Optimizing Fuel Consumption Prediction Model Without an On-Board Diagnostic System in Deep Learning Frameworks

Rıdvan Keskin ¹, Egemen Belge ¹, Senol Hakan Kutoglu ²

⁰Department of Electrical Electronics Engineering, Zonguldak Bulent Ecevit University, Zonguldak 67100, Türkiye.

Sensors (basel, Switzerland) +

|

November 27, 2025

View abstract on PubMed

Summary

We developed a new model for predicting vehicle fuel consumption rate (FCR) using only throttle, speed, and acceleration data. This data-driven approach improves accuracy and efficiency, reducing emissions without needing complex vehicle specifics.

Area Of Science

Automotive Engineering
Machine Learning
Environmental Science

Background

Accurate real-time fuel consumption rate (FCR) prediction is crucial for enhancing vehicle energy efficiency and reducing emissions.
Conventional methods often require specific vehicle parameters and environmental data, limiting their applicability.
On-board diagnostic (OBD) systems are typically used for FCR data acquisition but can be complex.

Purpose Of The Study

To propose a novel, data-driven FCR prediction model using readily available vehicle data.
To enhance prediction accuracy and reliability using deep learning and probabilistic methods.
To reduce reliance on complex vehicle-specific parameters and environmental conditions.

Main Methods

Development of a Bayesian optimization and Monte Carlo (MC) Dropout-based long short-term memory (BMC-LSTM) network.
Utilizing only vehicle throttle position, velocity, and acceleration data for prediction.
Integration of Bayesian optimization for hyperparameter tuning and MC-Dropout for uncertainty quantification.

Main Results

The BMC-LSTM model achieved superior prediction accuracy compared to conventional LSTM, Bidirectional LSTM, XGBoost, support vector regression, and polynomial models.
The proposed method demonstrated a significantly higher R-squared score and reduced error metrics (MSE, RMSE, MAE).
The model provides calibrated predictions and robustness against distribution drift.

Conclusions

The BMC-LSTM model offers a cost-effective and accurate solution for real-time FCR prediction.
This approach simplifies FCR prediction by requiring only basic vehicle kinematic data at inference.
The method enhances energy efficiency and emission reduction potential in vehicles.

Keywords:

Bayesian optimization deep learning fuel consumption prediction long short-term memory machine learning