Water Quality Prediction with Dilated 1D Convolutional Networks and Comprehensive Preprocessing Techniques

Water quality is essential for humans and the environment. One of the efforts to manage water resources, detect pollution, and ensure safe water access can be done by predicting it through computational methods such as deep learning algorithms. The framework for water quality prediction employs a systematic approach with comprehensive data preprocessing, including outlier detection, mean filtering, wavelet transform, linear interpolation, and Pearson correlation. MAE, MAPE, MSE, and RMSE are used to measure the performance of deep learning models employing Conv1D. The Conv1D model performed well in predicting pH (MAE = 0.0714) and salinity (MAE = 0.3374), but showed lower performance for DO (MAE = 1.1724) and water temperature (MAE = 0.5171), indicating the need to improve the accuracy of DO and water temperature predictions. The comparison of actual and predicted values for pH, DO, salinity and water temperature shows that the Conv1D model can follow the pattern of actual values, although it struggles with high spikes that cause deviations. Using 3203 test data taken every 5 minutes, the model can predict for about 11 days in ahead, demonstrating its ability to provide long-term water quality predictions.