Streamflow Prediction Using Machine Learning in the Aral Sea Basin
Author | : Kallibek Kazbekov |
Publisher | : |
Total Pages | : |
Release | : 2020 |
ISBN-10 | : OCLC:1204117849 |
ISBN-13 | : |
Rating | : 4/5 (49 Downloads) |
Book excerpt: Globally, streamflow is altered as a result of anthropogenic impacts and climate change, thus creating uncertainties in future planning for streamflow water users including hydropower plants and irrigation. These uncertainties are addressed using process-based hydrologic models. However, there are data-limited regions such as the Aral Sea Basin (ASB), for which calibration and validation of process-based models are challenging. In such cases, Machine Learning (ML) offers an alternative methodology to predict short-term future streamflow (days to months ahead) to optimize water management. Also, ML can be used to identify the main drivers of streamflow in different parts of the ASB, which will allow to prioritize and regionalize the investments into data availability. This will eventually lead to better performance of process-based models. The overarching objective of the project is to determine the dominant drivers of streamflow in different parts of the ASB. The first specific objective is to evaluate the predictive power of models when exposed to “future” data. Models were trained using three ML algorithms: Multi-layer Perceptron, Random Forest, and Gradient Boosted Decision Trees. The prediction accuracy of models on the testing set was evaluated using the Nash-Sutcliffe efficiency, percent bias, and relative root mean squared error. The best model for each streamflow gauging station was identified using a composite of these three metrics. Since empirical models require a large number of training observations, which is challenging in data-limited regions, we propose a novel method of increasing the number of training observations by grouping training data based on specific grouping rules. Therefore, the second objective is to validate the proposed method. This objective will assist in reaching the final objective, which is the determination of main streamflow drivers in different parts of the ASB. In total, 80 models out of 110 best models had acceptable prediction accuracy. The grouping frameworks improved the accuracy of models for 60% of gauging stations. The temperature was the main driver of streamflow at higher elevations vs. irrigation and evapotranspiration at lower elevations. The capacity of reservoirs was among the most important variables controlling the streamflow for gauging stations, watersheds of which include at least one reservoir.