The complexity of the series-fed microstrip antenna array (SFMAA) synthesis problem increases rapidly with increasing element number. When dealing with complex practical SFMAA synthesis problems, conventional ideal antenna-based and electromagnetic (EM) full-wave simulation-based methods are trapped in performance degradation and time-consuming iteratively “cut-and-try” processes. Machine-learning-assisted (MLA) methods face the “curse of dimensionality,” leading to significantly increased training and prediction times and reduced prediction performance. Here, base element (BE) modeling is used to develop a knowledge-guided and MLA synthesis method for SFMAAs. First, the similarity of BEs with different locations and dimensions is explored using knowledge-guided classification. Then, a cosine-domain learning method with low computational costs is introduced. Next, array synthesis, microwave network cascades, and MLA optimization approaches are used to design SFMAAs. This hybrid knowledge-guided and data-driven method greatly facilitates the synthesis process, achieving low training and prediction times and high accuracy. Furthermore, synthesis applications with three types of SFMAAs and various practical design goals are utilized to verify the effectiveness of the proposed method. Finally, the designed antenna prototypes are fabricated, and the measured results show excellent agreement with the simulation results.