<![CDATA[Accurate short-term electricity load forecasting is critical for planning power generation and maintaining cost efficiency. Since the amount of electricity generated significantly affects the cost-efficiency of power generation, a forecasting method with high accuracy is required. In response, this study developed a Recurrent Neural Network (RNN) model architecture trained using three different algorithms: Levenberg-Marquardt, Bayesian Regularization, and Scaled Conjugate Gradient. The model's performance was evaluated using the Mean Absolute Percentage Error (MAPE) metric. Historical load data were categorized by day type and divided into training and testing sets. The best-performing RNN model was used to carry out load forecasting, which was then compared to the forecasting results from PT PLN. Among the models tested, the RNN trained with Bayesian Regularization, configured with an 8-16-1 network architecture using a learning rate of 0.01 achieved the highest accuracy. In a two-week forecasting simulation, this model reached a MAPE of 1.4084%, significantly outperforming the 3.3160% error from PLN’s conventional forecasting method. These results underpin the effectiveness of RNNs, particularly when trained with Bayesian Regularization, for enhancing short-term electricity load forecasting within the scope of this dataset.]]>
Copyrights © 2025
