The converter steelmaking process is a crucial stage in iron and steel production, where effective control of oxygen supply significantly impacts the stability of the smelting process and the quality of molten steel. Traditional oxygen supply prediction models often focus on either mechanistic or algorithmic aspects but tend to overlook the high noise levels in the converter environment and the randomness in model training, leading to limitations in their practicality and reliability. To address these challenges, this study proposes a hybrid model based on multi-network optimization for predicting oxygen supply in converters. The model first applies the isolation forest algorithm to remove outliers, and then constructs a hybrid prediction model by combining elastic net with a backpropagation (BP) neural network. Five-fold cross-validation improves the model's generalization ability, and grid search ensures a globally optimal solution. The model is validated on data from a 150-ton oxygen converter in an industrial case study, and its performance is compared with three other models. Results show that the proposed model achieve a prediction hit rate of 76.54% within a ±200 Nm3 error range, and 94.61% within a ±300 Nm3 error range, with an R2 of 0.6512, RMSE of 159.7 Nm3, and MAE≤350 Nm3. This study demonstrates that integrating multiple network optimization methods can significantly improve prediction accuracy and model stability, highlighting the importance of MAE as a key metric for model usability.