Small angle x-ray scattering (SAXS) is a promising metrology technology for complex nanostructures in semiconductor manufacturing. However, parameter reconstruction based on SAXS measurement often faces challenges in achieving high precision and repeatability, due to the increasing complexity of structures and the demands for precise measurement. To address these challenges, a correlation learning-based method is proposed to enhance the accuracy and reduce the uncertainty of the profile reconstruction in SAXS measurement. This method leverages the Long Short-Term Memory (LSTM) mechanism to capture and learn inherent parameter correlation effectively. The precision and reliability of the proposed method are demonstrated through the simulations of synthetic Si gratings. Our method exhibits remarkable measurement accuracy with an improvement of at least 13.9% and the measurement repeatability is nearly 1.4 times higher compared to the previous learning-based methods. We expect that our approach will provide a novel solution for SAXS measurement, enabling accurate and reliable profile reconstruction of nanostructures.