Neural network-based diagnostics for PV plant

Abstract

The energy performance of PhotoVoltaic (PV) fields has to be constantly monitored after setting a PV plant up. For this aim the paper proposes a Neural Network (NN)-based diagnostic methodology, able to detect the performance losses due to typical PV plants failure modes like ageing, dust deposition on modules glass, and so on. The NN has a feedforward architecture, with one input layer, one hidden layer and one output layer, and is trained through a supervised learning approach. In the training phase, the NN learns the correlation between the power produced by a PV string and environmental parameters such as modules temperature and solar radiation. Moreover, a dependency on the previous monitored operating point is given in order to provide the NN with a sort of dynamicity. Then, the main parameter evaluated is the error computed by the algorithm in the estimation of the cumulative sum of produced energy for a fixed number of successive time instants. This quantity is analyzed through a suitable statistical test so that, in the end, information on the operative state of the PV string can be extracted. The procedure has been tested on an operating PV plant and results confirm the goodness of the proposed approach.