A Composite‐Kernel Mixture of Experts GPR Framework With Bayesian Hyperparameter Optimization for Accurate Mass Prediction in Greenhouse Solar Dryer

ABSTRACT

This study evaluated Gaussian process regression (GPR) models—Basic GPR, MoE‐GPR, and Bayesian‐optimized MoE‐GPR (BO‐MoE‐GPR)—for predicting product mass during greenhouse solar drying using solar radiation, temperature, and humidity as inputs. Pearson's correlation matrix (PCM) analysis revealed strong positive correlations between radiation and temperature, negative correlations between humidity and radiation, and moderate‐to‐strong negative correlations between environmental inputs and product mass, showing temperature to be the strongest linear correlate of mass loss, while noting that correlation does not establish causal dominance. Among 11 classical regression models, the Exponential model showed the best performance ( R ²  = 0.9936, MAPE = 5.5100%). The BO‐MoE‐GPR achieved superior training accuracy ( R ²  = 0.9996, MAPE = 0.68%) over Basic GPR ( R ²  = 0.9993, MAPE = 1.39%), corresponding to an approximately 51% reduction in training MAPE. Cross‐validation confirmed BO‐MoE‐GPR's robustness ( R ²  = 0.985 ± 0.009, MAPE = 1.07% ± 0.05%) versus Basic GPR ( R ²  = 0.939 ± 0.090, MAPE = 5.24% ± 1.89%), with tighter residual errors (−2% to +2%) and accurate tracking of all drying phases. The composite kernel and MoE framework effectively captured heterogeneous drying patterns, while Bayesian optimization efficiently tuned hyperparameters. The novelty of this work lies in demonstrating that the nonstationary, multiregime nature of greenhouse solar drying data can be explicitly exploited by a GMM‐based mixture of experts GPR architecture with composite RBF + Matérn 3/2 kernels and Bayesian‐optimized hyperparameters, rather than by applying conventional single global GPR models. The framework provides accurate mass predictions from environmental inputs alone and therefore shows promise for future non‐invasive monitoring and model‐based control applications, although these practical benefits remain to be demonstrated experimentally.