Abstract
A bottom-up energy demand estimation model considers changes in weather and equipment over time, so the model is useful for a quantitative evaluation of specific measures in the plan for global warming countermeasures. The model simulates the operation of equipment based on occupant behavior, equipment ownership, and so on. The input data is constrained by its survey methodology and frequency, and the accuracy of the input data leads to the difference between estimates and measured. We calibrated parameters related to occupant behavior and the operation of heating and cooling equipment by solving inverse problems with smart meter data. With the popularity of smart meters worldwide, the data of electricity consumption become available. However, in terms of the data capacity, management, and privacy protection, the resolution become rougher, so the development of methods to analyze such data have attracted attention. We utilize the very low-rate data of the order of 30 min by assuming the distribution of electricity consumption for each time and household type. The error measure was the Wasserstein distance between distribution of estimated and measured electricity consumption to reproduce with good accuracy not only the average electricity consumption of the target households, but the variation of the households.Amasyali et al. reviewed trends in the field of research that predicted data-driven energy demand estimation, most of which used machine learning methods such as support vector machine (SVM) and artificial neural network (ANN). Zhao et al. proposed the unsupervised learning method to decompose smart meter data into the electricity consumption of each equipment. In this study, we used a bottom-up residential energy end-use simulation model called TREES (Total Residential End-use Energy Simulation), in which occupant behavior is simulated by stochastic person-based activity model. This activity model probabilistically simulates the behavior of each resident, so the model can consider the variability in simulated activities among households and uses the Survey on time use and leisure activities, in which the data is collected from self-declaration of the activities on the limited surveyed days. Therefore, the data may deviate from the actual situation. We first calibrated the parameters related to occupant behavior in the intermediate period, and then calibrated the parameters related to the operation of heating and cooling equipment by Bayesian optimization to efficiently search for solutions. This allows us to clarify the trends in occupant behavior and the actual operation of heating and cooling equipment from electricity consumption.In the intermediate period, the sleep duration was shorter than that of the survey on time use and leisure activities, which is the input data of the activity model. However, the values were comparable to that of the highly reliable smartphone survey, indicating that the model reflects the actual occupant behavior. The error measure between distribution of estimated and measured electricity consumption was improved, while errors remained in the morning. This is because only parameters related to occupant behavior were calibrated, and it is necessary to add parameters, such as equipment performance, to the optimization target in the future. In the heating and cooling period, the distance between distribution of estimated and measured was greatly improved. The accuracy of TREES was improved, so we can quantitatively evaluate the effects of measures such as the Plan for Global Warming Countermeasures. Also, by interpreting the parameters, we can estimate the trend of occupant behavior from smart meter data, which is useful information for planning energy conservation measures in the residential sector.