The Economic Benefit of AI-Driven Day-Ahead Hydropower Production Forecasts

To reach the goal of net-zero emissions and carbon neutrality, the European power system is changing towards more variable renewable energy production. The increasing share of weather-dependent energy production however, makes it more challenging to maintain a stable grid frequency. This results in larger penalties for energy producers contributing to instability.

Norway is well connected to the European energy system, exposing it to market conditions in neighboring countries. To ensure a stable grid frequency, the national transmission system operator is responsible for balancing production and consumption volumes. To support the balancing operations, all power producers must submit day-ahead production forecasts. Deviations from the predicted volumes are subject to imbalance fees. In addition, power producers need to buy and sell energy in a dedicated market to balance deviations. To avoid large imbalance costs and support grid stability, accurate high-resolution day-ahead production forecasts are essential.

In Norway, the largest variable renewable energy source is run-of-river hydropower. Forecasting run-of-river hydropower production is equivalent to forecasting streamflow. The industry has expanded rapidly lately, resulting in many newly commissioned plants with limited streamflow observations. Thus, there is a need for a forecasting model that can make accurate predictions with limited training data.

In this study, we explore the potential of using a Long Short-Term Memory neural network to forecast hourly streamflow. The model is trained on historical data from 215 Norwegian gauging stations. To improve training efficiency, we adopt a multi-frequency approach in which earlier time steps are processed at a daily resolution, while more recent inputs retain their original hourly resolution. We explore two approaches of improving model performance: including data from 139 run-of-river hydropower plants during training and including streamflow estimated from production data through a data assimilation approach.

The results show that both approaches improve the performance of the model and the final model outperforms both a persistence model and one of the leading providers of run-of-river hydropower production forecasts in Norway. The potential economic value of the improved day-ahead forecast is estimated on the basis of both reduced imbalance fees and reduced exposure to volatile prices in the balancing market. This shows that the model we propose has the potential to improve upon existing models and contribute to overall grid stability.