Interval state estimation (ISE) can estimate state intervals of power systems according to confidence intervals of predicted pseudo-measurements, thereby analyzing the impact of uncertain pseudo-measurements on states. However, predicted pseudo-measurements have prediction errors, and their confidence intervals do not necessarily contain the truth values, leading to estimation biases of the ISE. To solve this problem, this paper proposes a pseudo-measurement interval prediction framework based on the Gaussian process regression (GPR) model, thereby improving the prediction accuracy of pseudo-measurement confidence intervals. Besides, a weight assignment strategy for improving the robustness of weighted least squares (WLS) ISE is proposed. This strategy quantifies the deviation between the pseudo-measurement intervals and their estimated intervals and assigns smaller weights to the pseudo-measurement intervals with larger deviations, thereby improving the estimation accuracy and robustness of the ISE. This paper adopts the data from the supervisory control and data acquisition (SCADA) system of the New York Independent System Operator (NYISO). It verifies the advantages of the GPR method for pseudo-measurement interval prediction by comparing it with the quantile regression and neural network methods. In addition, this paper demonstrates the effectiveness of the proposed weight assignment strategy through the IEEE 14-bus case. Finally, the differences in the estimation accuracy and the bad data identification between the robust interval state estimation and deterministic state estimation are discussed.