Estimating photovoltaic power generation: Performance analysis of artificial neural networks, Support Vector Machine and Kalman filter

Current energy policies are encouraging the connection to the grid of power generation based on low polluting technologies, mainly those using renewable sources with distribution networks. Photovoltaic (PV) systems have experienced a wide and high increase in their adoption as an energy source over the last years. Hence, it has become increasingly important to understand technical challenges, facing high penetration of PV systems on the grid, especially considering the effects of uncertainty and intermittency of this source on power quality, reliability and stability of the electric distribution system. On the other hand, the connections for distributed generators, by PV panels, changes the voltage profile on low voltage power systems. This fact can affect the distribution networks onto which they are attached causing over voltage, undervoltage, frequency oscillations and changes in protection design. In order to predict these disturbances, due to this PV penetration, this article analyzes seven training algorithms used in artificial neural networks, with NARX architecture, for the generated active power estimating, and thus the state of the distribution network onto which these micro generators are connected and then compare their best statistical results with the Support Vector Machine (SVM) and the Kalman filter (KF) techniques. The results show that the best training algorithm used for the ANN learning obtained a mean absolute percentage error (MAPE) of 0.02%, while the SVM and KF techniques obtained 0.33% and 3.41%, respectively. Taking in account the other statistical analysis, we concluded that artificial neural networks are more suitable for this type of problem than SVM and KF. In addition, performing the training process with cell temperature data improves the accuracy of the resulting estimations. (C) 2016 Elsevier B.V. All rights reserved.