A singular spectrum analysis-enhanced BiTCN-selfattention model for runoff prediction

To tackle the difficulties and challenges posed by the nonlinear and nonstationary characteristics of runoff sequences in hydrological prediction, this paper aims to provide a novel forecasting method for the field of runoff prediction by constructing an SSA-BiTCN-SelfAttention time series prediction model. This model consists of Singular Spectrum Analysis (SSA), Bi-directional Temporal Convolutional Network (BiTCN), and Self-Attention mechanism (SelfAttention). Firstly, the runoff sequence is decomposed and reconstructed by Singular Spectrum Analysis, and the reconstructed sequence removes the noise and reveals the periodicity, trend, and other information in the runoff data, to facilitate the learning of the subsequent model; after that, the BiTCN model is used for the bidirectional training of the new sequence to validate and combine with the self-attention mechanism to fully explore the dependency relationship within the long sequence, to further improve the performance of the model. To verify the effectiveness of the model, this paper uses multi-year measured runoff data from Jiayuguan Hydrological Station, Yingluxia Hydrological Station, and Manwan Hydrological Station for training and testing. It selects three evaluation metrics: RMSE, MAE, and R², and analyzes the performance of the SSA-BiTCN-SelfAttention model by comparing it with four models: LSTM, TCN, BiTCN, CNN-LSTM, and BiTCN-SelfAttention. The results show that the SSA-BiTCN-SelfAttention model has the smallest prediction error and the highest accuracy. Compared with the single TCN model, the model improves about 58.36%, 46.43%, and 38.27% in the RMSE index, 63.89%, 57.89%, and 61.88% in the MAE index, and 10.9%, 3.7% and 1.8% in the R² index. The proposed singular spectrum analysis method can be used for trend and periodicity analysis of runoff data, providing an important basis for hydrological management. The prediction results of the proposed model are the closest to the true values, indicating its strong hydrological prediction ability. It not only provides a new method for runoff prediction but also provides important data references for the rational utilization and scientific planning of water resources.