Interpreting the neural network model for HVAC system energy data mining

To enhance the energy efficiency of heating, ventilation and air conditioning (HVAC) systems, which is a nonlinear and complicated system, machine learning has been used intensively. However, traditional white-box machine-learning models with good interpretability often do not have satisfactory accuracy, while black-box machine-learning models that are more accurate could not be interpreted easily, which impedes its application. In this study, we propose a method to interpret a neural network (NN) model using gradients of the model, which quantifies the marginal influence of inputs to the output, based on the chain rule. Then the NN model is compared with other machine-learning models (the linear regression model and the XGBoost model) in accuracy, interpretability, and robustness. We then compare our result with the correlation analysis, a widely used method to extract the relation between the outputs (in this case, energy consumption) and inputs. Further, we perform feature selection based on gradients of the NN model, reducing 40% calculation time without sacrificing model accuracy. The feature importance given by the NN model is proved to be reasonable and informative compared with the other two models. The scope of this study is neither to verify the superiority of the NN model, nor to predict the energy consumption accurately. Instead, the goal of this study is to provide a method to interpret the results of NN models.

Automated summary

This study proposes a method to interpret neural network models using gradients, quantifying the influence of inputs on outputs. The method reduces calculation time and provides reasonable and informative feature importance compared to other models.