Embedded AI - Power Consumption Prediction

Outline

1. Overview
2. Features
3. Project Structure
4. Installation & Setup
5. Model Architecture
6. Hardware Requirements
7. Reference Video
8. Data Preprocessing & Model Training
9. Results & Performance
10. Future Improvements
11. License

---

Overview

Embedded AI is a project focused on implementing a machine learning model for power consumption prediction on an STM32F407G Discovery board. The workflow includes training a neural network using TensorFlow, converting the model to TensorFlow Lite (TFLite), and deploying it onto the microcontroller. The embedded system then predicts future power consumption based on historical data.

---

Features

• Trains a neural network model using TensorFlow
• Converts the model to TensorFlow Lite for deployment
• Implements AI inference on an STM32F407G Discovery Board
• Uses UART communication for debugging and data visualization
• Predicts power consumption trends for the upcoming month
• Optimized for TinyML and Edge AI processing

---

Project Structure

/project-root
│── /notebooks               # Jupyter Notebooks for data analysis & model training
│── /src                     # Source code for embedded implementation
│── /models                  # Trained models and TFLite conversions
│── /docs                    # Documentation and reports
│── /images                  # README images
│── README.md                # Project Documentation
│── requirements.txt         # Python dependencies
│── LICENSE                  # License information

---

Installation & Setup

1. Train the AI Model

• Ensure Python 3.8+ and TensorFlow 2.x are installed.
• Clone the repository and navigate to the project directory:

  git clone https://github.com/yourusername/EmbeddedAI.git
  cd EmbeddedAI

• Install the required dependencies:

  pip install -r requirements.txt

• Run the Jupyter Notebook:

  jupyter notebook

• Train the model and save the final .tflite file.

2. Convert Model to TFLite

• In the provided notebook, use:

  converter = tf.lite.TFLiteConverter.from_keras_model(model)
  tflite_model = converter.convert()

• Save the converted model for deployment.

3. Flash Model to STM32

• Install STM32CubeMX and X-CUBE-AI.
• Use STM32CubeIDE to compile and flash the firmware.
• Load the model onto the microcontroller and validate using UART communication.

---

Model Architecture

The neural network architecture used for power consumption prediction is based on a 1D CNN + Dense Layers model. The structure is optimized for edge deployment and efficiency in embedded environments.

---

Hardware Requirements

• STM32F407G Discovery Board

• USB-to-TTL Module (for UART debugging)

• Power Consumption Sensor (optional for real-time data)
• Computer with STM32CubeIDE installed

---

Reference Video

For a detailed step-by-step implementation, watch the following video:

---

Data Preprocessing & Model Training

The dataset consists of historical power consumption values.

• Feature Engineering:
  • Removed unnecessary columns
  • Extracted time-based features (day, month, year)
  • Normalized data between [0,1]
• Model Training:
  • Used a 1D CNN + Dense Layers architecture
  • Optimized using Adam optimizer
  • Evaluated using RMSE, MAE, and MAPE

Training Graphs:

---

Results & Performance

• Prediction RMSE: 0.000667201
• Memory Usage: 795.91 KiB
• Power Consumption: 14.68 KiB

---

Future Improvements

• Improve prediction accuracy by tuning the CNN architecture
• Integrate real-time power consumption monitoring
• Optimize memory usage using quantization techniques
• Expand dataset for more accurate trend analysis

---

License

This project is licensed under the MIT License - see the LICENSE file for details.

---