The attack of false data injection can contaminate the measurements acquired from the supervisory control and data acquisition (SCADA) system, which can seriously endanger the safety and stability of power system operations. The conventional machine learning attack detection methods use a single strong classifier and are difficult to solve the problem of overfitting, making them lack of generalization ability. On the other hand, most existing dimension reduction approaches based on feature extraction can change the original physical meanings of measurements. Here, a novel method is proposed based on feature selection and ensemble learning to solve the above problems. Squirrel search algorithm combines Latin hypercube sampling and opposition-based learning to form an improved algorithm with strong global search ability for feature selection. This avoids the problem of feature extraction changing the original physical meanings of measurements. Besides, the classifier based on adaptive boosting decision tree ensemble learning algorithm with stronger generalization ability is used to distinguish the false data injection. Simulation results using the IEEE 14-bus and IEEE 57-bus test systems verify the proposed method with higher performance of detection compared with other widely adopted methods.