GridOps Strategy Selector v7
India does not only need more solar panels. It needs intelligence between the panel, the battery, the grid, the diesel backup, and the people depending on power.
GridOps is our attempt to build that intelligence layer for community microgrids: a small learned strategy selector, a deterministic optimizer, and a simulated OpenEnv world where every decision is scored by cost, reliability, and diesel use.
GridOps is also a case study in the Capabl Machines thesis: climate-heavy problems need more than a checkpoint. They need a harness where models, optimizers, simulators, validators, rewards, and evaluation loops work together.
GridOps Strategy Selector v7 is the learned part of that system. It does not pretend to be the whole grid engineer. It reads a GridOps/OpenEnv observation and emits strict strategy JSON. A causal optimizer then converts that strategy into the final bounded dispatch action.
This is the central release lesson:
microgrid state
-> small AI chooses operating strategy
-> optimizer converts strategy into safe dispatch
-> OpenEnv scores cost, reliability, diesel, blackout
That split keeps the language model focused on contextual judgment while leaving constrained numerical dispatch to an optimizer.
What Capabl Machines Provides
This release should not be read as "a LoRA beat every baseline." It should be read as evidence for a stronger company pattern:
Capabl Machines builds climate AI operating harnesses.
The harness defines:
- the environment where decisions are tested;
- the schema for valid actions;
- the tools that handle physics and constraints;
- the critics and rewards that judge outcomes;
- the datasets that teach useful behavior;
- the model layer that selects intent or strategy;
- the evals that decide whether the system is actually better.
For some domains, the trained model will be the main breakthrough. For others, as GridOps shows, the harness and model interface may be the breakthrough. The customer still gets the thing that matters: an accountable AI operating system for a real climate workflow.
Why This Matters
Distributed solar is scaling quickly across Indian apartments, societies, campuses, factories, and local energy systems. That creates a second-order problem: who operates the system once it exists?
A local operator now has to answer questions like:
- Should the battery charge now because grid power is cheap?
- Should it preserve charge for the evening peak?
- Should diesel be allowed during a crisis, or conserved for an outage?
- Should demand response be used, knowing it rebounds later?
- How do we keep the lights on while still reducing cost and diesel?
Bad control turns clean infrastructure into higher bills, battery misuse, blackouts, or unnecessary diesel. Good control makes the same infrastructure more useful.
GridOps explores a practical pattern for sustainable infrastructure AI:
model for judgment
tools for physics
environment for truth
metrics for accountability
That is the product direction: not model magic, but tested climate-AI systems that can be adapted to energy, water, agriculture, logistics, robotics, and resilient infrastructure.
What We Built
The environment is a 72-hour community microgrid simulation with three regimes:
| Task | Situation | What it tests |
|---|---|---|
| Task 1 normal | Normal summer demand and solar | basic battery arbitrage |
| Task 2 heatwave | demand spike plus price stress | forecast-aware peak planning |
| Task 3 crisis | haze, heatwave, limited diesel, grid outage | islanding and reliability |
The action space remains the real OpenEnv dispatch contract:
{"battery_dispatch":0.0,"diesel_dispatch":0.0,"demand_shedding":0.0}
But the model is not asked to emit those floats directly. That was the wrong abstraction.
Directly asking a small model to output exact battery/diesel/shedding floats proved brittle. The model could learn JSON, but the optimization burden was too large for a small SFT policy. GridOps v7 turns the model into a strategy selector and lets a causal LP/MPC controller execute the details.
This is the mature pattern we arrived at:
LLM: choose operating intent
Optimizer: satisfy constraints and choose dispatch
OpenEnv: score the result
Key Results
100% valid strategy outputs from the learned selector
96.04% LP ceiling capture from the strategy-controller system
96.09% LP ceiling capture from an untuned 1.5B model using the same harness
3 tested operating regimes: normal, heatwave, crisis/outage
The strongest deployable system today is the deterministic v7 strategy-controller. The learned v7.3 selector nearly matches it while producing perfectly valid strategy JSON on the holdout set.
We also ran the most important sanity check: the untouched base model with the
same v7 strategy prompt and controller. It scored 0.7911, slightly above the
released adapter's 0.7888. That is not a failure of fine-tuning; it is the
central engineering result. The strategy abstraction and optimizer harness do
most of the heavy lifting. The adapter is the packaged, reproducible, audited
selector from the training pipeline, but the architecture is the real unlock.
This is exactly why Capabl Machines focuses on the harness and model together. If a base model is already strong once the interface is correct, we should use that. If a domain needs post-training, we should fine-tune. The job is not to force model training into every problem; the job is to deliver the most reliable AI operating loop for the climate system in front of us.
| Release highlight | Value |
|---|---|
| v7 deterministic controller average score | 0.7907 |
| untuned Qwen 2.5 1.5B + v7 harness average score | 0.7911 |
| v7.3 learned selector average score | 0.7888 |
| v7.3 valid strategy rate | 100.00% |
| v7.3 LP ceiling capture | 95.81% |
| v5.1 direct-action baseline average score | 0.7354 |
Output Schema
The model emits only strict JSON:
{
"mode": "cost_saving",
"risk_level": "low",
"battery_bias": "charge",
"diesel_policy": "avoid",
"shedding_policy": "never"
}
Allowed values:
mode: cost_saving | peak_shaving | outage_prepare | reliability | recovery | fuel_conservation
risk_level: low | medium | high | critical
battery_bias: charge | preserve | discharge | neutral
diesel_policy: avoid | allow_if_blackout | prewarm | conserve
shedding_policy: never | last_resort
Model Lineage
Base: Qwen/Qwen2.5-1.5B-Instruct
SFT: 77ethers/gridops-models/sft_qwen25_15b_gridops_strategy_v7
DPO v7.2: 77ethers/gridops-models/dpo_qwen25_15b_gridops_strategy_v72
DPO v7.3: 77ethers/gridops-models/dpo_qwen25_15b_gridops_strategy_v73_crisis
Release: capabl-machines/gridops-strategy-selector-v7
The released adapter is the v7.3 crisis-weighted DPO checkpoint. v7.3 remained stable and matched v7.2, but did not beat the deterministic controller. The recommended production policy is therefore the strategy-controller harness, with this model as the learned strategy selector.
Engineering Journey
The most important result was not one checkpoint. It was the discovery of the right interface.
v4: direct action SFT from reasoning traces
v5: causal LP teacher imitation
v5.1: crisis repair continuation
v6: tool-corrected action SFT, not promoted
v6.1: clean LP-critic action SFT, not promoted
v7: strategy-first harness
v7.1: SFT strategy selector
v7.2: DPO preference tuning
v7.3: crisis-weighted DPO release checkpoint
The lesson:
The model does not need to become the entire operator. It needs to learn the operating language that lets deterministic tools act safely.
Evaluation
Holdout seeds: 7001,7002,7003.
| System | Avg score | Valid strategy/action | Task 1 normal | Task 2 heatwave | Task 3 crisis | LP capture |
|---|---|---|---|---|---|---|
| v5.1 direct action model | 0.7354 | 0.9969 action | 0.7896 | 0.7681 | 0.6484 | - |
| v7 deterministic strategy-controller | 0.7907 | 1.0000 action | 0.7995 | 0.8224 | 0.7503 | 96.04% |
| untuned Qwen 2.5 1.5B + v7 harness | 0.7911 | 1.0000 strategy | 0.7993 | 0.8223 | 0.7517 | 96.09% |
| v7.1 SFT strategy selector | 0.7880 | 1.0000 strategy | 0.7994 | 0.8224 | 0.7421 | 95.71% |
| v7.2 DPO strategy selector | 0.7888 | 1.0000 strategy | 0.7993 | 0.8223 | 0.7449 | 95.81% |
| v7.3 DPO strategy selector | 0.7888 | 1.0000 strategy | 0.7993 | 0.8223 | 0.7449 | 95.81% |
| Full-episode LP ceiling | 0.8233 | - | 0.8372 | 0.8416 | 0.7912 | 100.00% |
Operational Footprint
The crisis task is the real stress test: haze reduces solar, demand rises, diesel is limited, and the grid outage forces islanded operation. The learned selector and the untuned base model both stay close to the deterministic controller, but the remaining gap to LP is mostly crisis blackout and cost.
Why This Is Useful
The learned model is small, stable, and schema-reliable. The controller is the stronger deployable policy. The base-model comparison makes the lesson sharper: we did not merely train a checkpoint, we found the right interface between an LLM and an optimization system. Together they show a practical pattern for domain AI systems:
Do not force the model to be the whole controller.
Teach it the decision language.
Use tools for physics, constraints, validation, and scoring.
That pattern is bigger than microgrids. The same structure can apply to:
- apartment and society energy systems;
- water pump scheduling;
- cold chains;
- EV charging depots;
- factory energy optimization;
- farm irrigation and storage;
- disaster-resilient local infrastructure.
GridOps is one case study in a broader Capabl Machines thesis: useful AI for physical systems should be trained and evaluated inside the world it claims to operate.
Usage
from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer
base_model = "Qwen/Qwen2.5-1.5B-Instruct"
adapter = "capabl-machines/gridops-strategy-selector-v7"
tokenizer = AutoTokenizer.from_pretrained(adapter)
model = AutoModelForCausalLM.from_pretrained(base_model, device_map="auto")
model = PeftModel.from_pretrained(model, adapter)
The model output should be parsed as GridOpsStrategy, then passed to the
GridOps controller. The final OpenEnv action remains:
{"battery_dispatch":0.0,"diesel_dispatch":0.0,"demand_shedding":0.0}
Intended Use
- Research and demos for strategy-conditioned microgrid operation.
- OpenEnv-style environment evaluation.
- Tool-assisted energy dispatch workflows where a validator/controller handles the final physical action.
Limitations
- This adapter is not a standalone power-system controller.
- It should not be used for real grid operation without hardware validation, safety review, and local regulatory checks.
- It was evaluated in the GridOps simulated 72-hour environment, not on live metered deployments.
- The deterministic strategy-controller remains the recommended runtime baseline until a learned selector beats it.
Links
- Demo Space: capabl-machines/gridops-demo
- Source repo: capabl-machines/gridops
- Earlier model archive: 77ethers/gridops-models
- Downloads last month
- 128