In industrial production, soft sensors play very important roles in ensuring product quality and production safety. Traditionally, global modeling methods, which use historical data to construct models offline, are often used to develop soft sensors. However, because of various complex and unknown changes in industrial production processes, the performance of global models deteriorates over time, and frequent model maintenance is difficult. In this study, locally weighted partial least squares (LWPLS) is adopted as a just-in-time learning method for industrial soft sensor modeling. In LWPLS, the bandwidth parameter h has an important impact on the performance of the algorithm, since it decides the range of the neighborhood and affects how the weight changes. Therefore, we propose a two-phase bandwidth optimization strategy that combines particle swarm optimization (PSO) and LWPLS. A numerical simulation example and an industrial application case were studied to estimate the performance of the proposed PSO-LWPLS method. The results show that, compared to the traditional global methods and the LWPLS with a fixed bandwidth, the proposed PSO-LWPLS can achieve a better prediction performance. The results also prove that the proposed method has apparent advantages over other methods in the case of data density changes.