Mono Lake Water Levels Forecasting using Machine Learning

Abstract

This thesis explores the application of advanced machine learning techniques to forecast water levels in Mono Lake, California, a critical ecological and hydrological resource. Given the complex interplay of factors influencing water levels, such as precipitation, evaporation, natural runoff, and diversions, accurately predicting these levels presents a significant challenge. Various machine learning models, including Support Vector Machine (SVM), Random Forest (RF), and Long Short-Term Memory (LSTM), were developed to forecast Mono Lake water levels. These machine learning models integrate historical water levels, multiple precipitation datasets using a Bayesian model, and outputs from the Large Lake Statistical Water Balance Model (L2SWBM) – an advanced Bayesian model. A novel contribution of this study is the development and application of the LSTM algorithm, training, and optimization process to develop an Ensemble model for forecasting water levels, which in simple terms develops a group of forecasts from multiple LSTM models to improve the prediction accuracy of Mono Lake water levels forecasts. By training on historical data from 1970 to 2009 and validating model predictions against historical data from 2009 to 2018, the study offers a comprehensive evaluation of the model’s performance, followed by forecasts from 2019 to 2023. The findings reveal that the LSTM Ensemble models can accurately predict future water level fluctuations, demonstrating the potential of machine learning in supporting Mono Lake water resources management. Notably, this thesis identifies a critical balance in model complexity, where neither overly simplistic nor excessively complex models yield the most accurate predictions. Instead, a balanced approach, incorporating nuanced model training and optimization methods emerges as crucial tools for minimizing model overfitting and capturing the nuanced patterns of Mono Lake’s water levels. The insights from these training and optimization exercises provided pivotal learning to improve accuracy. Hence, machine learning models like this can be used for informing water diversion strategies, ecological conservation efforts, and policy development, ensuring Mono Lake’s sustainability amidst changing environmental conditions. By providing a detailed analysis of Mono Lake’s water level dynamics and the predictive capabilities of LSTM Ensemble models, this thesis contributes valuable knowledge to the fields of hydrology, environmental management, and machine learning, offering a blueprint for leveraging machine learning in the stewardship of natural resources. The implications of this research extend beyond Mono Lake, suggesting a broader applicability of machine learning in hydrology, climate forecasting, and water resource management. As climate change and human activities increasingly impact water resources, the integration of predictive modeling like machine learning into natural resource management offers a path forward for balancing ecological preservation with human needs, ensuring the sustainable management of water bodies worldwide. The dataset for this thesis can be found at https://doi.org/10.7302/xbet-5212.