Temperature detection is now used for risk early warning in power system. Adopting wireless sensor network to collect temperature information is one of the most crucial applications. In order to accomplish temperature information gathering, hundreds of sensor nodes are placed in power system equipment, such as transformer, relay protection device, mutual inductor, to name a few. Studies have shown that, for nodes, message transmission consuming more energy than local data computing. It's unrealistic to configure batteries for sensor nodes or replace them on a large scale in power system. For the purpose of lengthening the lifetime of wireless sensor network and reducing energy consumption, the amount of message transmission should be minimized. In this paper, a new data collection strategy in power system based on Fourier transform is proposed. Because of periodic character of temperature detection data, Fourier transform is adopted to linearly fit these data in this paper, and least square strategy is used to optimize parameters. The sink nodes transmit model parameters to the base station instead of thousands of data, when the error is in the range of confidence level. Experiments show that the proposed strategy obtains greater fitness and less error between the measured data and predicted data.