LSTM model error distribution of battery electric vehicle charging load demand
Stability of electricity demand forecast accuracy from Battery Electric Vehicles (BEV) in production environment is studied. This study utilizes Long Short-Term Memory (LSTM) neural networks, incorporating a fixed origin procedure to split all training data to account for daily model updates, to sim...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Conference or Workshop Item |
| Language: | en |
| Published: |
Springer
2025
|
| Subjects: | |
| Online Access: | https://umpir.ump.edu.my/id/eprint/46391/1/LSTM%20Model%20Error.pdf https://umpir.ump.edu.my/id/eprint/46391/ https://doi.org/10.1007/978-981-96-8093-1_21 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Stability of electricity demand forecast accuracy from Battery Electric Vehicles (BEV) in production environment is studied. This study utilizes Long Short-Term Memory (LSTM) neural networks, incorporating a fixed origin procedure to split all training data to account for daily model updates, to simulate production environment. The mean absolute percentage error (MAPE) performance metric for LSTM forecasted observed data is computed over a specified epoch range, and the fluctuations in MAPE across this range are documented. Furthermore, the LSTM forecasted MAPE of decomposed Seasonal-Trend decomposition using Loess (STL), on both original, and positively normalized STL seasonality and residual components. The findings suggest that the forecasted MAPE fluctuations in the observed data is slightly lower in its mean and standard deviation, compared to both STL reconstructed MAPE. For LSTM iterations between 165 to 234 epochs, mean MAPE for observed data is 1.45% (standard deviation: 0.49), for STL reconstructed is 2.11% (standard deviation: 0.60), and for STL reconstructed (renormalized) is 2.04% (standard deviation: 0.70). These findings highlight the potential for refining models through combined LSTM, and other models, on the individual components of data using decomposition techniques. |
|---|