This research proposed a method of selecting an optimal combination of numbers of input (number of lagged values in the model) and of hidden nodes for modeling seasonal data using Artificial Neural Networks (ANN). Three data sets (rainfall, relative humidity and solar radiation) were used in assessing the proposed procedure and the resulting ANN models were compared with two traditional models (Holt-Winter’s and SARIMA). Models with large number of lagged values have shown tendency to outperform those with small number of lagged values. Selected ANN model was found to outperform the two traditional models on rainfall data; it performed better than SARIMA but worse than Holt-Winter’s model on relative humidity data and performed worse than the two methods on solar radiation data. The proposed procedure has hence, performed fairly well. Oscillatory performance recorded by ANN models that resulted from the proposed procedure in relation to the other two models only attests to the fact that no particular model is best on every data set. Rather than insist on elegance or sophistication, researchers should be guided by parsimony.
Keywords: Artificial Neural Networks, Forecasting, SARIMA, holt-winter, seasonal data
