37
$\begingroup$

I have four different time series of hourly measurements:

  1. The heat consumption inside a house
  2. The temperature outside the house
  3. The solar radiation
  4. The wind speed

I want to be able to predict the heat consumption inside the house. There is a clear seasonal trend, both on a yearly basis, and on a daily basis. Since there is a clear correlation between the different series, I want to fit them using an ARIMAX-model. This can be done in R, using the function arimax from the package TSA.

I tried to read the documentation on this function, and to read up on transfer functions, but so far, my code:

regParams = ts.union(ts(dayy))
transferParams = ts.union(ts(temp))
model10 = arimax(heat,order=c(2,1,1),seasonal=list(order=c(0,1,1),period=24),xreg=regParams,xtransf=transferParams,transfer=list(c(1,1))
pred10 = predict(model10, newxreg=regParams)

gives me: enter image description here

where the black line is the actual measured data, and the green line is my fitted model in comparison. Not only is it not a good model, but clearly something is wrong.

I will admit that my knowledge of ARIMAX-models and transfer functions is limited. In the function arimax(), (as far as I have understood), xtransf is the exogenous time series which I want to use (using transfer functions) to predict my main time series. But what is the difference between xreg and xtransf really?

More generally, what have I done wrong? I would like to be able to get a better fit than the one achieved from lm(heat ~ tempradiwind*time).

Edits: Based on some of the comments, I removed transfer, and added xreg instead:

regParams = ts.union(ts(dayy), ts(temp), ts(time))
model10 = arimax(heat,order=c(2,1,1),seasonal=list(order=c(0,1,1),period=24),xreg=regParams)

where dayy is the "number day of the year", and time is the hour of the day. Temp is again the temperature outside. This gives me the following result:

enter image description here

which is better, but not nearly what I expected to see.

$\endgroup$

3 Answers 3

37
+50
$\begingroup$

You're going to have a little bit of trouble modeling a series with 2 levels of seasonality using an ARIMA model. Getting this right is going highly dependent on setting things up correctly. Have you considered a simple linear model yet? They're a lot faster and easier to fit than ARIMA models, and if you use dummy variables for your different seasonality levels they are often quite accurate.

  1. I'm assuming you have hourly data, so make sure your TS object is setup with a frequency of 24.
  2. You can model other levels of seasonality using dummy variables. For example, you might want a set of 0/1 dummies representing the month of the year.
  3. Include the dummy variables in the xreg argument, along with any covariates (like temperature).
  4. Fit the model with the arima function in base R. This function can handle ARMAX models through the use of the xreg argument.
  5. Try the Arima and auto.arima functions in the forecast package. auto.arima is nice because it will automatically find good parameters for your arima model. However, it will take FOREVER to fit on your dataset.
  6. Try the tslm function in the arima package, using dummy variables for each level of seasonality. This will fit a lot faster than the Arima model, and may even work better in your situation.
  7. If 4/5/6 don't work, THEN start worrying about transfer functions. You have to crawl before you can walk.
  8. If you are planning to forecast into the future, you will first need to forecast your xreg variables. This is easy for seasonal dummies, but you'll have to think about how to make a good weather forecasts. Maybe use the median of historical data?

Here is an example of how I would approach this:

#Setup a fake time series
set.seed(1)
library(lubridate)
index <- ISOdatetime(2010,1,1,0,0,0)+1:8759*60*60
month <- month(index)
hour <- hour(index)
usage <- 1000+10*rnorm(length(index))-25*(month-6)^2-(hour-12)^2
usage <- ts(usage,frequency=24)

#Create monthly dummies.  Add other xvars to this matrix
xreg <- model.matrix(~as.factor(month))[,2:12]
colnames(xreg) <- c('Feb','Mar','Apr','May','Jun','Jul','Aug','Sep','Oct','Nov','Dec')

#Fit a model
library(forecast)
model <- Arima(usage, order=c(0,0,0), seasonal=list(order=c(1,0,0), period=24), xreg=xreg)
plot(usage)
lines(fitted(model),col=2)

#Benchmark against other models
model2 <- tslm(usage~as.factor(month)+as.factor(hour))
model3 <- tslm(usage~as.factor(month))
model4 <- rep(mean(usage),length(usage))

#Compare the 4 models
library(plyr) #for rbind.fill
ACC <- rbind.fill(  data.frame(t(accuracy(model))),
                    data.frame(t(accuracy(model2))),
                    data.frame(t(accuracy(model3))),
                    data.frame(t(accuracy(model4,usage)))
                )
ACC <- round(ACC,2)
ACC <- cbind(Type=c('Arima','LM1','Monthly Mean','Mean'),ACC)
ACC[order(ACC$MAE),]
$\endgroup$
5
  • $\begingroup$ What is the fitted() function. If I use that, I get way better results than with predict(model10, newxreg=regParams). $\endgroup$
    – utdiscant
    Nov 18, 2011 at 7:37
  • $\begingroup$ @utdiscant: predict() is used for forecasting, while fitted() returns the model fit over the historical period. If you want more specific help, you should post a reproducible example with some code. $\endgroup$
    – Zach
    Nov 18, 2011 at 13:18
  • $\begingroup$ @utdiscant: also, if you use dayy as an xreg, you run the risk of overfitting, as you only have 24 observations per day. You might get better forecasting results if you use month of the year. $\endgroup$
    – Zach
    Nov 18, 2011 at 13:22
  • $\begingroup$ @utdiscant: Furthermore, your time-based xregs need to be dummy variables. The way you have it modeled now is that you expect heat to linearly increase with hour of day, and then jump back down when the hour returns to 1. If you use dummy variables, each hour of the day will get it's own effect. Run through my example code, and pay careful attention to how I construct my xreg object. $\endgroup$
    – Zach
    Nov 18, 2011 at 13:24
  • 1
    $\begingroup$ One downside of the ARIMA functions in the stats and forecast packages is that they do not fit prober transfer functions. The documentation of the stats::arima function state the following: If an xreg term is included, a linear regression (with a constant term if include.mean is true and there is no differencing) is fitted with an ARMA model for the error term. So, if you actually need to fit transfer functions it seems like the TSA::arimax function is the way to go in R. $\endgroup$ May 3, 2018 at 13:37
11
$\begingroup$

I've been using R to do load forecasting for a while and I can suggest you to use forecast package and its invaluable functions (like auto.arima).

You can build an ARIMA model with the following command:

model = arima(y, order, xreg = exogenous_data)

with y your predictand (I suppose dayy), order the order of your model (considering seasonality) and exogenous_data your temperature, solar radiation, etc. The function auto.arima helps you to find the optimal model order. You can find a brief tutorial about `forecast' package here.

$\endgroup$
1
  • $\begingroup$ What is to be predicted is heat (the heat consumption of the house). $\endgroup$
    – utdiscant
    Nov 18, 2011 at 7:20
3
$\begingroup$

I personally don't understand transfer functions, but I think you got the xtransf and xreg reversed. At least in R's base arima it is xreg that contains your exogenous variables. It's my impression that a transfer function describes how (lagged data affects future values) rather than what.

I'd try using xreg for your exogenous variables, perhaps using arima if arimax demands a transfer function. The problem is that your model is daily, but your data has both daily and yearly seasonality, and I'm not sure right now if a first difference (the order=(*, 1, *)) will take care of that or not. (You certainly won't get magical year-round forecasts out of a model that only considers daily seasonality.)

P.S. What is the time that you use in your lm? Literal clock time or a 1-up observation number? I think you could get something by using a mixed-effect model (lmer in the lme4 package), though I haven't figured out whether doing that correctly accounts for the autocorrelation that will occur in a time series. If not accounted for, which anlm does not, you might get an interesting fit, but your concept of how precise your prediction is will be way too optimistic.

$\endgroup$
1
  • $\begingroup$ I have both the hour of the measurement, and the "day of the year" of the measurement. $\endgroup$
    – utdiscant
    Nov 16, 2011 at 21:50

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge that you have read and understand our privacy policy and code of conduct.

Not the answer you're looking for? Browse other questions tagged or ask your own question.