Abstract: Since chemical processes are highly non-linear and multiscale, it is vital to deeply mine the multiscale coupling relationships embedded in the massive process data for the prediction and anomaly tracing of crucial process parameters and production indicators. While the integrated method of adaptive signal decomposition combined with time series models could effectively predict process variables, it does have limitations in capturing the high-frequency detail of the operation state when applied to complex chemical processes. In light of this, a novel Multiscale Multi-radius Multi-step Convolutional Neural Network (MsrtNet) is proposed for mining spatiotemporal multiscale information. First, the industrial data from the Fluid Catalytic Cracking (FCC) process decomposition using Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN) extract the multi-energy scale information of the feature subset. Then, convolution kernels with varying stride and padding structures are established to decouple the long-period operation process information encapsulated within the multi-energy scale data. Finally, a reconciliation network is trained to reconstruct the multiscale prediction results and obtain the final output. MsrtNet is initially assessed for its capability to untangle the spatiotemporal multiscale relationships among variables in the Tennessee Eastman Process (TEP). Subsequently, the performance of MsrtNet is evaluated in predicting product yield for a 2.80 × 106 t/a FCC unit, taking diesel and gasoline yield as examples. In conclusion, MsrtNet can decouple and effectively extract spatiotemporal multiscale information from chemical process data and achieve a approximately reduction of 30% in prediction error compared to other time-series models. Furthermore, its robustness and transferability underscore its promising potential for broader applications.