The world faces a major global issue of increasing global warming and energy demands. There is an increasing need for renewable and eco-friendly energy sources that produce little greenhouse emissions. Hydro projects need a massive investment; likewise, wind energy is limited to coastal regions. Solar energy investments offer the same or even more benefits at a considerable cost. Tackling these issues, this chapter presents a comprehensive approach for predicting solar power generation using machine learning techniques. The study uses a dataset of 21 meteorological features, the critical being temperature ranges. Various visualization techniques are employed to understand the nature of variables. Preprocessing methods, such as removing constant and duplicate features, and handling data imbalance using SMOTE are applied. Three machine learning regression models—linear regression, elastic net regression, and random forest regression—are compared to identify the best-performing method. Through extensive testing, the study achieved an R2 score of 0.964.
TopIntroduction
Across the world, energy demand is increasing extensively, so there arises a need to supply energy in a sustainable way (Anuradha et al., 2021). Areas of human activities including global development are influenced by the supply and demand of energy. Due to this significant widespread issue and the confounding technical challenges involved in resolving it, a concerted national effort utilizing the most cutting-edge scientific and technological resources is necessary. Solar forecasting is a stepping stone to these challenges. A number of meteorological variables, including cloud cover, solar intensity, and site-specific variables have an impact on a photovoltaic (PV) system's energy output. Varied weather conditions have different effects on solar panels (Elsaraiti & Merabet, 2022). The solar panel consumes a lot more solar energy during the summer months. Yet, whether it is windy or raining, a significantly different quantity of energy is used. They take weather forecasts into account because weather conditions are a major factor in power generation. As a result, solar irradiance which is influenced by a variety of elements including location, time, and weather patterns determines how much energy is produced on any given day. Despite of having many advantages of using solar power, the cost of establishing large-scale solar plants is not so cheap. So, it has become important to forecast the solar power generation before establishing the plant based on the location, its zenith and azimuth angle with the sun, overall climatic conditions, etc.
The number of PV panel installations has surged recently due to the demand for solar PV panels as a sustainable energy source. Because of the increased data accessibility and computing power, machine learning algorithms can now make better predictions. Yet, predicting solar power is difficult since it is heavily reliant on climatic conditions which change over time. It's crucial to use innovative techniques to provide findings that are more reliable and accurate.
Researchers have used wavelet analysis (Gaizen et al., 2020), LSTM (Elizabeth et al., 2022), fuzzy clustering (Yoo et al. 2022), and bias correction (Deo et al., 2023) as prominent methods for solar power forecasting. Fentis et. al (2017) have discussed the wavelet decomposition method to decompose the solar power time series into different frequency components, which are then used as input features for the stacked autoencoder. The output of the stacked autoencoder is further processed by the long short-term memory neural network for final solar power forecasting.
Several hybrids and stacked extreme learning models (El Bourakadi et al., 2023) are explored along with a gravitational search algorithm (Pervez et al., 2019), spectrum analysis (Ahmed et al., 2020), and swarm optimization (Yang et al., 2020).
This chapter undergoes the following Research Objectives (ROs):
TopLiterature Review
Shi et al. (2012) have anticipated the weather classification based on four varieties of weather days: clear sky, clouds, fog, and rain. The correlation between PV power output (one day ahead) and regional forecasts using support vector machines (radial bias kernel) are performed. The forecasting errors are evaluated as Mean Relative Error (MRE): 8.64%, and Root Mean Square Error (RMSE): 2.10 MW respectively.
Andrade & Bessa (2017) have deployed a forecasting framework using a gradient boosting tree in combination with principal component analysis, and smoothing techniques. The information is explored from a grid of numerical weather predictions having Mean Absolute Error (MAE) for solar (16.09%) and wind (12.85%) power to enhance forecasting.