This study improves forecasting of water flows through hydraulic structures, such as freshwater diversions, even when they are not actively operating. These non-operational flows (e.g., leakage through the structure), can still impact downstream conditions but are often left out of prediction models. This study applies machine learning to forecast flows at downstream locations to inform flows at a diversion structure, in this case the Bonnet Carré Spillway, a diversion on the lower Mississippi River (USA). This model uses publicly available, upstream discharge and precipitation data to predict both the timing and amount of flows, including the likelihood of diversion events, up to several weeks in advance. While developed for the Bonnet Carré Spillway, this method can be applied to other diversions and water gauges to support flood risk mitigation, water quality management, and resource planning where engineered diversions are present.