[Purposes] In order to obtain accurate and reliable ultra-short term power load forecasting results to meet the needs of rapid response and real-time dispatching of power system, considering the characteristics of non-linear and timely ordering of power grid load data, this paper proposeed a method of ultra-short term interval forecasting of power load based on curve feature clustering and information aggregation. [Methods] Firstly, considering the local fluctuation characteristics and overall trend characteristics of the load curve, the power load curvewas divided into different categories. Then, the Gaussian process regression model was used as the prediction model to characterize the overall trend of load, and the quantile based bidirectional long-term and short-term memory neural network was used as the prediction model to characterize the local fluctuation of load. Finally, the aggregation idea was introduced, and the Choquet integral algorithm was used as the aggregation function to aggregate the results of the two prediction models. [Findings] This method effectively realized the clustering of daily load curves considering multiple characteristics. Meanwhile, the prediction results of a single model were aggregated, which obtained the interval prediction results of each confidence level under different scenarios. Under example analysis, the reliability index of the prediction method in this paper was 14.70% and 10.81% higher than that of the two single models, and the comprehensive performance was 3.14% and 15.55% higher than that of the two single models. [Conclusions] The results of the example showed that the forecasting accuracy and reliability of the model proposed in this paper are significantly improved comparing with the common load probability forecasting methods and the single forecasting model. In addition, the introduction of clustering method and information aggregation idea is useful to improve the accuracy of the forecasting model.