Local AI Agents in Power Electronics Simulation: Deployment, Measurements and Costs
Organizations: ELITEGrid Research Laboratory, Department of Electrical and Computer Engineering, University of Alberta, Edmonton, AB, Canada
Abstract
Agentic artificial intelligence (AI) is emerging in power electronics, where it drives the simulation loop that engineers used to close largely by hand. In this loop, a converter design, or a study of the network it connects to, is built, run, measured and revised until it satisfies its specification. Such an agent can be served from a cloud API or from an inference device on the premises. For organizations whose designs, customer data or export-controlled material cannot leave the site, and whose simulators are licensed there, local deployment is the natural choice. In such a deployment, the agent becomes an electrical load on the organization's own equipment, i.e., every run consumes power and energy on a local device, generates tokens and takes time. However, existing work focuses mainly on what the agent can achieve with cloud-hosted models. What a local deployment costs in power, energy, tokens and time on real simulation tasks, and what its choices trade against one another, has received little quantitative investigation under comparable conditions. To address this gap, a systematic evaluation framework is presented in this paper. It applies locally served models to agentic power electronics tasks, from the converter level to the system level, and measures their power, energy, token and time costs. Based on these measurements, the capability boundary of a local agent on such tasks and the cost per solved task are characterized. The results are expected to provide practical insight for the use of agentic AI in power electronics and the power systems it serves.
Figures & tables
| Category of work | Simulator in loop | Open weights | Energy measured | Local vs. cloud |
|---|---|---|---|---|
| Agents for converter design [ 3 , 4 , 1 , 5 , 7 , 6 ] | yes | no | no | no |
| Agents for other engineering tools and power system solvers [ 8 , 9 , 10 , 11 , 12 , 13 , 14 , 15 ] | yes | some | no | capability only |
| Energy of agents on generic benchmarks [ 16 , 2 , 17 , 18 , 19 ] | no | yes | yes | by estimate [ 20 , 21 ] |
| Capability indices, open-weight vs. closed [ 22 , 23 , 24 ] | no | yes | no | implied |
| This paper | yes | yes | yes, per run | yes, same loop |
| Term | Meaning here |
|---|---|
| Executor | The agent software that runs the loop: assembles the context, calls the model, runs the tools and enforces the token budget. |
| Context | Everything sent to the model on one request: the task text, the tool descriptions and the full history of the run. |
| Prefill, decode | The two phases of a request on the inference device: reading the context (short, high power), then generating output one token at a time (long, near-constant power). |
| Output tokens | What the model generates, reasoning and visible message together; the token budget counts these. |
| Reasoning tokens | Output the model produces for itself before its visible message; decoded, and metered, like any other output. |
| Test condition | One task or task family, one model and one configuration, with its repeated runs. |
| Device class (example) | Memory (GB) | (GB/s) | Rated power (W) |
|---|---|---|---|
| Desktop CPU, Ryzen 9000 [ 42 ] | 128 a | 90 | 65–170 |
| Unified, Ryzen AI Max+ 395 [ 43 ] | 128 | 256 | 45–120 |
| Unified, GB10 [ 44 ] | 128 | 273 | 140 |
| Server CPU, EPYC 9005 [ 45 ] | 6,000 | 614 | 450–500 |
| Unified, Apple M5 Ultra [ 46 ] | 512 | 1,200 | 480 |
| Graphics card, RTX 5070 [ 47 ] | 12 | 672 | 250 |
| Family | Deliverable | Verified on |
|---|---|---|
| Defect repair | Repaired model and a diagnosis | Diagnosis; behaviour; minimality of the change |
| Controller tuning | Settings, or a statement that the specification cannot be met | Behaviour; interface; footprint; stated limit |
| Construction | A simulating model | Behaviour; interface; structure |
| Development stage | Script or attribution, with simulation | Agreement with a reference analysis |
| Extension | Extended model | Behaviour; interface; footprint |
| Implication | Measurement | Sec. | |
|---|---|---|---|
| 1 | Decode power is a plateau that the task does not alter | 33–38 W across families, budgets and outcomes | V-B |
| 2 | Prefill is higher in power and short | About 1.4 pu; 8.8% of busy time | V-B , V-C |
| 3 | Nominal input tokens overstate the computation | 6.0% of nominal tokens computed | V-D |
| 4 | Reasoning is decoded and metered as output | About three quarters of local output; no lower power with reasoning disabled | V-B , V-D |
| 5 | Energy follows duration, then output tokens | Median error 2.4% from duration alone, 38 held-out runs | V-C |
| 6 | The model does not see its token consumption | With a notice, cap stops fell from 10 of 24 to 2 of 16 | V-D |
| Converter and data-centre sets, valid evaluated runs | 2,041 |
|---|---|
| converter set | 2,017 |
| data-centre set, cloud only | 24 |
| passes | 617 |
| two earlier cloud models, pooled in the cloud column | 183 |
| with an energy record (70,392 kJ, 19.6 kWh) | 882 |
| no energy record: Flash-Next tuning | 789 |
| Role | Instance |
|---|---|
| Executor | Codex command-line agent [ 52 ] , pinned per deployment: 0.153.3 local, 0.159.3 cloud |
| Live-session coupling | MATLAB/Simulink through the vendor’s agent toolkit: thirteen MCP tools (five read a model, one edits it, three run code, four check), one MATLAB session per agent, 600 s per tool call |
| Service coupling | PSCAD/EMTDC service on a Windows host with one licence, run one request at a time; 10 to 170 s per simulation |
| Inference node | NVIDIA DGX Spark: GB10 single-package device, 140 W thermal design power, 128 GB unified memory shared by weights and context, 273 GB/s; vLLM [ 53 ] ; two nodes with tensor parallelism over a direct 200 Gb/s link for the larger model |
| Main local model | Flash-Next [ 54 ] , mixture of experts in NVFP4: 176B total, 6B active per token, about 126 GB of weights; a second NVFP4 conversion of 135 GB for the power system set |
| Other local models | DeepSeek V4 Flash [ 55 ] , 284B (13B active), on two nodes; Qwen3.8-27B [ 56 ] and Muse Glimmer 30B [ 57 ] (dense); Nemotron 3.5 Lightning 30B [ 58 ] ; gpt-oss-120b [ 59 ] |
| Base-model scale | Blocks | Runs | Pass | Median kJ/run | kJ per solved task |
|---|---|---|---|---|---|
| Bridges | 10 2 | 98 | 81 | 30.2 | 46.0 |
| PV boost, relay | 563/753 | 48 | 33 | 95.5 | 139.0 |
| EV grid-tied node | 3,025 | 47 | 27 | 105.8 | 191.9 |
| Model | Runs | p10 (W) | Median (W) | p90 (W) | |
| Flash-Next | 384 | 28.6 | 33.7 | 37.5 | |
| gpt-oss-120b | 90 | 32.8 | 39.3 | 43.0 |
| G1 construction | S12 fault location | |||
| Local | Cloud | Local | Cloud | |
| Time | ||||
| Duration (min) | 80 | 3 | 177 | 11 |
| Model time, transcript (min) | 77 | 2 | 152 | 3 |
| Tool time (min) | 2.4 | 0.5 | 25.3 | 7.5 |
| Busy / idle, power trace (min) | 78 / 2 | – | 151 / 26 | – |
| Cloud reference | Local, attempt 1 | Local, attempt 2 | |||||||||||||
| ID | min | sims | tok | ctx | min | sims | tok | ctx | min | sims | tok | ctx | |||
| S9 | P | 11 | 13 | 6 | 33 | P | 94 | 52 | 70 | 127 | P | 89 | 57 | 73 | 124 |
| S12 | P | 11 | 3 | 7 | 30 | P | 94 | 15 | 83 | 131 | P | 177 | 24 | 142 | 204 |
| S8 | P | 21 | 18 | 8 | 33 | P | 105 | 56 | 59 | 110 | P | 83 | 34 | 63 | 120 |
| S4 | P | 46 | 80 | 17 | 63 | P | 135 | 137 | 85 | 126 | P | 160 | 216 | 92 | 152 |
| S11 | P | 11 | 19 | 14 | 42 | P | 151 | 35 | 124 | 204 | P | 155 | 54 | 133 | 198 |
| Setting | Effect on the fixed workload |
|---|---|
| GPU clock locked at half its default | Task energy %; wall-clock time essentially unchanged |
| Speculative decoding | Energy per output token % |
| Four concurrent agents | Energy per output token %: power W, throughput |
| Model kept resident | Break-even against a cold start at about 18 min of idle |
Appendix figures & tables5 assets
Supplementary material from the paper’s appendix.
Appendix
| ID | Layer | Reference model | Deliverable |
|---|---|---|---|
| S1 | System | IEEE 39-bus system | Critical clearing times at several buses |
| S2 | System | Single machine, infinite bus, with PSS | PSS parameters for damping and gain margin over several operating points |
| S3 | System | 132/33 kV OLTC with a feeder | Tap-changer control settings for voltage regulation |
| S4 | EMT | 400 kV line, statistical switching | Pre-insertion resistor from a catalogue |
| S5 | EMT | 230 kV capacitor banks | Inrush reactor from a catalogue |
| S6 | EMT | 230/13.2 kV transformer on an open line | Fault-record diagnosis and a ferroresonance damping resistor |