关键词: 核事故, 源项反演, 深度学习, 超参数优化

Abstract: During nuclear accidents, accurate estimation of source term parameters is a critical component of nclear accident consequence assessment and emergency response. However, during the transient phase of an accident, timely acquisition of monitoring data is frequently hindered by conditions including high temperature, intense radiation, and instrumentation failure. Therefore, this paper proposes two deep learning-based integrated models for source term inversion to enhance both accuracy and robustness. The first model is an LSTM-Transformer hybrid architecture, which fully utilizes the advantages of the Long Short-Term Memory (LSTM) network in handling long-term dependencies, non-stationarity, and gradient stability in time series,thereby effectively capturing the dynamic evolution of radioactive nuclide dispersion. It also introduces the Transformer structure to leverage self-attention to modelglobal dependencies between different time steps in the sequence, significantly enhancing long-range contextual awareness . The second model is a dual-branch fusion LSTM model (Dual-Branch LSTM with Fusion, DBL-LSTM), which employs two parallel LSTM networks to independently extract features from gamma dose rate measurements and meteorological parameters, followed by feature-level integration via a fusion LSTM model for comprehensive modeling. This approach improves the capability for joint analysis of multi-source data and enhances adaptability to meteorological variability and noise. The Optuna automated hyperparameter optimization algorithm is integrated to further improve inversion accuracy and reduce uncertainties in manual parameter tuning. Based on the simulated data generated by the International Radioactive Assessment System (InterRAS), the release rates of three short-lived nuclides, Sr-91, Te-132, and I-131, were estimated, with a single LSTM model serving as the baseline comparator and the Mean Absolute Percentage Error (MAPE) as the evaluation indicator. Results indicate that the single LSTM achieved MAPE values of 15.10%, 8.20%, and 27.33% for ⁹¹Sr, ¹³²Te, and ¹³¹I respectively; the LSTM-Transformer model reduced these to 13.90%, 5.63%, and 25.63%; and the DBL-LSTM achieved further improvements, yielding MAPEs of 13.34%, 4.93%, and 24.21%.The integrated models demonstrate consistent improvements in both inversion accuracy and robustness relative to the single LSTM model, highlighting their potential for application in nuclear accident emergency scenarios.

Key words: nuclear accident, source term inversion, deep learning, hyperparameter optimization