Continuous casting technology plays a crucial role in steel production. Variations in the liquid level of the crystallizer during this process can influence both the stability and the yield of the final product, so predicting the fluctuation of the crystallizer liquid level is of great significance. However, traditional pre-diction of crystallizes liquid level fluctuations faces problems such as difficulty in coordinating multiple factors for joint prediction and inaccurate handling of outliers. In this article, we propose a HECNN-LSTM hybrid model to learn and model key time series data of crystallizer operation. In addition, Hampel filters are used to improve data processing. The experimental results present that the HECNN-LSTM model has more accurate predictions compared to traditional deep learning methods, and the predicted values closely match the amplitude of liquid level fluctuations. When the amplitude deviation is +0.03, the hit rate reaches 90%, which demonstrates the significant performance of the proposed model.