Comparative analysis of inflow forecasting using machine learning and statistical techniques: case study of Mangla reservoir and Marala Headworks

Abstract

This study, under the context of a global perspective, focuses on the Indus Basin Irrigation System (IBIS) of Pakistan specifically the Jhelum and Chenab rivers inflows. The IBIS operation relies on seasonal planning strategies, informed by forecasts of river inflows at key stations by the Indus River System Authority (IRSA). In this study, Artificial Intelligence (AI) models including Generalized Regression Neural Network (GRNN), and Multi-Layer Feedforward Neural Network (MLFN) along with the statistical model Autoregressive Integrated Moving Average (ARIMA) were used to forecast the inflows of both rivers for 5 years (2020–2024) with a lead time of 1 year. Historic flow data of 59 years (10 daily from 1966 to 2024) were collected from IRSA. The collected data from 1966 to 2014 are used for calibration/training and from 2015 to 2020 are used for validation/testing of selected models for both study locations. The results of correlation and error estimation depicted that Artificial Neural Network (ANN) models predicted better inflows than the ARIMA model. The average RMSE and R 2 of ANN models is 9.68 and 0.92 and the average RMSE and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="m1"><mml:mrow><mml:msup><mml:mi mathvariant="normal">R</mml:mi><mml:mn>2</mml:mn></mml:msup></mml:mrow></mml:math> of ARIMA Model is 10.17 and 0.88, this results in improvement of average RMSE and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="m2"><mml:mrow><mml:msup><mml:mi mathvariant="normal">R</mml:mi><mml:mn>2</mml:mn></mml:msup></mml:mrow></mml:math> by 4.82% and 4.35% in case of ANN Models when compared with ARIMA Model. Qualitative analysis shows that ANN techniques better predicted the high and low flows when compared with statistical methods. Specifically, the application of the ANN models has enhanced the precision of forecasted inflows assessment compared to the probabilistic inflow forecasting methods used by IRSA. The average RMSE and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="m3"><mml:mrow><mml:msup><mml:mi mathvariant="normal">R</mml:mi><mml:mn>2</mml:mn></mml:msup></mml:mrow></mml:math> in case of IRSA forecast is 11.47 and 0.88 and the average RMSE and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="m4"><mml:mrow><mml:msup><mml:mi mathvariant="normal">R</mml:mi><mml:mn>2</mml:mn></mml:msup></mml:mrow></mml:math> in case of ANN Models is 10.30 and 0.92, this results in improvement of average RMSE and <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" id="m5"><mml:mrow><mml:msup><mml:mi mathvariant="normal">R</mml:mi><mml:mn>2</mml:mn></mml:msup></mml:mrow></mml:math> by 10.20% and 4.35% in case of ANN Models when compared with IRSA forecast. This study highlights the need for utilization of ANN models in place of probabilistic inflow forecasting methods to improve the accuracy of time series inflow forecasts.