Suspended sediment is of importance in river and dam engineering. Due to its high nonlinearity and stochasticity, sediment prediction by conventional methods is a challenging task. Consequently, this paper establishes a new hybrid model for an improved forecast of suspended sediment concentration (SSC). It is a nonlinear autoregressive network with exogenous inputs (NARX) integrated with a data pre-processing framework (thereafter INARX). In this model, wavelet transformation (WT) is used for time series decomposition and multigene genetic programing (MGGP) for details scaling. The two incorporated modules improve time and frequency domain analysis, allowing the network to unveil the embedded characteristics and capture the non-stationarity. At a hydrological station on the upper reaches of the Yangtze River, the records of daily water stage, flow discharge and suspended sediment are collected and refer to a nine-year period during 2004-2012. The data are used to evaluate the models. Several wavelets are explored, showing that the Coif3 leads to the most accurate prediction. Compared to the sediment rating curve (SRC), the conventional MGGP, multilayer perceptron neural network (MLPNN) and NARX, the INARX demonstrates the best forecast performance. Its mean coefficient of determination (CD) increases by 7.7%-38.6% and the root mean squared error (RMSE) reduces by 15.1%-54.5%. The INARX with the Coif3 wavelet is further evaluated for flood events and multistep forecasts. Under flood conditions, the model generates satisfactory results, with CD > 0.83 and 84.7% of the simulated data falling within the +/- 0.1 kg/m3 error. For the multistep forecast, at a one-week lead time, the network also yields predictions with acceptable accuracy (mean CD = 0.78). The model performance deteriorates if the lead time becomes larger. The established framework is robust and reliable for real-time and multistep SSC forecasts and provides reference for time series modeling, e.g. streamflow, river temperature and salinity.