Arid Zone Journal of Engineering, Technology and Environment, December, 2018; Vol. 14(4): 583-592 Copyright © Faculty of Engineering, University of Maiduguri, Maiduguri, Nigeria. Print ISSN: 1596-2490, Electronic ISSN: 2545-5818, www.azojete.com.ng 583 THREE-TIER NEURAL NETWORK FORECAST OF POWER OUTPUT FROM A MINI PHOTOVOLTAIC PLANT IN OGUN STATE, NIGERIA M. O. Osifeko 1 , O. Folorunsho 1 , O. I. Sanusi 1 , P. O. Alao 2 , O. O. Ade-Ikuesan 2 and O. G. Olasunkanmi 2 ( 1 Department of Computer Engineering, Olabisi Onabanjo University, Ago-Iwoye, Nigeria 2 Department of Electrical and Electronic Engineering, Olabisi Onabanjo University, Ago-Iwoye, Nigeria) Corresponding Author’s email: osifeko.martins@oouagoiwoye.edu.ng Abstract The unreliability of solar energy as an alternative source of electricity is a source of concern to stakeholders. To mitigate this challenge, researchers have proposed photovoltaic (PV) power output forecasting which is aimed at predicting the power output of a PV plant. This study develops and validates a three-tier neural network model for forecasting the output of a mini PV plant located in Ifo, Ogun State, Nigeria. The result of the developed model was compared with a state-of-the-art mathematical model using three statistical tools of mean bias error (MBE), root mean square error (RMSE) and mean average percentage error (MAPE) over a period of three months. From the monthly evaluation, results reveal that the MBE values of the three-tier model were lower than that of the mathematical model with a difference of 0.08, 0.03, and 0.09. In terms of the RMSE, the difference between the three-tier and mathematical model values are 0.07, 0.01 and 0.02. The MAPE differences between the two models were 0.05, 0.00 and 0.02. In all the obtained results, the three-tier model showed a consistently better performance than the mathematical model which validates it as a reliable tool for forecasting the power output of a PV plant. Keywords – PV cell power output, Mathematical modelling, Artificial neural network, PV power forecasting 1. Introduction Inadequate supply of electricity in Nigeria is a problem that has defied many solutions for decades despite several theoretical and practical attempts made by successive governments to solve the challenge. However, it is believed that a systematic investment in renewable energy would solve this problem of inadequate supply of electricity faced by Nigeria. Also, it was predicted that future power systems would have an increased share of variable renewable energy (wind and solar PV) (Ueckerdt et al., 2015). Nigeria is known for her abundant possession of one of such sources, solar. Thus, the country is expected to tap into this resource to tackle her electricity challenge. However, despite the excess availability of solar radiation in Nigeria, the country has not benefited much from the use of photovoltaic system as a means of generating electricity. Apart from high cost, two reasons for this include variability and uncertainty, that is, photovoltaic power output exhibits variability at all timescales and that variability is not easy to predict (Pelland et al., 2013). To reduce the impact of the two factors on solar systems, experts have suggested the use of forecasting methods to predict the power output of solar systems (Tuohy et al., 2015, Voyant et al., 2017). Solar forecasts are used by power sector stakeholders like system operators to schedule generation and ensure adequate flexibility to manage changes in power output (Tuohy et al., 2015). Forecasting methods can be classified as physical or statistical. The physical approach uses solar and PV models to generate PV forecasts. On the other hand, the statistical approach uses historical data to “train” models, with little or no reliance on solar and PV models (Pelland et al., 2013). Some of these methods include Time Series Prediction with Statistical Learning, Sky Imagers, Satellite Imaging, Numerical Weather Prediction and Ensemble Forecasting. Forecast can be carried on different time horizons which includes intra-hour, intra-day and day-ahead. This study uses a blend of both the physical and statistical approach to design, develop and