Comparison of predictive models for time series forecasting : Power demand prediction of DC charging network
Veikkolainen, Juho-Eemeli (2024)
2024
Tieto- ja sähkötekniikan kandidaattiohjelma - Bachelor's Programme in Computing and Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2024-12-12
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-2024120210688
https://urn.fi/URN:NBN:fi:tuni-2024120210688
Tiivistelmä
The widespread adoption of electric vehicles has made DC charging networks a crucial part of the infrastructure. These networks experience fluctuating power requirements influenced by factors such as traffic patterns and charging behavior, which has brought new challenges to the power grid and energy management. Consequently, accurate forecasting of power consumption of DC charging networks has become crucial to support grid stability and infrastructure planning. This thesis aims to develop a robust and practical forecasting model capable of predicting power demand in a DC charging network up to 24 hours in advance.
Time series forecasting can be performed using various different method. This thesis explores 4 different models from 3 different categories. These models are seasonal autoregressive integrated moving average (SARIMA) from the statistical approaches, exterme gradient boosting (XGBoost) model and Facebook's Prophet model from the machine learning techniques, and long short-term memory (LSTM) from the deep learning methods. All of the models were trained on multiple years of data from charging network consisting of over 850 DC chargers.
The models are evaluated on couple of different metrics. The primary evaluation metric is prediction error, which there is 3 different flavors: mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean squared error (RMSE). The models are also evaluated on runtime efficiency and how well they fit in to practical use. Based on these aspects, the XGBoost model architecture is clear outperformer across all prediction error metrics as well as runtime efficiency and practicality.
Just as DC charging infrastructure is in its early stages, so are the applications of machine learning in this domain. The vast amounts of data generated daily by chargers present significant opportunities for innovation and development. Leveraging this data source enhances the operational efficiency of charging networks, optimizes grid management, and improves the overall charging experience for end-users.
Time series forecasting can be performed using various different method. This thesis explores 4 different models from 3 different categories. These models are seasonal autoregressive integrated moving average (SARIMA) from the statistical approaches, exterme gradient boosting (XGBoost) model and Facebook's Prophet model from the machine learning techniques, and long short-term memory (LSTM) from the deep learning methods. All of the models were trained on multiple years of data from charging network consisting of over 850 DC chargers.
The models are evaluated on couple of different metrics. The primary evaluation metric is prediction error, which there is 3 different flavors: mean absolute error (MAE), mean absolute percentage error (MAPE), and root mean squared error (RMSE). The models are also evaluated on runtime efficiency and how well they fit in to practical use. Based on these aspects, the XGBoost model architecture is clear outperformer across all prediction error metrics as well as runtime efficiency and practicality.
Just as DC charging infrastructure is in its early stages, so are the applications of machine learning in this domain. The vast amounts of data generated daily by chargers present significant opportunities for innovation and development. Leveraging this data source enhances the operational efficiency of charging networks, optimizes grid management, and improves the overall charging experience for end-users.
Kokoelmat
- Kandidaatintutkielmat [8907]