Enhancing water quality index forecasting: a novel approach integrating signal decomposition and deep learning models for multi-step ahead predictions

The water quality index (WQI) is a popular indicator for evaluating surface water quality and reliable multi-step forecasts are essential for effective management decisions. Existing studies primarily focus on WQI point estimation or single-parameter prediction, with limited progress in long-term forecasting. This study proposes a multi-step ahead WQI forecasting framework integrating variational mode decomposition (VMD), convolutional neural network (CNN) and long short-term memory (LSTM). Approximately 30 years of bimonthly-interval water quality data from Malaysian rivers were used for training, enabling six-step-ahead (one-year) forecasts. VMD improved the CNN-LSTM hybrid model by eliminating rapidly oscillating noise components with a 20% filtering threshold yielding the highest stability. The model achieved a Kling-Gupta efficiency (KGE) of 0.62 and a mean absolute percentage error (MAPE) of 4.24%. Notably, performance remained satisfactory (KGE = 0.42 and MAPE = 6.12%) when half of the input parameters were omitted, demonstrating reliable one-year WQI forecasting even under incomplete data condition.