NWP Models Combined with Ensemble Models for Hourly Estimation of Global Solar Irradiance

Abstract

The paper presents a 8-months-long dataset containing meteorological variables day of the year, sunrise and sunset time, maximum and minimum temperature, along with numerical weather forecasting model (NWP) dataset factors such as hourly solar irradiance, precipitation, humidity, wind direction, wind speed to train machine learning (ML) models and ensemble models for hourly global solar radiation prediction for City of lakes Bhopal, in the Central region of India. Bhopal has a very diverse climate. During the dry seasons of winter and summer, the sky is clear, and the sun shines brightly for most part of the day leading to solar radiation in a finite range every season. Thus, the climate of this city depicts the climate of the country and is suitable as solar radiation-based case study for further analysis of renewable energy source. The ensemble models were developed using Python scilearn kit code and trained over the dataset, and neural network model of machine learning is implemented through Keras library. The experimental setup identified two strong correlated factors wind direction and day of the year that affects the radiation the most, and results were obtained in the form of smaller values of root mean square error (RMSE), mean bias error(MBE) and r²_score by varying kernel initializer and optimizer, to demonstrate good accuracy of the model. When compared with mathematical model for the given case study of the city, the experimental neural network model in the paper produced better results. Combination of NWP model and ensemble model is found to be fruitful for solar radiance prediction.