Abstract: Conventional just-in-time (JIT) learning-based soft sensors only employ a single similarity measure and cannot efficiently deal with the nonlinear characteristics of complex industrial processes, resulting in poor prediction performance. To tackle this issue, we propose a soft sensor-modeling method based on ensemble locally weighted partial least squares (ELWPLS) using diverse weighted similarity measures (DWS). First, we create a set of diverse training subsets by repeatedly performing random subspace and Gaussian mixture model clustering. Then, we determine the weights of input variables using PLS regression, thereby allowing us to define a set of diverse weighted similarity measures. During the online implementation phase for an arbitrary query sample, a group of diverse LWPLS models can be built and further combined via ensemble learning to provide the final prediction. The effectiveness and superiority of the proposed DWS-ELWPLS soft sensor method is demonstrated through a numerical example and an industrial debutanizer column process.