Now-casting Photovoltaic power with wavelet analysis and Extreme Learning Machines

High penetration of Photovoltaic (PV) systems, a variable resource, poses challenges to the stability and power quality of electrical grids. Forecasting accurately the PV power has been recognized as a way to ease this problem. This work addresses now-casting of PV power with Extreme Learning Machines (ELMs) without exogenous inputs. Wavelet decomposition and multi-resolution analysis is the most effective way to achieve high accuracy for 5 min-ahead forecast up to 70% greater than the persistence model. A neural network evaluation algorithm based on multiple initializations and incremental hidden nodes is applied and ELMs performance and computation efficiency is evaluated versus Time Delay Neural Networks (TDNNs) for time and time-frequency domain forecasting.