Improving reservoir water quality by optimizing weir operations with reinforcement learning and SWAT
Published in Ecohydrology & Hydrobiology, 2025
Recommended citation: Pyo, J. C., Baek, S. S., Abbas, A., Kim, H. G., Lee, J., Kim, S., Chun, J. A., & Cho, K. H. (2025). Improving reservoir water quality by optimizing weir operations with reinforcement learning and SWAT. Ecohydrology & Hydrobiology, 25(4), 100679. https://doi.org/10.1016/j.ecohyd.2025.100679
Software agents that learn by trial and error cut algal pigment concentrations by an average of 35% across six weirs on Korea’s Nakdong River, with the largest gains during the summer peaks that matter most. Weirs — low barriers built across a river to hold water back — can be operated to discourage algae, but deciding how much to release and when involves trade-offs that are difficult to reason through directly. A watershed model stood in for the real river, letting the agents rehearse release schedules and observe the water quality consequences before any strategy was judged. Because the improvement comes purely from re-timing releases at structures that already exist, it requires no new construction.