In order to improve the classification accuracy of transient stability assessment of power systems, a novel method based on local learning machine and an improved bacterial colony chemotaxis (BCC) algorithm is proposed, where local learning machine(LLM) is used to build a TSA model. Considering the possible real-time information provided by PMU, a group of system-level classification features extracted from the power system operation parameters are employed as inputs, and the stability result is used as output of the LLM model. The relation ship between input and output is trained and the ideal model is obtained by applying the improved BCC combined with chaotic search strategy to determine the optimal parameters of LLM automatically. The effectiveness of the proposed method is shown by the simulation results on the New England 10-unit-39-bus power system.