PV Generation Prediction Using Multilayer Perceptron and Data Clustering for Energy Management Support

Accurate PV power generation forecasting is critical to enable grid utilities to manage energy effectively. This study presents an approach that combines machine learning with a clustering methodology to improve the accuracy of predictions for energy management purposes. First, various machine learning models were compared, and multilayer perceptron (MLP) outperformed others by effectively capturing the complex relationships between weather parameters and PV power output, obtaining the following results: MSE: 3.069, RMSE: 1.752, and MAE: 1.139. To improve the performance of MLP, weather characteristics that are highly correlated with PV power outputs, such as irradiation and sun elevation, were grouped using K-means clustering. The elbow method identified four optimal clusters, and individual MLP models were trained on each, reducing data complexity and improving model focus. This clustering-based approach significantly improved the accuracy of the predictions, resulting in average metrics across all clusters of the following: MSE: 0.761, RMSE: 0.756, and MAE: 0.64. Despite these improvements, further research on optimizing the MLP architecture and clustering methodology is required to address inconsistencies and achieve even better performance.