Abstract & Overview
This paper proposes a Taguchi-based optimization framework for short-term load forecasting (STLF) using single-layer feedforward neural networks (SLFNs). Although SLFNs are computationally efficient, their accuracy strongly depends on proper hyperparameter configuration, which is often selected through inefficient trial-and-error procedures. The proposed approach applies orthogonal arrays and signal-to-noise analysis to identify robust and reproducible SLFN settings. A MATLAB-based load forecasting interface is developed to support data preprocessing, model selection, parameter tuning, forecasting, and performance evaluation. The methodology is validated using real historical load and meteorological data from the distribution network supplying the municipalities of Kosovska Mitrovica, Zvečan, Leposavić, and Zubin Potok. The Taguchi-optimized SLFN achieves a mean absolute percentage error of 4.32% and a coefficient of determination of 0.991, outperforming all reference methods. More than one year of operational use confirms that forecasting errors consistently remain in the 3–5% range. These results demonstrate that lightweight neural architectures, when systematically optimized, provide a practical, accurate, and computationally efficient solution for real-world STLF applications.