A practical temporal transfer learning model for multi-step water quality index forecasting using A CNN-coupled dual-path LSTM network

Study region

Klang River, Malaysia.

Study focus

This study presents a multi-step ahead water quality index (WQI) forecasting framework in Klang River to address persistent challenges such as missing data and seasonally variable hydrological patterns. A hybrid deep learning architecture was developed by combining a 1d-Convolutional Neural Network (CNN) with a dual-path Long Short-Term Memory (LSTM) network to capture long-term hydrological memory and site-specific temporal variability. To enhance adaptability in data-scarce environments, the architecture incorporates transfer learning, allowing knowledge from historical water quality (WQ) data to be effectively applied to current conditions for robust forecasting.

New hydrological insights for the region

WQ time series are often incomplete and exhibit non-linear interdependencies, which pose significant challenges for accurate WQI forecasting. The CNN-LSTM model effectively extracts inter-parameter features and learns temporal patterns, achieving strong five-step ahead forecasting performance. Despite challenges like missing data and non-stationary WQ patterns, the dual-path LSTM tuning approach effectively transfers and fine-tunes knowledge from historical records to improve prediction accuracy across different temporal domains. The model maintains a MAPE below 5 % and KGE values between 0.36 and 0.67, demonstrating robust performance in multi-step WQI forecasting. These results highlight its potential to support regional WQ assessments by capturing seasonal flow dynamics and pollutant transports in tropical monsoon catchments, thereby reinforcing the role of WQI in river classification and water management, especially under data-scarce conditions.