In a microgrid system, a phasor measurement device (PMU) is used to measure the electrical quantities of nodes, which can provide accurate data for system stability control. How to use the data measured using a PMU to improve the stability of a microgrid is an important practical problem. The mismatch between generation power and load power in a microgrid system will cause oscillation in the system. To ensure accurate and rapid load forecasting in a microgrid system and the reliable and safe operation of the microgrid, deep learning is introduced into microgrid load prediction, and a method of predicting the short-term load for a microgrid based on multivariable and multistep long short-term memory (MM-LSTM) is proposed in this paper. The method considers the effects of meteorological factors on load data and forecasts the current load situation from the load data and the temperature and humidity data of the previous period. A Keras-based model of the short-term load for microgrid prediction based on MM-LSTM is built and its parameters are optimized. Then, the load of a microgrid is predicted using the power consumption and meteorological data. The average absolute percentage error between the experimental results and the actual power consumption is 8.827%, demonstrating the effectiveness of the method.