Winter jujube is favored by consumers for its rich nutritional content, sweet taste, and excellent flavor. With the increasing demands of consumers for fruit quality, Soluble Solids Content (SSC) has become a key factor in fruit quality evaluation, serving as an important indicator for measuring fruit ripeness and taste quality. Therefore, efficient and non-destructive prediction of winter jujube SSC has significant practical value and importance. This paper proposes an LSTM-TE model, which integrates Long Short-Term Memory (LSTM) networks with a Transformer Encoder, aiming to achieve rapid and non-destructive prediction of winter jujube SSC. By collecting hyperspectral data from 900 winter jujube samples and determining their SSC values, multiple spectral data preprocessing methods (including Multivariate Scatter Correction (MSC), Vector Normalization (VN), Savitzky-Golay (SG) filtering, first derivative (D1), and second derivative (D2)) were applied to process the data. The effects of 10 preprocessing combinations were compared through five models: PLSR, SVR, VGG16, ResNet18, and LSTM, to determine the optimal preprocessing scheme, MSC-SG-D1. Based on this preprocessing method, a multi-model comparison system was further constructed, including PLSR, SVR, VGG16, ResNet18, LSTM, and LSTM-TE, and their performance was analyzed on the test set. Experimental results showed that the LSTM-TE model achieved a coefficient of determination of 0.959 8 and a root mean square error (RMSE) of 1.269 0 on the test set, an improvement of 17.4% compared to the traditional machine learning modelPLSR ($R_{\text{p}}^2$=0.817 3) and 10.9% compared to the single LSTM model ($R_{\text{p}}^2$=0.865 2). This model effectively explored the nonlinear feature relationships in hyperspectral data by capturing temporal features through LSTM and leveraging the global dependency modeling advantages of the Transformer encoder. This study provides a new technical solution for the online detection and grading of winter jujube quality, offering important reference value for applying hyperspectral technology in precision agriculture.