Development of Early Prediction Model of a CRUD-Induced Power Shift by Applying Autoencoder ML Architecture

This study presents a deep learning framework for the long-horizon forecasting of the axial shape index (ASI) to support a few-months early detection of the CRUD-induced power shift (CIPS) in pressurized water reactors. CRUD accumulation and boron hideout, which drive the axial offset anomaly (AOA), are difficult to observe directly in operating plants, motivating the use of data-driven inference models trained on synthetic multiphysics simulations.Using STREAM/RAST-K with CRUD deposition and power-transient modeling, an extensive spatial-temporal data set was generated to represent both full-power and seasonal power decrement operations. Three architectures, a long-short-term memory (LSTM)-based autoencoder (LBAE), a transformer-based autoencoder (TBAE), and a hybrid convolutional neural network–LSTM-transformer (CLT), were trained using staged learning, scheduled sampling, and autoregressive rollout to improve long-sequence stability.The CLT model achieved a mean absolute error of 0.000264 for a 300-day ASI prediction under full-power conditions and 0.000320 under 10-day power decrement scenarios, indicating moderate sensitivity to transient operations. Comparative evaluations showed that the TBAE and CLT models generalized effectively to long-range predictions, maintaining errors within approximately 1% of the allowable ASI margin.These results suggest that hybrid spatial-temporal deep learning models can function as digital twin components for monitoring CRUD evolution and providing a proactive CIPS risk assessment during reactor operation, pointing toward future extensions with multiphysics coupling and uncertainty quantification.