Reducing market risk in run-of-river hydropower: The economic value of improving predictions of day-ahead production

Accurate streamflow forecasts are essential for optimizing hydropower production and reducing market imbalance costs, particularly for small-scale hydropower plants with limited storage capacity. This study evaluates the potential of an LSTM neural network for hourly streamflow prediction at operationally monitored hydropower systems with limited historical observations and assesses its economic value in a hydropower production context. The effect of two aspects on prediction accuracy is investigated: (i) including back-calculated streamflow data from 110 run-of-river hydropower plants during training, and (ii) implementation of a data assimilation (DA) approach using back-calculated streamflow as input. The performance of the model is assessed in a perfect weather forecast context using regional reanalysis data and compared with both a persistence model and the operational model currently used by the largest small-scale hydropower operator in the Nordics, Småkraft AS. The results show that including data from hydropower plants in training provides the model with valuable information, while data assimilation yields the greatest performance gain. The proposed model has the potential to decrease annual imbalance fees for Småkraft AS by 23.6% (143 kEUR) compared to the model currently used and reduces exposure to extreme prices in the balancing market by reducing the CVaR 75 with 16% (380 kEUR). The findings demonstrate that enhanced hydrological forecasting reduces imbalance cost and market risk, while supporting more stable grid operations.