochastic database is generated through time-domain simulations covering a wide range of operating conditions and contingency scenarios. Oscillatory parameters are estimated using Matrix Pencil method to systematically label critical SSO events based on damping criteria. The proposed 1D-ConvLSTM model is trained using multivariate time-series measurements to enable early SSO identification. Simulation results obtained from a realistic offshore multi-energy system demonstrate that the proposed approach achieves competitive performance in identifying critical SSO events, outperforming standalone Conv1D and LSTM models. These results highlight the strong potential of datadriven artificial intelligence methods for real-time and predictive identification of SSO in multi-energy systems.

Although several data-driven approaches for short-term voltage stability (STVS) assessment have been proposed, most of them do not extend to corrective control actions nor consider the joint dynamics of generation and load. To address this gap, this work introduces a real-time adaptive load shedding scheme (ALSS) driven by an integrated assessment of the short-term stability state (STSS) and the identification of critical induction motors (CIM) as the mechanism driving STVS instability. The methodology employs two recurrent convolutional neural network (RCNN) models operating in parallel: i) the STSS-RCNN, which classifies the system state as stable, unstable by transient stability (TS), or unstable by STVS; and ii) the CIM-RCNN, which identifies the critical motors responsible for instability, thereby inherently recognizing STVS-related problems. The joint operation of these models ensures that the ALSS is activated only when both responses consistently recognize an STVS event. This enables not only the correct activation of the load shedding scheme but also its accuracy and adaptive parameterization based on the identified CIMs. Validation on the IEEE 39-bus test system demonstrates that the proposed approach achieves robust real-time performance, outperforms single deep learning baselines, and significantly overcomes traditional load shedding schemes in efficiency and reliability.