LSTM vs xLSTM: A Comparative Analysis.
Project Collaborator: Melika Moayer - GitHub | LinkedIn
Project Supervisors
- Rafael Martins - Academic Supervisor.
- Tibo Bruneel - Company Supervisor.
Project Overview
Predicting electricity prices is a tricky game, the market moves fast, reacts to weather, demand, and policy shifts. In this thesis, we wanted to see whether a newer variant of a deep learning architecture called xLSTM-TS could outperform the classic LSTM model in this challenging task. While xLSTM-TS claims better long-term memory and stability, both models were put through the same tests to see which one really shines when predicting real-world energy data.
Research Goals
- Compare the LSTM and xLSTM-TS models in terms of forecasting accuracy and reliability.
- Evaluate how model configuration (small vs. large) and forecast horizon (1h, 12h, 24h) impact performance.
Methodology
Both models were trained on the energy_dataset.csv using consistent preprocessing steps including MinMax normalization and Principal Component Analysis (PCA). Each configuration (small and large) was evaluated using standard metrics such as MAE, MSE, and R². Experiments covered short-, medium-, and long-term horizons with repeated runs for reliability.
Key Findings
Across all configurations and forecast horizons, the results clearly showed that LSTM consistently outperformed xLSTM-TS. It achieved lower error values across metrics such as MAE and MSE and demonstrated higher R² scores, indicating stronger correlation with actual electricity prices. While xLSTM-TS occasionally performed competitively in mid-range forecasts, its results were inconsistent and often accompanied by higher error variance. LSTM models, on the other hand, exhibited greater stability, smoother predictions, and better generalization-particularly in long-term forecasts where xLSTM-TS tended to overfit. These findings reinforce the idea that architectural complexity alone does not guarantee better performance, especially in time series domains characterized by volatility and noise.
Visualizing Volatility
The difference in model behavior becomes especially clear when looking at the prediction patterns of the large models over the long term horizon. The chart below illustrates how LSTM produces smoother, more stable forecasts that closely follow the actual price curve, while xLSTM-TS introduces sharper jumps and occasional overreactions.
These fluctuations may reflect the underlying volatility of the electricity market, where rapid changes can occur due to demand spikes, renewable variability, or market constraints. While such volatility is expected, the sharp jumps from xLSTM-TS suggest greater sensitivity to noise and less effective temporal smoothing, particularly at longer forecast horizons. However, both models demonstrate reliable short-term predictions when using smaller configurations, where limited time dependency reduces the risk of overfitting and model instability.
Discussion
Architectural complexity alone does not ensure better results. LSTM’s established gating mechanisms manage temporal dependencies more efficiently, whereas xLSTM-TS’s added layers can introduce instability and overfitting in volatile datasets such as electricity markets.
Conclusion & Future Work
The research concludes that LSTM remains the more reliable and accurate model for electricity price forecasting. Future work should test both models across multiple markets, integrate external data such as weather or policy indicators, and explore hybrid or ensemble models for improved generalization.
Significance
This study contributes one of the first direct comparisons of xLSTM-TS in the energy forecasting field. The findings emphasize that simpler architectures like LSTM still provide superior performance and stability for practical forecasting applications.