This vignette explains how to setup and use an onlineforecast model. This takes offset in the example of building heat load forecasting and assumes that the data is setup correctly, as explained in setup-data vignette. The R code is available here. More information on onlineforecasting.org.
Start by loading the package:
Set the scoreperiod
as a logical vector with same length
as t
. It controls which points are included in score
calculations in functions for optimization etc. It must be set.
Use it to exclude a burn-in period of one week:
# Print the first time point
D$t[1]
## [1] "2010-12-15 01:00:00 UTC"
# Set the score period
D$scoreperiod <- in_range("2010-12-22", D$t)
# Plot to see it
plot(D$t, D$scoreperiod, xlab="Time", ylab="Scoreperiod")
Other periods, which should be excluded from score calculations, can
simply also be set to FALSE
. E.g.:
# Exclude other points example
scoreperiod2 <- D$scoreperiod
scoreperiod2[in_range("2010-12-30",D$t,"2011-01-02")] <- FALSE
would exclude the days around new year (must of course be set in
D$scoreperiod
, not in scoreperiod2
to have an
effect).
A simple onlineforecast model can be setup by:
# Generate new object (R6 class)
model <- forecastmodel$new()
# Set the model output
model$output = "heatload"
# Inputs (transformation step)
model$add_inputs(Ta = "Ta",
mu = "one()")
# Regression step parameters
model$add_regprm("rls_prm(lambda=0.9)")
# Optimization bounds for parameters
model$add_prmbounds(lambda = c(0.9, 0.99, 0.9999))
# Set the horizons for which the model will be fitted
model$kseq <- 1:36
Let’s go through the steps of setting up the model.
First a new forecastmodel object is generated and the model output is
set (per default it is "y"
):
The output is simply the variable name from D
we want to
forecast.
The model inputs are defined by:
So this is really where the structure of the model is specified. The
inputs are given a name (Ta
and mu
), which
each are set as an R expression (given as a string). The expressions
defines the transformation step: they will each be
evaluated in an environment with a given data.list
. This
means that the variables from the data can be used in the expressions
(e.g. Ta
is in D
) - below in Input transformations we will detail
this evaluation.
Next step for setting up the model is to set the parameters for the
regression step by providing an expression, which
returns the regression parameter values. In the present case we will use
the Recursive Least Squares (RLS) when regressing, and we need to set
the forgetting factor lambda
by:
The expression is just of a function, which returns a list - in this
case with the value of lambda
(see onlineforecasting). The
result of it begin evaluated is kept in:
# The evaluation happens with
eval(parse(text="rls_prm(lambda=0.9)"))
## $lambda
## [1] 0.9
# and the result is stored in
model$regprm
## $lambda
## [1] 0.9
We will tune the parameters, for this model it’s only the forgetting factor, so we set the parameter bounds (lower, init, upper) for it by:
Finally, we set the horizons for which to fit:
The horizons to fit for is actually not directly related to the model, but rather the fitting of the model. In principle, it would be more “clean” if the model, data and fit was kept separate, however for recursive fitting this becomes un-feasible.
We have set up the model and can now tune the lambda
with the rls_optim()
, which is a wrapper for the
optim()
function:
# Call the optim() wrapper
rls_optim(model, D, kseq = c(3,18))
## ----------------
## lambda
## 0.99
## k3 k18 sum
## 0.831 0.826 1.657
## ----------------
## lambda
## 0.991
## k3 k18 sum
## 0.831 0.827 1.658
## ----------------
## lambda
## 0.989
## k3 k18 sum
## 0.831 0.826 1.657
## ...output cropped
Note, how it only calculated a score for the 3 and 18 steps horizons
- since we gave it kseq
as an argument, which then
overwrites model$kseq
for the optimization only. The
parameters could be optimized separately for each horizon, for example
it is often such that for the first horizons a very low forgetting
factor is optimal (e.g. 0.9). Currently, however, the parameters can
only be optimized together. By optimizing for a short (3 steps) and a
long horizon (18 steps), we obtain a balance - using less computations
compared to optimizing on all horizons.
The optimization converge and the tuned parameter was inserted:
Now we can fit with the optimized lambda
on the horizons
in model$kseq
over the entire period:
# Fit for all on entire period in D
fit1 <- rls_fit(model$prm, model, D)
## ----------------
## lambda
## 0.985
See the summary of the fit:
# See the summary of the fit
summary(fit1)
##
## Output: heatload
## Inputs: Ta = Ta
## mu = one()
##
## Regression parameters:
## lambda = 0.985
##
## Scoreperiod: 1656 observations are included.
##
## RLS coeffients summary stats (cannot be used for significance tests):
## mean sd min max
## Ta -0.14 0.044 -0.26 0.19
## mu 5.20 0.280 4.50 6.80
##
## RMSE:
## k1 k2 k3 k4 k5 k6 k7 k8 k9 k10 k11 k12 k13 k14 k15 k16 k17 k18 k19 k20 k21 k22 k23
## 0.82 0.83 0.83 0.83 0.83 0.84 0.84 0.84 0.84 0.84 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.82 0.83 0.83 0.83
## k24 k25 k26 k27 k28 k29 k30 k31 k32 k33 k34 k35 k36
## 0.84 0.85 0.86 0.86 0.86 0.86 0.86 0.86 0.86 0.86 0.86 0.86 0.86
See ?summary.rls_fit
for details.
Plot the forecasts (Yhat
adheres to the forecast matrix
format and in plot_ts()
the forecasts are lagged
k
steps to be aligned with the observations):
# Put the forecasts in D
D$Yhat1 <- fit1$Yhat
# Plot them for selected horizons
plot_ts(D, c("^heatload$|^Y"), kseq = c(1,6,18,36))
We clearly see the burn-in period, where the forecasts vary a lot,
Plot a forecast for a particular time point and forward in time:
# Select a point
i <- 996-48
# and kseq steps ahead
iseq <- i+model$kseq
# The observations ahead in time
plot(D$t[iseq], D$heatload[iseq], type = "b", xlab = "t", ylab = "y")
title(main=pst("Forecast available at ",D$t[i]))
# The forecasts
lines(D$t[iseq], D$Yhat1[i, ], type = "b", col = 2)
legend("topright", c("Observations",pst("Predictions (",min(model$kseq)," to ",max(model$kseq)," steps ahead)")), lty = 1, col = 1:2)
The inputs can be transformations of the variables in the data,
i.e. D
in this example. The function one()
generate a forecast matrix of 1 for the needed horizons. It cannot be
called directly:
(the code above was not executed)
however we can see the result of the evaluation by:
# Evaluate input expressions
datatr <- model$transform_data(D)
# See what came out
summary.default(datatr)
## Length Class Mode
## Ta 36 data.frame list
## mu 36 data.frame list
# In particular for the mu = "one()"
head(datatr$mu)
## k1 k2 k3 k4 k5 k6 k7 k8 k9 k10 k11 k12 k13 k14 k15 k16 k17 k18 k19 k20 k21 k22 k23 k24 k25 k26 k27 k28 k29 k30
## 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
## 2 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
## 3 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
## 4 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
## 5 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
## 6 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
## k31 k32 k33 k34 k35 k36
## 1 1 1 1 1 1 1
## 2 1 1 1 1 1 1
## 3 1 1 1 1 1 1
## 4 1 1 1 1 1 1
## 5 1 1 1 1 1 1
## 6 1 1 1 1 1 1
If we wanted to debug we could:
# Set the function to debug (uncomment the line)
#debug(one)
# Run the input transformation now and it will stop in one()
datatr <- model$transform_data(D)
# Set to undebug
#undebug(one)
(the code above was not executed).
Let’s extend the model by adding a low-pass filter transformation of the ambient temperature forecasts. We could just update the input by:
but let’s just repeat the whole model definition for clarification - including the new transformation:
# Define a new model with low-pass filtering of the Ta input
model <- forecastmodel$new()
model$output = "heatload"
model$add_inputs(Ta = "lp(Ta, a1=0.9)",
mu = "one()")
model$add_regprm("rls_prm(lambda=0.99)")
model$add_prmbounds(Ta__a1 = c(0.5, 0.9, 0.9999),
lambda = c(0.9, 0.99, 0.9999))
model$kseq <- c(3,18)
Note how also a new set of parameter bounds were added in
add_prmbounds()
following a neat little syntax:
Ta__a1
indicates that the first appearance of
a1
in the Ta
input expression, will be changed
in the optimization.
We can see the parameter bounds with:
To inspect the result of low-pass filtering:
# Low-pass filter Ta (with a1=0.9 as defined above)
datatr <- model$transform_data(D)
# Actually, lp() can be called directly (although two warnings are thrown)
Talp <- lp(D$Ta, a1=0.99)
## Warning in state_getval(initval = yInit): In state_getval() the object of class input was not found in the parent
## environments. The initval was returned.
## Warning in state_setval(val[nrow(X), ]): In state_setval() the object of class input was not found in the parent
## environments, so the state value could not be updated.
and to see the result we could:
# Plot the Ta$k1 forecasts
plot(D$t, D$Ta$k1, type="l")
# Add the filtered with a1=0.9
lines(D$t, datatr$Ta[ ,"k1"], col=2)
# Add the filtered with a1=0.99
lines(D$t, Talp[ ,"k1"], col=3)
hence with a low-pass coefficient a1=0.99
, which is very
high (max is 1), the Ta forecast is really smoothed, which models a
system with a time constant (i.e. slow dynamics, e.g. well insulated and
building with lots of concrete).
There are quite a few functions available for input transformations:
one()
generates an matrix of ones (for including an
intercept).fs()
generate Fourier series for modelling harmonic
functions.bspline()
wraps the bs()
function for
generating base splines.pbspline()
wraps the pbs()
function for
generating periodic base splines.AR()
generates auto-regressive model inputs.and they can even be combined, see more details in onlineforecasting and in
their help description, e.g. ?fs
.
Tuning the two parameters: the low-pass filter coefficient
a1
and the forgetting factor lambda
, can now
be done:
# Optimize the parameters
model$prm <- rls_optim(model, D)$par
## ----------------
## Ta__a1 lambda
## 0.90 0.99
## k3 k18 sum
## 0.823 0.817 1.640
## ----------------
## Ta__a1 lambda
## 0.901 0.990
## k3 k18 sum
## 0.823 0.817 1.640
## ----------------
## Ta__a1 lambda
## 0.899 0.990
## k3 k18 sum
## 0.823 0.817 1.639
## ...output cropped
Plot the forecasts (Yhat adheres to the forecast matrix format and in
plot_ts
the forecasts are lagged k
steps to
sync with the observations)
# Fit for all horizons
model$kseq <- 1:36
# Fit with RLS
fit2 <- rls_fit(model$prm, model, D)
## ----------------
## Ta__a1 lambda
## 0.850 0.991
# Take the forecasts
D$Yhat2 <- fit2$Yhat
# Plot all
plot_ts(D, c("^heatload$|^Y"), kseq = c(1,18))
We can see the summary:
summary(fit2)
##
## Output: heatload
## Inputs: Ta = lp(Ta, a1=0.85)
## mu = one()
##
## Regression parameters:
## lambda = 0.991
##
## Scoreperiod: 1656 observations are included.
##
## RLS coeffients summary stats (cannot be used for significance tests):
## mean sd min max
## Ta -0.17 0.036 -0.29 0.016
## mu 5.20 0.180 4.80 6.000
##
## RMSE:
## k1 k2 k3 k4 k5 k6 k7 k8 k9 k10 k11 k12 k13 k14 k15 k16 k17 k18 k19 k20 k21 k22 k23
## 0.81 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82 0.82
## k24 k25 k26 k27 k28 k29 k30 k31 k32 k33 k34 k35 k36
## 0.82 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83 0.83
but more interesting is it to see if an improvement was achieved with the low-pass filtering, so calculate the RMSE for both models:
# Calculate the score
RMSE1 <- summary(fit1, printit=FALSE)$scoreval
RMSE2 <- summary(fit2, printit=FALSE)$scoreval
Now, this is calculated for the points included in the
scoreperiod
, so it’s important to make sure that exactly
the same values are forecasted. A check can be done by:
# Check that all NAs in the scoreperiod are at the same positions
all(is.na(fit1$Yhat[fit1$data$scoreperiod, ]) == is.na(fit2$Yhat[fit2$data$scoreperiod, ]))
## [1] TRUE
Finally, plot the RMSE for the two models:
# Plot the score for the two models
plot(RMSE1, xlab="Horizon k", ylab="RMSE", type="b", ylim=range(RMSE1,RMSE2))
lines(RMSE2, type="b", col=2)
legend("topleft", c("Input: Ta","Input: Low-pass Ta"), lty=1, col=1:2)
We can see, that we obtained improvements. Around 3-4% for the longer horizons.
For more on evaluation, see the vignette forecast-evaluation.
See more on how to extend this model even further in building heat load forecasting.
Often we need to have the time of day as an input to a forecastmodel:
make_tday(D$t, kseq=1:3)
## k1 k2 k3
## 1 2 3 4
## 2 3 4 5
## 3 4 5 6
## 4 5 6 7
## 5 6 7 8
## 6 7 8 9
## 7 8 9 10
## 8 9 10 11
## 9 10 11 12
## 10 11 12 13
## 11 12 13 14
## 12 13 14 15
## 13 14 15 16
## 14 15 16 17
## 15 16 17 18
## 16 17 18 19
## 17 18 19 20
## 18 19 20 21
## 19 20 21 22
## 20 21 22 23
## 21 22 23 0
## 22 23 0 1
## 23 0 1 2
## 24 1 2 3
## 25 2 3 4
## 26 3 4 5
## 27 4 5 6
## ...output cropped
So we can use it like this:
See the help ?make_tday
for more details.
Working with time consuming calculations caching can be very valuable. The optimization results can be cached by providing a path to a directory, by setting the argument ‘cachedir’ to e.g. “cache”. See the vignette nice-tricks for an example with code.