COMPARISON OF OCCUPANT BEHAVIOR MODELS APPLIED TO A HOUSEHOLD
暂无分享,去创建一个
This paper presents a preliminary study on occupant behavior modeling for estimating high-resolution energy demand profile of residential buildings. The authors developed two occupant behavior models based on the existing modeling approaches. We then applied the models to a four-member household living in Osaka, Japan, to evaluate the capability of the behavior models. Time use diary for nine days was collected for each household member to develop time use data as the input of the models. Several indicators showing the duration and transition of behaviors were calculated for the actual time use and simulation results were calculated and compared. Based on the comparison, the authors discussed the strength and weakness of the modeling approaches and proposed methods to improve the modeling approaches. INTRODUCTION Usually, in a household energy demand model, the occupant behavior is given by a pattern that represents an average occupant’s behavior. Although this approach is easy to set up and useful to estimate the total energy consumption of households or the average pattern of energy consumption, it does not provide useful inputs to replicate a high-temporal resolution energy demand. To replicate such profile, stochastic occupant behavior must be directly simulated. Our examination of the existing occupant behavior models revealed three models proposed by Richardson et al. (2010), Widén et al. (2010), and Tanimoto et al. (2008). These models employ time use data (TUD), which show how people spend their time. The data is usually developed on the basis of one-day diaries recorded by several thousands of people. If the raw data of a diary are available, information on the number of respondents who performed an activity i and who changed their behavior from behavior i to j (Nij) can be generated. The approaches proposed by Richardson and Widén model such behavior transitions as a Markov Chain (MC) process, which is a stochastic process in which transitions of the state (i.e., change of behavior in the occupant behavior model) only depend on the state at the previous time step. In Widén’s model, there are nine states (or behaviors) that occupants can undertake. Behaviors of all the examined household members are individually simulated. The probability of the transition of state for each time step at the corresponding time in TUD is given by Nij. In Richardson’s model, the MC process is used to determine the number of “active occupants” who are in the house and are not sleeping. The states that can be undertaken are each number of occupants up to the household size (the number of family members). The transition probability is developed considering four transition states using Nij: from active to active, from active to inactive, from inactive to active, and from inactive to inactive. After the number of active occupants is determined, the operation of home appliances by them is determined. One of the limitations of the modeling approach is that it requires raw data of the time use survey, while only statistical data are publicly available in many countries. Tanimoto’s model overcomes this limitation. His approach only uses the following statistic information: Average ongoing minutes (AOM) of the activity that occurs in a day, Standard deviation of AOM (SDOM), and Percentage of respondents who adopt the behavior (PB) at a specific time of a day. First, the duration time of all considered behaviors is determined while assuming a logarithmic Gauss distribution defined by AOM and SDOM. The list of discrete behaviors with a determined time period is located on a day by considering PB in the following manner. First, a time step is randomly selected. One behavior is selected using PB at the time step. After the first behavior is located, the behavior beginning from the time at which the first behavior ends is selected using PB at that instant. This process is repeated until all the discrete behaviors in a day are placed. Although the duration of behaviors can be modeled by using AOM, SDOM, and PB, in the Tanimoto’s approach, the transition of behaviors are unnecessarily well replicated. This is because the number of transitions depends on the number of discrete behaviors. This might be the vital weakness to well replicate the actual stochastic behavior of the energy demand of a household, while the MC model well replicates it. In order to evaluate the strength and weakness of the two modeling approaches, the authors developed two occupant behavior models based on the MC approach and Tanimoto’s approach and applied the models to a residential household. By comparing the actual time use data of the household members and the simulated time use, the authors evaluate the performance of the modeling approaches. In the remaining parts of the paper, we first introduce the household to which the models were applied. The method used to develop TUD is also described. Then, we explain the occupant behavior models. The simulation result is then compared with the actual time use. We finally discuss the strength and weakness of the modeling approaches to model high-temporal resolution energy demand for households. COLLECTION OF TIME USE OF A HOUSEHOLD and TUD The household to be modeled is a four member family living in Osaka, Japan, consisting of a working male, a working female, a high school student, and a junior high school student. The authors handed a format of time use diary shown in Figure 1 to collect time use of each household member. The time use was classified into 11 categories in addition to two lists for free description. The diary was filled by all the family members for two weeks. Then 9 weekdays of time-use diary was collected. Based on the time use diary, TUD necessary to perform behavior models (AOM, SDOM, and PB) was developed by following the definition from 9 days data. Figure 1. The format of the time use diary (the respondents were asked to draw a line on the cell of corresponding time and behavior listed on the left side). OCCUPANT BEHAVIOR MODELS Two behavior models were developed based on the abovementioned approaches. As previously mentioned, the Markov Chain model is based on a stochastic process in which transitions of the state only depend on the state of the previous time step. For this model, a 2 dimension matrix (13x13) was generated for each time step (Bijt) indicating the probability of transition from behavior i, into behavior j, at time step t. On the start of each simulation, the behavior of the first time step is determined by using PB at time step 0:00 and a random number. For the next time step, another random number is generated and by the probabilities inputted on the transition matrix Bijt, the behavior is determined. All time steps are filled with behaviors 1 to 13 following the above mentioned process using transition matrix Bijt. Figure 2 shows the simulation procedure developed based on the Tanimoto’s approach mentioned above. Several modifications were added. We call this model the Roulette Selection (RS) model in this paper. The model was originally developed to model occupant behavior by using the TUD developed by the Broadcasting Culture Research Institute of Japan (2001). 27 types of behaviors such as sleep, work, and watching TV are defined. 0:00 0:10 0:20 0:30 0:40 0:50 1:00 1:10 1:20 1 Sleeping 2 Personal care 3 Meal 4 Bathing 5 Cooking 6 Houseworks 7 Caring of children/elderly persons 8 TV/radio/music 9 Resting 10 Work/study at home 11 Outing 12 Free description 1 ( ) 13 Free description 2 ( ) Behavior First, the duration of considered behavior per day is determined by assuming that it follows the Gauss distribution defined by AOM and SDOM. Then, discrete behaviors made for the following five behaviors, sleeping, outing for work or school, eating, and bathing, are placed on a day prior to the rest of behaviors. We call these five behaviors “routine behaviors” since these behaviors are undertaken routinely by occupants. The method to place the discrete behaviors is same as in the Tanimoto’s model. A random number is generated to determine a time step from those with positive PB. Then, the time at which the occupant begins the behavior is determined by using PB as the duration time before and after the selected time step is determined by PB before and after the selected time step. If a behavior has been occupied to the time steps, the probability is assumed to be zero. If the duration of the behaviors is larger than the time steps selected for the behavior, the behavior is allocated to the selected time steps and the duration of the behavior is shortened so that the remaining amount can be selected at other time steps. Next, the remaining non-routine behaviors are placed on the day. An unoccupied time step is randomly selected. Then, a random number is assigned to PB at the time step to select a behavior. The selected behavior is allocated to the selected time step. The times at which the occupant starts and stops the behavior are determined using the method mentioned above. This process is repeated until all time steps are occupied by a behavior. The initially determined time period for each behavior can be divided in this process, resulting in more frequent changes in behavior. RESULT AND DISCUSSION The simulation result was evaluated by using the following indicators: Duration time for the behaviors per day, Time at which the routine behaviors start and end, Number of behavior transitions per day, Probability distribution showing percentage of behaviors at each time step, and Number of different pattern of occupant behavior transition in 500 simulations START Input of occupants’ attribute and day classification Select corresponding TUD (AOM, SDOM, and PB) Determine the time for sleep Day classification is weekday & People attribute is working person/student Determine time for work/study and commuting Determine time for dining and bathing Randomly select an unocc
[1] Y. Shimoda,et al. Evaluation of city-scale impact of residential energy conservation measures using the detailed end-use simulation model , 2007 .
[2] Aya Hagishima,et al. Validation of methodology for utility demand prediction considering actual variations in inhabitant behaviour schedules , 2008 .
[3] David Infield,et al. Domestic electricity use: A high-resolution energy demand model , 2010 .
[4] J. Widén,et al. A high-resolution stochastic model of domestic activity patterns and electricity demand , 2010 .