In the industrial internet, time series forecasting plays a central role in enhancing the level of intelligent manufacturing. However, current mainstream deep forecasting models still encounter significant challenges, such as overfitting and inadequate exploitation of dependencies in time dimensions when dealing with complex industrial time series prediction tasks. These issues largely arise from the non-stationary nature of industrial time series data distributions and the limitations of methods that identify dependency relationships through point-wise mappings. To address above challenges, the paper proposes FrePDL, a progressive decomposition learning framework based on frequency domain MLP. FrePDL mitigates overfitting by eliminating redundant information through implicit decomposition. Additionally, it enhances global dependency mining by incorporating a complex MLP designed for spectral analysis. Employing gated mechanisms and progressive refinement, the model exhibits improved forecasting performance. Experiments conducted on six time series datasets derived from diverse industrial internet scenarios demonstrated that FrePDL improved forecasting accuracy, mitigated overfitting, and enhanced information extraction efficiency compared to state-of-the-art forecasting models.