COMPARATIVE ANALYSIS OF LONG SHORT-TERM MEMORY (LSTM) AND MULTI-LAYER PERCEPTRON (MLP) MODELS FOR RIVER RUNOFF PREDICTION IN THE HINDU KUSH HIMALAYAN REGION

Abstract

Hydrological forecasting in the Hindu Kush Himalayas (HKH) presents special challenges because of the complex interplay between climatic and environmental factors. The quantitative predictive capabilities of two well-established models, Long Short-Term Memory (LSTM) and Multi-Layer Perceptron (MLP), chosen for their proven performance in previous studies, are meticulously compared in this thesis. The analysis uses comprehensive data spanning 2001 to 2013, including discharge records from the Department of Hydrology and Meteorology (DHM), precipitation data from APHRODITE, temperature data from APHRODITE, and snow cover area information from Google Earth Engine with MOD09A1 V6.1. The study employs rigorous evaluation metrics, revealing nuanced insights into the hydrological processes. Contrary to expectations, the MLP model exhibited slight superiority, showcasing a nuanced understanding of the region's complexities. The quantitative assessment, including RMSE (LSTM: 0.2396, MLP: 0.1733), MAE (LSTM: 0.1698, MLP: 0.0841), R2 Score (LSTM: 0.9976, MLP: 0.9987), and NSE (LSTM: 0.9976, MLP: 0.9987), emphasizes the indispensable role of robust predictive models, showcasing the necessity of reliable models for enhancing accurate river runoff predictions crucial for effective water resource management and flood preparedness in challenging terrains like the HKH.