Cellulose from lignocellulosic biomass is the most promising renewable feedstock which may become a substitute for petrochemical products. However, it is challenging to extract cellulose from biomass because of the structural resistance of lignocellulose. Phosphoric acid plus hydrogen peroxide (PHP) pretreatment is an efficient approach that might be applied to get the cellulose-enriched fraction (CEF) from biomass. This study employed the artificial neural network (ANN) to predict the PHP pretreatment efficiency. The critical conditions, including pretreatment time (t), temperature (T), H3PO4 concentration (Cp), and H2O2 concentration (Ch), were employed as input variables for the ANN model to predict the output variables: cellulose content (C-C), cellulose recovery (C-Ry), hemicellulose removal (H-Rl), and lignin removal (L-Rl). The key parameters of ANN models are selected depending on the root mean square errors (RMSE). ANN models' final optimal topological structure contains one hidden layer with 9, 10, 10, and 12 neurons for C-C, C-Ry, H-Rl, and L-Rl, respectively. The actual testing data fit the predicted data with an R2 of 0.8070–0.9989. Additionally, we computed the relative importance (RI) of input variables on output variables using the Garson equation with net weight matrixes. And the results revealed that Cp and Ch (RI 12.0–62.6%) impacted the effectiveness of PHP pretreatment primarily. T (RI 78.6%) dominates the removal efficacy of hemicellulose, and t (RI 9.5–24.6%) has less influence compared to the other conditions. The study provides insights into the optimization of biomass pretreatment.