0

I have a timeseries with hourly gas consumptions of some buildings. I need to forecast them with ARIMA and R, but I'm not an expert of R.

I tried to see if there are periodicities during one week.

Data format

Source:

fmt <- '%Y-%m-%d %H:%M:%S'
dat <- read.zoo('740.csv', format=fmt, header=TRUE, sep=',', tz='GMT', stringsAsFactors=FALSE)
#suppress warning of duplicates index in zoo
dat.ts <- dat2$gas..m3.[duplicated(index(dat2$gas..m3.)) == FALSE]
dat.ts <- as.xts(dat2.ts)
eats.week <- dat.ts["2008-02-04::2008-02-08"]
fit <- stl(eats.week, s.window="periodic")

Head:

2008-02-04 00:00:00   53
2008-02-04 01:00:00   54
2008-02-04 02:00:00   55
2008-02-04 03:00:00   53
2008-02-04 04:00:00   54
2008-02-04 05:00:00   53
2008-02-04 06:00:00   66
2008-02-04 07:00:00   55
2008-02-04 08:00:00  112
2008-02-04 09:00:00   54
2008-02-04 10:00:00  113
2008-02-04 11:00:00   55
2008-02-04 12:00:00  108
2008-02-04 13:00:00   55
2008-02-04 14:00:00  101
2008-02-04 15:00:00   54
2008-02-04 16:00:00   99
2008-02-04 17:00:00   57
2008-02-04 18:00:00   92
2008-02-04 19:00:00   65

Plot

Why isn't it periodic? Do I need to specify the frequency? (How in case?)

marcodena
  • 550
  • 2
  • 5
  • 15
  • Please provide some sample data for others to test and adapt your code, for example by adding the output of `dput(head(dat, 20))` to your question. – talat Jul 18 '14 at 16:29
  • What is the time span of your data? This typically happens when you have a time series that is less than two times the length of one complete observation window. For example, if I have data sampled once per week and make a `ts` object with `frequency = 52`, but only have 100 weeks of data, `stl()` will produce an error. Also, yes, you should explicitly specify the frequency / period in `ts()`. – nrussell Jul 18 '14 at 16:33
  • @nrussell data is sampled per hour. In the sample I was analyzing one week as you can see – marcodena Jul 18 '14 at 16:46

1 Answers1

1
d0 <- as.POSIXct(
  "2014-01-01 00:00:00",
  format="%Y-%m-%d %H:%M:%S",
  tz="America/New_York")
##
dWeek <- d0 + seq(
  from=0,
  to=(3600*167),
  by=3600
)
##
set.seed(1234)
x <- rnorm(168,50,5)
tsData <- data.frame(
  Time = dWeek,
  Value=x,
  stringsAsFactors=FALSE)
x.ts <- ts(
  tsData$Value,
  frequency=24,
  start=1)
##
fit <- stl(x.ts,s.window="periodic")
plot(fit)

enter image description here

> summary(fit)
 Call:
 stl(x = x.ts, s.window = "periodic")

 Time.series components:
    seasonal             trend            remainder         
 Min.   :-2.033404   Min.   :46.34685   Min.   :-12.191364  
 1st Qu.:-1.349070   1st Qu.:48.56738   1st Qu.: -2.674591  
 Median :-0.595044   Median :50.09490   Median : -0.273946  
 Mean   : 0.000000   Mean   :49.55673   Mean   :  0.017022  
 3rd Qu.: 0.974903   3rd Qu.:50.49250   3rd Qu.:  2.196548  
 Max.   : 5.250292   Max.   :51.51990   Max.   : 11.635797  
 IQR:
     STL.seasonal STL.trend STL.remainder data 
     2.324        1.925     4.871         6.401
   %  36.3         30.1      76.1         100.0

 Weights: all == 1

 Other components: List of 5
 $ win  : Named num [1:3] 1681 37 25
 $ deg  : Named int [1:3] 0 1 1
 $ jump : Named num [1:3] 169 4 3
 $ inner: int 2
 $ outer: int 0
nrussell
  • 18,382
  • 4
  • 47
  • 60