Data-driven dryout prediction in helical-coiled once-through steam generator: A physics-informed approach leveraging the Buckingham Pi theorem

Abstract

A dimensionally consistent physics-informed neural network, named DimNet, has been developed to predict dryout quality in helical coils. Central to its design is the automated and optimizable dimensionality reduction technique, leveraging the Buckingham Pi theorem, which transforms 11 dimensional physical quantities into 8 dimensionless groups. Rigorous 5-fold cross-validation and cross-fluid testing affirm its performance, with an mean absolute error of 0.0540 and 0.198, respectively. The strategic incorporation of noise during training elucidates pronounced improvements, accentuating the model's adaptability. In comparison to three other neural network architectures, DimNet consistently displays superior accuracy. Ablation experiments have underscored the efficacy of each module within the model's design. Ultimately, while DimNet presents as a promising tool for enhancing thermal efficiency in helical-coiled steam generators, its design principles also shed light on the broader potential and versatility of dimensionally consistent neural architectures.