Sinter ore is the main raw material of the blast furnace, and burn-through point (BTP) has a direct influence on the yield, quality, and energy consumption of the ironmaking process. Since iron ore sintering is a very complex industrial process with strong nonlinearity, multivariable coupling, random noises, and time variation, traditional soft-sensor models are hard to learn the knowledge of the sintering process. In this article, a multistep prediction model, called denoising spatial-temporal encoder-decoder, is developed to predict BTP in advance. First, the mechanism analysis is carried out to determine the relevant-BTP variables, and the BTP prediction is defined as a sequence-to-sequence modeling problem. Second, motivated by the random noises of industrial data, a denoising gated recurrent unit (DGRU) is designed to alleviate the impact of noise by adding a denoising gate into the GRU. In this case, the encoder with DGRU can better extract the latent variables of original sequence data. Then, spatial-temporal attention is embedded into the decoder to simultaneously capture the time-wise and variable-wise correlations between the latent variables and the target variable BTP. Finally, the experimental results on the real-world dataset of a sintering process demonstrated that the integrated multistep prediction model is effective and feasible.