Household daily-peak electricity load

forecasting with statistical models

Luciano Guivant Viola

Master Student in Energy Systems Planning; Department of

Mechanical Engineering, University of Campinas, Campinas, SP,
Brazil (e-mail: [email protected]).

Abstract: This article proposes to obtain a statistical model of the daily peak electricity
load of a household located in Austin-TX,USA. The Box-Jenkins methodology was followed
to obtain the best fit for the time-series. Four models provided a good fit: ARIMA(0,1,2),
ARIMA(1,1,2), SARIMA(0,1,2)(0,1,1) and SARIMA(1,1,2)(0,1,1). The model with the highest
Akaike Information Criteria was the ARIMA(1,2,2). However, the model with the highest
forecast accuracy was the SARIMA(1,1,2)(0,1,1), which obtained an RMSE of 0.296 and a
MAPE Of 15.00.

Keywords: Time-series Forecast, ARIMA, SARIMA, Statistical Modeling, Peak Load Forecast

1. INTRODUCTION Φ(B S ) = 1 − Φ1 B − Φ2 B 2S − ... − ΦP B P is the sea

autoregressive operator (AR) of order P,
This article proposes to obtain a statistical model of
the daily peak electricity load of a household located in Θ(B) = 1−Θ1 B −Θ2 B 2 −...−Θq B q is the regular moving
Austin-TX,USA. The data was gathered from the Pecan average operator (MA) of order q,
Street Project Database of the University of Texas at ϑ(B S ) = 1 − ϑ1 B S − ϑ2 B 2S − ... − ϑQ B Q is the seasonal
Austin. They account for minute-to-minute electricity use moving average operator (MA) of order Q,
during the year of 2015 of the house (a total of 525600
observations) The programming language Python was d is the number of regular differences; D is the number of
used to create an algorithm to find the maximum minute- seasonal differences; t is the residual at time t, identically
load of a day, for all the 365 days of the year. This data and independently distributed with mean equal to zero
then served as input for for the statistical modeling, which and constant variance.
was performed in the programming language R. The data The stages in the Box-Jenkins Methodology are the fol-
was separated between training data (first 10 months and lowing:
test data (last two months). The training data was used
to identify the model and the test data for the accuracy of (1) Identification: verify the stationarity of the series
the forecast. The next section will explain the methodology and select the most relevant combination of auto-
and then apply it to the case study. regression and moving-average. Several plausible mod-
els may appear;
2. THE BOX-JENKINS METHODOLOGY (2) Estimation: the parameters of each of the models are
estimated;
The Box-Jenkins Methodology is a three stage method for (3) Diagnostic checking: ensure that the residuals of the
selecting an appropriate model for the purpose of estimat- estimated model are white noise.
ing and forecasting a univariate time series. The model The methodology only applies to time series that are
forecasts future values based on a combination of past stationary. Stationary series are series that have constant
observations (AR) and past shocks (MA). A generalization mean, constant variance, and an auto-correlation that only
of this type of model is called SARIMA, which stands for depends on the lag of two periods. Therefore, the first step
Seasonal Autoregressive Integrated Moving Average. is to test if the series is stationary.
A time series {Xt |t = 1, 2, ..., N } is generated by a
SARIMA(p, d, q)(P, D, Q) process if:
2.1 Identification

ϕ(B)Φ(B S )(1 − B)d (1 − B S )D Xt = Θ(B)ϑ(B S )t Stationary series are series that have constant mean, con-
where N is the number of observations, p, d, q, P, D and Q stant variance, and an auto-correlation that only depends
are integers; B is the lag operator; s is the seasonal period on the lag between two periods. To check if the series is sta-
length. tionary we can make a visual inspection by looking at its
plot, analyzing the ACF (Auto-correlation Function) plot,
ϕ(B) = 1 − ϕ1 B − ϕ2 B 2 − ... − ϕp B p is the regular and performing unit root tests (such as the Augmented
autoregressive operator (AR) of order p, Dickey-Fuller (ADF) and the Phillips-Peron).
The ACF plot of the series (Figure 1) shows auto- First order difference The first order difference (1 − B)
correlations persisting over time, a signature of non- of the series (Figure 2) is taken to try to eliminate non-
stationary series. Moreover, the series appears to have an stationarity. Moreover, from now on we will be working
upward trend up to the month of August and a downward with the log of the series to account for the change in
one from September on. This is the result of the annual variance across time. After the first order difference is
seasonality which is not captured by this series. The con- taken, the series appears to have constant mean. To verify
sequence is that the mean has different values across time this hypothesis the ADF and the Phillips-Perron tests are
(i.e. it is not constant). What does appear in the series is performed: (Dickey−F uller = −10.901; p−value = 0.01),
its weekly seasonality, typical of electricity use series. The (Dickey − F ullerZ(alpha) = 276.18; p − value = 0.01).
series can usually be detrended by applying its first differ- Both test indicate that the first order difference of the
ence. We will then be looking at changes in peak electricity series is stationary at 1% significance level.
consumption per day. As for the weekly seasonality, this
could be captured by the first order difference If its not,
then a seasonal difference may have to be applied.

Fig. 2. Plot and ACF of the series first order difference

Adding a seasonal difference Even though the first order

Fig. 1. Plot and ACF of the raw data difference of the series is stationary, it might not be
capturing the weekly seasonality of the data. Therefore,
Unit root test Even though the plot and the ACF a seasonal difference (1 − B 7 ) with a period equal to seven
appoint to a non-stationary series, we will check if it is applied in addition to the first order difference. This
contains a unit root by performing the Augmented Dickey way it will have a first difference and a seasonal difference.
Fuller Test (ADF) and Phillips-Perron test. What we are looking at now is the variation of the change
in electricity consumption per week. The ADF and the
The ADF test (Said and Dickey, 1984) evolved from the Phillips-Perron tests for the seasonally differenced series
Dickey Fuller Test (DF), in which the null-hypothesis is also appoints to a stationarity at 1% significance level
that α < 1 for the model xt = αxt−1 + ut in which ut is (Dickey −F uller = −13.035, p−value = 0.01), (Dickey −
white noise. The ADF allows the differenced series ut to F ullerZ(alpha) = −291.47, p − value = 0.01)
be any stationary process rather than only white noise. If
α > 1 the series explodes and does not return to the mean Identification of the terms that fit the series Now that we
(i.e. non-stationary). The test resulted in a Dickey-Fuller have the series stationary, it is possible to identify the the
value of -1.446 with a p-value equal to 0.8098. The p-value order of the terms (AR(p),MA(q),SAR(P) and SMA(Q))
is not statistically significant, therefore the null-hypothesis that best describe the behavior of the series. To accomplish
cannot be rejected. 39.906 this, let’s start by looking at the auto-correlations of the
series with only the first order difference (Figure 2)
The Phillips-Perron test is different than the the ADF be-
cause it estimates the auto-correlations directly instead of The ACF shows significant auto-correlations at lags 1 and
approximating it with an auto-regressive model. The result 2. The fact that they are negative and that they cutoff after
of the test is a Dickey-Fuller Z(alpha) equal to -39.906 lag 2 is an indication that is the series is overly-differenced
with a p-value less than 0.01. The ADF and Phillips- (i.e., a pattern of changes of sign from one observation
Perron appoint different results for the stationarity of the to the next) and is suited for an MA model. Adding MA
series. Since at least one of them (ADF) indicates non- terms can partially cancel an order of difference in the
stationarity, the null-hypothesis cannot be rejected and series, correcting for over-differentiation. Moreover, this
the series is assumed to be non-stationary. kind of model has short term memory. Specifically, in an
MA-2 process the forecast in of a given period is only (1 − B)Xt = (1 + 0.39B + 0.34B 2 )t ;
dependent on the last two forecast errors (shocks). The
last forecast is adjusted in the direction of the error it AIC = −63.03; BIC = −51.89; σ 2 = 0.047
made. If the error was positive, i.e. if the previous forecast ARIMA(1,1,2):
was too low, then the next forecast is adjusted upward
by a fraction of that error. Therefore a potential model (1 − 0.37B)(1 − B)Xt = (1 + 0.72B + 0.13B 2 )t ;
for the series is an ARIMA(0, 1, 2). This kind of model is AIC = −67.78; BIC = −52.92; σ 2 = 0.04
equivalent to a Simple Exponential Smoothing.
SARIMA(0,1,2)(0,1,1)[7]:
Let’s now look at the series with the additional seasonal
difference to see what type of information can be extracted. (1 − B)Xt = (1 + 0.39B + 0.33B 2 )(1 − B 7 )t ;
The plot of the seasonally differenced series (Figure 3) AIC = −53.44 BIC = −38.6; σ 2 = 0.48
resembles a moving average process due to a rapid mean
reverting behavior of the series. The ACF plot confirms SARIMA(1,1,2)(0,1,1)[7]:
this. The negative auto-correlations at lags 1 and 2 and
(1−0.36B)(1−B)Xt = (1+0.71B +0.14B 2 )(1−0.99B 7 )t ;
significant partial auto-correlation at lag 3 are signature
of an MA(2) process. Moreover, the presence of a signifi- AIC = −31.64; BIC = −13.19; σ 2 = 0.047
cant negative auto-correlation and negative partial auto-
correlation at multiples of the seasonal period are signa- The most parsimonious is ARIMA(1,1,2) and the least,SARIMA(1,1
ture of and SMA(1). Therefore, a good fit for the series
could also be a SARIMA(0, 1, 2)(0, 1, 1)7 . 2.3 Validation

This stage of the methodology is about diagnostic checking

the residuals of the fitted models. The closer they resemble
a white noise process, the better they are at capturing the
the signal from the original series and hence forecasting.
We are looking to find a model fit with the following
characteristics:
(1) An ACF of the residuals that has virtually no signif-
icant values;
(2) Normality of the residuals; and
(3) Non-significant values of the Llung-Box Statistic

ARIMA(0,1,2) The ACF (Figure 4) shows only two

significant auto-correlations, They could be duo to chance
since at least one in 20 of the auto-correlations are ex-
pected to be significant out of randomness. We could say
the residuals are white noise, but the p-values for the Lung-
Box Statistic tell a different story. The null-hypothesis that
the combined auto-correlation up to lag 12 is zero cannot
be rejected. This indicates that there is still information
to extract from the series. The series seems to have a
persistent memory up to lag 12 that the MA terms cannot
capture. This could be taken into account by adding an
Fig. 3. Plot and ACF of the seasonal and first difference AR term. Also, the Q-Q plot indicates that the residuals
follow a normal distribution expect for extreme values.
2.2 Estimation ARIMA(1,1,2) Adding an AR term to the previous
model compensates for a possible under-differentiation.
In this stage of the methodology, the parameters ϕ, φ, ϑ, Θ It will make the forecast of a period t dependent on
for each model are estimated by maximizing the likely- a fraction of t−1 (including the error term). While an
hood function (the probability of obtaining the data given MA(1) will be only of the forecast error (the shock). In
the model). The fit of the models can be compared on essence and AR model, a shock will have permanent effects
the basis of the Akaike Information Criteria (AIC) that on the series, while in an MA(k) model only for the k
penalizes models with too many parameters: periods. When we add AR terms we are adding memory
AIC = 2k − 2 ln(L) to the series. Adding an AR term lowered the standard
error of the residuals and the AIC of the model. The
In which k equals the log-likelihood and L the number of residuals have only 1 significant auto-correlation (Figure
parameters. The lower the value of AIC the better the fit 5), which is expected to happened by chance in a 25 period
of the model to the series. ACF. The residuals follow a normal distribution expect on
The estimation of the parameters for the chosen models extreme values (outliers). The significant difference with
are the following: the ARIMA(0,1,2) is in the Lung-Box Statistic, which now
shows no significant auto-correlation up to any lag. This
ARIMA(0,1,2): means that by adding an AR(1) we were able to capture
 not recognizing. It is possible that by including an AR term
in the non-seasonal part this signal could be captured.

Fig. 4. Residual analysis of ARIMA(0,1,2)

more signal from the series and leave only the noise. In
other words, the ARIMA(0,1,2) was trying to model a Fig. 6. Residual analysis of SARIMA(0,1,2)(0,1,1)[7]
rapid mean reverting behavior, while the original series
has a slower one.
SARIMA(1,1,2)(0,1,1) This model has only one signif-
icant value of auto-correlation, which is to be expected
in 20 lags. The residuals resemble a normal distribution,
again except outliers. The significant difference that the
addition of an ordinary AR(1) term brought is the values
for the Llung-Box statistic. The values are all bellow the
significance level and thus the residuals of this model fit
are purely white noise. This makes it a good candidate for
forecasting.

Fig. 5. Residual analysis of ARIMA(1,1,2)

SARIMA(0,1,2)(0,1,1) The AIC value is larger than for

the model with only a first difference. Lets take a look at
the residuals. The ACF shows (Figure 6)two significant
spikes, they could be due to chance as they are small
in value (less than 0.2). The problem lies in Ljung-Box
statistic, telling us that the null hypothesis that all the
auto-correlation up to lag 15 are zero cannot be rejected. Fig. 7. Residual analysis of SARIMA(1,1,2)(0,1,1)[7]
This implies that there is some pattern that the model is
2.4 Forecasting

With the fitted models validated, the next step in the

methodology is to use them for forecasting and measure
their accuracy. The forecasts are dynamic,in the sense
that out-of-sample values are updated. The forecast plots
(shown only the out-of-sample values) are in Figure 8. The
gray line represents the observed data and the black, the
forecast.
Two forecast measures are used, Root Mean Square Er-
ror(RMSE) and Mean Absolute Percentage Error(MAPE).
Their values for each model is are in Table 1.
Table 1. Forecast accuracy
Model RMSE MAPE
ARIMA(0,1,2) 0.327 17.25
ARIMA(1,1,2) 0.321 16.789
SARIMA(0,1,2)(0,1,1)[7] 0.304 15.63
SARIMA(1,1,2)(0,1,1)[7] 0.296 15.00

The model with the lowest RMSE and MAPE is the

SARIMA(1,1,2)(0,1,1)[7].

REFERENCES
Cowpertwait, P.S. and Metcalfe, A.V. (2013). Introductory
Time Series with R. doi:10.1007/978-0-387-88698-5.

Dougherty, C. (2007). Introduction to Econometrics. 464.

doi:10.2307/2552778.

Enders, W. (2004). Applied Econometric Time Series,

volume 46. doi:10.1198/tech.2004.s813.

Fan, C., Xiao, F., and Wang, S. (2014). Develop-

ment of prediction models for next-day building en-
ergy consumption and peak power demand using data
mining techniques. Applied Energy, 127, 1–10. doi:
10.1016/j.apenergy.2014.04.016.

James, G., Witten, D., Hastie, T., and Tibshirani, R.

(2013). An Introduction to Statistical Learning. doi:
10.1007/978-1-4614-7138-7.

Studenmund, A.H. (2014). Using econometrics : a practi-

cal guide.

Wooldridge, J. (2012). Introductory Econometrics. A

Modern Approach.
Fig. 8. Forecast of the models