@inproceedings{Nijhum2026,
 title = {Evaluating the Role of Input Window Size and Horizon-Specific Models in Multi Horizon Solar Irradiance Forecasting},
 booktitle = {International {{Conference}} on {{Image Processing}} and {{Pattern Recognition}} ({{IC-IPPR}} 2026)},
 author = {Nijhum, Ifran Rahman and Lillis, David and Dev, Soumyabrata},
 editor = {Tian, Dongping and Cao, Changqing and Pu, Yifei and Liu, Gang},
 year = {2026},
 month = {September},
 pages = {21},
 publisher = {SPIE},
 address = {Xianyang, China},
 doi = {10.1117/12.3124578},
 url = {https://www.spiedigitallibrary.org/conference-proceedings-of-spie/14388/3124578/Evaluating-the-role-of-input-window-size-and-horizon-specific/10.1117/12.3124578.full},
 urldate = {2026-09-28},
 abstract = {Solar irradiance forecasting is key to reliable smart grid operation under high photovoltaic penetration. However, the interaction between historical input window length, forecasting horizon, and learning strategy remains insufficiently understood. In this study, we investigate the coupling between temporal window size and multi-horizon forecasting at different lead times using lightweight recurrent models. We investigate the performance of single-task models on predicting individual horizons with multi-task variants of similar models to forecast multiple horizons at once. Observations show that gated recurrent units perform best in single-task settings regardless of the window size and forecasting horizon. A simple multi-task RNN obtains comparable RMSE with reduced model redundancy. The results provide practical guidance for engineering efficient and scalable short-term solar forecasting systems.},
 isbn = {979-8-90232-888-9 979-8-90232-889-6},
}