**5. Conclusions**

We proposed a novel MCRNN architecture for the quality prediction of CCS. The major contributions of the new architecture are the transformations of time series input and feature extraction with LSTM and FCN. The proposed architecture can automatically extract the long-term trend and short-term change of time series, which greatly enhances feature learning ability and abnormal slab detecting performance. Extensive experimental results show that traditional methods are more incapable when dealing with messy and unbalanced data, and multiscale convolution and recurrent neural networks outperform other state-of-the-art baseline methods in quality prediction. Accordingly, a real-time quality prediction system based on MCRNN architecture has also been developed. The mold level fluctuation collected by the data module in the system is fed into the trained model. The continuous casting process will be adjusted in real-time based on expert knowledge if there is a high probability of prediction that it is an abnormal slab. The system greatly enhances steelmaking efficiency, improves slab quality, and reduces costs. Due to class imbalance caused by a few abnormal slabs, we use a random sampling method to generate training sets with three different sampling ratios to help mitigate class imbalance. Experimental results demonstrated that the proposed method can detect more abnormal slabs and reduce the misjudgment of normal slabs when the sampling ratio is 1:2.

For future research, although the established quality system has achieved certain results, it is still insufficient in several aspects such as interpretability of prediction and root cause analysis, the sampling method of dealing with the problem of unbalanced data is still worthy of our continued study. In recent years, the interpretability of deep learning is an important research field. In the future, we will utilize the interpretable method and root cause analysis to find out the cause of the abnormal slab, which will further improve the performance of intelligent steelmaking.

**Author Contributions:** X.W., H.J., X.Y. and J.W. (Jianjia Wang) are the main authors of this manuscript. All the authors contributed to this manuscript. Conceptualization, X.W.; data curation, J.W. (Jianjia Wang); methodology, X.W.; software, H.J. and X.Y.; validation, X.Y. and J.W. (Jianjia Wang); writing— original draft preparation, H.J. and X.Y.; writing—review and editing, X.W., H.J., X.Y., J.W. (Jianjia Wang), Z.L., Y.L., J.W. (Jie Wang) and Y.G.; supervision, X.W. All authors have read and agreed to the published version of the manuscript.

**Funding:** This work was supported by the Natural Science Foundation of Shanghai, China (Grant No. 20ZR1420400), the State Key Program of National Natural Sc hasience Foundation of China (Grant No. 61936001).

**Institutional Review Board Statement:** Not applicable.

**Informed Consent Statement:** Not applicable.

**Data Availability Statement:** Data is contained within the article.

**Acknowledgments:** We appreciate the High Performance Computing Center of Shanghai University, and Shanghai Engineering Research Center of Intelligent Computing System (No. 19DZ2252600) for providing the computing resources.

**Conflicts of Interest:** The authors declare no conflict of interest.
