Using the hybrid Taguchi experimental design method – TOPSIS to identify the most suitable artificial neural networks used in energy forecasting

The use of artificial neural networks (ANN) in forecasting has many applications. Appropriate design of ANN parameters enhances the performance and accuracy of neural network models.  Most studies use a trial and error approach in setting the value of ANN parameters. Other methods used to determine the best structure of a neural network only use a single evaluation criterion to determine the appropriate structure. In this study, the authors provide a new method to design the network structure. In this method, we use a combination of Taguchi experimental design and TOPSIS methods, to determine he optimal ANN structure, taking into account three evaluation criteria simultaneously. The  estimated demand for gasoline in the Hormozgan province produced using this method, confirms its efficiency and effectiveness. Analysis of variance (ANOVA) of the ANN variables indicates that contribution of the number of neurons in the first hidden layer to the changes in the network performance is about 54% while the contribution of the learning algorithm is about 27%.