Accurate prediction of key parameters in tobacco primary processing plays a key role in its precise optimization and control. Existing prediction methods usually do not consider time dynamic characteristics, and the performance of multi-step prediction is not good, which cannot meet the actual index needs of tobacco primary processing. In response to the above problems, a multi-step prediction method for key parameters of tobacco primary processing based on the time-varying attention-temporal convolutional network (TVA-TCN) is proposed. Firstly, for the key information in the input variables, the attention mechanism is introduced to capture the information. Then, a multi-channel temporal convolutional network is used to extract the long-term dependence of the data. Finally, by extracting the hidden layer information and output information of the previous time step, the model can be dynamically updated over time, thereby improving the performance of multi-step prediction. Multi-step prediction experiments are carried out using real data of tobacco primary processing, and the results show that the proposed method has obvious advantages compared with traditional methods.