Research on a Modular Intelligent Prediction Framework Based on Structural Recognition and Temporal Steady-State Fusion

Abstract

In multivariate predictive modeling, traditional models are often difficult to balance prediction accuracy, robustness and interpretability in the face of challenges such as high dimensionality, strong noise, nonlinear coupling and structural uncertainty. To this end, this paper proposes a new modular intelligent modeling framework - KFPS-Net. The framework integrates four key functional modules: temporal denoising, nonlinear feature modeling, causal structure exploration and integration fusion, corresponding to sequence signal stabilizer, feature perception modeler, structure explorer and integrated prediction engine, respectively. Through the collaborative work between sub-modules, KFPS-Net effectively realizes the closed-loop optimization of the system from data preprocessing to structural interpretation to the final prediction output. The experimental part was evaluated by using multiple high-dimensional multivariate time series datasets, and multiple sets of baseline models were set up for comparative analysis with ablation experiments. The results show that KFPS-Net is significantly better than the existing mainstream methods in terms of MAE, RMSE and R², and still maintains strong stability under noise disturbance and structural change scenarios. At the same time, the causal structure output further enhances the interpretability of the model and provides a theoretical basis for reliable prediction and systematic decision-making. This work provides a unified and scalable solution for building an intelligent prediction system with robustness and structural cognition.