MPF-Net: A computational multi-regional solar power forecasting framework

Short-term solar irradiance forecasting plays a pivotal role in the effective integration of significantly fluctuating solar power into power grids. Existing computational approaches lack to investigate which climate parameter/s influence the most in attaining the optimal forecasting performance. The paper in hand utilizes diverse feature selection approaches to find the optimal subset of features. Using selected subset of features, a rigorous experimentation is performed with 12 adopted machine learning and 10 newly developed deep learning based regressors for most reliable global horizontal irradiance measurements of 9 different regions of Pakistan using 4 evaluation measures. Further, to attain better predictive performance of solar irradiance, we reap the benefits of different individual regressors and present a robust multi regional meta-regressor. Among machine and deep learning based regressors, proposed meta-regressor along with optimal subset of feature/s achieves the best R-2 score of 98% for 6 regions and 97% for other 3 regions of Pakistan. MPF-Net as web service is accessible here.

Automated summary

This paper investigates the importance of short-term solar irradiance forecasting and proposes a multi-regional meta-regressor for accurate prediction using diverse feature selection methods and a variety of regressors. The proposed meta-regressor, along with an optimal subset of features, achieves the best R-2 scores for different regions in Pakistan.