Structure-Aware Mapping with Dynamical Sources for Informer Weight Initialization in Time-Series Forecasting

Weight initialization determines the starting configuration of a neural network and can influence its subsequent optimization. In Transformer-based forecasting models, different parameter groups also have distinct structural roles that conventional initialization schemes do not explicitly capture. This study develops a structure-aware weight initialization framework for Informer in which deterministic dynamical sequences are transformed into initial weights according to the structural roles of different parameter groups. The proposed procedure combines sequence processing with tensor-specific normalization, clipping, fan-aware scaling, and module-dependent gains to control initialization on feed-forward, convolutional, and projection weights. We evaluate the framework on ETTh1, ETTh2, ETTm1, and ETTm2 benchmark datasets. Compared with the Standard Informer, the proposed configuration achieves lower mean MSE and MAE across the evaluated datasets and forecasting horizons, with the largest reductions observed on ETTh2 at horizon 48, reaching 20.73% and 11.05% for MSE and MAE, respectively. The attribution ablation further shows that the observed gains arise from the combined effect of the structure-aware mapping and the dynamical source. The mapping controls the scale and organization, while the dynamical sources introduce deterministic structural characteristics that persist after processing and are associated with dataset-dependent differences in forecasting behavior. These findings support integrating dynamical sequence generation with structure-aware parameter mapping as an effective initialization strategy for time-series forecasting.