Accurate load forecasting is recognized as a critical requirement for the secure and economical operation of modern power systems. In this study, two advanced data-driven neural architectures are presented and compared for forecasting the load behavior of Iran’s power grid. In the first approach, a Time-Delay Cerebellar Model Articulation Controller (TD-CMAC) is employed to capture temporal dependencies in hourly load series, and its performance is systematically benchmarked against a conventional Multilayer Perceptron (MLP), Long Short-Term Memory (LSTM), and Extreme Gradient Boosting (XGBoost). In the second approach, a rough Gaussian Neural Network (RGNN), trained by the Levenberg–Marquardt algorithm, is developed, in which rough set concepts are combined with Gaussian activation functions to enhance robustness under uncertainty. This RGNN is evaluated on power load time-series data, so that its generalization capability across regression tasks can be assessed. Both neural frameworks are trained on real load data from the Iranian power network. Their performance is quantified using cost functions, regression plots, and accuracy metrics, demonstrating that while the proposed RGNN provides exceptional robustness against fluctuations, both architectures offer complementary advantages depending on the forecasting horizon. The empirical findings indicate that the TD-CMAC achieves competitive and often superior accuracy compared with the MLP, and XGBoost for short-term load forecasting, whereas the Levenberg–Marquardt-trained RGNN provides effective and robust modeling for prediction in the presence of uncertainty. Overall, it is demonstrated that the combination of time-delay cerebellar structures with rough Gaussian neural modeling constitutes a powerful data-driven toolkit for power load forecasting in complex power systems.