It is difficult to establish a precise model for the distributed parameter systems (DPSs), which is affected by spatiotemporal coupling characteristic, strong nonlinearity, complex energy exchange and unknown factors. Aiming at this problem, a data-driven based low-dimensional constrained embedding modeling method was proposed. The data local nonlinearity and the global nonlinearity were considered based on the data manifold distribution. By nonlinear mapping and manifold learning methods, the nonlinear connection of local manifold structure was guaranteed and the nonlocal manifold structure was constrained to avoid data chaos in the low-dimensional space. The least squares support vector machine (LS-SVM) was used to establish the temporal series model to obtain the dynamic features in time direction. A complete predictive model of the system was reconstructed by spatiotemporal integration. Experimental results of thermal process show that the proposed method can effectively establish a model of strongly nonlinear DPS. Compared with the traditional method, the proposed method has stronger modeling performance and predictive ability.