Abstract: Accurate and timely multivariate prediction of critical parameters during small modular reactor (SMR) accidents is essential for enhancing operational safety and enabling proactive intervention. Current physical models struggle to capture complex nonlinear dynamics in real-time, while conventional data-driven methods typically focus on single-parameter predictions and lack robust long-term, multi-variable forecasting capabilities. To address these limitations, an integrated prediction framework based on long short-term memory (LSTM) neural networks was developed in this study, combining SHAP-based feature selection with a rolling prediction mechanism for three typical accident scenarios: loss of coolant accident (LOCA), steam generator tube rupture (SGTR), and station black-out (SBO). High-fidelity transient data were generated via RELAP5/SCDAPSIM simulations on the ACP100 reactor, capturing 197-198 seconds post-accident at 1 Hz sampling frequency. The XGBoost-SHAP algorithm was employed to select the four most influential parameters for each accident from 24 thermal-hydraulic candidates, ensuring physical relevance and model interpretability. For LOCA, the selected parameters were pressurizer liquid temperature, water level, pressure, and steam generator heat transfer; for SGTR: pressurizer liquid temperature, total heat transfer, main loop 2 steam outflow, and secondary pressure; and for SBO: hot-leg coolant temperature (loop 1), main loop 1 flow rate, pressurizer pressure, and core power. Multi-input-multi-output LSTM architectures were constructed with dual stacked layers of 60 units each, ReLU-activated fully connected layers, Adam optimization, batch size of 256, and mean squared error loss function. The model used 10-second historical sequences (4 parameters×10 time steps) to predict parameter values at the next time step (4×1). This single-step predictor was then embedded in an iterative rolling scheme to achieve long-term forecasting over nearly 190 seconds. Results demonstrate excellent single-step accuracy, with R² values exceeding 0.983 for all parameters and reaching 0.999 9 for LOCA pressurizer pressure. In rolling prediction, LOCA achieves outstanding performance with mean absolute error (MAE) between 0.004 8 and 0.014 0 across eight test conditions, completing forecasts in under one second. SGTR and SBO scenarios show higher MAE (0.016 2-0.111 7) due to sharp parameter gradients and fluctuations, but maintained reliable trend prediction for situational awareness. This framework provides a robust, computationally efficient tool for SMR accident management, supporting predictive operator intervention and enabling a shift from detection-response to prediction-intervention paradigms. Future work will expand scenario diversity, incorporate longer time-series data, and explore enhanced neural architectures to further improve model generalization and practical applicability in nuclear power plant settings.