Due to limitations such as sensor accuracy, data storage capacity and the efficiency of manual labelling, coarse-grained discrete labels are often used instead of the original continuous values in industrial scenarios. However, discrete labels typically result in information loss, making it challenging to accurately model the underlying continuous dynamic processes. Additionally, differences in distribution characteristics and information granularity exist between discrete and continuous variables, preventing unified prediction. To address these issues, a hybrid time series prediction method called coarse-grained discretized labels-hidden continuity recovery (CDL-HCR) is proposed. Using the hidden continuity recovery strategy, the method infers unobservable continuous variables from observable discrete labels in order to estimate originally unmeasurable continuous variables. Then, combined with a self-supervised learning mechanism, the recovered hidden continuous variables and the original continuous variables are fused in a unified space to improve the prediction accuracy of mixed time series data. During the online application stage, the model can input real-time mixed data directly for prediction. Finally, the proposed method is verified using real drum drying process data, and the results show that it is more accurate than traditional methods for predicting mixed time series data.