Multi-Granular Embedding and Hybrid Encoder-Decoder Architecture with Temporal Linear Regression Bypass for...

In this work, we propose a robust deep encoder-decoder neural network that integrates multiple deep sequence models for time series forecasting in sensor-based applications. The proposed approach features a sophisticated input embedding layer that integrates scalar features, position encodings, and multi-granularity time embeddings. To improve computational efficiency, we employ a generative-style decoder inspired by the Informer architecture, which enables full-length sequence prediction in a single iteration, avoiding the latency associated with traditional autoregressive models. The architecture is highly flexible, supporting either Gated Recurrent Unit (GRU) or Transformer blocks as foundational layers to process latent temporal dependencies. Recognizing that purely nonlinear deep learning models often struggle to capture sharp, short-term local variations due to inherent input scale insensitivity, we augment our model with a parallel temporal linear regression unit. This bypass mechanism captures linear local trends, effectively combining the predictive power of nonlinear deep feature extraction with the sensitivity of linear regression. Experimental evaluations on the Electricity Consumption (EC) and Speed Index California (SIC) datasets demonstrate that the GRU-based implementation reduces the average Mean Squared Error (MSE) by approximately 22% and 16%, respectively, compared with the best-performing baseline across different output lengths. Ablation studies further confirm that the temporal linear regression bypass substantially improves forecasting accuracy, particularly on the EC dataset, where it provides an average MSE reduction of approximately 55%.