Abstract: Permanganate index (COD_Mn) is one of the most important parameters to measure water quality and could reflect the degree of water pollution by reducing substances. A novel COD_Mn forecast model (EWT-BLSTM) by combining empirical wavelet transform (EWT) and bidirectional long short-term memory (BLSTM) neural network was proposed. First, the original COD_Mn time series was decomposed into several components by EWT. Next, BLSTM neural network was employed to predict each decomposed component. Finally, the predictions of all components were reconstructed to obtain the new hybrid model EWT-BLSTM for the final COD_Mn predictions. COD_Mn data of Poyang Lake was used to evaluate the proposed forecast model. The results showed that EWT-BLSTM model had a powerful forecast capacity. For 1, 7-day ahead forecasting, the mean absolute percentage error (MAPE) of the forecast by EWT-BLSTM was 2.25% and 8.36%, respectively. The MAPE reduced by EWT-BLSTM over BLSTM was 10.53% for 1-day ahead forecasting and 16.16% for 7-day ahead forecasting. Furthermore, the proposed model showed highly stable forecasting performance for COD_Mn peak points, indicating that the proposed method was still applicable in the case of relatively complex data with extreme points.