Long and Short-term Memory Map Attention Incinerator Temperature Prediction Model Based on Hyperparameter Optimization

The incinerator in a nuclear fuel reprocessing plant is used for the controlled combustion and harmless treatment of nuclear fuel waste. Technologically, real-time monitoring of key nodal temperatures within the furnace during the combustion process is essential to guide feed control. This ensures the thoroughness of waste combustion while effectively preventing excessively high temperatures that could damage equipment. Compared to reaction equipment in conventional chemical plants, the incinerator in a nuclear fuel reprocessing plant operates in a high-radiation environment with difficult maintenance access. Consequently, real-time temperature monitoring at key furnace nodes presents significant challenges. To address the challenge of being unable to directly measure key temperatures in the nuclear fuel reprocessing plant incinerator using sensors, an ABCHB-LSTM-GAT model (hyperparameter-optimized long short-term memory graph attention network) was proposed in this paper. The data preprocessing module constructs dynamic graph structures using k-nearest neighbors (KNN) graphs. The LSTM-GAT module employs long short-term memory (LSTM) neural networks integrated with graph attention networks (GAT) to handle spatiotemporal dependencies within sequential data. The optimizer module determines the model’s optimal hyperparameters by combining a Bayesian sampling algorithm with an ultra-wideband dynamic pruning mechanism, enhancing both training efficiency and prediction accuracy. Experimental results demonstrate that this model exhibits significant advantages over baseline LSTM and GAT models in terms of both training efficiency and prediction accuracy. During operational engineering processes, it enables precise prediction of internal furnace temperatures, assisting operators in controlling feed rates and effectively monitoring abnormal conditions such as furnace overheating.