Mobile Batteries Cut Outage Impacts by 41% — While Earning Revenue Waiting for Disasters
In a simulated 33-bus distribution network hit by extreme weather, a new scheduling framework for mobile energy storage systems (MESSs) reduced expected energy not served by 41% and cut average restoration time by 3.2 hours compared to static deployment strategies — all while generating revenue during normal operations through energy arbitrage.
This isn’t science fiction. It’s the result of a three-stage optimization framework developed by Ali Abbasi and Kyri Baker at the University of Colorado Boulder, designed to transform how power grids withstand disasters. As hurricanes, wildfires, and ice storms grow more frequent and intense, the resilience of our electricity infrastructure is no longer a technical footnote — it’s a matter of public safety, economic stability, and climate adaptation. The stakes are high: in 2023 alone, U.S. power outages cost an estimated $150 billion, with extreme weather responsible for over 80% of major disruptions.
The innovation lies in treating mobile batteries not as emergency-only assets, but as dynamic, revenue-generating tools that shift roles across time. During sunny afternoons, they buy low and sell high in wholesale markets. When storm warnings flash, they relocate to strategic roadways, poised to sprint toward blackout zones. And once lines go down, they adapt in real time — rerouting around flooded roads or new fault points, all coordinated through a blend of optimization and machine learning.
The implications extend far beyond one simulation. For cities from Miami to Mumbai, this approach offers a blueprint for smarter, faster recovery — not by building more infrastructure, but by using what we already have, more intelligently.
The Science
The researchers modeled a radial 33-bus distribution grid coupled with a transportation network, simulating how MESS units — essentially large batteries on trucks — could be scheduled across three distinct operational phases: normal operation, proactive positioning, and dynamic relocation.
In the first stage, Normal Operation, MESS units are stationed at fixed buses and optimized for economic arbitrage. The objective is simple: minimize generation costs by charging when electricity prices are low and discharging when prices spike. This stage uses a Mixed-Integer Second-Order Cone Program (MISOCP) to model power flow constraints, including voltage limits, line thermal capacities, and battery dynamics like state-of-charge (SoC) and charging efficiency. Crucially, the model ensures each MESS unit is located at exactly one bus and cannot charge and discharge simultaneously.
When a warning signal arrives — say, a hurricane forecast with probabilistic damage scenarios — the system shifts to the second stage: Proactive Positioning. Here, the goal is no longer profit, but readiness. The model evaluates multiple outage scenarios, each with an associated probability, and determines where MESS units should stage on the transportation network to minimize expected time of arrival at critical loads.
This is where the problem gets complex. The optimal location isn’t just the geographic center of likely outage zones — it’s a balance between proximity, road accessibility, and scenario likelihood. To solve this, the authors propose a three-step heuristic: first, identify the best restoration buses for each scenario; second, cluster these buses into groups that minimize internal travel distances; third, search the transportation graph for the node that minimizes expected travel time to each cluster.
Finally, in the Dynamic Relocation stage, the system responds to real-time updates — new line failures, repair progress, or changes in MESS SoC. Instead of re-solving the full optimization from scratch, which would be too slow during a crisis, the authors introduce a graph neural network (GNN) that maps current grid and transportation states to near-optimal dispatch decisions in milliseconds.
The GNN is trained offline on thousands of simulated scenarios, learning to recognize patterns in grid topology, damage likelihood, and traffic conditions. Once deployed, it acts as a fast policy engine, enabling continuous adaptation without the computational lag of traditional solvers.
What They Found
The simulation, based on a standard 33-bus test feeder with 5 MESS units (each 500 kWh, 250 kW), was run over a 24-hour period with realistic load profiles, price signals, and a set of 10 probabilistic outage scenarios reflecting different storm paths and intensities.
During normal operation, the MESS units generated an average of $1,240 in arbitrage revenue per day — enough to offset a significant portion of their operational costs. This economic viability is critical: without it, utilities would struggle to justify investing in mobile storage for rare events.
When the warning phase activated, the proactive positioning algorithm relocated MESS units from their arbitrage locations to staging points on the transportation network. On average, these new positions reduced the expected time to reach critical loads by 68% compared to staying in place.
Once outages occurred, the dynamic relocation phase kicked in. The GNN-based controller made dispatch decisions every 15 minutes, responding to simulated updates like secondary line failures or delayed repairs. Compared to a static post-event plan, this adaptive approach reduced total energy not served (ENS) by 41% and accelerated full system restoration by 3.2 hours on average.
Reduction in Expected Energy Not Served
Expected energy not served (MWh) averaged across 10 outage scenarios
| Label | Value |
|---|---|
| Baseline (Static) | 12.4 |
| Three-Stage Framework | 7.3 |
The model also revealed a key insight: the most critical factor in resilience isn’t battery size or speed — it’s timing. Units that arrived within the first 2 hours after an outage restored 3.7 times more critical load than those arriving after 4 hours. This underscores the value of the proactive stage: getting batteries into position before the storm hits, not after.
Another finding was the importance of clustering. By grouping likely target buses into spatial clusters, the algorithm ensured that even if one MESS unit was delayed, others could cover overlapping zones. This redundancy improved system-wide reliability, especially in scenarios with multiple simultaneous faults.
Economic Value of Mobile Storage
Daily value of mobile storage under the three-stage framework
| Label | Value |
|---|---|
| Arbitrage Revenue | 1,240 |
| Resilience Value | 3,860 |
The GNN’s performance was particularly striking. While a full MISOCP re-solve took an average of 8.7 minutes per decision cycle, the trained network delivered comparable solutions in just 120 milliseconds — fast enough to support real-time control during rapidly evolving outages.
Why This Changes Things
Today, most grid resilience strategies are either passive or reactive. Passive approaches — like hardening poles or adding stationary storage — are expensive and inflexible. Reactive ones — like sending diesel generators after a storm — are slow and often too little, too late.
This three-stage framework offers a third way: anticipatory resilience. It treats extreme events not as unpredictable shocks, but as probabilistic threats that can be prepared for in advance. And it does so without sacrificing economic efficiency.
Consider Puerto Rico after Hurricane Maria. Over 3 million people lost power, some for months. Relief efforts were hampered by damaged roads and fuel shortages. Had mobile storage units been pre-positioned based on storm forecasts — staged not at substations, but at accessible road junctions — they could have reached isolated communities faster, especially where fixed infrastructure was destroyed.
Or consider California’s wildfire season, where utilities preemptively de-energize lines to prevent fires. These public safety power shutoffs (PSPS) affect millions annually. A system like this could deploy MESS units to high-risk zones days in advance, ensuring critical facilities — hospitals, shelters, water pumps — remain powered even during intentional outages.
The transportation-power nexus is key. Most previous models assumed MESS units move in straight lines or ignored traffic entirely. This framework explicitly couples the two networks, using real shortest-path distances on roads, which can change if bridges flood or highways close. That realism makes the solutions actually deployable.
And the use of GNNs marks a shift in how we think about grid control. Traditionally, optimization and machine learning have been seen as competing paradigms: one exact but slow, the other fast but approximate. Here, they’re combined — the heavy lifting done offline, the real-time decisions made instantly. It’s a template that could apply to other domains: flood response, wildfire evacuation, or pandemic supply chains.
What’s Next
The model has limitations. It assumes accurate probabilistic forecasts — a big if in many regions. It also simplifies traffic dynamics and doesn’t model competition for road access among multiple response vehicles. And while the GNN is fast, its decisions are only as good as the training data; rare or unprecedented events might fall outside its learned patterns.
Future work could integrate real-time weather updates, repair crew locations, or even drone-based damage assessment. The framework could also be extended to include other mobile resources — portable solar arrays, mobile microturbines, or even electric school buses repurposed as storage during emergencies.
Perhaps most importantly, this approach challenges the siloed thinking that separates “normal” operations from “emergency” response. In a world of climate volatility, that boundary is blurring. Grids must be designed not for average conditions, but for adaptability — able to shift roles, reconfigure on the fly, and serve multiple purposes across time.
The 41% reduction in unserved energy isn’t just a number. It’s refrigerators staying cold. It’s dialysis machines running. It’s children doing homework under electric light. Resilience isn’t about preventing all outages — that’s impossible. It’s about ensuring that when the lights go out, they come back on faster, smarter, and more equitably.