Modeling and forecasting water inflow time series: A comparative study of classical and deep learning techniques

Abstract

Anticipating how rivers will behave is crucial in overseeing water resources, especially as the climate changes rapidly. This forecasting holds considerable economic importance, as it aids in managing water for farming, preventing water scarcity, and minimizing potential flood destruction. The major rivers within the Indus River system depend on the melting of snow and glaciers. Their water levels shift significantly during different times of the year. Researchers are engaged in a study to either examine or predict how much water will enter the Indus River. The input of water, known as inflow, plays a vital role in how we manage water resources. Therefore, being able to accurately predict this inflow is essential for effectively handling water resources. For inflow forecasting and modeling, we used a comparative study of classical and deep learning techniques. This study uses 5 years of data on Indus Tarbela Inflow ranging from January 2018 to September 2022. The data is collected from the Water and Power Development Authority (WAPDA). The first four years of the data is utilized for the estimation of the models and its subsequent one year is used for one-month-ahead out-of-sample forecast purpose. In this research work, we apply the, five different forecasting techniques that have been used for forecasting one-month ahead INdus Tarbela Inflow. These include the AutoRegressive Integrated Moving Average (ARIMA), Seasonal AutoRegressive Integrated Moving Average(SARIMA), Autoregressive neural network (ARNN), Seasonal-Naive, and Long Short-Scond Memory (LSTM). Three error measures have been used for assessing the forecasting accuracy of the above models that includes, mean absolute percentage error (MAPE), root mean square error (RMSE), and mean absolute error (MAE). The findings indicate that the LSTM approach is effective in forecasting Indu’s table inflow with lower forecasting measures