Multi-Model and Multi-Scale Operational Streamflow Forecasting in Arid and Semi-Arid Regions Using Machine Learning and Data Synergy

Abstract Submission: Accurate streamflow forecasting plays a crucial role in operational management within arid and semi-arid landscapes by enabling effective resource allocation, optimizing hydropower production, and mitigating flood and drought events. Large-scale hydrological datasets, such as CAMELS-US, often fail to adequately represent arid/semi-arid regions like Arizona. To address this, we employ multiple neural network models—including Gated Recurrent Units (GRU), Long Short-Term Memory (LSTM), RNN-CNN hybrids, and Transformer architectures—to generate operational streamflow forecasts at 1-day, 3-day, 7-day, and 30-day horizons.

We then leverage Caravan datasets to evaluate data synergy and integrate earth observations using Google Earth Engine to enhance model inputs. Additionally, we implement physical processes governing streamflow dynamics to embed hydrological constraints directly into the models. We utilize Google Cloud-Vertex AI infrastructure to train our models on river streamflow data using 8 catchment characteristics, and 40 hydro-meteorological covariates.

Our preliminary results show that data synergy from Caravan datasets significantly improves forecasting accuracy across all models and forecast horizons. Among the evaluated models, LSTM and Transformer architectures exhibit superior performance. LSTMs effectively capture long-term dependencies in the data, while Transformers excel in handling peak flow events and sudden hydrological changes due to their attention-based mechanisms. These models benefit from the enriched input provided by the Caravan and EO datasets, leading to more accurate and reliable streamflow forecasts.

Keywords: Operational forecasting, Data synergy, LSTM, Transformer, Arid and semi-arid landscapes

Learning Objectives/Expected Outcome (Optional) :