When it comes to hydropower, water is both the resource and the challenge. Predicting how much water will be available weeks or months in advance has historically been fraught with uncertainty. But a new generation of machine learning-based tools is beginning to change that. Among those involved in this shift is Upstream Tech, whose HydroForecast platform has gained recognition across the hydropower sector for its data-driven, real-time forecasting capabilities. With the launch of HydroForecast Seasonal-3 , the company is pushing the frontier of seasonal streamflow prediction, helping dam operators and water managers better plan for what lies ahead.
“There’s a clear need for reliable seasonal forecasts,” says Marshall Moutenot, CEO and co-founder of Upstream Tech. “But it has always been one of the most difficult challenges in the field of hydrology.”
HydroForecast Seasonal-3, launched earlier this year, builds on five years of work by the Upstream Tech team, whose interdisciplinary background spans remote sensing, software engineering, environmental science, hydrology, and applied machine learning. The latest release represents a step-change in predictive power, particularly over multi-week to seasonal timescales.
A fundamental challenge
Forecasting river flows is fundamentally difficult because of the complex, nonlinear interplay of variables involved. Snowpack, soil moisture, precipitation patterns, temperature trends, vegetation cover, and upstream infrastructure all influence how much water will make its way to a given point in a river system. While short-term forecasts can often rely on recent observations and weather predictions, seasonal forecasting requires integrating long-range climate data, hydrologic memory, and historical patterns.
“In many ways, streamflow forecasting is harder than weather forecasting,” Moutenot says. “You’re dealing with chaos, noise, and uncertainty. But the stakes
are high, especially for hydropower operators who need to make decisions about water storage, turbine dispatch, and flood control weeks or even months
in advance.”
Traditional models used for seasonal forecasting tend to be either physically based, requiring detailed hydrologic and land surface modelling, or statistical, relying on historical analogues. Each approach has limitations. Physical models can be cumbersome and require detailed calibration. Statistical models may struggle under nonstationary climate conditions. HydroForecast aims to blend the strengths of both, using machine learning to learn hydrologic behaviour from vast datasets without being locked into a rigid physics-based framework.
Learning from the data
At the core of HydroForecast Seasonal-3 is a machine learning architecture that ingests terabytes of global data, including satellite imagery, weather forecasts, reanalysis products, and river gauge observations. This allows the model to learn spatial and temporal patterns in hydrology without needing manual calibration for each watershed.
“We use global data sources so we can train a single model that works across diverse conditions,” explains Moutenot. “We want it to be scalable and transferrable – something that can give you a good forecast whether you’re in the Rockies or the Alps.”
This is critical for hydropower operators with assets spread across multiple sites. A single unified model reduces the burden of customization and supports consistent, portfolio-wide forecasting. According to Moutenot, Seasonal-3 represents a major step forward in this direction, improving predictive skill, extending the forecast horizon, and increasing trust among
end users.
“We’ve focused a lot on making the uncertainty quantification more useful,” he notes. “Operators need to understand not just the most likely outcome, but the range of possible scenarios. We’re helping them turn forecast uncertainty into risk-aware decisions.”
From forecast to action
HydroForecast is designed to deliver value in multiple operational contexts. Some customers use it for day-ahead energy market participation, others for seasonal reservoir planning or flood risk management. What unites these use cases is the need for timely, reliable streamflow information.
“We work with large utilities that need forecasts for dozens of sites,” Moutenot says. “They can’t afford to spend weeks calibrating models for each one. That’s where a machine learning approach really shines – it can generalise without sacrificing accuracy.”
With Seasonal-3, Upstream Tech has also improved how the model communicates its predictions. Enhanced visualisations, clearer probabilistic outputs, and integration with existing data pipelines all help customers put forecasts into practice.
“We often hear that our model helps teams have better conversations,” Moutenot says.
“When everyone’s looking at the same forecast with the same level of confidence, it makes coordination easier.”
Trust and transparency
While machine learning has immense promise, it can also be a black box – an opaque system that users struggle to understand. To address this, Upstream Tech has invested heavily in model interpretability and transparency.
“We publish our methods, we validate our performance, and we make sure the model’s predictions make sense,” says Moutenot. “We’ve open-sourced components of our work and provided detailed documentation so users can see how the model works.”
This has helped build trust with partners in both the public and private sectors. HydroForecast is currently used by utilities, energy traders, government agencies, and hydropower producers across 15 countries. The team has collaborated with industry leaders to evaluate and refine the tool. Key examples include grant awards from the US Department of Energy as well as leading results in a streamflow forecasting competition alongside hydropower organizations, like Hydro-Québec (H-Q), the Tennessee Valley Authority (TVA), the U.S. Bureau of Reclamation (USBR), and Southern Company.
“We don’t expect people to take the forecasts on faith,” Moutenot says. “We want to earn that trust through rigorous validation and continuous improvement.”
The need for accurate seasonal streamflow forecasts is only growing. Climate change is disrupting historical patterns, leading to more frequent droughts and floods. Meanwhile, hydropower operators are under increasing pressure to balance energy generation with environmental flows, recreation, and flood risk mitigation.
“There’s a lot riding on how well we understand our water systems,” says Moutenot. “That’s why we see this as more than just a technical challenge. It’s about making better decisions for people and the planet.”
For Upstream Tech, the launch of Seasonal-3 is part of a broader mission to bring modern data science to water management. The company was founded in 2016 with the goal of applying cutting-edge tools to environmental challenges. In addition to HydroForecast, it offers Lens, a remote sensing platform for land and habitat monitoring.
“Our team comes from a range of backgrounds – NASA, Google, conservation NGOs – and we’re united by a shared mission,” Moutenot says. “We believe technology can help us steward natural resources more wisely.”
Beyond the model
Looking ahead, Upstream Tech sees room for further innovation. While Seasonal-3 focuses on streamflow, future versions could incorporate water temperature, sediment load, and other hydrologic variables. The team is also exploring how to integrate local observations and user feedback to continually refine the model.
“We’re just scratching the surface of what’s possible,” says Moutenot. “There’s so much potential in combining global data with local expertise.”
For now, the focus is on getting Seasonal-3 into the hands of more users and demonstrating its value across diverse applications. That includes helping dam operators plan for spring melt, utilities manage energy markets, and conservationists protect aquatic ecosystems.
“We’re proud of the progress we’ve made, but we’re even more excited about what’s next,” Moutenot says. “This is a long-term effort. Water forecasting is never ‘solved,’ but we’re getting better with each season.”
Operational impact and user engagement
The practical benefits of HydroForecast are already being felt by users in the field. One utility, for example, used the tool to plan early reservoir drawdowns ahead of a heavy snowmelt season. By anticipating high inflows, they avoided spill events that could have compromised both safety and energy production. In another case, a conservation group used seasonal forecasts to time fish habitat restoration work, aligning efforts with predicted low-flow conditions.
Moutenot notes that user feedback has been crucial to the tool’s evolution. “We listen closely to what our partners need,” he says. “That’s how we’ve improved the forecast lead times, tailored the outputs to their operational systems, and added flexibility in how the data are delivered.”
HydroForecast’s delivery methods are also designed with operators in mind. Data can be accessed via API, web interface, or direct integration into energy management platforms. This allows users to ingest the forecasts into their existing workflows without added friction.
“It’s not enough to have a good forecast,” Moutenot says. “It has to be actionable. It has to fit into the way people already make decisions.”
Climate adaptation and a role for policy
As climate change continues to reshape hydrology, tools like HydroForecast are likely to become increasingly indispensable. Changing snowpack patterns, altered storm tracks, and higher evapotranspiration are all shifting the timing and magnitude of flows. Traditional rule curves and planning assumptions may no longer hold.
“We see HydroForecast as part of the climate adaptation toolkit,” Moutenot says. “Better foresight allows for more flexible, resilient management.”
He also sees a potential role for policy and regulation in supporting the adoption of
advanced forecasting. Just as grid operators now
rely on probabilistic load forecasts, water managers may eventually be required – or incentivised – to use probabilistic inflow forecasts to inform decision-making.
“We’re not there yet, but I think we’ll get there,” he says. “The more we can quantify uncertainty and turn it into useful guidance, the more we can support proactive rather than reactive management.”
The flow ahead
As water managers grapple with increasing complexity, tools like HydroForecast Seasonal-3 offer a powerful new way to peer into the future. By combining global data, modern machine learning, and a commitment to transparency, Upstream Tech is helping bring water forecasting into the 21st century.
“We’re incredibly motivated by the opportunity to make a difference,” says Moutenot. “Whether it’s helping a hydropower operator optimize generation or supporting a community facing drought, we believe better forecasts can lead to better outcomes.”
With Seasonal-3, that vision is becoming a reality. And as climate pressures grow, the ability to anticipate and adapt will be more valuable than ever.
References
https://www.upstream.tech/posts/hydroforecast-outperforms-the-competition
https://www.upstream.tech/posts/2021-07-21-upstream-tech-awarded-phase-2-sbir-grant-from-department-of-energy
HydroForecast Seasonal-3
HydroForecast Seasonal-3 is the latest generation of seasonal streamflow forecasting, built on a newly developed AI-based hydrology architecture. This model is designed to support long-term water resource planning, medium- and long-range risk assessment, and seasonal water management.
Key features of Seasonal-3 include:
- Daily timestep forecasts with flexible horizons of up to one year.
- Integration of multiple ensemble weather forecasts, satellite observations, and snowpack datasets.
- Sub-basin level data ingestion to better reflect local hydrological variability.
- Enhanced performance in cold climates, aided by training on Canadian gauge data.
The new architecture brings together global observations and meteorological inputs to generate probabilistic forecasts that are both scalable and adaptable. Seasonal-3 has demonstrated strong performance across a range of hydrological conditions and regions.
