Railway in-train forces are vital for evaluating multiple aspects of rolling stock. Conventional methods necessitate substantial investment in manpower and expertise, while only gathering data for specific service conditions one at a time. This paper presents a predictive machine learning approach for railway in-train forces, using data collected from automatic train operation (ATO) systems. The proposed method employs longitudinal train dynamics simulations (LTSs) to establish the correlation between ATO measurements and in-train forces. Subsequently, a self-attention-based convolutional neural network (SA-CNN) is trained to forecast the histories of in-train forces. Results demonstrate that the well-trained SA-CNN model exhibits superior compatibility with arbitrarily combined inputs while reducing computational time. The proposed approach holds promise for quick and reliable in-situ monitoring of railway in-train forces, benefiting research on in-train forces and industrial applications.