Steam drum water level is recognized as a key factor associated with industrial boiler operations. Therefore, it is important to predict the changes in the steam drum water level to ensure safe operation of the boilers. Due to the complexity of the boiler processes, multiple correlated process variables are found to be responsible for the variations in the steam drum water level. In this study, the key variables that significantly influence the changes in the steam drum water level were successfully extracted by combining the approaches of kernel principal component analysis and maximal information coefficient. Subsequently, a long-term prediction method for the steam drum water level based on gated recurrent unit network was created. To verify the effectiveness, this method was applied to actual historical operating data for steam cylinders. In comparison to long short-term memory neural networks, traditional recurrent neural networks, convolutional neural networks, and deep neural networks, the proposed approach was proved to be more effective.