Application of Machine Learning for Daily Forecasting Dam Water Levels

Abstract

The evolving character of the environment makes it challenging to predict water levels in advance. Despite being the most common approach for defining hydrologic processes and implementing physical system changes, the physics-based model has some practical limitations. Multiple studies have shown that machine learning, a data-driven approach to forecast hydrological processes, brings about more reliable data and is more efficient than traditional models. In this study, seven machine learning algorithms were developed to predict a dam water level daily based on the historical data of the dam water level. Multiple input combinations were investigated to improve the model’s sensitivity, and statistical indicators were used to assess the reliability of the developed model. The study of multiple models with multiple input scenarios suggested that the bagged trees model trained with seven days of lagged input provided the highest accuracy. The bagged tree model achieved an RMSE of 0.13953, taking less than 10 seconds to train. Its efficiency and accuracy made this model stand out from the rest of the trained model. With the deployment of this model on the field, the dam water level predictions can be made to help mitigate issues relating to water supply.