Increased water variability means that historical streamflow patterns no longer hold, extreme droughts are becoming commonplace, and critical ecosystems are nearing tipping points. At the same time, transforming the energy grid away from fossil fuels will require stable, dispatchable power sources. In this context, data-informed decision making is crucial.
Increased water variability means that historical streamflow patterns no longer hold, extreme droughts are becoming commonplace, and critical ecosystems are nearing tipping points. At the same time, transforming the energy grid away from fossil fuels will require stable, dispatchable power sources. In this context, data-informed decision making is crucial.
Earlier this month we rolled out a new generation of HydroForecast Seasonal. This service enables water users to confidently make time-sensitive decisions on longer, relevant, and critical timelines. Read on to learn about HydroForecast’s new seasonal streamflow forecasts. This post will cover:
Water management occurs across many time horizons, from hourly to annual to decadal scales. Granular, hour-by-hour information is vital for making decisions for the immediate days ahead. But only when combined with a seasonal frame of reference can these short-term decisions account for variability in water supply. Researchers in Germany recently estimated that “seasonal forecast-based action for droughts achieves potential economic savings up to 70%...[and] savings of at least 20% occur even for forecast horizons of several months” for very warm months and droughts.
High-quality forecasts for the upcoming months can answer questions like:
Traditionally, water users might approximate answers to these questions by relying on a combination of their extensive experience in a basin, extrapolated historical trends, and/or coarse resolution outlooks that focus primarily on precipitation and temperature. Yet as climate and weather patterns change, historical trends often no longer hold true, and more factors than just the weather influence streamflow. A forecast that adapts and anticipates these changing basin conditions is critical.
HydroForecast Seasonal can provide quantitative answers to these questions.
With HydroForecast, users no longer need to rely on historical trends or review rough weather forecasts to make critical decisions. HydroForecast incorporates meteorological and historical information, plus thousands of near real-time global and localized data points, into a consolidated streamflow prediction that allows users to make decisions informed by high-quality and forward-looking data. The forecasts quantify and validate experienced users’ intuitions and alert them to potentially abnormal conditions as soon as possible. The model identifies the range of possibilities, depicts how confident the model is that those ranges will happen, and compares those likelihoods to historical data so users can make decisions sooner and with more confidence.
HydroForecast Seasonal’s streamflow predictions are presented using box and whisker plots. The model predicts both the mean and median flow, as well as confidence intervals that bound the range of predicted flows during the month. Those confidence intervals illustrate the expected range of variability. For instance, when looking at a plot of historical streamflow, the box would showcase the middle 50% of historical flows, while the whiskers would extend to capture 90% of flow ranges during that month historically. A typical monthly prediction for streamflow (in cubic feet per second or cubic meters per second) might look like the plot below:
HydroForecast Seasonal creates a prediction like this for each month, illustrating: 1) historical observations where available, 2) HydroForecast’s predictions, and 3) alternative forecasts. Since HydroForecast predictions are reissued daily, users can easily visualize how the predictions have evolved and the confidence bounds have changed as time progresses.
Let’s walk through plots from a sample HydroForecast Seasonal dashboard before returning to the planning questions from the introduction. We’ll start with the streamflow forecast itself. The image below highlights a sample streamflow forecast for March 2021 to May 2022, depicting:
Since water managers have historically used a few key variables to estimate seasonal flows, such as precipitation and temperature, the HydroForecast Seasonal dashboard includes visualizations for them. This suite of plots helps users make sense of what they’re seeing and provides additional context on the streamflow predictions.
Next is an example of one of the precipitation inputs used for a forecast at the same site. The precipitation plot illustrates:
Note that NOAA doesn’t produce probabilistic forecasts, so the orange and green predictions are lines rather than ranges. Hovering over a data point, as shown below for May 2022, brings up a chart with the exact values.
We also include a satellite-derived Normalized Difference Vegetation Index (NDVI) - which measures ‘vegetation vigor’ or how much foliage is in an area, and Normalized Difference Water Index (NDWI) - which measures the presence of surface water. Both of these are inputs into HydroForecast Seasonal. Tracking vegetation trends over a year - for example, whether the annual green-up is more intense or earlier than normal - sheds light on both streamflow patterns and on the broader conditions in a basin, like fire risk and habitat.
As the gray boxes in the plot below show, vegetation in this basin typically increases over the course of the spring and through the summer, then drops off in the fall as winter approaches. For the first three months of this water year though, you can see that the observed vegetation is on the high side: the May value is just above the 75th percentile of historical observations. This tells us that this basin is ‘greening’ slightly more and earlier than the historical median.
Using what we’ve learned from the HydroForecast Seasonal model, let’s return to the questions we posed earlier to glean some insights for the basin we’ve been exploring:
What are streamflow volumes likely to look like in the upcoming months?
Let’s look at May 2021. In April, HydroForecast expected 50% of flows would be in the 5,000 to 8,500 CFS range during May, with potential but unlikely flows ranging from 4,000 to 9,000 CFS. By May, the model predicted a smaller band of flows: 90% confidence between 5,600 and 7,800 CFS, with potential extreme flows from 5,000 to 8,250 CFS. This increasing confidence is common as time periods in the future approach. For more granular predictions for the upcoming days, many of our partners use HydroForecast’s hour-by-hour short-term forecasts.
Will this year be dry or wet compared to what we’ve seen in the past?
For this basin it looks like most of 2021 will be an average year. Early predictions for the fall show potentially wetter and warmer than average conditions, with October and November precipitation in the 75th percentile and November flows potentially up by ~300 CFS from average. For hydropower asset owners, this information might suggest a 3-6 month power generation and revenue plan similar to or just above prior years. For environmental planners, this could suggest generally viable flows for fish growth and passage.
Is our basin at risk of a drought? How severe could it be?
At no point are flows predicted to be anomalously low. While they appear likely to drop below 2,000 CFS in 2021 from August to October, those ranges are within the historical averages of 1,000 to 2,000 CFS.
Do we expect the spring snowmelt period and green-up to happen earlier or later than normal?
While the predictions are still early, it appears that flows in April and May 2022 may be higher than average. Note that the model incorporates thousands of data points beyond just temperature and precipitation. So, even though the NOAA precipitation forecasts only extend 6 months ahead, HydroForecast is able to utilize climatological data to predict higher than normal flows next spring. According to our NDVI plot, this year’s green-up is higher than it has been in the past, so it will be interesting to follow how the rest of the year plays out.
Will this stream run too low to support fish passage? When will conditions be most suitable for environmental restoration projects, in-water maintenance, and construction?
While forecasting absolute maximum and minimum flows is a challenge, the model’s 5% and 95% confidence intervals can be illuminating. The lowest 5th percentile prediction is in October, with a forecast of 524 CFS as of the May 12th forecast. The stream therefore doesn’t appear to be at risk of low flows that wouldn’t support fish passage. The lowest overall periods that might be conducive to any in-stream work (or more likely, near-stream shore-based work, since flows stay high throughout this year) would be August to October 2021 and potentially December to February 2022. An expected November 2021 spike in flows could disrupt long-term construction projects on dams or environmental restoration projects.
Before deploying the new models for production, our team thoroughly benchmarked their prediction skill in each customer location. We evaluated the model in the case study above in this way, and do so for each partner’s model. We typically train our models on a subset of available historical data, then evaluate performance against a different historical period that the model has never seen before. When we evaluate them, we examine the models’ prediction accuracy, or ‘goodness of fit’ (using metrics like Nash-Sutcliffe Efficiency and Kling-Gupta Efficiency), as well as difference from observations (using metrics like Mean Absolute Error and percent error).
For one customer, HydroForecast Seasonal is predicting flows one year ahead more accurately than using the long-term average. The customer’s previous model was not able to make predictions so far into the future, so the model’s time horizon, coupled with its accuracy, represented a significant improvement for them. In another location, the new seasonal model outperformed both historical averages and an existing forecast product our customer used in-house.
HydroForecast’s models are physically-guided statistical models. They use approaches like machine learning to make sense of a large volume of data and learn the relationships between various inputs (e.g., precipitation and temperature) and streamflow. Unlike many purely statistical models though, Upstream Tech’s team of hydrologists, meteorologists, and climatologists validate that HydroForecast’s models align with what we know to be true about physical hydrologic processes. For example, our team constrains the models so that rain on one day cannot influence streamflow on the day before, when it has not yet fallen. This constraint may seem obvious, but innovative approaches must be taken to apply machine learning techniques to the complex relationships of the natural world.
Each customer’s seasonal model starts with Upstream Tech’s core forecasting approach and is tuned specifically for the customer’s location. Where inputs improve forecasting skill and align with hydrological theory, they are incorporated. Where they don’t, they are removed. The model might include inputs such as:
A recent Nature article summed up seasonal forecasts’ value:
With rising demands on water management, switching towards beneficial seasonal-forecast-based early actions saves expenses and aids climate proofing. Consequently, we stress the advantage and necessity of considering seasonal forecasts in hydrological decision making.
We’re excited by the new HydroForecast Seasonal’s performance and interested to explore ways in which it can help water users make more informed decisions. If you’d like to learn more, share feedback on what would be most helpful for your organization, or just talk through water-based decision making challenges, we’re eager to collaborate. You can reach us at team@hydroforecast.com or on Twitter at @upstream_tech or @hydroforecast.