Article Preview
TopForecasting techniques can be categorized into deterministic and probabilistic forecasting (Letendre et al., 2014). Deterministic forecasting yields a single solution, while probabilistic forecasting gives a distribution of possible outcomes. The majority of solar PV forecasting uses a deterministic approach. Furthermore, forecasting models can be either physical or statistical, where physical methods take meteorological data as input to predict PV power, and statistical methods determine a relationship between power generation and meteorological data to make predictions (Huang et al., 2010). In addition, solar PV forecasting can be technically classified into three categories based on forecast horizons: intra-hour, intra-day and days ahead forecasting based on the available data (Kostylev & Pavlovski, 2011). The intra-hour is essential for system maintenance and hazards prevention, while intra-day and days ahead are critical for system management and operation.
In the process of predicting solar PV, machine learning-based methods have proven to be very effective and accurate. Numerous approaches have been carried out and summarized in (Voyant et al., 2017) which reviews the successes of the previous researches. Popular learning methods are Artificial Neural Network (ANN), Support Vector Machine (SVM), Autoregressive Integrated Moving Average (ARIMA) and K-mean, whereas less popular methods include Boosting, Regression Tree and Random Forest. In term of prediction accuracy, ANN, SVM and ARIMA methods give promising results, but the flexibility of ANN and SVM to approximate nonlinearity makes them more preferable than classical ARIMA. In general, the accuracy of these methods depends on the quality of training data.