Project Overview
Simulations for thermal energy storage charging and discharging are widely used to predict costs, time requirements, and energy outcomes for various systems. However, these simulations are computationally intensive and slow. By leveraging time-sequential machine learning models, we aim to develop faster and more efficient alternatives while maintaining accuracy.
Models Used
- Long Short-Term Memory (LSTM): A type of recurrent neural network (RNN) designed to handle long-term dependencies by mitigating the vanishing gradient problem, making it effective for sequential data modeling.
- Gated Recurrent Unit (GRU): A simplified variant of LSTM that reduces computational complexity while maintaining performance, particularly well-suited for time-series forecasting.
- Recurrent Neural Network (RNN): A fundamental neural network architecture for sequential data, though prone to vanishing gradients, limiting its ability to capture long-range dependencies.
- Radial Basis Function Recurrent Neural Network (RBF-RNN): A hybrid approach integrating radial basis functions into RNNs to enhance pattern recognition and approximation capabilities for time-dependent data.
Results
We tested five cases with thermal input patterns that differ significantly from the training data to assess whether the models capture the underlying physical behaviors of thermal energy dynamics rather than merely interpolating from training data.
Case 1: Comparison of actual vs. predicted temperature values over time
Case 4: Comparison of actual vs. predicted temperature values over time
Performance Comparison
| Model | GRU | LSTM | RBF-RNN | RNN |
|---|---|---|---|---|
| Training Time (s) | 60 | 58 | 122 | 49 |
| Average Testing Time (s) | 0.084 | 0.067 | 0.13 | 0.07 |
Conclusion
This research is ongoing, with discharge patterns yet to be trained or tested. Expanding the dataset with a broader range of features will be essential for developing models that generalize well across different simulation scenarios. As the feature space grows, dimensionality reduction techniques will become crucial for maintaining efficiency.
Preliminary results indicate that GRU-based predictions are the most accurate, and all models can process thermal input patterns in under a second with adequate hardware—dramatically faster than the current 10-minute simulation runtime.