Determining the thermohydraulic characteristics in pillow-plate channels (PPCs) is challenging due to their intricate structure. This work introduces novel geometric parameters specific to a flow cross-sectional area in PPCs and employs a two-stage optimized algorithm for artificial neural networks (ANNs) to enhance its thermohydraulic feature predictions significantly. These parameters enable practical consideration of various PPC structures and help develop a new mean hydraulic diameter (MHD) correlation. Besides the general parameters in PPCs, the novel parameters are incorporated into two two-stage optimized ANNs to predict thermohydraulic characteristics. The traditional method for approximating MHD neglects the actual channel structure and only considers the maximum channel height, leading to deviations of −16.93% to +52.00% from finite element simulated PPCs. Our new MHD correlation and the first ANN, which consider the channel structure, yield more accurate results with deviations of −6.97% to +13.07% and −7.85% to +13.21%, respectively. The second ANN predicts Nusselt numbers and Darcy friction factors with errors of less than 13.04% and 15.83%, respectively, across a wide range of geometric and thermohydraulic conditions. These promising results and presenting the maximum and minimum possible MHD in PPCs by the novel correlation pave the way for more accurate future studies and industrial development of PPCs.