Charging Station Edge Power Dispatch with AIoT Gateway

Charging Station Edge Power Dispatch: Local Closed-Loop Control with an AIoT Edge Gateway

The Problem: 120 Guns, One Transformer

A city fast-charging station has 120 DC fast-charging guns, each up to 180 kW, on a 2000 kVA transformer. Full simultaneous output would demand 21,600 kW—more than 10× transformer capacity. Power must be limited.

Cloud dispatch has three hard problems:

• Latency: 200–500 ms round trip, sometimes over 2 seconds; charging modules need response within 100 ms.
• Network dependence: When offline, the cloud cannot issue commands—risking transformer trips or full shutdown.
• Bandwidth cost: 120 guns reporting power, voltage, current, SOC, and temperature every second is expensive to upload.

Power dispatch must therefore close locally: collection, calculation, and command issuance inside the station, at millisecond response.

Hardware: BL370 Interface Allocation

We use the BL370 edge gateway: Rockchip RK3562J, quad-core Cortex-A53 + single-core M0, 1.8/2.0 GHz, 1 TOPS NPU.

Configuration: 1× X board + 2× Y board slots

• X20: 4× RS485/RS232
• Y63: 4× RS485
• Total: up to 8× RS485

The gateway does not poll all 120 guns directly. It connects to station meters and charging pile group controllers, which manage individual piles.

Allocation:

• RS485-1: Main incoming meter
• RS485-2/3/4: Zone meters A/B/C
• RS485-5/6: Charging pile group controllers
• RS485-7: Energy storage PCS (if PV-storage-charging)
• RS485-8: Spare or environmental monitoring

If no group controller exists, the gateway can poll pile metering modules, but 120-gun polling is too slow for real-time dispatch. A hierarchical architecture is preferred: gateway manages zones; group controllers manage individual piles.

Software: Local Closed-Loop Logic

The gateway runs a configurable collect → calculate → issue loop, typically every 100–500 ms.

Collect: Modbus RTU master polls main and zone meters for total active power, phase currents, voltage, and power factor. Eight channels take about 400–800 ms. Modbus TCP can be used in parallel if supported.

Calculate: A constrained power allocation algorithm uses transformer capacity, current total power, zone power, pile count and demand, and storage SOC/capability. Output: power upper limit per zone. The 1 TOPS NPU accelerates optimization solving from tens of milliseconds to a few milliseconds; simple proportional allocation can run on CPU.

Issue: The gateway writes zone power limits to group controller registers via RS485, or directly to pile power-setting registers. Group controllers then allocate downstream to each gun.

The whole loop is local—no cloud dependency.

Offline Autonomy

When the network fails, cloud dispatch is useless, but local closed-loop control continues:

• Data collection: meters and controllers directly connected via RS485
• Calculation: local algorithm with NPU acceleration
• Command issuance: RS485 writes to controllers or piles
• Data caching: local eMMC stores history and retransmits after recovery

Even fully offline, the transformer does not overload, piles do not all shut down, and charging service continues. Cloud-only solutions cannot do this.

What the 1 TOPS NPU Is For

Basic power allocation can run on CPU. The NPU adds predictive dispatch:

• Load forecasting: Lightweight LSTM/GRU models predict 15–30 minute charging demand using history, time, weather, and traffic—enabling proactive adjustment.
• Fault prediction: Analyzes voltage, current, and temperature curves to detect abnormal patterns and warn of power module aging.
• Storage optimization: Optimizes charge/discharge based on price curves and demand forecasts, reducing demand charges.

For multi-channel video or heavier AI, choose the BL450 with a 6 TOPS NPU.

Cloud-Edge Collaboration

Local closed-loop does not mean no cloud:

• Edge: real-time power allocation, local control, offline autonomy, data caching
• Cloud: cross-site optimization, long-term analytics, OTA upgrades, billing

They communicate via MQTT or HTTP. The cloud issues policy parameters (safety factor, time-of-use pricing), not real-time commands. If the cloud is unavailable, the edge keeps running on the last policy.

Remote O&M

With BLRAT, engineers establish encrypted tunnels from the office to check status, debug parameters, and upgrade firmware—no public ports needed. EdgeCoder allows browser-based changes to allocation parameters, polling cycles, and meter address mapping.

Selection Logic

Core needs: multiple RS485, local closed-loop, offline autonomy, industrial reliability.

BL370 + X20 + Y63 — 8× RS485, quad-core A53, 1 TOPS NPU, -40°C to 85°C, EMC Level III — is sufficient for most charging stations.

For more zone meters or PV-storage-charging systems requiring complex optimization, choose BL440/BL450: 12× RS485 + 6 TOPS NPU.

Summary

Charging station power dispatch requires local closed-loop control: collection, calculation, and issuance inside the station, millisecond response, and uninterrupted operation during network outages. The 1 TOPS NPU adds load forecasting and fault warning. Eight RS485 ports manage 120 guns through a hierarchical architecture—main and zone meters provide real-time power data; group controllers execute allocation. BL370 delivers the right mix of interfaces, compute, and industrial reliability.