Determination of drinking water quality using Recurrent Neural Network (RNN) And Long-Short Term Memory (LSTM): a case study of Kano State waterworks

Abstract

This study evaluates the drinking water quality in Kano State, Nigeria, using Long Short-Term Memory (LSTM) and Recurrent Neural Network (RNN) models. In forecasting the important water quality characteristics across time. A 76-months (2018–2024) dataset from Kano State Waterworks was digitized, preprocessed, and evaluated.  A thorough depiction of the state's water quality profile was produced by combining the dataset, that was gathered from three different treatment facilities: Challawa, Tamburawa, and Watari.  In order to determine the most significant factors associated with water potability, data pretreatment included cleaning, normalization, resampling to daily means, and feature selection.  By resolving class imbalance and eliminating noisy samples, the SMOTEEEN (Synthetic Minority Over-sampling Technique with Edited Nearest Neighbors) approach was adopted to balance the dataset and enhance quality during cleaning.  To provide a fair and trustworthy model evaluation across different data distributions, stratified K-Fold cross-validation was used. The goal of this research is to enhance the accuracy and reliability of water quality assessment by capturing nonlinear and temporal dynamics within our dataset. Potability levels were determined by comparing the created models to drinking water requirements set by the World Health Organisation (WHO).  The LSTM produced MAE of 0.1400, RMSE of 0.1483, MAPE of 14%, and R² of 0.9083, whereas the RNN produced Mean Absolute Error (MAE) of 0.3500, Root Mean Square Error (RMSE) of 0.3536, Mean Absolute Percentage Error (MAPE) of 35%, and R² of 0.4792.  While both models worked well, LSTM did slightly better.