IOSR Journal of Applied Physics (IOSR-JAP) e-ISSN: 2278-4861.Volume 9, Issue 4 Ver. III (Jul. – Aug. 2017), PP 01-07 www.iosrjournals.org DOI: 10.9790/4861-0904030107 www.iosrjournals.org 1 | Page The Use of Autoregressive Moving Average and Artificial Neural Network as Short Term Wind Speed Forecasting Tools for Lagos, Nigeria * Joseph Aidan 1 , Issa Abdourahamane Ide 2 1 Department of Physics, Modibbo Adama University of Technology, P. M. B. 2076, Yola, Adamawa State,Nigeria 2 Departmwent of Physics, Faculty of Sciences, Gombe State University, Gombe. Gombe State, Nigeria. Corresponding Author: Joseph Aidan Abstract: Autoregressive moving average (ARMA) and Artificial neural network (ANN) models were applied to a univariate response series of hourly wind speed data to forecast future wind speeds. After model identificatio n, parameter estimation and diagnostic checking; and according to Bayesian Information Criterion (BIC), ARM A (2, 2)(1, 1) model was found suitable. The ANN design and training sampling revealed that a 5, 5, 1 neuron st ructure with sigmoid transfer functions at both the hidden and the output layers was the most suitable for the wi nd speed forecast for Lagos. The models were compared and evaluated on the basis of their performance for 1 t o 6 hours ahead forecast. The maximum values of MAE, RMSE and MRE were found to be 0.098, 0.131 and 0.1 00 respectively for the ANN model; and 0.318, 0.403 and 0.733 respectively for the ARMA model. Hence ANN m odel has proven a better forecasting model for Lagos than the ARMA model. Keywords: Wind speed forecast, Autoregressive moving average, Artificial neural network. --------------------------------------------------------------------------------------------------------------------------------------- Date of Submission: 15-06-2017 Date of acceptance: 26-07-2017 -------------------------------------------------------------------------------------------------------------------------------------- I. Introduction Owing to wind variability, forecasting has become a necessary tool to locate sites for optimum wind tur bine performance. To this end, several forecasting methods have evolved over the years among them are the Aut o Regressive Integrated Moving Average (ARIMA) and the Artificial Neural Networks (ANN). These forecastin g models have been used severally by authors [1, 2, 3, 4] to forecast short and long term wind speeds for differe nt locations. In Nigeria, these methods have not been made popular. Only one or few others [5] have used them f or wind speed forecasting. What is lacking is a short term forecast. Short term (1to 6 hours ahead) forecasting is a veritable tool that can be used to determine the reliability of candidate sites for wind power development. So metimes, wind speed forecast do help power system operators to dispatch more economically. However,some of the attempts for short term forecast though very few, lack detailed and proper inte rpretations of processes hence difficult to be understood and applied. In this work, ARMA and ANN models for short term hours ahead wind speed forecast have been developed and applied to the wind speed data of Lagos. S imple and more detailed interpretations of the processes involved have been made and results from the models c ompared using some goodness of fit tests. II. Data And Data Analysis 2.1 Data source and processing A six months (184 days) secondary wind speed data, recorded at 5 minute intervals at UNILAG CR100 0 weather station in Lagos State has been used in this study. The data which was converted to hourly steps using Microsoft Excel were divided into two sets: 179 day for training (model fitting) and 5 day for validation. Figures 1 (a & b) Shows the time series plots of the observed and quartic root wind speeds of Lagos for the first ten days. The wind speed series plot (Figure 1a) exhibited seasonal pattern consisting of one major pea k and several minor peaks daily. The major peaks appeared to be bigger or smaller over the days, indicating the non constancy of the seasonal variation. Data preprocessing has been done to stabilize the seasonal variation and the quartic root of the wind speed plot (Figure 1 b) has indicated just that and has shown a more uniform season al variation with reduced difference in major peaks hence fitted for stationarity test.