<p>Reservoir petrophysical parameters are critical for evaluating the quality of deep tight sandstone reservoirs. To address the challenge of inaccurate petrophysical property prediction due to high heterogeneity, this paper proposes a multichannel spatio–temporal stacking ensemble method constrained by lithofacies (LCMC-stacking). This method was applied to the Sanzhao Sag in the Songliao Basin, China, in the following three steps: (1) well-log data were preprocessed to obtain baseline porosity and permeability datasets; (2) six typical lithofacies were classified on the basis of core and well-log data; and (3) an LCMC-stacking model was established, in which multiple base learners were integrated under lithofacies constraints, feature extraction was enhanced via a time series attention (TSA) mechanism, and a gradient boosting decision tree model was employed as the meta-learner for accurate petrophysical prediction. The results revealed that the LCMC-stacking model achieved prediction accuracies of 94.61% for porosity and 92.99% for permeability, representing significant improvements over all the individual models and 15 ensemble methods, with a maximum accuracy gain of up to 14%. The model also performed well in predicting new wells in the Sanzhao Sag, with prediction accuracies of 93.38% (porosity) and 90.07% (permeability). Additionally, the effects of the base learner combination, TSA module, meta-learner selection, and geological characteristics were analyzed, and fracture development, dissolution, and cementation were identified as the primary geological factors. These findings suggest that the incorporation of lithofacies constraints and multichannel features significantly improves the prediction performance, thereby providing reliable support for the precise evaluation of the quality of deep tight sandstone reservoirs.</p>