ESP32 TinyML Power Virtual Sensor (121 µs)

Overview & Motivation

In heavy industrial facilities such as steel mills, installing revenue-grade power meters across every secondary branch or remote feeder is cost-prohibitive. However, auxiliary parameters—like reactive power (inductive and capacitive) and power factors—are already monitored by local power-factor correction units and sub-panels.

This project implements an edge neural soft sensor running on a bare-metal ESP32. By feeding 6 auxiliary telemetry parameters into an INT8-quantized Multilayer Perceptron (MLP), the microcontroller predicts real-time active power consumption (Usage_kWh) with laboratory-grade accuracy, zero cloud dependencies, and minimal silicon footprint.

Hardware & Key Performance Metrics

• Runtime Framework: TensorFlow Lite for Microcontrollers (TFLM)
• Model Size: 4.23 KB Flash (compared to 34.1 KB Float32 baseline)
• SRAM Tensor Arena: 772 Bytes
• On-Device Inference Latency: 121 microseconds
• Model Accuracy (Hold-out test set, N=7, 008): R^2 Score: 0.9991 | Mean Absolute Error (MAE): 0.608 kWh

System Architecture

The pipeline takes 6 industrial telemetry inputs:

1) Lagging_Reactive_Power_kVarh

2) Leading_Reactive_Power_kVarh

3) CO2(tCO2)

4) Lagging_Power_Factor

5) Leading_Power_Factor

6) NSM (Number of Seconds from Midnight)

Edge Pipeline Workflow:

• MinMax Normalization: Input features are normalized using precomputed training ranges directly on the MCU.
• INT8 Quantization: Features are scaled to integer representation (q = x / scale + zero_point).
• TFLM Inference: The INT8 neural network executes inside a dedicated 772-byte memory arena.
• Dequantization: Model output is translated back into engineering units (Usage_kWh).

Quantization & Memory Optimization

Running floating-point neural networks on microcontrollers without hardware FPUs introduces significant latency. By applying Post-Training Quantization (PTQ) with a representative telemetry dataset:

• Size Reduction: ~87.6% drop in binary weight footprint.
• Speedup: Inference drops from ~650 µs (emulated float) down to121 µs per sample.
• Accuracy Trade-off: Negligible performance degradation (Float32 MAE: 0.463 kWh vs INT8 MAE: 0.608 kWh).

Source Code & Pre-compiled Binaries

The full source code, Python training pipelines, test suite, and pre-compiled ESP32 binaries (firmware.bin and .tflite model) are available under the MIT license on GitHub:

👉 steel-power-TinyML on GitHub