Mar 12, 2026
AI for grid-aware thermal storage
A simulation note on coordinating thermal storage dispatch with local grid conditions, forecast uncertainty, and building-level comfort constraints.
Beyond Price Arbitrage
Thermal storage is often modeled as a simple buffer: charge when energy is cheap, discharge when it is expensive. That misses the operational detail that makes the problem useful. Local congestion, demand ramps, weather-driven load, and equipment recovery all change the value of a storage decision minute by minute.
Our current simulation stack treats each building as a local agent connected to a feeder-level model. The model is intentionally lightweight, but it preserves enough topology to test whether a control policy moves stress around the network or actually reduces it.
The Control Surface
The useful policy is not the one that optimizes a single price signal. It is the one that can absorb noisy forecasts and still produce stable behavior under changing demand. We use surrogate models for comfort response, equipment lag, and temperature recovery so the control surface remains fast enough for repeated search.
The control agent receives weather forecasts, building state, feeder constraints, and a limited telemetry window. It must decide when to pre-cool, when to coast, and when to preserve headroom for a likely peak. We score it on cost, comfort, grid stress, and sensitivity to forecast error.
Why Simulation Comes First
Field deployment is expensive because a bad control policy can create comfort complaints or new peak loads. Simulation lets us test policies against rare conditions before they touch a building. Heat waves, telemetry gaps, compressor lockouts, and coincident feeder peaks can all be generated as stress cases.
The goal is not to find a perfect controller in simulation. The goal is to reject brittle controllers before they reach a pilot, and to identify which sensors would most improve the policy. Sometimes the best simulation result is a measurement requirement.
Next Benchmark
The next benchmark will compare direct optimization, learned dispatch policies, and hybrid controllers against the same feeder traces. Each controller will receive the same imperfect forecast data and the same equipment constraints.
We will publish performance curves rather than a single score. A controller that performs slightly worse in the median case but degrades gracefully under bad forecasts may be more valuable than one that wins the clean test and fails under realistic noise.