Abstract
Short and long range reservoir inflow forecast is essential for efficient real time operational planning, scheduling of hydroelectric power system and management of water resources. Large-scale climate phenomenon indices have a strong influence on hydrological processes under complex weather conditions, and it should be considered to forecast reservoir inflow for efficient dam operation strategies. This study aims to explore the relevance of large-scale climate phenomenon indices in improving the reservoir inflow prediction at short-term time scales. This paper presents a simple and effective framework to combine various data-driven machine learning (ML) algorithms for short-range reservoir inflow forecasting. Random Forest (RF), Gradient Boosting Regressor (GBR), K-Nearest Neighbors Regressor (KNN), and Long Short-Term Memory (LSTM) were employed for predicting daily reservoir inflows considering various climate phenomenon indices (e.g., Arctic Oscillation, North Atlantic Oscillation, and Southern Oscillation Index) and hydroclimatic variables (precipitation), accounting for time-lag effects. After training the individual ML algorithm, a framework was developed to create an ensemble model using a robust weighted voting ensemble method to quantify forecasting uncertainty and to improve the model performance by combining the inflow results of the single ML model and the highest vote is chosen based on the weights assigned to the single ML model. The developed framework was examined in two distinct reservoirs located in India and California, USA. The ensemble model consistently outperformed the standalone RF, GBR, KNN, and LSTM in predicting high (flood control and monsoon seasons) and low flows (runoff and non-monsoon seasons) of both study reservoirs. The demonstrated short-term reservoir forecasting model allows reservoir operators to adapt and add regional hydrological and large-scale climate indices in real-time decision-making. The presented framework can be applied for any reservoir inflow forecasting. Introduction Reservoirs are important hydraulic structures in terms of collection, storage, and diversion of freshwater for purposes such as drinking, irrigation, hydropower, flood management operations, and drought relief (Simonovic, 2020). Many developing countries (such as India) have made efforts towards sustainable reservoir operations to account for the adverse effects of dams for regional water resources management (Bhadoriya et al., 2020, Goel et al., 2020, Rehana and Mujumdar, 2014). Reservoir inflows play a major role in water allocations fulfilling various water demands and balancing the flows from upstream catchments to downstream regions during floods (Rehana et al., 2020). The operation of reservoir is complex, involving multiple time scales, multi-flow regimes, and unpredictable emergencies (Zhang et al., 2018). Forecasting reservoir inflows is a preliminary steps in reservoir operation, providing guidelines and rule curves for optimal water allocations to satisfy water supply, irrigation, industrial, hydropower, and environmental conservation requirements (Kasiviswanathan et al., 2020). In this context, long and short-term reservoir inflow forecasts are vital for sustainable water resources planning and management (Rehana et al., 2020). Short-term seasonal and monthly operations are relevant for optimal economic benefits, water supply, and downstream augmentation operations (Zhang et al., 2018). Daily time scale operations also relate to controlling floods, power grid loads, and emergency operations (Noorbeh et al., 2020). In this context, the prediction of daily reservoir inflows is prominent for real-time reservoir operation under hydroclimatic extremes such as low and high flows. Conventionally, real-time reservoir operations are associated with high variability and rapid changes, often deviating from the operation rule curves developed based on physical models (Oliveira & Loucks, 1997). Further, physically based models are subject to various uncertainties and need to account for complex natural and human-influenced hydrological conditions and water demands. Although hydrological models are best tools for predicting inflows, and providing details of precipitation, evapotranspiration, soil, and land use characteristics of the upstream catchment area, these models require a rigorous amount of landscape information at various spatiotemporal scales and several pre-processing efforts (Awol et al., 2019). Alternatively, data-driven algorithms based on Machine Learning (ML)
