International Journal of Science and Technology (IJST)

An Improved Temporal Convolutional Network with Residual Multi-head Attention for Long-term Water Quality Prediction

Accurate prediction of river water quality based on time-series data is essential for understanding variations and protecting river environments. However, forecasting accuracy is severely influenced by strong seasonality, nonlinearity, and periodicity. Traditional statistical methods suffer from low accuracy, high temporal complexity, and poor long-term prediction capability, particularly in dynamic and constantly changing surroundings in open waters. To address these deficiencies, a novel deep learning framework based on a Temporal Convolutional Network (TCN) integrated with a residual multi-head attention mechanism is presented to forecast dissolved oxygen (DO). Initially, the raw data of water quality parameters were smoothed by using Savitzky Golay (SG) filter, which is then decomposed into three components: trend, seasonal, and residual components. The trend and residual components are supplied with the enhanced TCN model for training and prediction. Dilated causal convolutions enable effective extraction of long-term temporal dependencies, while the multi-head attention mechanism improves the capability to capture both local and long-range dependencies in time series data, thereby enhancing feature representation and prediction accuracy. Moreover, when compared with deep learning models such as LSTM, GRU and traditional statistical approaches, the proposed model consistently outperformed all baseline methods with R² values of 0.9906 and 0.9651 for Short-term(1 day) and Long-term (7 days) respectively . The findings confirmed that the proposed model provides a robust and effective solution for complex river water quality prediction tasks and offers valuable support for sustainable water resource management and environmental protection.