Traditional methods for predicting the stability of cosmetic antioxidants suffer from lengthy cycles and difficulties in modeling nonlinear coupled factors. This study proposes a hybrid time series prediction model, Att-LSTM-Prophet, which integrates an attention mechanism. This model constructs a high-dimensional feature space by collecting multi-source time series data from accelerated stability tests and real storage environments. It employs an LSTM encoder to extract local temporal dependencies, combines a temporal attention mechanism to automatically capture critical decay nodes, and finally utilizes a Prophet decoder to achieve precise long-term trend prediction. Experiments on a real cosmetic formulation dataset (12 months, 3 lotions, 52 time points) demonstrate that this model reduces the root mean square error (RMSE) and mean absolute error (MAE) for predicting antioxidant activity retention to 2.01% and 1.52%, respectively, with an R² of 0.978. This significantly outperforms single Prophet, LSTM, and non-attention hybrid models. Visualization of attention weights further validated the model's sensitivity to historical steep decline events and its interpretability. These findings provide an efficient, low-cost, and transferable intelligent prediction tool for assessing the stability of antioxidant systems in cosmetics.
Deng et al. (Sun,) studied this question.