Energy consumption prediction for electric buses using machine learning
Ladda ner
Typ
Examensarbete för masterexamen
Master's Thesis
Master's Thesis
Program
Infrastructure and environmental engineering (MPIEE), MSc
Publicerad
2024
Författare
Wise, Antonia
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
With the increased adoption of electric buses, understanding their energy consumption (EC) has become crucial. For stakeholders such as city planners and bus company owners, having accurate predictions of energy consumption is essential for
effective planning and resource allocation. Thus, identifying the relevant data to be
collected for accurate predictions is of high importance. Machine learning models
have emerged as the most promising tools for predicting energy consumption, offering the precision and reliability needed by stakeholders. Hence, this report aims
to forecast energy consumption in electric buses by finding the important features
in energy consumption and then exploring a suitable machine learning technique
for the given data. Additionally, the report compares the selected model of Multilayered Perceptron Nueral Network (MLPNN) with two other models and assesses
the impact of temporal factors on energy consumption predictions. To achieve this
purpose, first, feature selection is conducted using correlation analysis and multicollinearity checks via the Variance Inflation Factor (VIF). The base MLPNN model
is constructed using the Keras library in Python, with hyperparameter optimisation
performed using GridSearch from the sklearn library. Afterward, the performance
of the MLPNN model is compared to that of two other models: Random Forest (RF) and Extreme Gradient Boosting (XGB), using standard metrics such as
Mean Square Error (MSE) and Mean Absolute Error (MAE). Feature importance
is evaluated for each model, with the MLPNN model assessed using SHapley Additive exPlanations (SHAP). Temporal effects on features are also analysed. The
features deployed in the model are: ’total mileage’, ’speed’, ’AC switch’, ’outside
temperature’, ’inside temperature’, ’run mileage’, ’run duration’, ’bus ID’ and ’time
category’. The optimal hyperparameters for the MLPNN model are: batch size of
20, 100 epochs, Stochastic Gradient Descent (SGD) optimizer, Rectified Linear Unit
(ReLU) activation function, learning rate of 0.01, 2 hidden layers, 32 neurons per
layer, and no regularisation. The evaluation shows that the MLPNN model, using
the selected features and optimised hyperparameters, does not outperform the RF
and XGB models in terms of MAE and MSE. Feature importance analysis reveals
that while MLPNN provides stable importance measures, RF and XGB models are
dominated by a single feature: a run mileage (the Euclidean distance between the
origin and destination of trips) of over 50%. And secondly, run duration with 20%.
SHAP analysis suggests that Run duration and run mileage are most significant for
MLPNN as well. When examining the temporal impact on features, no features
are impacted by time, contrary to initial expectations that speed would show a
substantial temporal effect. The study concludes that the MLPNN model, as constructed, is not significantly better than simpler models in predicting the energy
consumption of electric buses for the given dataset. However, there is potential for
improvement with additional features or more training data. Future research should
explore the inclusion of other relevant features and larger datasets to enhance model
performance.
Beskrivning
Ämne/nyckelord
Energy consumption , Electric Buses , Machine Learning , Multi Layered Perceptron Neural Network , Random Forest , eXtreme Gradient Boosting , Correlation , Validation Inflation Factor , SHapley Additive exPlanations