Accurate short-term power load forecasting is essential for the formulation of daily power generation plans and the real-time dispatching of new power systems. In order to obtain accurate and reliable load forecasting results, this paper proposes an Attention-GAT-LSTM intelligent algorithm targeting the temporal and uncertain characteristics of real power load data, and applies it in practical new power systems. In raw data processing, this algorithm innovatively integrates a self-attention mechanism, introduces a data processing unit to assign weights, and employs a skip-connection mechanism to prevent overfitting. The processed data is transmitted to a graph attention network (GAT) for spatial node feature extraction, then passed to a long short-term memory (LSTM) network for temporal feature extraction. Through forward propagation, back propagation, and gradient descent methods, the weights and biases of the LSTM layer are iteratively updated, effectively minimizing information loss during iterations and highlighting key time-point information. Comparative analyses with various models demonstrate the method’s high prediction accuracy in short-term (hour-level) power load forecasting, providing data support for the operational dispatching, planning and construction of new power systems.