Abstract: Addressing the challenges of multi-source features(geologic parameters, engineering parameters, and production performance) nonlinear coupling and difficult quantification of production fluctuation uncertainty in shale oil reservoir production forecast, this study proposes a dynamic prediction framework that integrates feature importance screening with Bayesian deep learning. A multi-dimensional feature matrix is constructed, covering matrix permeability, porosity, fracturing parameters, and historical production data. Random Forest Mean Decrease Impurity(MDI) algorithm is utilized to quantify the contributions of parameters to production. And, core time-series, with greatest contributions to production forecast, are screened on the basis of importance sequencing. Furthermore, the Bayesian Neural Network is introduced to represent the weight parameters which describe the nonlinear mapping relation between the input features and production as probability distribution, enabling the quantification of forecast uncertainty. P₁₀, P₅₀, and P₉₀ quantile predictions are generated along with confidence intervals. 532 days of production data from 12 horizontal wells in Daqing Gulong shale oil reservoir demonstrate that the predicted values by the model proposed are extremely consistent with actual values, yielding a P₅₀ prediction MAE of 0.56±0.12 t/d. The average coverage of P₁₀-P₉₀ confidence interval reaches 88.7%±4.3%, and the error distribution exhibits low skewness characteristics. Compared with traditional LSTM and Bayesian Neural Network, this method significantly improves the peak production prediction error and response speed to decline trend breakover. This research offers a dual-dimensional decision-making tool for high-precision forecast and risk early warning for single well production scheme optimization of shale oil. This advancement facilitates the development of intelligent decision-making systems for unconventional reservoirs, moving toward probabilistic and dynamic methodologies.