The development of robotic exoskeletons for human motion assistance and rehabilitation has intensified the need for efficient methods to discern users' motion intentions, particularly for complex upper limb movements. The shoulder joint poses the most significant challenge due to its anatomical complexity, multi-degree-of-freedom (DoF) movements, and variable muscle function. Consequently, real-time estimation of shoulder joint motion intentions remains the least explored. Existing prediction methods often fail to address the shoulder’s range of motion, as many models are restricted to below the shoulder level or rely on computationally intensive physiological parameters unsuitable for real-time applications. The shoulder's varying moment arm further complicates signal normalization and accuracy. To overcome these challenges, we propose a novel approach combining a Neuromusculoskeletal Model with a Multilayer Perceptron (MLP). This method leverages biological insights from the model to improve neural network performance, focusing exclusively on regression for real-time predictions. The method inputs include sEMG signals from shoulder muscles and shoulder angle data. These sEMG signals are preprocessed and integrated into the Neuromusculoskeletal Model, which incorporates activation dynamics, contraction dynamics, and musculoskeletal measurements. By correlating skeletal measurements with humeral elevation, the method reduces computational complexity and calculates muscle torques, which are then used by the MLP to estimate humeral acceleration. This approach effectively addresses normalization issues by correlating shoulder angle, sEMG data, and estimated torques with humeral acceleration.