Abstract:To address the issues of insufficient prediction accuracy and slow convergence in sinter mixture moisture content forecasting, this study proposes a hybrid prediction model based on an Improved Northern Goshawk Optimization algorithm (UNGO) integrated with a CNN-BiLSTM-Self Attention architecture. The model first employs a Convolutional Neural Network (CNN) to extract spatial correlation features from the input data, then utilizes a Bidirectional Long Short-Term Memory network (BiLSTM) to capture the temporal dependencies of the sinter mixture moisture content. Meanwhile, a Self-Attention mechanism is incorporated to strengthen key feature weighting and long-range dependency modeling, thereby improving prediction stability and accuracy. To reduce model complexity and enhance computational efficiency, the Spearman correlation coefficient is applied to screen input features and remove variables with low relevance. To overcome the tendency of traditional optimization algorithms to fall into local optima, the proposed UNGO introduces chaotic initialization and a nonlinear sine-cosine decay strategy, significantly enhancing both global exploration and local exploitation capabilities. This enables adaptive optimization of key hyperparameters, including the learning rate, number of BiLSTM neurons, regularization coefficient, and attention key dimension. Experimental results based on production data from a large steel plant in Hebei Province demonstrate that, under identical data conditions, the CNN-BiLSTM-Self Attention model achieves MSE = 0. 057 5, RMSE = 0. 239 8, MAE = 0. 192 3, and R2 = 0. 790 8, outperforming CNN, LSTM, CNN-LSTM, and CNN-BiLSTM models. This confirms the model's superior capability in feature representation and temporal dependency learning. After optimization with the UNGO algorithm, the performance metrics are further improved to MSE = 0. 022 4, RMSE = 0. 149 8, MAE = 0. 118 5, and R2 = 0. 918 4, showing faster convergence and higher prediction accuracy than models optimized by NGO, PSO, and WOA. Furthermore, compared with other mainstream prediction models such as UNGO-Transformer, UNGO-BP, and UNGO-XGBoost, the proposed model exhibits superior prediction accuracy, generalization, and robustness. In summary, the UNGO-CNN-BiLSTM-Self Attention model demonstrates excellent overall performance in both structural design and parameter optimization, enabling high-precision prediction of sinter mixture moisture content. It provides a reliable technical foundation for intelligent control and energy efficiency optimization in the sintering process.
