Global Forecasting Model for LED Lumen Degradation: An Optimal Cluster Estimation Method

The degradation process of Light-Emitting Diodes (LEDs) is considerably slow, making lifespan estimation through traditional testing impractical and cost-ineffective. Data-driven methods are also challenged by this slow degradation. Testing an LED for 10,000 hours only results in 11 data points, a very short time series for the effective application of machine-learning methods. This study introduces a novel approach utilizing Global Forecasting Models (GFMs) that learn across time series, in contrast to local methods which fit separate models to individual time series. Leveraging an LM-80 dataset of 4,831 samples, each tested for 10,000 hours, we compare our GFM approach with the standard TM-21-11 method. Our results demonstrate significantly improved accuracy over the traditional method. GFMs offer flexibility in integrating additional stress conditions, device information, and feature extractions, promising further advancements in LED lifespan prediction. Additionally, this work introduces a new clustering algorithm that aims to estimate the group of series that gives the best model accuracy without an iterative process. Compared to the only existing algorithm, the suggested method is much faster and yields better results. Across all series, a global LightGBM (no clustering or exogenous/categorical inputs) reduces error versus TM-21-11 by 44.5% (SMAPE) and 36.9% (MASE); applying GDMC clustering further improves accuracy.