<![CDATA[The Electricity Supply Business Plan (RUPTL) prepared annually by PLN still shows a high error rate in predicting electricity consumption, exceeding 10% in various provinces, such as North Sumatra (36.92%), DKI Jakarta (24.87%), West Kalimantan (40.24%), and South Sulawesi (31.56%), due to the limitations of the linear regression method used in the RUPTL. This study aims to evaluate and recommend the best electricity consumption forecasting model based on artificial intelligence using a Feed Forward Backpropagation Neural Network (FFBP-NN) combined with six training algorithms: Bayesian Regularization (BR), Conjugate Gradient (CG), Levenberg-Marquardt (L-M), Gradient Descent (GD), Quasi-Newton (Q-N), and Resilient Backpropagation (RB), resulting in a total of 13 algorithmic combinations. The data used consists of RUPTL indicators for DKI Jakarta from 2018 to 2023. Testing results of the 13 training functions on the FFBP-NN demonstrate that the TRAINOSS (Quasi-Newton) algorithm achieves the best performance with the lowest Mean Square Error (MSE) of 0.0000065546 and Mean Absolute Percentage Error (MAPE) of 0.06696%. This algorithm outperforms the linear regression method currently used in PLN’s RUPTL, which has a MAPE of approximately 21.14%. The second and third best algorithms are TRAINSCG and TRAINLM, with MAPE values of 0.09455% and 0.10020%, and MSE values of 0.0012160450 and 0.0012229340, respectively. The FFBP-NN model trained with TRAINOSS is highly recommended as the primary alternative to support long-term electricity load planning such as in PLN’s RUPTL.]]>
