Impact of Lookback Window Length on Transformer-Based Multi-Step Forecasting of Chlorophyll-a in an Agricultural Reservoir

초록

In this study, a Transformer-based Sequence-to-Sequence (S2S) model for multi-step (72 h) chlorophyll-a (Chl-a) forecasting was developed, and the impact of the lookback window (i.e., historical input sequence length) on the forecasting performance was investigated. By evaluating various input sequence lengths ranging from 12 to 96 h, the results showed that a 48 h lookback window achieved the most robust performance, yielding an average Nash-Sutcliffe Efficiency (NSE) of 0.811 across the entire forecasting horizon and maintaining a stable NSE of 0.762 even at t+72 h. In contrast, the forecasting accuracy at t+72 h declined to NSE values of 0.730 and 0.705 when the 72 h and 96 h input sequences were used, respectively. Cross-attention analysis revealed that the model assigned the highest weights to the most recent past interval (t-1 h－t-7 h). These results suggest that an excessively long lookback window introduces redundant or irrelevant information during training, leading to a degraded forecasting performance. Overall, these findings have highlighted that the optimal configuration of the lookback window is a critical factor for Chl-a forecasting in agricultural reservoirs. This study provides a basis for the design and optimization of future time-series-based water quality forecasting models.

키워드