Development-of-predictive-AI-model-for-evtol-energy-profile

This project is created for my MSc Thesis course in Autonomous Vehicle Dynamics and Control at Cranfield University. Supervised by Dr. Yang Xing and Prof. Argyrios Zolotas
The project consists of a dataset downloaded from Klib website which is a battery dataset which has been simulated and modelled for an eVTOL.
The data is collected by a BioLogic : BCS-815 Modular Battery Cycler
The battery used here is Sony-Murata 18650 VTC-6 cell

Link to the dataset: - https://kilthub.cmu.edu/articles/dataset/eVTOL_Battery_Dataset/14226830
Link to the paper: - https://www.nature.com/articles/s41597-023-02180-5
Link to the Kaggle Notebook (Main): -https://www.kaggle.com/code/dineshkrishnamurthy/evtol-energy-consumption-pred-s425895
Link to the Kaggle Notebook (Trial): -https://www.kaggle.com/code/dineshkrishnamurthy/predictive-ai-model-energy-consumption-s425895

The dataset consists of 14 Experiments.
For this project i have used 4 mission profile as mentioned below : - 1) Baseline Flight, 2) Extended cruise (1000 sec), 3) 10% power reduction during discharge (flight) and 4) 20% power reduction during discharge (flight)

Mission Profile 1: Baseline Mission

• No changes to the variables, as this is the reference mission profile.

Mission Profile 2: Cruise time, tc=1000 s

• time_s: Increased cruise time, so values will be higher during the cruise phase.
• Ecell_V, I_mA, EnergyCharge_W_h, QCharge_mA_h: Unchanged, as charge parameters remain the same.
• EnergyDischarge_W_h, QDischarge_mA_h: Possibly increased, due to longer cruise time and discharge.
• Temperature__C: Possibly increased, due to longer cruise time and discharge.
• cycleNumber, Ns: Unchanged, as cycle structure remains the same.

Mission Profile 3: Power reduction of 10% for takeoff, cruise, and landing

• Ecell_V, I_mA: Decreased by 10% during takeoff, cruise, and landing phases.
• EnergyCharge_W_h, QCharge_mA_h: Unchanged, as charge parameters remain the same.
• EnergyDischarge_W_h, QDischarge_mA_h: Decreased by 10% during discharge.
• Temperature__C: Possibly decreased, due to reduced power and discharge.
• cycleNumber, Ns: Unchanged, as cycle structure remains the same.

Mission Profile 4: Power reduction of 20% for takeoff, cruise, and landing

• Ecell_V, I_mA: Decreased by 20% during takeoff, cruise, and landing phases.
• EnergyCharge_W_h, QCharge_mA_h: Unchanged, as charge parameters remain the same.
• EnergyDischarge_W_h, QDischarge_mA_h: Decreased by 20% during discharge.
• Temperature__C: Possibly decreased, due to reduced power and discharge.
• cycleNumber, Ns: Unchanged, as cycle structure remains the same

The below are the variables used in the dataset
** Variable List**

A. Name: time_s
   Description: Time since beginning of experiment in seconds

B. Name: Ecell_V
   Description: Cell Voltage

C. Name: I_mA
   Description: Cell current in milliamperes

D. Name: EnergyCharge_W_h
   Description: The amount of energy supplied to the cell during charge in watt-hours

E. Name: QCharge_mA_h
   Description: The amout of charge supplied to the cell during charge in milliampere-hours

F. Name: EnergyDischarge_W_h
   Description: The amount of energy extracted from the cell during discharge in watt-hours

G. Name: QDischarge_mA_h
   Description: The amount of charge extracted from the cell during discharge in milliampere-hours

H. Name: Temperature__C
   Description: Cell surface temperature in degrees celcius

I. Name: cycleNumber
   Description: Cycle number as recorded by the cell tester (see above note)

J. Name: Ns
   Description: Cycle segment (varies, useful for discerning between segments)

I have selected the initial 3 cycles as they include both charging and dicharging characteristics of the battery which helps us in understanding the flight characteristics

I have observed a pattern in the Ns Column with respect to each number how other parameter differ and what it means-

Revised eVTOL Battery Cycle (Ns) Explanation:

0-Ns: Pre-flight Charging

Full charging of the battery
Temperature rise (27°C to 32°C) due to charging heat generation

1-Ns: Pre-flight Battery Management

Controlled cooling phase (25°C to 27°C)
Ensures optimal battery temperature before flight

2-Ns & 3-Ns: Pre-flight Systems Check**

Battery at rest, mimicking pre-flight checks
Slight self-discharge and voltage stabilization
Minimal temperature change (21°C to 23°C)

4-Ns: Take-off Phase

High power discharge simulating vertical take-off
Significant temperature increase (21°C to 38°C) due to high current draw

5-Ns: Cruise Phase

Moderate power discharge for forward flight
Steady temperature rise (30°C to 34°C) indicating sustained power output

6-Ns: Landing Phase

Highest power demand for vertical landing
Critical temperature spike (29°C to 62°C) showing maximum stress on the battery

7-Ns: Post-flight Cooling

Rapid initial cooling followed by gradual temperature decrease (27°C to 44°C)
Prepares battery for next cycle or maintenance

8-Ns: Battery Recovery and Balancing

No active charging (I_mA & QCharge_mA_h constant at zero)
Voltage recovery (Ecell_V increases from 2.60V to 2.80V)
Gradual decrease in stored charge (EnergyCharge_W_h & QDischarge_mA_h reduces)
Slight reduction in discharge energy (EnergyDischarge_W_h goes from -11 to -9.401 Wh)
Mild temperature increase (22°C to 25°C)

For the Energy prediction Model, I have chosen 3 different ML Algorithms as mentioned below: -
1) Random forests
2) XGBoost Regression
3) linear regression

I have created 3 separate methods with respect to 3 different models to predict the target variable Energy Prediction in Watt hour
Method 1: Training Models Experiment-wise Using Pipelines
Method 2: Ensemble Combined Model Training and Saving
Method 3: Walk-Forward Validation and Model Training

The final conclusion from the whole thesis study is as mentioned below: -

Model Comparison: Across different training methods, including walk-forward validation, ensemble combined training, and experiment-wise pipelines, walk-forward validation method has been displaying high accuracy in the prediction of the target variable and when it comes to ML algorithm, Random Forest consistently emerged as the top performer, followed by XGBoost. Linear Regression was consistently outperformed, demonstrating its limitations in this context.

Overall Conclusion: The study concludes that Random Forest is the most reliable and effective model for predicting energy discharge in eVTOL batteries, with XGBoost as a viable alternative. These models provide a strong foundation for optimizing battery management in urban air mobility solutions, contributing to more efficient and sustainable operations.