Solar PV Power Forecasting For Smart Grid System (A Case Study of Solar PV Power Plant at Singh Durbar K3 Substation Kathmandu, Nepal)

Abstract

The Photovoltaic (PV) systems are considered as clean and efficient sources of renewable energy with its rapid growth in market and technology over the past years. However, the power output of the PV system is an intermittent and non-stationary random process because of the variability of solar irradiation and weather characteristics. To ensure stable, secure operation and economic integration of solar PV system in smart-grids which is growing in the world as a major source of energy and development in solar technology this power is going to be economical. Data analysis and power forecasting of solar PV power is an important issue for reliability and energy management in the smart grid. In this research, the real power data of Singhdarbar substation (K3), Kathmandu, Nepal has been analysed. The data proposed here has been analysed first to ensure the correlation of solar PV power with different metrological parameters such as Irradiance, temperature and wind. After analysis, the mathematical expression of power in terms of irradiance has been developed by 4th order polynomial regression modelling. The use of multilayer neural network (MNN) and long-short term memory recurrent neural network (LSTM-RNN) to forecasting the output power of PV systems. The use of LSTM further reduces forecasting error compared with the other methods. The proposed forecasting method can be a helpful tool for planning and energy management in the smart grid. This study presents, solar Photovoltaic (PV) power modelling using polynomial regression and artificial neural deep learning techniques. This method has been developed and validated using a 35.58 kWp Solar PV system installed inside the K3 substation of Singhdarbar, Kathmandu Nepal. In this study, first, the PV power data is statistically analyzed and modelled by using the polynomial regression technique and further modelled by using two different neural network namely multilayer neural network (MNN) and long short – term memory networks (LSTM) are examined. For PV power data modelling using deep learning techniques, all recorded data is used and parted into two as 90% for training and 10% for prediction in the various structure of both deep learning techniques in order to find the best deep learning structure in terms of low loss error. With these best models, prediction results show the long short term memory network has a better performance compared to the multilayer neural network and polynomial regression technique.