<![CDATA[Short-term electricity load forecasting (STLF) is important for supporting reliable and efficient power system operation. Bayesian Regularization Neural Network (BRNN) can model nonlinear relationships in electricity load data while controlling model complexity. This study analyzes the performance of a previously developed BRNN model by examining the effects of network architecture and forecasting horizon. Four input variables, namely hour, air temperature, air humidity, and previous-day electricity load, were used to forecast the following-day electricity load. Five BRNN configurations with 5, 10, 15, 20, and 25 hidden units were evaluated using Mean Squared Error (MSE) and correlation coefficient (R), while forecasting performance was assessed using Mean Absolute Percentage Error (MAPE) for horizons h+1 to h+7. The 4-10-1 architecture achieved the lowest training MSE of 0.000933 and was selected as the reference configuration based on the predefined minimum-training-MSE criterion. The testing R values ranged from 0.930 to 0.952 across the evaluated architectures. For the selected configuration, MAPE values across h+1 to h+7 were 8.46%, 7.76%, 8.16%, 5.28%, 2.45%, 8.02%, and 6.11%, respectively. The lowest MAPE of 2.45% occurred at h+5. These results indicate that BRNN performance varies with network architecture and forecasting horizon, while the observed performance should be interpreted within the evaluated dataset and experimental conditions.]]>
