Horizon-Adaptive Residual Learning for Multi-Horizon Phycocyanin Forecasting

Published in 2027 IEEE Symposium Series on Computational Intelligence (SSCI), 2027

Phycocyanin fluorescence serves as a continuous indicator of cyanobacterial biomass, with multi-horizon forecasting aiding early bloom detection and water treatment planning. Long-term prediction presents challenges due to diminishing dependence between current and future states as lead time increases. This study redefines multi-horizon phycocyanin forecasting as a residual estimation problem: each model predicts the deviation of the future value from the most recent observation, inherently maintaining persistence. We evaluate direct and residual variants of Random Forest, XGBoost, and LightGBM models; introduce the Horizon Conditioned Residual (HCR)-LightGBM, which shares residual structure across different lead times by incorporating lead time as an input; and propose the Horizon Adaptive Residual Persistence (HARP) Ensemble, combining persistence, seasonal climatology, and HCR-LightGBM using non-negative weights optimized on validation data. Experiments utilize 15-minute interval records from two USGS stations near Houston, Texas, at six forecast horizons ranging from 12 hours to 10 days. Relative to the best-performing direct models, residual and hybrid methods decrease the mean root-mean-square error (RMSE) by 19.57% at Station A and 27.03% at Station B. Additionally, they maintain positive $R^2$ values at longer horizons where direct models often exhibit negative values. The best forecasting method remains station- and horizon-dependent.