Abstract: Under severe strata pressure, the output signals of hydraulic support pin shaft sensors exhibit strong nonstationarity and time-varying characteristics. A single neural network architecture is unable to simultaneously account for multiscale spatial feature extraction and long-term temporal dependency modeling, and it lacks an adaptive weight allocation mechanism during multiscale feature fusion, which limits the generalization performance of error prediction models. To address these issues, a hybrid neural network CNN-LSTM-SAtt that integrated a Convolutional Neural Network (CNN), a Long Short-Term Memory Network (LSTM), and a self-attention mechanism (SAtt) was proposed and was applied to load error prediction of hydraulic support pin shaft sensors. First, a combined method of Variational Mode Decomposition (VMD), Fast Fourier Transform (FFT), and Hilbert Transform (HT) (VMD-FFT-HT) was adopted to construct multidomain features. Then, CNN was used to extract deep spatial morphological features in the frequency domain and time–frequency domain, while LSTM was employed to capture the long-term temporal evolution patterns of time-domain signals. Finally, SAtt was introduced to dynamically assign weights to multidomain features according to signal fluctuation characteristics, thereby establishing a high-precision nonlinear mapping between the sensor load response signal and the excitation signal. The results of five typical loading experiments of pin shaft sensors conducted using a force standard machine indicated that the predicted values of the CNN-LSTM-SAtt model can effectively correct the error components in the load response signals of pin shaft sensors. Compared with traditional models and single neural network models, this model exhibits significant advantages in both prediction accuracy and generalization capability, enabling effective prediction of load errors of hydraulic support pin shaft sensors under complex working conditions.