Solar Photovoltaic Power Forecasting

Table 4

Summary of the literature review in the short-term time horizon.

References	Methods	Inputs	Best results
[62]	Quantile regression forest (QRF) method and 3 selecting methods, which are previous, KT, and Kolmogorov–Smirnov distance (KS). The result classification is based on the daily clearness index (KTd). At the same time, 3 classes are cloudy, partially cloudy, and clear days.	The past values of power, POA, temperature, wind, and NWP data.	NRMSE = 3.29%.
[63]	Prediction interval centred on the maximum likelihood estimation method, SVR for analysing the relationship between the input data and the NWP data (mesoscale model, GPV-MSM).	The past values of power and NWP of temperature, RH and cloud cover (CC), and extraterrestrial irradiance (EI).	The annual forecast error coverage with prediction intervals = 85–95% and the error aggregation of 1.5%.
[64]	Machine learning with functional analysis of variance (FANOVA), North American mesoscale model (NAM), (NOAA), rapid refresh (RAP), and high-resolution rapid refresh (HRRR).	GHI, DNI, temperature, and wind speed taken from NWP. However, the vertical atmospheric and cloud profiles and surface albedo are used to calculate the DNI.	RAP/HRRR/NAM: MAE is less than 2 MW.
[39]	The gradient boosting (GB) technique for the deterministic prediction technique and K-nearest neighbour (KNN) regression for probabilistic forecasts.	The NWP variables taken from ECMWF and past values of the PV system and from the adjacent PV power plants.	—
[16]	ANN and SVR techniques.	Inverter historical power data, NWP of temperature, wind direction (WD), and solar geometry (SG).	RMSE = 182.6 kWh.
[51]	Probabilistic forecasting based on the voted set of QRF and fixed random forest (RF) methods.	The NWP data and earlier values of power.	—
[65]	The prediction bands based on time series equations and algebraic viewpoint and the test of normality based on the algebraic setting of Jarque–Bera, Kolmogorov–Smirnov, and Lilliefors theories.	The data for one day collected from the rent of two PV systems based in France country.	The mean interval length (MIL), the prediction interval coverage probability (PICP), and the best cooperation between MIL and PICP obtained according to the clear sky index.
[66]	MLP, PHANN, and clear sky radiation model (CSRM) for sunny and cloudy conditions.	Irradiance, temperature, day, and clear sky index.	MAPE = 10%.
[67]	Adaptive-network-based fuzzy inference system (ANFIS) and PSO-ANN models.	One year of input data including actual recorded PV power from the PV system rent in the northeast of Thailand country, solar irradiance, module temperature, and air temperature.	RMSE = 0.1184%.