Machine learning approaches for PV power predicting, which one is better?


Photovoltaic (PV) technologies have continued to attract significant research attention as a promising approach for combating environmental pollution and the worsening global energy crisis. To this end, PV installation capacity has tremendously increased in recent years. Despite being promising renewable energy, the intermittence and high instability output of PV power are the greatest challenges for grid utility. This is often caused by fluctuation of the PV output power can be attributed to many factors such as PV module temperature and solar radiation intensity. These factors often make balancing the power supply and demand very difficult for the utility grid, and require effective prediction methods to address.

The available methods for predicting the PV output power can be classified as either direct or indirect. Indirect methods use predefined mathematical and existing PV generation models to predict the environmental factors related to PB generation and PV power output, respectively. Direct methods have been proven to be generally better because they use historical data to predict PV output power using different techniques like support vector regression. Although these methods are useful for predicting PV power output, they have several limitations that may produce inaccurate results, hindering their practical applications. Machine learning methods are feasible solutions for enhancing the stability of PV power. To expand their practical application, it is essential to systematically compare the prediction performance of different machine learning approaches based on bid data to determine the most appropriate one.

On this account, Dr. Xiaoyang Wang from Beijing Institute of Technology, Mr. Yunlin Sun from Guangdong Huaju Testing Technology Company, Mr. Duo Luo from China Shuifa Singyes Energy Holdings Limited and led by Professor Jinqing Peng from Hunan University conducted a comparative study of the performance of different machine learning methods used for predicting short-term PV power output. The impacts of an extremely randomized trees classification (ETC) model for weather type classification on the prediction results was also established. The work is currently published in the journal, Energy.

In their approach, the datasets used were derived from a commercial PV power station in Yangjiang, China. The algorithms were trained to learn the relationship between the PV power output and the weather conditions while considering the historical data for weather and power output. Meteorological data for daytime from 0730 to 1800 were divided into six at 2-hour intervals, which were further divided into four categories using the ETC model according to the generated PV power for individual intervals. Based on weather type classification, nine different machine learning architectures were established and their feasibility in predicting the PV power output was explored by analyzing a large dataset of a commercial power station.

The research team demonstrated the feasibility of the ETC model in predicting weather type and the machine learning models in predicting the PV power output with improved accuracy, both based on the historical data. Among the nine models, the Gradient Boosting Regressor, Lasso Regressor, Support Vector Regressor and Random Forest Regressor models, all based on the ETC, showed better performances than the widely used models under similar conditions. All the models exhibited relatively higher accuracy during periods with stable weather, such as October, November and December, which respectively recorded mean relative error values of 2.07, 1.07 and 1.73. During unstable weather, however, only the Support Vector Regressor model produced better performance. Considering each day period, the prediction accuracy for morning and evening periods was higher and with small mean relative errors.

In summary, the authors successfully demonstrated the use of a novel approach combining artificial intelligence techniques and weather classification methods for accurate short-term PV power output prediction. The results revealed the importance of weather type classification in selecting the most appropriate machine learning model for accurate PV power output prediction due to the changes in the characteristic relationship between the PV power and meteorological data. The proposed models addressed the limitations of the conventional machine learning models. In a statement to Advances in Engineering, Professor Jinqing Peng, the lead and corresponding author stated that their findings provided valuable insights for the appropriate selection of machine learning models for accurate prediction of PV power generation in different conditions.

About the author

Jinqing Peng, professor at Hunan University, China. He received his Ph.D. degree from the Department of Building Services Engineering of The Hong Kong Polytechnic University in 2014. He had been worked at Lawrence Berkeley National Laboratory as a post-doctoral fellow from 2015 to 2017. He joined in Hunan University at 2017. His research interests focus on the R&D of BIPV windows, distributed energy storage, building energy demand response and energy flexibility. He has made a series of academic achievements in the fields of theoretical analysis and numerical modelling of BIPV windows, R&D of high-efficiency BIPV modules, and the modelling and rating of window shading systems. He won the second prize of Science and Technology Progress Award of the Ministry of Education in 2019. Prof. Peng has published more than 150 peer-review publications, including 100 SCI indexed papers. He is the youngest Most Cited Chinese Researchers in the field of building and construction. He is also an active scholar, acting as the editorial board member of the top international journal Applied Energy and Engineering, the subject editor of the International Journal of Building Simulation, the associate editor of the Journal of Thermal Science etc.


Wang, X., Sun, Y., Luo, D., & Peng, J. (2022). Comparative study of machine learning approaches for predicting short-term photovoltaic power output based on weather type classificationEnergy, 240, 122733.

Go To Energy

Check Also

Development of an innovative process for post-combustion CO2 capture to produce high-value NaHCO3 nanomaterials - Advances in Engineering

Development of an innovative process for post-combustion CO2 capture to produce high-value NaHCO3 nanomaterials