Predicting Solar Energy Production Using Incremental Machine Learning Techniques

Abstract

Energy is a significant part of life and the economy, with an aggressively increasing demand due to population growth. Non-renewable sources, such as fossil fuels, are rapidly depleting and cannot meet the demand, leading to a reliance on different energy sources. Considering the environmental effects of fossil fuels, many individuals are leaning towards cleaner and renewable energy sources, such as solar and wind power. Solar power holds an important share due to its abundance and ease of implementation. The amount of solar energy produced depends on various factors, such as temperature, photovoltaic radiation, cloud cover, and location. Predictive models considering those factors for solar energy play a crucial role in creating efficient production and distribution networks. Machine learning models are becoming increasingly popular among other predictive approaches thanks to technological advancements. Machine learning is an area of programming that creates mathematical algorithms and models, enabling computers to learn and make predictions without explicit programming. There are different training approaches for machine learning models. The traditional approach divides data into training and testing sets and uses all training data at once. Online (Incremental) learning is the principle of feeding the prediction model with one data point from a training set at a time, often used in sectors where data patterns are variable. This principle can be adapted to various data mining algorithms, including supervised and unsupervised learning. In this study, the suitability of the incremental training approach is tested on solar energy production using six different machine learning models (Linear Regression, Lasso Regression, Ridge Regression, Decision Tree, Random Forest, and Artificial Neural Network). An open-source competition on the Kaggle platform, provided by the American Meteorological Society, is utilized to assess whether online models can outperform traditional models in solar energy predictions. Incremental training methods found to perform better than traditional methods in terms of Mean Absolute Error.