To address the challenges posed by complex perturbation environments and noise interference affecting line-of-sight (LOS) micro-angle measurement systems onboard remote sensing satellites during in-orbit operations, this study proposes a denoising method that integrates Bidirectional Long Short-Term Memory (BiLSTM) and Convolutional Neural Networks (CNN) to enhance the accuracy and reliability of measurement data. By combining physical modeling of micro-angle measurement with Monte Carlo simulations, the noise distribution and spatiotemporal correlations in the data are systematically analyzed and validated. A high-quality labeled dataset is then constructed from existing in-orbit measurements to ensure robust and generalizable model training. In the proposed BiLSTM-CNN architecture, BiLSTM captures bidirectional temporal dependencies, while CNN extracts local spatial features. A gradient balancing mechanism is incorporated to mitigate issues such as gradient vanishing and overfitting, thereby improving model stability under complex conditions. Experimental comparisons with typical neural network models show that, on the a1-axis, the proposed model reduces MSE, RMSE, and MAE by 7.9%, 4.3%, and 16.4% respectively compared to the best-performing BiLSTM, and on the b1-axis, achieves respective reductions of 4.6%, 2.3%, and 6.4% compared to the best-performing GRU. These results demonstrate the robustness, generalizability, and effective noise suppression capability of the proposed method, offering a practical and promising solution for high-precision satellite attitude measurement.