In the noninvasive load identification algorithm of extreme learning machine, the input weight and hidden layer threshold are generated randomly, which leads to misjudgment. An improved genetic algorithm is proposed to optimize the extreme learning machine. The selection operator in genetic algorithm is improved by solving the fitness value of each individual and completing the sorting in the order from small to large, dividing the sorted population into four parts, selecting the best from these four parts to form a new population according to the proportion, and selecting operators with greater fitness from this new population. The optimized weights and thresholds are obtained by hillclimbing method, and the optimized extreme learning machine network is constructed to identify the load. Through a large number of simulations on MATLAB, the simulation test results show that the accuracy of load identification is improved by about 7.41% compared with the results of non optimization algorithm, presenting better classification performance and verifying the effectiveness of the algorithm for load identification.