Attention-based interpretable neural network for building cooling load prediction

Abstract

Machine learning has gained increasing popularity in building energy management due to its powerful capability and flexibility in model development as well as the rich data available in modern buildings. While machine learning is becoming more powerful, the models developed, especially artificial neural networks like Recurrent Neural Networks (RNN), are becoming more complex, resulting in “darker models” with lower model interpretability. The sophisticated inference mechanism behind machine learning prevents ordinary building professionals from understanding the models, thereby lowering trust in the predictions made. To address this, attention mechanisms have been widely implemented to improve the interpretability of deep learning; these mechanisms enable a deep learning-based model to track how different inputs influence outputs at each step of inference. This paper proposes a novel neural network architecture with an attention mechanism for developing RNN-based building energy prediction, and investigates the effectiveness of this attention mechanism in improving the interpretability of RNN models developed for 24-hour ahead building cooling load prediction. To better understand, explain and evaluate these neural network-based building energy prediction models, the obtained attention vectors (or metric) are used to visualize the influence of different parts of model inputs on the prediction result. This helps the users to understand why predictions are made by the model, as well as how input sequences proportionally influence the output sequences. Further analysis of attention vectors can provide interesting temporal information for understanding building thermal dynamics, like the thermal inertia of the building. The proposed attention-based architecture can be implemented in developing optimal operation control strategies and improving demand and supply management. The model developed based on this architecture is assessed using real building operational data, and shows improved accuracy and interpretability over baseline models (without adopting attention mechanisms). The research results help to bridge the gap between building professionals and advanced machine learning techniques. The insights obtained can be used as guidance for the development, fine-tuning, explanation and debugging of data-driven building energy prediction models.