Impacts of data preprocessing and selection on energy consumption prediction model of HVAC systems based on deep learning

Accurate energy consumption prediction is the basis of predictive control for heating, ventilation and air conditioning (HVAC) systems. Data-driven models are widely used for energy consumption prediction. The prediction accuracy can be affected by data preprocessing or selection, which are not well studied yet. This paper attempts to study the impacts of different data processing methods and feature selection on the HVAC energy consumption prediction based on data-driven models. Long and short-term memory models are developed based on historical data to predict the day-ahead hourly energy consumption of HVAC systems. Two data smoothing methods, Gaussian kernel density estimation and Savitzky-Golay filter, are selected and compared. The impacts of feature selection, training set volumes and update frequency of models are analyzed and compared. To obtain general results of the above problems, three office buildings are selected. Results show that the smoothing methods can not ensure the improvement in accuracy by removing the exceptional data in the raw data which arise reasonably in practice. The inputs with raw 1-day historical energy consumption, the dry-bulb temperature and dew-point temperature in the next day, holiday type and day type is recommended to predict day-ahead HVAC energy consumption. It is recommended to use a larger training set if computing cost is acceptable. The model should be updated after being used for over 7 weeks. This study would guide the development of prediction models in practice.