Modelling and forecasting of photovoltaic power output based on machine learning technique

The sun is one of the primary sources of energy for most of the processes on the earth. It is the energy source for heat, wind, and rain. Currently, human kind is using fossil fuels as its primary energy source, but the fossil fuels are non-renewable and they contribute to also air pollution, environmental hazards and etc . This leads to the usage of sunlight as a renewable source of energy to produce electricity, called solar energy. Moreover, solar energy is green energy and it is environmentally friendly. Despite the advantages of solar energy, the generated power in photovoltaic panels depends directly on solar radiation, temperature, humidity, and other factors. These parameters are not constant during the day and therefore the amount of power generated, is typically unknown. Therefore, efficient power planning is impossible unless power generation can be predicted. In this regard, this research aims to propose a method for photovoltaic generated power forecasting. This method consists of neural network and PSO, which finds the optimum structure based on the defined cost function to maintain both accuracy and complexity of the network. It has to be noted that the data which are used in this thesis are collected from KLIA Sepang meteorological data and therefore it does not contain any power values. Therefore, two neural network structures, namely General Regression Neural Network and Feedforward Back Propagation, have been used to model a photovoltaic panel and approximate the generated power. As the results show, the prediction of power using FFBP was more accurate comparing with GRNN. The results then can be used for forecasting part. Such networks are selected due to their popularity, effectiveness, and high level of accuracy. These systems are then validated by the meteorological data of KLIA Sepang at latitude 02°44’N, a longitude of 101°42’E to evaluate their performances. For the modelling part, the five years of data are divided into two parts. From 2006 until 2009 the data used for training and the whole year of 2010 used for testing. However, due to the discontinuity and inconsistency of data because of fault or maintenance in the system, only 814 data have been selected for the forecasting part in which half of data are used for training and the rest for testing. The cost function includes a parameter k which maintains the complexity versus accuracy. In this regard, the increment of k and selection of a value near 1 for this parameter causes the network to become more accurate while the evolved structure will be more complex.

Text 115605.pdf Download (1MB) Official URL or Download Paper: http://ethesis.upm.edu.my/id/eprint/18240

Actions (login required)

View Item