“Soft sensors” with time series prediction models can reduce cost for devices and be used for controlling and monitoring systems instead of hardware sensors. They have been actively studied in the field of machine learning. In particular, Echo State Network (ESN) is known to a promising method to realize soft sensors in terms of both maintaining accuracy and reducing training cost. On the other hand, since most of the weights of network are fixed to random values, ESN carry an inherent risk of generating models with unstable prediction result due to variability in weight distribution. In this paper, we introduce several mechanisms, related to hyperparameter optimization and network stability, to eliminate the risk for a typical ESN training procedure and propose a method to automatically and efficiently design a robust ESN. In addition, the proposed method is applied to a soft sensor modeling task in a real-world dataset. As a result of the evaluation, it is clarified that the introduced mechanisms significantly reduce accuracy variation and the risk of divergence of the prediction.