52
Internet of Things (IoT)
3.6.6.2 Residential Energy Consumption Survey Data
The Residential Energy Consumption Survey (RECS) data (Kavousian et al., 2012) were
collected in 2005 through a national survey on residential energy-related data, where 4,381
households in housing units were randomly selected to represent the 111.1 million housing units in the United States. These data were obtained from residential energy suppliers
with the consumption and expenditures per unit sample. The consumption and expenditures and intensities data are divided into two parts: In the first part, the data provide
energy consumption and expenditure by census region, population density, climate zone,
type of housing unit, year of construction, and ownership status, whereas in the second
part, the same data are provided according to household size, income category, race, and
age. The next update to the RECS survey (2009 data) made available in 2011, summarized
in Table 3.2, has been used in this study.
3.6.6.3 The Reference Energy Disaggregation Data
The Reference Energy Disaggregation Data set (REDD) (Kolter and Johnson, 2011) contains
home electricity data: one is with high-frequency current/voltage waveform data of the
two power mains and the other is lower-frequency power data with mains and individual
labeled circuits. The low_freq directory contains average power readings for both the power
mains and the individual circuits of the house with plug loads and plug monitors. The data
are logged at a frequency of about once a second for mains and once every three seconds for
the circuits. AC waveform data for power mains and a single-phase voltage for household
purposes are present in the high-frequency directory. In order to reduce the data to a manageable size, the waveform may be compressed using lossy compression. This is mainly
because the voltage signal in most homes is approximately sinusoidal (unlike the current
signals, which can vary substantially from a sinusoidal wave), and zero-crossings of the
voltage signal to isolate a single cycle of the AC power are found. For the time spanned by
this single cycle, both the current and voltage signals are recorded, and the entire waveform
is then reported. However, because the waveform remains approximately constant for long
periods of time, the current and voltage waveform at “change points” in the signal are only
reported. The high-frequency raw directory finally contains raw voltage and current waveform without alignment and compression, for a small number of sample points throughout
the data as the entire dataset consists of more than a terabyte of data.
3.7 Experimental Results and Discussion
All the experiments are conducted in an Intel Pentium 2.8 GHz CPU with 200 GB HDD
and 2 GB RAM and Microsoft XP professional. At first, we use RECS data (Residential
Energy Consumption Survey-2010, PART-1 and 2) as per the statistics provided below: with
household size in Figure 3.5 and Table 3.1, with census region and division in Table 3.2, and
then with a most populated state in Table 3.3.
3.7.1 Experiment 1: Prediction and Forecasting Using Deep
Convolutional Neural Network with RECS Dataset
All the experiments are conducted with the full training set to predict and forecast the
model with deep convolutional neural network, with 95% confidence interval and 10 years
Internet of Things (IoT)
3.6.6.2 Residential Energy Consumption Survey Data
The Residential Energy Consumption Survey (RECS) data (Kavousian et al., 2012) were
collected in 2005 through a national survey on residential energy-related data, where 4,381
households in housing units were randomly selected to represent the 111.1 million housing units in the United States. These data were obtained from residential energy suppliers
with the consumption and expenditures per unit sample. The consumption and expenditures and intensities data are divided into two parts: In the first part, the data provide
energy consumption and expenditure by census region, population density, climate zone,
type of housing unit, year of construction, and ownership status, whereas in the second
part, the same data are provided according to household size, income category, race, and
age. The next update to the RECS survey (2009 data) made available in 2011, summarized
in Table 3.2, has been used in this study.
3.6.6.3 The Reference Energy Disaggregation Data
The Reference Energy Disaggregation Data set (REDD) (Kolter and Johnson, 2011) contains
home electricity data: one is with high-frequency current/voltage waveform data of the
two power mains and the other is lower-frequency power data with mains and individual
labeled circuits. The low_freq directory contains average power readings for both the power
mains and the individual circuits of the house with plug loads and plug monitors. The data
are logged at a frequency of about once a second for mains and once every three seconds for
the circuits. AC waveform data for power mains and a single-phase voltage for household
purposes are present in the high-frequency directory. In order to reduce the data to a manageable size, the waveform may be compressed using lossy compression. This is mainly
because the voltage signal in most homes is approximately sinusoidal (unlike the current
signals, which can vary substantially from a sinusoidal wave), and zero-crossings of the
voltage signal to isolate a single cycle of the AC power are found. For the time spanned by
this single cycle, both the current and voltage signals are recorded, and the entire waveform
is then reported. However, because the waveform remains approximately constant for long
periods of time, the current and voltage waveform at “change points” in the signal are only
reported. The high-frequency raw directory finally contains raw voltage and current waveform without alignment and compression, for a small number of sample points throughout
the data as the entire dataset consists of more than a terabyte of data.
3.7 Experimental Results and Discussion
All the experiments are conducted in an Intel Pentium 2.8 GHz CPU with 200 GB HDD
and 2 GB RAM and Microsoft XP professional. At first, we use RECS data (Residential
Energy Consumption Survey-2010, PART-1 and 2) as per the statistics provided below: with
household size in Figure 3.5 and Table 3.1, with census region and division in Table 3.2, and
then with a most populated state in Table 3.3.
3.7.1 Experiment 1: Prediction and Forecasting Using Deep
Convolutional Neural Network with RECS Dataset
All the experiments are conducted with the full training set to predict and forecast the
model with deep convolutional neural network, with 95% confidence interval and 10 years