Performance of RF regression model.

The direct current (DC) in optical orthogonal frequency division multiplexing (DCO-OFDM) scheme is commonly adopted in light fidelity (Li-Fi) technology as it offers a spectrally efficient solution. A prior study adopted a machine learning (ML)-based solution to predict the optimum DC bias using key parameters, including the statistical properties of the OFDM transmitted signal and a polynomial regression model. However, the model’s robustness decreased when the data structure was shuffled, indicating limited generalization to unseen data. This study builds upon that work by utilizing the same dataset and improving the prediction model with advanced ML tools, such as the LazyPredict algorithm (LPA), to systematically evaluate and select a regression model. A robust ML regressor selection process is proposed to ensure the reliability of predictions. Additionally, a comprehensive data analysis is conducted to assess the importance of features affecting the optimum DC bias. The results demonstrate that the ensemble learning algorithm, Random Forest (RF), outperforms other models with an R-squared of 0.953 and an RMSE of 0.233. A Friedman statistical test was applied to validate the results over five iterations of model training. Furthermore, hyperparameter tuning and bootstrap sampling were employed to conduct a deeper investigation into the model’s performance and stability. The proposed model significantly enhances the accuracy and robustness of DC bias prediction compared to previous approaches, ensuring consistent performance across different data distributions.