Photovoltaic source modeling and prediction of Maximum power point using neural networks

Kamath, Ravishankar H (2008) Photovoltaic source modeling and prediction of Maximum power point using neural networks. Phd. Thesis thesis, Manipal Institute of Technology, Manipal. PDF ravishankar kamath.pdf - Submitted Version Restricted to Registered users only Download (112MB) | Request a copy

## Abstract

Use of photovoltaic (PV) technology to generate electricity is increasing worldwide. Over the past two decades PV has become well established in remote area power supply, where it can be the most cost-effective choice. PV is also becoming more common in grid connected applications, motivated by concerns about the contribution of fossil fuel use to the enhanced greenhouse .effect and other environmental issues. In designing any power generation system that incorporates photovoltaic (PV) there is a basic requirement to accurately estimate the output from Currently PV panel efficiency is only about 12 - 20 % in their ability to convert sunlight to electrical power, and again the efficiency can drop further with other factors such as solar panel temperature, illumination level and load conditions which are highly non linear in nature. Therefore, accurate identification of optimal operating point and real time continuous control are required to achieve the maximum power output. Non-linear equations, representing the I-V characteristics, are usually utilized to identify the optimal point yielding maximum power, as well as the corresponding voltage and current at any given time. The conventional solar-array mathematical model requires detailed knowledge of physical parameters relating to the solar-cell material, weather condition, solar trajectory, illumination factor, temperature, and load conditions. At times due to lack of this information, the derived mathematical model may be inaccurate the proposed PV array.under varying operating conditions. Good system design is essential to provide reliable systems. An appropriately sized PV array enables consumers, especially of remote area systems, to receive a reliable energy supply at a reasonable cost The main objective of the work is to model a PV source and predict the maximum power point using two methods of artificial neural network namely Back-propagation algorithm and Radial Basis Function and to develop a comprehensive approach with regard to selection of functions and architecture to obtain an optimized trained network. The work is further extended for tuning and optimization of sampling error between the developed models with the help of Kalman estimator by estimating the zero error sampling steps The real time data base is developed by taking the readings on two types of PV panels i.e. poly crystalline silicon and amorphous silicon with fixed tilt angle of 13.2° (latitude) for different seasons of the year i.e. summer, winter and rainy seasons at different environmental conditions The conventional mathematical model is developed using Matlab coding. And the effects on I-V characteristics of cell under different circuit and environmental parameters i.e. effect of ideality factor, series resistance, illumination level and temperature. Results of mathematical model are compared with real time data. The neural network is modeled using BP and RBF with illumination level, temperatureand load voltage as input and average load current, maximum voltage and maximum current as output. The network is optimized using proper functions and architecture. An extensive analysis is done to obtain source modeling and MPPT in one structure. The work is further extended by implementation of an error estimator usmg Kalman estimator in simulink domain to estimate the sampling error difference between the two ANN models as compared to the standard model for zero error reduction. To predict the maximum power from PV panel, training results of back propagation and radial basis function gives similar results with proper selection of function, epochs, error goal and neurons architecture in back-propagation training and with proper selection of spread constant, error goal and hidden layer neurons in radial basis function training. Training using TRAINSCG function should be preferred in Back-propagation algorithm implementation. To get the optimized RBF structure there should be proper selection of SC as well as the number of hidden layer neurons and also the error goal. Spread constant is chosen in such a manner that its value should be larger than the distance between adjacent input vectors, so as to get good generalization,but it should be smaller than the distance across the whole input space. As Radial Basis networks can be optimally designed and also it takes lesser time than Back-propagation for similar results. Hence it is concluded that Radial Basis function is best suited for maximum power point prediction. The implementation of Kalman estimator helps in estimating the error convergence duration of different models to estimate the model tracking accuracy. From simulation results it is found that RBF model gives an error convergence in minimum sampling rate so it will be the most suitable model for estimating tracking error if Model Reference Adaptive Controller (MRAC) scheme is used for solar tracker.

Item Type: Thesis (Phd. Thesis) Photovoltaic,Maximum power point tracking, source modeling, Artificial neural network, Radial basis function, Back propagation network, Kalman filter Engineering > MIT Manipal > Electrical and Electronics MIT Library 06 Jan 2015 09:15 06 Jan 2015 09:15 http://eprints.manipal.edu/id/eprint/141416

### Actions (login required) View Item