A genetic programming-based ensemble method for long-term electricity demand forecasting

This study introduces a novel genetic programming-based ensemble method for forecasting long-term electricity consumption in Ethiopia. The technique utilizes a two-stage ensemble approach to project Ethiopia's electricity consumption through 2031. In the initial stage, genetic algorithms, particle swarm optimization, and simulated annealing methods are applied to various regression models (linear, quadratic, and exponential). The preliminary forecast values generated in this stage were further refined in the second stage. Here, the genetic programming method was utilized to develop a formula based on the initial forecast values, which then provided the final forecast results. The most accurate predictions in the first stage were obtained using the GA_Quadratic, PSO_Quadratic, and SA_Quadratic methods, resulting in mean absolute percentage error (MAPE) values of 3.61, 3.63, and 4.68, respectively. In the second stage, the GP-based prediction achieved an even lower MAPE value of 2.83. Other error metrics, including MSE, root mean square error (RMSE), and R², were also evaluated, with the proposed model outperforming all methods from the first stage on these metrics. The study projected Ethiopia's total annual electricity consumption through 2031 under two different scenarios. Both scenarios indicate that by 2031, electricity consumption will have tripled compared to 2021 levels.