In real-world industrial IoT environments, sensor data is often contaminated by external noise, making it difficult to obtain clean data, which affects the accuracy of data-driven time series prediction tasks. To address this issue, this paper proposes a novel time series prediction framework that combines an improved contrast blind denoising autoencoder (CBDMAE) and a TCN-Transformer network, called MoCo-CBDAE-TCN-Transformer. The framework enhances its ability to capture dynamic features of time series data by introducing additional momentum encoders, dynamic queues, and information noise-contrastive estimation (InfoNCE) regularization. It effectively utilizes historical negative sample information by capturing and comparing temporal correlations and noise features. Without prior knowledge of noise and clean sensor data, the network achieves blind denoising of sensor data by capturing and comparing temporal correlations and noise features. The denoised data is then predicted using a TCN-Transformer network, which combines the advantages of TCN"s residual connections and dilated convolutions, as well as Transformer"s attention mechanism, significantly improving prediction accuracy and efficiency. Finally, the experimental results show that, compared with traditional noise removal methods and time series prediction models, the model designed in this paper can achieve better noise removal effects and higher prediction accuracy through simulation verification on a public four-cylinder process dataset. Its real-time processing capability is suitable for deployment in actual industrial environments and provides an effective technical solution for data processing and analysis in industrial Internet of Things.