An improved firework algorithm is proposed to achieve load distribution optimization of parallel chillers. Achieving lowest energy consumption of the parallel chiller system is the optimization goal and the partial load rate of each chiller is used as an optimization parameter. In the improved firework algorithm, a variable definition based on chaotic initialization is proposed to solve the problem of non-uniformity of initial solutions. To solve the problem of Gaussian mutation which can not jump out of the local optimization, the Levy flight variation method, which has a larger variation range, is used to improve the searching ability of the basic firework algorithm. To verify the feasibility and effectiveness of the proposed algorithm, two cases using parallel chiller systems are tested and compared with other algorithms. The experimental results show the improved fireworks algorithm can search for a better operating strategy and save more energy.