Least Squares Support Vector Machines(LSSVM) regression principle and sparsity configuration were introduced. In this paper online dynamic modeling based on Sparse LSSVM(SLSSVM) was proposed for wood drying process with strong coupling and nonlinear characteristics. The sample data of Fraxinus mandshurica in the speed-down drying stage were gathered in the experiments of a downscaled industrial wood drying kiln. According to the actual needs of predictive control, an online model of drying process was established for online predicting wood moisture content. Results of simulation and comparison experiments showed that the SLSSVM online model updated learning data based on basic sparse method to rolling optimize model structure so as to predict next system output, and could reflect current state of wood drying process more effectively. The model had a high predict precision, strong generalization ability and simple structure, which could be further used in online predictive control of practical wood drying process.