Front of meter (FTM) energy storage is the fastest-growing segment of the storage market today. The U.S. Energy Information Administration (EIA) recently predicted 10 gigawatts of large-scale storage will be added to the grid between 2021 and 2023 – 10 times the capacity added in 2019. Already, these kinds of projects are helping Hawaii move past fossil fuels, Massachusetts to achieve cleaner peaks, New York to compensate distributed resources, and California to become more resilient.
Customers and partners often ask me about the role of software and machine learning or AI in maximizing ROI for their FTM projects, specifically in capturing wholesale market revenues. They’re particularly interested in forecasting: How does Stem do it? How can they know a provider is good at it? What kind of data is being analyzed? As new technology evolves in this market, they understandably want every benefit available.
In this blog, I’ll explain how Stem’s advancements across the three basic parameters of AI-driven energy storage and optimization software – forecasting, optimization, and scheduling (“FOS,” to insiders) – are driving value creation for FTM projects in wholesale markets. I’ll also introduce AthenaⓇ Bidder, Stem’s new wholesale market bidding software and share some of Stem’s takeaways from having deployed several first-of-kind FTM projects that are currently earning multiple wholesale market value streams.
Accurately predicting short-term market prices, solar generation, and battery state of charge (SOC) at different points in the future is Athena Bidder’s first step in forecasting for our customers and partners. Storage optimization can only be as good as the forecasts that go into it! In a day-ahead market, for example, bidding a certain amount of stored solar energy into tomorrow’s evening peak relies on accurately forecasting tomorrow’s solar generation. Furthermore, in markets where settlements occur at the nodal level, price predictions must be granular with respect to specific transmission nodes.
But how are these forecasts actually generated? In Stem’s case, Athena uses machine learning to pick the type of model that best fits a given circumstance. Athena does this by ingesting terabytes of historical data and then evaluates multiple machine learning frameworks or models based on which best predicts the outcomes that actually occurred.
For example, for a specific forecasting task Athena might evaluate “deep neural network,” “gradient boosting,” and “augmented naïve” models. Gradient boosting models and others that can generate a confidence band are sometimes particularly helpful in wholesale price forecasting because they can be tuned to dial up or down the aggressiveness of a market participant’s bidding strategy.
Once Athena selects a model, we know it will be accurate within a certain margin of error, and after we deploy a project, we fine-tune the model based on actual results. We also increase the training features used by the base model – like feeding in outage data and market demand forecasts – to customize it for a project’s specific location or node. And Athena continuously evaluates the model performance and may update model selection based on seasonality and other factors. The longer the Athena software operates the project, the more accurate the forecasts become.
As challenging as wholesale market price forecasts can be, they’re easier than predicting site loads at individual buildings for behind the meter (BTM) projects. Shyam Sivaramakrishnan, our Head of Data Science, is fond of reminding me that by successfully deploying hundreds of BTM projects in different markets over more than a decade, Stem actually solved the hardest problem first. Athena’s sophisticated load forecasting automation and vast machine learning library developed operating BTM projects gave us a robust, adaptable structure for solar and price forecasting serving our FTM projects.
Optimization is often perceived as less complex than forecasting, and to some extent, that’s true. Optimization is a mathematical equation and solving it uses techniques, like linear programming, that have been around for decades, then applying them with all the horsepower of today’s computer processors.
But not all optimizations are created equal. The key is in how you design the equation. Energy optimization software needs to understand opportunity cost – how capturing certain value streams now might affect the ability to capture others later.
A good example of this in wholesale markets is selecting which products to sell into the market at what time. If a 50MW storage asset bids 50MW of frequency regulation, that asset cannot buy or sell energy at the same time (assume for this example a market where frequency regulation is one product, so “reg up” and “reg down” are bid together). Furthermore, participating in frequency regulation typically results in a lot of battery cycles, creating more wear and tear on the battery.
So providing frequency regulation might be the right decision, but only if the optimization accurately represents all tradeoffs. Getting this right is part art and part science – a mix of technical modeling and quantitative skills and deep market knowledge.
Athena Bidder’s approach to optimization uses time-series data – for example, how much solar generation will occur in one market interval, and the next, and the next. Or, what the real-time price for energy will be at a given node in different intervals. Or, the value of an incentive payment for solar energy delivered in late afternoon and evening intervals.
Once all value streams are accurately represented, Athena considers constraints such as grid export/import limits, solar charging, and other factors. The modeling uses forecasted data and projects optimal operating strategies several days into the future. After representing the problem this way, the optimization engine evaluates hundreds of scenarios and tradeoffs – of providing frequency regulation instead of energy, of charging the battery in the middle of the night versus the afternoon, and of prioritizing battery health and minimizing degradation, and then delivers an optimal and feasible operating strategy.
Because we’ve long since automated our data engineering processes that prepare time-series data for these optimizations, we can rapidly provide Athena Bidder with data for different wholesale markets and new market products with market intervals of different durations. And as with the forecasting example, Stem’s wholesale market optimization capabilities are rooted in our extensive experience operating systems with over 20 million hours of cumulative runtime. Customer loads and dynamic energy pricing were among our first time-series data.
The main thing to understand about scheduling is that it’s a very different solution for energy storage and hybrid solar plus storage assets than it is for traditional assets like natural gas plants or even renewable assets like solar and wind. This is related to forecasting and optimization, but I’m categorizing it under “scheduling” because it’s fundamentally about an asset’s readiness and availability to be bid into a market.
Scheduling storage relies on understanding the physical nature of energy storage systems. For one thing, unlike generation resources, storage is both a seller and buyer of power. And when storage is co-located as part of a hybrid plant it may be scheduled separately from the renewable asset or as a single resource. And DC-integration of solar and storage assets offers further opportunities and challenges to scheduling the hybrid asset. The specific physical characteristics of battery storage must be accounted for as well: for example, different battery chemistries may be kept at different states of charge to prolong battery life and maximize asset performance. There are also complex calculations that need to be performed to capture inverter efficiencies when converting between AC and DC power.
All of this is specific to storage, and distinct from scheduling other asset types. Stem has been doing this for so long that it’s baked into the DNA of the company, and it’s what Athena Bidder and the rest of our Athena applications have been specifically built for. Having a deep library of highly specialized, easily scalable software is a key reason we’re able to support our customers and partners in getting to market quickly and seizing new opportunities when they’re available.