The Java–Madura–Bali (Jamali) power system, which includes Java Island and the neighboring islands of Madura and Bali, represents Indonesia’s most populous and economically significant region, resulting in highly complex electricity demand. Accurate daily electricity load forecasting is thus critical for reliable power system operation in the Jamali power system, yet remains challenging when relying solely on traditional methods. Forecasting with 30-min intervals is adopted to capture rapid fluctuations in load that are driven by industrial activities on weekdays, thereby providing greater temporal resolution for operational decision-making. This paper aims to identify the most efficient deep learning model for 30-min interval weekday load forecasting using Jamali dataset, with the goal is to minimize forecasting errors, consistently maintaining prediction errors below 2% of the annual peak load, or approximately 630 MW based on our dataset, thereby supporting frequency stability 50 Hz ± 0.2 Hz. Several advanced deep learning models are evaluated, including backpropagation neural network (BPNN), convolutional neural network (CNN), long short-term memory (LSTM), hybrid CNN-LSTM, and convolutional LSTM architectures. Experimental results show that all models satisfy the testing criteria, with the one-dimensional CNN model attaining the lowest test MAE of 191.99±33.22 MW for one-step-ahead forecasting based on a 300-step input window. These findings indicate that deep learning models provide robust solutions for operational load forecasting in the Jamali power system. © The Author(s), under exclusive license to Springer Nature Switzerland AG 2026.