Threshold-based methods are extensively applied for performance assessment of photovoltaic (PV) modules. When applying the conventional methods with fixed normal thresholds, the main problem is the low detection rate of relatively small power losses. As a result, taking the uncertainties in the PV data due to the dynamic environment into account when designing the normal thresholds is a real challenge. To address this issue, improved threshold-based methods for PV performance assessment based on probabilistic models are proposed in this paper. This is motivated by the availability of probabilistic models which allow to quantify the likely uncertainty in the deterministic point estimation. Two probabilistic models are developed using the quantile regression forests (QRF) method and the Bayesian regression (BR) method. The developed models provide confidence intervals for PV output efficiency which act as the normal thresholds for PV performance assessment. The proposed methods are tested on a grid-connected PV plant located in East China, and compared against the traditional methods based on deterministic models. They are also compared with an adaptive threshold-based method which updates the adaptive thresholds by a moving window. Results indicate that the proposed methods are able to identify outliers with 10% power losses almost twice as many as the ones detected by the other methods.