A Northwest Pacific typhoon intensity prediction scheme has been developed based on multiple neural networks with the same expected output and an evolutionary Particle Swarm Optimization (PSO) algorithm. Typhoon samples during June-September spanning 2001-2012 are used for model development and Climatology and Persistence (CLIPER) factors are used as potential predictors. The new model input is constructed from potential predictors by employing both a stepwise regression method and a Locally Linear Embedding (LLE) algorithm. The LLE algorithm is able to learn and identify the underlying structure of a high dimensional vector space, and then perform dimensionality reduction and feature extraction. In this scheme, the new developed model, which is termed the PNN LLE model, is used for monthly typhoon intensity prediction at 24 and 48 h lead time. Using identical modeling samples and independent samples, predictions of the PNN LLE model are compared with the widely used CLIPER method. According to the statistics, the PNN LLE model shows reductions of the mean absolute errors of 23.34%, 24.46%, 19.41% and 27.45% relative to the CLIPER method for June-September 24 h forecasts, respectively, being 23.10% for the 4 months averagely. From June-September the mean absolute errors of the PNN LLE model are 44.82%, 16.73%, 0.89% and 49.26% more skillful than homogeneous CLIPER intensity forecasts for 48 h forecast, respectively, being 25.54% for the 4 months averagely. By adopting different numbers of nearest neighbors in the LLE algorithm, sensitivity experiments further show that the prediction results of the ensemble model are stable and reliable, and the forecast skill level of the ensemble model is better than that of the CLIPER method, potentially providing operational forecast tool and modeling method for the objective prediction of typhoon intensity.