Integrating data-driven methods with a complex hydrodynamic modelling system to assess sensitivity, uncertainty, and forecast performance

For nuclear power plants sourcing water from a nearby river or lake, water temperature forecasting is critical. With environmental and safety constraints related to both cooling water releases and withdrawals, accurate forecasting is needed for optimal nuclear plant operations. Low-quality forecasts can result in unnecessary generation curtailment, inefficient nearby hydropower operations (resulting in lost low-carbon generation replaced by carbon-intensive generation), and insufficient lead time for mitigating or eliminating high water temperatures. Moreover, with future projected climate change, threats to plant cooling or environmental water temperature limits are expected to increase – especially in the seasonally warm, dry months. To predict and mitigate such temperature concerns, a calibrated Delft3D hydrodynamic model, coupled with a nuclear cooling system model, is used. The Tennessee Valley Authority’s Browns Ferry Nuclear Plant is used as a case study, as it provides a large amount of generation regionally and the Tennessee River has recently been subject to increased hydrometeorological variability. While the model is used to predict and mitigate water temperature issues, in its current state it lacks a historical dataset from which to fully validate its assumptions and it uses suboptimal statistical approximations as potentially sensitive inputs. Due to computational expense and lacking historical data, model sensitivity is poorly understood. To overcome these issues, we utilize machine learning (ML) to accurately reconstruct the historical upstream boundary’s timeseries, develop a water temperature forecast model (at the boundary) that improves over current approximations, cluster joint, hydropower generation schedules to understand the historical diversity of operations, and design and execute a computationally efficient sensitivity analysis that assesses both natural forecasts and decision-based inputs. We then use analysis of variance (ANOVA) to determine contributions of various components to the variance among hindcasts including meteorology (GEFS Reforecast V2), joint hydropower schedules, system state, location, and forecast leadtime. We show that, in different meteorological regimes, a specific range of joint hydropower generation scenarios can be scheduled, thereby allowing for the possibly of more optimal generation in the wake of water temperature concerns for nuclear operations.