Multiple digital models of a two-stage proportional cartridge valve are developed using deep learning architectures for digital twin (DT) and digital shadow (DS) applications intended for an automated manufacturing system (AMS). Experimental data are collected from a closed-loop steel manufacturing plant, and input-output relationships are mapped using neural network (NN) regression models. A set of deep temporal models, including recurrent neural network (RNN), long short-term memory (LSTM), gated recurrent unit (GRU), temporal convolutional network (TCN), and multi-scale TCN (MS-TCN), are employed for system modeling. The models are further used for state detection in health monitoring and leakage identification. A comparative analysis is conducted in terms of detection performance, i.e., detecting leaks, predictive accuracy, and computational metrics (floating-point operations (FLOPs), training time, and inference latency). The results indicate that all models are capable of capturing system dynamics and meeting real-time constraints for edge deployment, while the GRU model demonstrates the best overall performance across the considered evaluation criteria.