Energy management systems for islanded microgrids often rely on predictions of energy availability and usage. Such predictions can be used to plan actions, such as shedding non-essential loads, so that critical loads continue to be served. However, uncertainties in the prediction models may lead to incorrect decisions, and subsequently jeopardize reliable operation of the microgrid. For a photovoltaic (PV) and battery based microgrid, uncertainties in the PV rating and the battery capacity model parameters can lead to otherwise avoidable outages. In this paper, techniques have been developed to identify and compensate for such model uncertainties. The approach uses differences between the actual and predicted data sequences to determine compensation factors to improve prediction accuracy. The developed techniques account for operating condition changes automatically, and no additional sensors are needed for their implementation. The method has been evaluated using data from rooftop irradiance and temperature sensors and the corresponding forecasts. It has been shown that the proposed techniques can improve the accuracy of the predictions and hence lead to more effective energy management decisions. Together with a pre-emptive load shedding strategy, the total outage time of the microgrid can be shortened by as much as 11% for the chosen scenario.