AI-Controlled Hybrid Energy Storage System for Regenerative Braking

ECS Mini Project | VIT Chennai | April 2026

---

Overview

This project simulates an AI-controlled Hybrid Energy Storage System (HESS) for regenerative braking in electric vehicles. A Random Forest ML classifier decides in real time how to split recovered braking power between a lithium-ion battery and a supercapacitor, optimising energy recovery while protecting battery health.

In conventional vehicles, all kinetic energy is wasted as heat during braking. EVs can recover this energy — but smart storage management is critical. This project solves that using machine learning.

---

Results

Metric	Value
Total energy recovered	1,304 Wh
Braking events handled	4,363
ML decision accuracy	~95% (5-fold CV)
Battery cycles accumulated	0.175
Simulation duration	20 minutes
Training samples	27,615

---

System Architecture

%%{init: {
  "theme": "base",
  "themeVariables": {
    "fontSize": "20px"
  },
  "flowchart": {
    "nodeSpacing": 120,
    "rankSpacing": 100,
    "curve": "basis"
  }
}}%%

flowchart LR

A["Drive Cycle"]
--> B["Motor Gen"]

B --> C["DC-DC Conv"]

C --> D{"Power Split"}

D --> E["Battery"]

D --> F["Supercap"]

E --> G["ML Agent"]

F --> G

G --> D

style A fill:#065A82,color:#ffffff,stroke:#ffffff,stroke-width:2px
style B fill:#065A82,color:#ffffff,stroke:#ffffff,stroke-width:2px
style C fill:#065A82,color:#ffffff,stroke:#ffffff,stroke-width:2px

style D fill:#1C7293,color:#ffffff,stroke:#ffffff,stroke-width:2px

style E fill:#0F6E56,color:#ffffff,stroke:#ffffff,stroke-width:2px
style F fill:#534AB7,color:#ffffff,stroke:#ffffff,stroke-width:2px

style G fill:#BA7517,color:#ffffff,stroke:#ffffff,stroke-width:2px

Loading

---

Component Details

Component	Description
Drive Cycle	Generates vehicle driving conditions
Motor Generator	Simulates regenerative braking
DC-DC Converter	Controls energy conversion
Power Split	Distributes recovered energy
Battery Model	Thevenin battery model
Supercap Model	RC supercapacitor model
ML Agent	Random Forest optimization agent

Project Structure

File	Description
storage_model.py	Battery (Thevenin) + Supercapacitor (RC) models with SoC tracking
drive_cycle.py	Vehicle dynamics + synthetic drive cycle generator
train_agent.py	Dataset generation + Random Forest ML training pipeline
simulate.py	Full closed-loop simulation + 6-panel results dashboard
cosim_controller.py	Python to Simulink bridge via MATLAB Engine API
regen_model.slx	MATLAB Simulink power electronics model
requirements.txt	Python dependencies

---

How to Run

1. Clone the repository

git clone https://github.com/tanisha-ahl/regen-braking-hess.git cd regen-braking-hess

2. Install dependencies

pip install -r requirements.txt

3. Run the full simulation

python simulate.py

This will automatically:

• Generate a labelled training dataset from drive cycle simulations
• Train the Random Forest classifier and save it to model/
• Run a 20-minute closed-loop simulation (ML agent vs threshold baseline)
• Print a results summary in the terminal
• Save a 6-panel dashboard to results/simulation_dashboard.png

4. Train the model separately (optional)

python train_agent.py

5. Run co-simulation with Simulink (requires MATLAB R2022a+)

python cosim_controller.py

---

ML Agent — Decision Logic

The Random Forest classifier makes real-time power split decisions at every braking event:

Class	Decision	Split Ratio	Trigger Condition
0	Supercap-heavy	10% battery / 90% SC	Jerk > 1.5 m/s3, OR battery SoC > 85%
1	Balanced	50% / 50%	Normal SoC range, moderate braking
2	Battery-heavy	80% battery / 20% SC	Supercap SoC > 75% OR supercap SoC < 25%

Feature Importances

Feature	Importance	Description
jerk	38%	Rate of change of acceleration
soc_supercap	22%	Supercapacitor state of charge
bat_cycles	18%	Accumulated battery charge cycles
velocity_kmh	12%	Vehicle speed
soc_battery	10%	Battery state of charge

Decision Distribution (20-min simulation)

• Supercap-heavy (Class 0): 2,613 events — 59.9%
• Balanced (Class 1): 838 events — 19.2%
• Battery-heavy (Class 2): 912 events — 20.9%

---

Storage Models

Battery — Thevenin Equivalent

• Capacity: 50 Ah at 48V nominal (2.4 kWh)
• Series resistance R0: 0.01 ohm
• Polarisation branch: R1 = 0.005 ohm, C1 = 2000 F
• SoC range: 20% to 95%
• SoC tracking: Coulomb counting

Supercapacitor — RC Equivalent

• Capacitance: 3000 F
• Voltage range: 20V to 48V (usable ~100 Wh)
• ESR: 0.002 ohm
• SoC definition: (V - Vmin) / (Vmax - Vmin)

---

Tech Stack

• Python 3.13 — simulation, ML training, visualisation
• scikit-learn — Random Forest classifier (200 trees, balanced class weights)
• MATLAB R2025a + Simulink — power electronics model
• MATLAB Engine API for Python — co-simulation bridge
• NumPy / Pandas — numerical computation and data handling
• Matplotlib — results dashboard and SoC plots

---

Requirements

• Python 3.9 to 3.13
• pip install -r requirements.txt
• MATLAB R2022a or later with Simulink (only needed for cosim_controller.py)
• matlab.engine: pip install matlabengine==25.1

---

References

1. Cao & Emadi (2012) — Battery/UltraCapacitor HESS for EVs. IEEE Trans. Power Electronics
2. Song et al. (2014) — Energy Management Strategies for HESS. Applied Energy, Elsevier
3. MathWorks — MATLAB Engine API for Python Documentation

---

License

MIT License — free to use, modify, and distribute with attribution. Copyright 2026 Tanisha Ahlawat