# Comparing Long Short-Term Memory (LSTM) and Bidirectional LSTM Deep Neural Networks for power consumption prediction

**Authors:** Davi Guimarães da Silva, Anderson Alvarenga de Moura Meneses  
**arXiv:** [2305.16546](https://arxiv.org/abs/2305.16546)  
**Date:** 2023-05-26  
**Summary page:** https://www.summarizepaper.com/en/paper/2305.16546v1/

## Key points

- The study compares LSTM and BLSTM models for short-term forecasting of electric consumption time series.
- Four different datasets with varying contexts and scales were used, including household power consumption in France, a university building in Brazil, city zones in Morocco, and aggregated electric demand in Singapore.
- Performance evaluation metrics such as RMSE, MAE, MAPE, and R2 were calculated using time series cross-validation scheme.
- Friedman's test was applied to normalized RMSE results to determine statistical significance between LSTM and BLSTM.
- Results show that BLSTM outperforms LSTM with a statistically significant difference (p = 0.0455).
- Bidirectional weight updating significantly improves the performance of LSTM when dealing with different scales of electric power consumption.
- The study provides insights into the application of deep learning models for electric consumption prediction.
- BLSTM is more effective compared to LSTM in this context.
- Bidirectional weight updating is important for forecasting tasks related to energy efficiency and market dynamics.

## Summary

The study focuses on comparing two deep learning models, Long Short-Term Memory (LSTM) and Bidirectional LSTM (BLSTM), for short-term forecasting of univariate electric consumption time series. The objective is to assess the robustness of these models by using four different datasets with varying contexts and scales. The datasets include the power consumption of a household in France, a university building in Santarém, Brazil, the Tétouan city zones in Morocco, and the aggregated electric demand in Singapore. To evaluate the performance of the models, several metrics such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and R-squared (R2) were calculated using a time series cross-validation scheme. Additionally, Friedman's test was applied to normalized RMSE results to determine if there is a statistically significant difference between LSTM and BLSTM. The results show that BLSTM outperforms LSTM with a statistically significant difference (p = 0.0455). This indicates that bidirectional weight updating significantly improves the performance of LSTM when dealing with different scales of electric power consumption. Furthermore, this study provides valuable insights into the application of deep learning models for electric consumption prediction and highlights the effectiveness of BLSTM compared to LSTM. It also emphasizes the importance of considering bidirectional weight updating in forecasting tasks related to energy efficiency and market dynamics.