Getting to Know the Model Inputs
A neural network based model allows HydroForecast™ to take inputs such as near real-time satellite snow and forest greenup observations covering the entire basin. Hydrologic theory indicates these observations are highly relevant to inflow forecasting, yet successfully integrating them into a traditional physical model, especially the operationalization of a data processing pipeline capable of collecting and processing remote data for forecasts at scale, is complex and time consuming. All together our model uses the following inputs, each of which are provided as a time series of dynamically changing values.
Weather Forecast Data
HydroForecast™ uses data from both medium and long term weather datasets including NOAA’s Climate Ensemble Forecast System (CEFS) and NOAA’s Global Ensemble Forecast System (GEFS). We also use the European Center For Medium-Range Weather Forecasts’ (ECMWF) High Resolution (HRES), Atmospheric Ensemble (ENS), and Seasonal (SEAS) forecasts. We and others have found ECMWF weather forecast skill to produce more accurate inflow forecasts, particularly at shorter forecast horizons. Critically, we have found that by providing inputs from multiple forecasts simultaneously, the neural network is able to learn to use the similarities and differences between weather forecasts to produce an inflow forecast with higher accuracy than can be achieved with either forecast on their own. Today, many sophisticated utilities address the problem of single-forecast reliability by having teams that create bespoke hybrid forecasts from multiple forecasts based on intuition. By providing multiple forecasts to the neural network directly, we are able to replicate this process with a data-driven, reliable, and flexible (i.e. add more forecasts as they become available) approach. From each of our weather forecasts we use precipitation, air temperature, solar radiation, humidity, and wind speed data as well as additional input variables as needed. We regularly review current and prospective inputs as part of our ongoing research and development.
Satellite Land Surface Data
Near real-time satellite observation data provide the model with hydrologically relevant observations of a range of key land surface conditions spanning the full extent of the basin in question. The satellites selected provide up to daily updates of conditions, ensuring the models have an always-up-to-date understanding of both gradual and abrupt changes to a landscape. The analyses provided to the model include Vegetation Vigor from NASA’s MODIS Normalized Difference Vegetation Index (NDVI) and Snow Extent from NASA’s MODIS. Land Surface Temperature also processed from MODIS during both the day and night time provides direct observations of snow accumulation and melt conditions as well as soil moisture (dry land has larger fluctuations between day and night time surface temperatures). We also include high resolution elevation data to derive basin boundaries and river paths, elevation, and slope information.
Visualizations of satellite vegetation and snow observations computed from NASA MODIS satellite data on three selected days during a spring snowmelt period. Healthy vegetation (dark green), bare ground (brown), and snow (white) demonstrate the spatially detailed and dynamic information these observations provide.
In Situ Sensor Data
Our models additionally integrate in situ sensor data such as meteorological gauge data, stream gauge observations, and snow sensor information via a novel technique of data assimilation. During the initial phase of this pilot we will identify ground sensors in the basin which can provide hydrologically relevant data and integrate them as model inputs.