Comparison of Time Series Forecasting Techniques Applied for Water Quality Prediction in Southwest Iran

Aims: The main objective of the current study was to assess the efficiency of fourtime series prediction methods to forecast the values of total dissolved solids (TDS) using a time series of over sixteen years. Materials and Methods: The applied methods comprised of autoregressive integrated moving average (ARIMA) as the most traditional method, two neural network based techniques including multilayer perceptron (MLP) along with extreme learning machines (ELM) and a novel approach known as temporal hierarchies (TH) which was applied for the first time in water resources and water quality researches. Findings: It was found that with respect to the forecasting accuracy, the MLP outperforms the ARIMA model for the training series where the MAPE (%) and MASE (mg/l) were reduced from 5.109 to 3.146 and 0.553 to 0.323, respectively. On the other hand, the forecasting accuracy of ELM was lower than that of MLP however the respective outofsample generalization ability of this model was higher with MAPE and MASE values of 6.526 and 0.683. Conclusion: Meanwhile, it was concluded that temporal hierarchies gave the best results for the test part of time series. The main shortcoming of neural network based approaches was their reduced outofsample prediction due to overfitting. Based on the results, TH is a viable alternative for conventional time series forecasting techniques.