.png){kind=icon}
.png){kind=icon}
Return
A Reliable Framework for Batch Reactor State Response Forecasting
.png){kind=icon}
OA
AI
DOI：10.1109/ACCESS.2026.3668708.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Reliable automation of industrial nitrile butadiene rubber (NBR) batch reactors remains challenging because operator interventions can abruptly change future temperature trajectories and conventional surrogate models rarely indicate when forecasts are trustworthy. This paper proposes Reliable State Response Forecasting (RSRF), a model-agnostic reliability layer that unifies data preprocessing, offline training, online forecasting, Confidence Index (CI) reporting, and periodic model refreshing. RSRF formalizes such interventions as Unrecorded State Change Events (USCEs) and excludes samples whose forecast horizons contain a USCE to prevent spurious input–label mappings. Forecast trustworthiness is summarized by a calibrated 0–1 CI derived from ensemble dispersion. The framework is evaluated on 54 industrial NBR batches using eight representative surrogate architectures. Excluding USCE-contaminated samples reduces mean absolute percentage error (MAPE) by 10%–46%, and CI shows a consistent inverse relationship with forecast error, thereby enabling confidence-aware decision support via error–retention analysis. ResNet-family forecasters (ResNet, TANet, and ResTCN-DAM) achieve the lowest MAPEs and significantly outperform the Transformer-inspired Autoformer (batch-level one-tailed Wilcoxon signed-rank test, $p\ll 0.01$ ); Autoformer even underperforms the DLinear baseline in this non-periodic, short-horizon setting.
Keywords:
Batch reactor
confidence index
deep learning
process control
surrogate model
time-series forecasting
uncertainty quantification
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.