LSTM model optimization based on the epoch numbers for forecasting battery electric vehicle charging
To address the scarcity of high-resolution BEV charging data, a novel feature engineering technique was applied, transforming start-stop electricity charging data sourced from the My Electric Avenue project into the count of concurrent charging events. Recognizing the nonlinear, dynamic, and noisy n...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | en |
| Published: |
UPM
2024
|
| Subjects: | |
| Online Access: | https://umpir.ump.edu.my/id/eprint/46394/1/LSTM%20Model%20Optimization%20Based%20on%20the%20Epoch%20Numbers%20for%20Forecasting.pdf https://umpir.ump.edu.my/id/eprint/46394/ https://www.persama.org.my/images/Menemui_Matematik/2024/MMv463_75_85.pdf |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | To address the scarcity of high-resolution BEV charging data, a novel feature engineering technique was applied, transforming start-stop electricity charging data sourced from the My Electric Avenue project into the count of concurrent charging events. Recognizing the nonlinear, dynamic, and noisy nature of BEV charging patterns, a Long Short-Term Memory (LSTM) network was chosen to model the electricity demand arising from multiple concurrent BEV charging events. The selected LSTM network comprises a single layer with 125 units of LSTM cells employing a tanh activation function and a single dense output layer. This study investigates the error distribution of LSTM networks across a range of epochs beyond the point of initial convergence, focusing on the Mean Absolute Percentage Error (MAPE) as the primary error metric. Unlike previous analyses that often concentrate on specific epochs or report loss values without considering the variability of performance metrics across epochs, this study examines MAPE values at intervals between 10 and 50 epochs, with increments of 10 epochs. The LSTM network converges earlier than epoch 10, and the lowest MAPE was achieved at epoch 20. The lowest recorded MAPE was 1.19%, with a corresponding Root Mean Squared Error (RMSE) of 0.51 . The findings contribute to optimizing LSTM training and improving the generalization of the model to unseen data. |
|---|