Severe transients in nuclear power plants (NPPs) are strongly coupled and highly nonstationary, which makes reliable short-horizon multivariate forecasting difficult for conventional sequence models. To address this challenge, this study develops a hybrid LSTM-Transformer forecasting framework for severe nuclear accident time series and uses the Sparrow Search Algorithm (SSA) as a task-oriented joint hyperparameter optimization tool for nuclear accident forecasting. In this framework, the self-attention mechanism captures long-range temporal dependencies and cross-variable interactions, while the LSTM component strengthens the modeling of short-term dynamics and local temporal memory. SSA is employed as a task-oriented joint hyperparameter optimization tool to adapt key model settings, including the number of attention heads, encoder depth, model dimension, LSTM hidden units, and dropout rate, for severe nuclear accident forecasting. In addition, a regularized training strategy combining dropout and validation-based early stopping is adopted to alleviate overfitting and improve training stability. The main comparison results are reported as mean ± standard deviation over 20 independent runs with the same data split and different random seeds. Experiments on high-fidelity PCTran/APR1400 simulations covering LOCA, LACP, and SLBIC scenarios, together with a severity-shifted LOCA test, demonstrate strong and statistically stable predictive performance. Across the three representative accident scenarios, the proposed framework achieves mean values of 0.943 ± 0.009, 0.951 ± 0.007, and 0.946 ± 0.010, while maintaining about 30% lower mean nRMSE and nMAE than the strongest LSTM-Transformer baseline. A 2 × 2 ablation study shows that regularization mainly improves training efficiency, reducing the required epochs by a range of about 36–41%, whereas SSA primarily improves predictive accuracy through better hyperparameter selection. Their combination provides the best overall generalization. Cross-severity LOCA evaluation further confirms the robustness of the proposed model, yielding mean = 0.885 ± 0.017 and mean nRMSE = 0.100 ± 0.010. The model also achieves low inference latency (P50 = 7.6 ms per sample), indicating its computational potential for near-real-time multivariate forecasting in safety-critical transient monitoring.