TLDR: A new neural network, the Physics-Guided Memory Network (PgMN), combines deep learning and physics-based models like EnergyPlus to improve building energy forecasting. It addresses challenges such as limited historical data in new buildings or missing sensor data by intelligently integrating both data sources, dynamically correcting biases, and adapting to various real-world scenarios for more accurate predictions.
Accurate energy consumption forecasting is crucial for managing resources efficiently and promoting sustainability in buildings. Traditionally, two main approaches have been used: deep learning (DL) models and physics-based models (PBMs), such as EnergyPlus. While deep learning models excel at prediction when ample historical data is available, they struggle significantly when data is limited or non-existent, as is often the case with newly constructed buildings. On the other hand, physics-based models can simulate energy consumption without relying on historical data, but they demand extensive building parameter specifications and considerable time for setup.
A new research paper introduces the Physics-Guided Memory Network (PgMN), a novel neural network designed to overcome these limitations by integrating predictions from both deep learning and physics-based models. This hybrid approach aims to combine the strengths of both methodologies, offering a more robust and adaptable solution for energy forecasting.
How the Physics-Guided Memory Network Works
The PgMN architecture is composed of three key components:
- Parallel Projection Layers: These layers process inputs from both deep learning models and physics-based simulations. They are designed to handle incomplete inputs and transform them into a format that the network can learn from effectively. This ensures the model can adapt even when some data sources are missing.
- Memory Unit: This is a learnable component that acts as a global repository of past experiences, specifically recording and correcting persistent biases in energy forecasts. It allows the model to dynamically adjust predictions by learning from previous errors, making it more accurate over time.
- Memory Experience Module: This module is responsible for optimally combining the processed information from both deep learning and physics-based predictions. It determines the best balance between these two sources to minimize prediction errors. Crucially, it can even extend forecasts beyond the range of the individual input models if doing so leads to a more accurate prediction.
Also Read:
- Enhancing Weather Forecasting with Dynamic Graph Learning for Earth Observations
- Dynamic Mesh Optimization: Introducing UGM2N for PDEs
Addressing Real-World Challenges
The PgMN was rigorously evaluated across various challenging scenarios, demonstrating its versatility and accuracy:
- When Both Predictions Are Available: In situations where both deep learning and EnergyPlus predictions are accessible (e.g., in operational buildings with infrastructure changes like HVAC upgrades), PgMN slightly outperforms standalone deep learning models, showing enhanced accuracy by leveraging the combined insights.
- With Sparse Historical Data: When actual energy consumption data is partially missing due to sensor failures or other issues, PgMN can still provide reliable forecasts. It intelligently imputes missing values and adapts its learning process, proving more robust than traditional deep learning models in such conditions.
- In Newly Constructed Buildings (No Historical Data): For new buildings where historical data is entirely absent, PgMN can operate by using physics-based simulations as its primary input and even as a proxy for ground truth during training. This allows it to make meaningful predictions where deep learning models would be unusable.
- Solely Using Physics-Based Data: If only EnergyPlus forecasts are available (e.g., due to lack of historical data for DL), PgMN can still refine these predictions, demonstrating its ability to improve upon physics-based simulations alone.
- Solely Using Deep Learning Data: Conversely, if physics-based models are not feasible, PgMN can still enhance existing deep learning forecasts by leveraging its internal memory and input transformation mechanisms.
The research paper provides theoretical proofs for PgMN’s capabilities, including its ability to approximate any continuous function, correct biases, and produce unbounded outputs when necessary to minimize errors. Experimental validation on a student residence facility in London, Ontario, Canada, confirmed its superior applicability and accuracy compared to standalone approaches.
While setting up physics-based models like EnergyPlus can be time-consuming and requires expert knowledge, the PgMN’s inference time is minimal, making it computationally efficient for real-world deployment once configured. This innovative approach offers a promising solution for energy consumption forecasting in dynamic building environments, significantly enhancing model applicability in scenarios where historical data is limited or unavailable, or when physics-based models alone are inadequate. For more details, you can refer to the full research paper here.
