Optimization of effluent COD prediction using LSTM based on improved fluctuation-aware positional attention mechanism

Urban wastewater treatment processes exhibit complex characteristics such as uncertainty, nonlinearity,and time delay, which pose significant challenges to water quality control and energy consumption optimization. To address these issues, this study proposed an SSA-LSTM-Attention model enhanced with fluctuation-aware positional encoding for the prediction of effluent chemical oxygen demand(COD). The model integrated an attention mechanism to extract key information from water quality data and employed an improved fluctuation-aware positional encoding to dynamically identify regions of data fluctuation, thereby enhancing the model's focus on critical positions. A long short-term memory(LSTM) network was utilized to capture the long-term dependencies in COD data, while the sparrow search algorithm(SSA) was applied to optimize the LSTM-Attention model by identifying the optimal network parameters. Historical water quality data from a wastewater treatment plant were used to validate the predictive performance of the model. Experimental results showed that the proposed model achieved a mean squared error(MSE) of 4.469 3 (mg/L)2, mean absolute error(MAE) of 1.382 3 mg/L, mean absolute percentage error(MAPE) of 7.592 4%, and a coefficient of determination(R²) of 0.910 8, all of which outperformed those of the comparison models. Furthermore, comparative experiments demonstrated that the fluctuation-aware positional encoding outperforms relative and absolute positional encoding in capturing data fluctuation points. These findings indicated that the SSA-LSTM-Attention model was effective in predicting effluent COD in wastewater treatment and provided valuable support for the operation and management of wastewater treatment plants.