A Dynamic Gated Fusion Diffusion Model (DGFDiff) for Multivariate Time Series Imputation in Aluminum Electrolysis Process

Data missing significantly impairs the performance of data-driven models in the aluminum electrolysis industry. The process parameters in aluminum electrolysis are characterized by complex patterns, noise interference, and strong coupling among variables. Traditional time series imputation methods exhibit insufficient information capture capabilities and are sensitive to noise when dealing with such complex dynamic datasets. To address these issues, this paper proposes a Dynamic Gated Fusion Diffusion imputation model (DGFDiff) that takes into account the correlations among multivariate variables. This method is mainly based on the diffusion transformer model. The inverse denoising network adopts an encoder-decoder architecture and is synergistically combined with diffusion-based probabilistic sampling. Specifically, the encoder network is a spatio-temporal decoupling encoder incorporating a global-channel dynamic weighting (GDW) mechanism. It combines the attention mechanism with MLPs based on the GDW mechanism to capture temporal patterns and cross-variables dependencies respectively. The decoder uses a cross-attention mechanism to fuse the local conditional information and the global spatio-temporal information from the encoder. Experimental results on the aluminum electrolysis dataset show that, compared with traditional deep learning methods, the DGFDiff model has certain advantages in solving the imputation problem of multivariate time series with complex patterns.