Time Series Forecasting for Outdoor Temperature Using Nonlinear Autoregressive Neural Network Models

Weather forecasting is a challenging time series forecasting problem because of its dynamic, continuous, data-intensive, chaotic and irregular behavior. At present, enormous time series forecasting techniques exist and are widely adapted. However, competitive research is still going on to improve the methods and techniques for accurate forecasting. This research article presents the time series forecasting of the metrological parameter, i.e., temperature with NARX (Nonlinear Autoregressive with eXogenous input) based ANN (Artificial Neural Network). In this research work, several time series dependent Recurrent NARX-ANN models are developed and trained with dynamic parameter settings to find the optimum network model according to its desired forecasting task. Network performance is analyzed on the basis of its Mean Square Error (MSE) value over training, validation and test data sets. In order to perform the forecasting for next 4,8 and 12 steps horizon, the model with less MSE is chosen to be the most accurate temperature forecaster. Unlike one step ahead prediction, multi-step ahead forecasting is more difficult and challenging problem to solve due to its underlying additional complexity. Thus, the empirical findings in this work provide valuable suggestions for the parameter settings of NARX model specifically the selection of hidden layer size and autoregressive lag terms in accordance with an appropriate multi-step ahead time series forecasting.