Artificial Neural Network-Based All-Sky Power Estimation and Fault Detection in Photovoltaic Modules

The objective of this study is to develop an Artificial Neural Network (ANN) based system for output power estimation and fault detection in PV modules in a minute by minute basis. Such system, after being trained on sample data paths in a supervised manner, is ought to be capable of real-time output power estimation and fault detection in PV modules. The process of data acquisition is carried out during a three-month interval, from Nov. 1st, 2015 to Jan. 31st, 2016 using sensitive measurement equipment and precise mathematical formulas. Resultantly, around 30,000 healthy and faulty per-minute accurate data paths containing the solar altitude and azimuth angles, incident angle, irradiance level (W/m2), PV module output power (W/min) and PV module surface temperature (°C) are acquired and normalized in the range 0 to 1 in order to be fed as input to different ANNs. In order to simulate the faulty operation conditions, the PV module is coated with glasses of different shades of gray color. Two different ANNs of Multi-Layer Perceptron type, namely being the Estimation Artificial Neural Network (EANN) and Detection Artificial Neural Network (DANN) are developed for PV power estimation and fault detection purposes, respectively. Both ANNs are of three-layer fully-connected feed-forward architecture, with input, hidden and output layers. Log-sigmoid activation function is deployed in both ANNs’ hidden layers, while softmax transfer function is utilized in the DANN output layer to solve binary classification problems which lead to fault detections in a PV module. The training process of the ANNs is carried out based on Bayesian Regularization (BR) back-propagation algorithm, using the mentioned collected data paths in a supervised way by providing the ANNs with training output targets in each training epoch. After the training goal for both ANNs is satisfied, the ANNs go through a rigorous testing process with new and unseen input data and no more output targets in order to measure their generalization capabilities. At the end of the testing process, the ANNs are ready to be implemented in real life situations. The ANNs’ implementation is carried out during a 15-day interval from Feb. 1st, to Feb. 15th, 2016. The mentioned interval contains highly meteorologically fluctuating wintry days, providing context for a rigorous performance examinations of the mentioned ANNs. During the implementation period, six different fault simulations namely being the lightgray, dimgray and darkslategray shadings as well as the light, moderate and heavy dirt and dust coverings are homogeneously applied to the surface of the PV module. The first two of the mentioned fault simulations had been used during the ANNs training data acquisition process, but the rest four faults are only demonstrated in the implementation period in order to measure the generalization capabilities of the ANNs for unseen situations. Expectedly, the lightgray shading – light dirt and dust covering, the dimgray shading – moderate dirt and dust covering and the darkslategray shading – heavy dirt and dust covering fault simulation pairs led to almost similar amounts of drops in the PV module output power, whilst the homogeneous fault application technique made the power drops independent of the internal architecture of the PV module. The 15-day per-minute implementation period resulted in 6222 PV module power estimation and fault detection acts carried out by the EANN and DANN. The overall EANN average MAPE between the estimated and the measured PV module output power values is 4.44%, and the DANN sensitivity, specificity, and overall accuracy rates are 97.6%, 99.7%, and 98.6%, respectively. The results are promising and the developed and verified PV module-level output power estimation and fault detection system is expected to be deployed in broader PV fleets after taking the developmental requirements into consideration, thus increasing the efficiency and decreasing the support and maintenance costs of the PV systems in long term. Keywords: Renewable Energy, Solar Energy, Photovoltaic, Artificial Intelligence, Artificial Neural Network, Output Estimation, Fault Detection