[Paper Review] Comparative Analysis of Machine Learning Algorithms for Solar Irradiance Forecasting in Smart Grids

Saman Soleymani, Shima Mohammadzadeh|arXiv (Cornell University)|Oct 20, 2023

Energy Load and Power Forecasting4 citations

TL;DR

This study evaluates state-of-the-art machine learning models—random forests, XGBoost, LightGBM, CatBoost, and MLP-ANN—for solar irradiance forecasting in smart grids, using Bayesian optimization for hyperparameter tuning. Results show that random forests outperform other models, and MLP-ANN performance improves significantly with prior feature selection, highlighting the importance of preprocessing in neural networks.

ABSTRACT

The increasing global demand for clean and environmentally friendly energy resources has caused increased interest in harnessing solar power through photovoltaic (PV) systems for smart grids and homes. However, the inherent unpredictability of PV generation poses problems associated with smart grid planning and management, energy trading and market participation, demand response, reliability, etc. Therefore, solar irradiance forecasting is essential for optimizing PV system utilization. This study proposes the next-generation machine learning algorithms such as random forests, Extreme Gradient Boosting (XGBoost), Light Gradient Boosted Machine (lightGBM) ensemble, CatBoost, and Multilayer Perceptron Artificial Neural Networks (MLP-ANNs) to forecast solar irradiance. Besides, Bayesian optimization is applied to hyperparameter tuning. Unlike tree-based ensemble algorithms that select the features intrinsically, MLP-ANN needs feature selection as a separate step. The simulation results indicate that the performance of the MLP-ANNs improves when feature selection is applied. Besides, the random forest outperforms the other learning algorithms.

Motivation & Objective

• To address the challenge of solar power variability in smart grid operations by improving solar irradiance forecasting accuracy.
• To evaluate the performance of modern ensemble and neural network models in predicting solar irradiance.
• To investigate the impact of feature selection on MLP-ANN performance relative to tree-based models.
• To apply Bayesian optimization for hyperparameter tuning across all models to ensure fair comparison.

Proposed method

• The study employs five machine learning models: random forests, XGBoost, LightGBM, CatBoost, and Multilayer Perceptron Artificial Neural Networks (MLP-ANNs).
• Bayesian optimization is used to tune hyperparameters across all models, ensuring optimal configuration for each algorithm.
• Feature selection is applied as a separate preprocessing step for MLP-ANNs, while tree-based models perform intrinsic feature selection.
• Model performance is evaluated using standard regression metrics on a real-world solar irradiance dataset.
• The comparative analysis includes both short-term forecasting and model robustness assessment across different data conditions.
• The study uses a consistent experimental setup with cross-validation and standardized data preprocessing to ensure reliability.

Experimental results

Research questions

• RQ1 Which machine learning model delivers the highest accuracy in solar irradiance forecasting among random forests, XGBoost, LightGBM, CatBoost, and MLP-ANNs?
• RQ2 How does the inclusion of feature selection affect the performance of MLP-ANNs compared to tree-based ensemble models?
• RQ3 To what extent does Bayesian optimization improve the predictive performance of each model in the forecasting task?
• RQ4 Do tree-based ensemble models inherently outperform feedforward neural networks in solar irradiance prediction due to their built-in feature selection?
• RQ5 How do the models compare in terms of generalization and robustness across different solar irradiance conditions?

Key findings

• Random forests achieve the highest forecasting accuracy among all evaluated models, outperforming XGBoost, LightGBM, CatBoost, and MLP-ANNs.
• MLP-ANNs show significant performance improvement when feature selection is applied prior to training, indicating that input quality strongly influences their output.
• Tree-based ensemble models such as random forests, XGBoost, LightGBM, and CatBoost demonstrate superior performance due to their intrinsic feature selection capability.
• Bayesian optimization effectively tunes hyperparameters across all models, enabling a fair and reliable comparison.
• The performance gap between tree-based models and MLP-ANNs widens when feature selection is applied to the latter, reinforcing the advantage of built-in feature importance in ensemble methods.
• The study confirms that preprocessing, particularly feature selection, is critical for maximizing the predictive power of neural networks in solar irradiance forecasting.