Providing accurate forecasts of human modified streamflow is critical for applications ranging from natural resource management to hydropower generation. In this study we evaluate the performance of Long Short Term Memory (LSTM) based neural networks, which maintain an internal set of states and are therefore well-suited to modeling dynamical processes. This research builds on previous work demonstrating that an LSTM model can predict streamflow in out of sample basins with similar or greater accuracy than traditional forecast models specifically calibrated on those basins [1]. Using meteorological data from NOAA’s Global Forecasting System (GFS) and North American Land Data Assimilation System (NLDAS), remote sensing data including snow cover, vegetation and surface temperature from NASA’s MODIS sensors and streamflow data from USGS, we first train an LSTM model on 100 unmodified river basins, and evaluate its predictions on previously unseen human-altered basins. We then train models on a combination of natural and human modified basins and experiment with the effects of new data sources and additional model architecture in predicting human altered streamflow. By training on multiple basins with dynamic climate, land surface and human inputs, we can test the model’s understanding of general hydrologic relationships and human use patterns. We evaluate our models on “out of sample” rivers (previously unseen by the model) that have been altered by dam operations and agricultural withdrawals in northern California. We find that the models trained on natural and modified basins capture human modified flows better than our baseline model trained on natural basins. [1] Kratzert, F., Klotz D., Shalev, G., Klambauer, G., Hochreiter, S. & Nearing, G. Benchmarking a Catchment-Aware Long Short-Term Memory Network (LSTM) for Large-Scale Hydrological Modeling. Preprint at https://arxiv.org/abs/1907.08456 (2019)