Abstract: The support load of hydraulic supports in fully mechanized mining faces of shallow-buried coal seams exhibits strong spatial non-stationarity and regional differences. At present, most studies only predict the load of a single support, and the experimental design ignores the correlations with adjacent supports, making it difficult to accurately predict support loads in different regions of the working face. To address this issue, a hydraulic support load zoning prediction method based on spatiotemporal feature fusion was proposed. The distribution characteristics of roof deflection and support load in the working face were analyzed based on elastic foundation beam theory. Combined with field-measured support load data, the K-means++ clustering algorithm was used to divide hydraulic supports into regions and to verify the rationality of the obtained zoning. Through feature engineering, the load data of six adjacent supports were introduced as exogenous variables into an XGBoost model. By integrating the advantages of LSTM in capturing long-term dependencies and nonlinear responses, an XGBoost–LSTM combined prediction model was constructed. By optimizing the input window length and output step size parameters, accurate prediction of representative support loads in each region was achieved. The results showed that the support load exhibited a distribution trend of "larger in the middle and smaller at both ends", and the supports were divided into five typical regions. When the window length was 60 and the output step size was 6, the model achieved the best performance in multi-step prediction tasks. The coefficient of determination of the prediction results in each region increased to 0.947, while the root mean square error (RMSE) and mean absolute error (MAE) decreased to 0.303 MPa and 0.152 MPa, respectively. Compared with the suboptimal XGBoost model, RMSE and MAE were reduced by 44.4% and 35.4%, respectively, and the prediction accuracy was higher during critical stages such as periodic weighting.