When the echo state network (ESN) based on the least squares method randomly selects the input weights and the neuron thresholds in hidden layers, the convergence speed is slow and the prediction accuracy is not stable. We propose a modified ESN based on ant colony algorithm (ACO-ESN). The optimization of the initial input weights and neuron thresholds in hidden layers can be changed into a problem of finding the best path by the ants in the ant colony algorithm. The output weights are calculated by the least squares method. The ESN is trained through update, variation and genetic operations of the ant colony algorithm, and the input weights and thresholds which have the minimum ESN prediction error are selected to increase the ESN's prediction performance. Compared with other 4 ELM neural networks, the simulation results show that the ESN optimized by the ant colony algorithm can accelerate the convergence speed, and improve its prediction performance and the sensitivity of the neurons in hidden layers.