Accurate prediction of key parameters in manufacturing process plays a key role in its precise 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 needs of manufacturing process. In response to the above problems, a multi-step prediction method for key parameters of manufacturing process based on Time-Varying Attention-Temporal Convolutional Network (TVA-TCN) is proposed. First, in view of the limitations of the receptive field of ordinary convolutional network, multi-channel temporal convolutional network is used to extract the long-term dependence of data, and the Softplus activation function is used to reduce the sensitivity of data outliers. Then, a time-varying model structure is proposed, by extracting the hidden layer information and output information of the previous time step, the model can not only be dynamically updated over time, but also can alleviate the disappearance of gradients, thereby improving the performance of multi-step prediction. Finally, multi-step prediction experiments were carried out using the real data of food processing process, and the results showed that this method has obvious advantages compared with the traditional method.