Monotonic Neural Network: combining Deep Learning with Domain Knowledge for Chiller Plants Energy Optimization

Fanhe Ma, Faen Zhang, Shenglan Ben, Shuxin Qin, Pengcheng Zhou, Changsheng Zhou, Fengyi XuarXiv 2021 · OpenReview

Abstract: In this paper, we are interested in building a domain knowledge based deep learning framework to solve the chiller plants energy optimization problems. Compared to the hotspot applications of deep learning (e.g. image classification and NLP), it is difficult to collect enormous data for deep network training in real-world physical systems. Most existing methods reduce the complex systems into linear model to facilitate the training on small samples. To tackle the small sample size problem, this paper considers domain knowledge in the structure and loss design of deep network to build a nonlinear model with lower redundancy function space. Specifically, the energy consumption estimation of most chillers can be physically viewed as an input-output monotonic problem. Thus, we can design a Neural Network with monotonic constraints to mimic the physical behavior of the system. We verify the proposed method in a cooling system of a data center, experimental results show the superiority of our framework in energy optimization compared to the existing ones.

Introduction

Chiller plants—the massive cooling systems used in data centers, factories, and airports—are significant energy consumers, accounting for roughly 20% of building electricity usage worldwide. Optimizing these systems is difficult because the relationships between variables (like fan speed, water temperature, and power consumption) are highly non-linear and evolve over time as equipment ages.

While deep learning is effective at modeling non-linear data, it faces two major hurdles in physical systems: data scarcity and physical inconsistency. Standard neural networks often 'learn' relationships that violate physics, creating 'wiggles' in the model due to noise in small datasets. These wiggles result in jagged optimization surfaces with false local minima, making it nearly impossible for optimization algorithms to find the most efficient settings.

Figure 2: Comparison of non-monotonic surfaces in standard models, showing jagged landscapes that hinder effective optimization.

Methodology

We propose two methods to enforce physical monotonicity—where an increase in input always leads to a consistent increase or decrease in output—within a neural network architecture.

Soft-MNN (Loss-based)

This approach incorporates domain knowledge through a pairwise rank loss function. During training, the model is penalized if its predicted output order contradicts known physical laws, effectively teaching the network to maintain monotonic relationships without strictly constraining the architecture.

Hard-MNN (Architecture-based)

Inspired by Input Convex Neural Networks (ICNN), this version bakes monotonicity into the network's structure. By constraining weights to be non-negative and using a custom activation function called PTRelu (Parametric Truncated ReLU), the model is mathematically guaranteed to be monotonic. To prevent the model from becoming too simple, we include pass-through layers that connect the input directly to hidden layers, allowing the network to capture complex, higher-order relationships while maintaining physical constraints.

Results and Discussion

The proposed models were validated using real-world operational data from a large data center's cooling system.

The Smoothness Advantage

A critical insight from this work is that smoothness is often more valuable than raw accuracy for control systems. While a standard Multi-Layer Perceptron (MLP) might achieve a lower error on a static test set, it often produces a 'jagged' prediction surface with non-physical wiggles. In contrast, the MNN produces a smooth, often convex surface, which allows gradient-based solvers to reliably find the true global minimum.

Figure 6: Standard neural networks often exhibit non-physical wiggles in sensitivity curves due to noise in small datasets, violating domain knowledge.

Energy Optimization Performance

In real-world trials, the Hard-MNN demonstrated superior stability across varying weather conditions. The optimization framework achieved an average reduction in Power Usage Effectiveness (PUE) of approximately 1.5% compared to standard deep learning methods. For a large-scale data center, this represents significant cost savings and a meaningful reduction in carbon footprint.

Conclusion

By using physics to define the shape of the solution and AI to learn the parameters from data, we bridge the gap between black-box models and traditional engineering. This framework can be extended to other physical systems, such as power grids or chemical processes, where monotonicity is a fundamental characteristic of the underlying physics.

Summary updated

Abstract

In this paper, we are interested in building a domain knowledge based deep learning framework to solve the chiller plants energy optimization problems. Compared to the hotspot applications of deep learning (e.g. image classification and NLP), it is difficult to collect enormous data for deep network training in real-world physical systems. Most existing methods reduce the complex systems into linear model to facilitate the training on small samples. To tackle the small sample size problem, this paper considers domain knowledge in the structure and loss design of deep network to build a nonlinear model with lower redundancy function space. Specifically, the energy consumption estimation of most chillers can be physically viewed as an input-output monotonic problem. Thus, we can design a Neural Network with monotonic constraints to mimic the physical behavior of the system. We verify the proposed method in a cooling system of a data center, experimental results show the superiority of our framework in energy optimization compared to the existing ones.
BibTeX
@misc{ma2021monotonic,
  author = {Fanhe Ma and Faen Zhang and Shenglan Ben and Shuxin Qin and Pengcheng Zhou and Changsheng Zhou and Fengyi Xu},
  title = {{Monotonic Neural Network: combining Deep Learning with Domain Knowledge for Chiller Plants Energy Optimization}},
  year = {2021},
  eprint = {2106.06143},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2106.06143},
}
Figure