Performance Evaluation of NARX-CG Model for Electricity Forecasting: Bali Blackout Case Study

Abstract

Bali experienced a widespread blackout in May 2025 that disrupted economic and social activities across the island, revealing weaknesses in electricity demand forecasting and system resilience. This study evaluates the performance of a Hybrid Nonlinear Autoregressive with Exogenous Inputs-Conjugate Gradient (NARX-CG) model as an advanced electricity forecast. The dataset covers the 2018-2023 period and includes six variables: electricity energy, connected capacity, number of customers, tariffs, Gross Regional Domestic Product (GRDP), and population, aligned with the national electricity planning framework. The NARX-CG model was developed using a 6-12-6-1 network architecture and trained with tansig transfer function. Forecasting performance was evaluated using Mean Squared Error (MSE) and Mean Absolute Percentage Error (MAPE) metrics. Results show that the NARX-CG model achieved an MSE of 0.09853 and an average MAPE of 8.12%, outperforming conventional projections with a MAPE of 28.48%. Yearly evaluations show consistent model stability, with the lowest MAPE values of 1.93% and 5.86% in 2023 and 2022, respectively. The NARX-CG model effectively captures nonlinear temporal dependencies, enhances predictive accuracy, and contributes to improved power system reliability and resilience, providing valuable insights for adaptive energy planning following the 2025 Bali blackout.