Improving the accuracy and interpretability of multi-scenario building energy consumption prediction considering characteristics of training dataset

This paper proposed an innovative method, which adopts training dataset characteristics, to improve the accuracy and interpretability of multi-scenario building energy consumption prediction. It adopts the pattern recognition, time-spectrograms, and quantitative training data properties (discrete level, daily cycle intensity, and feature similarity), and investigates and analyzes the data from both time and frequency domain perspectives. The prediction predictability and interpretability are enhanced by combining correlation analysis with the input features. The results show that the energy consumptions for both air conditioning (AC) system and the whole building are significantly impacted by the opening rate of AC terminal equipment. The proposed method can achieve a reduction of 68% for the coefficient-of-variation of root mean square error of the total energy consumption and a decrease of 67% for the AC energy consumption. Training data for AC and total energy consumptions show that the discrete level and feature similarity affect prediction performance, with correlation values ranging from 0.4 to 0.6. Moreover, there is a negative correlation (−0.37) between the daily cycle's intensity and the AC energy use, indicating that prediction performance improves with increasing daily cycle intensity. For building energy consumption prediction, the three indicators, namely discreteness, daily cycle intensity, and feature similarity of training data, can be used to effectively analyze the prediction effect of the target task.