In maxikellerbauer/stAirPol: Spatio-temporal Modelling of Air Pollution Data

knitr::opts_chunk$set(
  collapse = TRUE,
  fig.width = 8,
  fig.height = 6,
  comment = "#>"
)

require(stAirPol)
require(spTimer)
require(data.table)

Now we want to check the temporal predictions of the GPP Model. Therefore, we specify a training set and specify the data up to the 2017-12-15 as the training set for the model. We want to make a prediction for the next 9 days

# Read the data -----------------------------------------------------------

Read the data

data("muc_airPol_p2")
data <- clean_model_data(muc_airPol_p2)
data.fit <- data[timestamp < quantile(timestamp, 0.5)]
data.predict <- data[timestamp >= quantile(timestamp, 0.5) &
                       timestamp < quantile(timestamp, 0.8)]

We specify a trainings set for crossvalidation

training_set <- get_test_and_training_set(data, sampel_size = 0.75,
                                          random.seed = 220292)

For speeding up the computation we only use the GPP model

formula = value ~ humi + temp + rainhist + windhist +
  trafficvol + log(sensor_age)
priors.gpp <- spT.priors(model = "GPP", inv.var.prior = Gamm(a = 2, b = 1),
                         beta.prior = Norm(0, 10^4))
spatial.decay = spT.decay(distribution = Gamm(a = 2, b = 1), tuning = 0.25)
scale.transform = "SQRT"
cov.fnc = "exponential"
model.gpp.random <- fit_sp_model(data = data.fit,
                                 formula = formula,
                                 model = 'GPP',
                                 priors = priors.gpp,
                                 cov.fnc = cov.fnc,
                                 knots_count = 20,
                                 knots_method = 'random',
                                 knots_plot = TRUE,
                                 knots_seed = 220292,
                                 training_set = training_set,
                                 scale.transform = scale.transform,
                                 spatial.decay = spatial.decay)
pred <- predict(model.gpp.random, newdata = data.fit,
                    training_set = training_set)
forecast <- predict(model.gpp.random, newdata = data.predict,
                    type = 'temporal', training_set = training_set,
                    foreStep = length(unique(data.predict$timestamp)))

evaluate the predictive performance

evaluate_prediction(forecast)
evaluate_prediction(pred)

visulatise with a plot

gridExtra::grid.arrange(grobs = list(
  plot(pred, time_dimension = TRUE) + ggtitle('spatio prediction') +
    theme(legend.position = 'bottom'),
  plot(forecast, time_dimension = TRUE) + ggtitle('spatiotemporal prediction') +
    theme(legend.position = 'bottom')
), ncol = 2)





# forcasting intraday data ------------------------------------------------

forcasting intraday data

The same is possible for dataset which contains 8h timeintervalls

data("muc_airPol_p2_8h")
data <- clean_model_data(muc_airPol_p2_8h)
data.fit <- data[timestamp < quantile(timestamp, 0.5)]
data.predict <- data[timestamp >= quantile(timestamp, 0.5) &
                       timestamp < quantile(timestamp, 0.75)]

We add an intercept for every distinct hour

formula = value ~ humi + temp + rainhist + windhist +
  trafficvol + log(sensor_age) + stAirPol::hour(timestamp) - 1

priors.gpp <- spT.priors(model = "GPP", inv.var.prior = Gamm(a = 2, b = 1),
                         beta.prior = Norm(0, 10^4))
spatial.decay = spT.decay(distribution = Gamm(a = 2, b = 1), tuning = 0.25)
scale.transform = "SQRT"
cov.fnc = "exponential"
model.gpp.random <- fit_sp_model(data = data.fit,
                                 formula = formula,
                                 model = 'GPP',
                                 priors = priors.gpp,
                                 cov.fnc = cov.fnc,
                                 knots_count = 20,
                                 knots_method = 'random',
                                 knots_plot = TRUE,
                                 knots_seed = 220292,
                                 training_set = training_set,
                                 scale.transform = scale.transform,
                                 spatial.decay = spatial.decay)

spatial prediction

pred <- predict(model.gpp.random, newdata = data.fit,
                training_set = training_set)

spatial and temporal prediction

forecast <- predict(model.gpp.random, newdata = data.predict,
                    type = 'temporal', training_set = training_set,
                    foreStep = length(unique(data.predict$timestamp)))

evaluate_prediction(forecast)
evaluate_prediction(pred)

gridExtra::grid.arrange(grobs = list(
  plot(pred, time_dimension = TRUE) + ggtitle('spatial  prediction') +
    theme(legend.position = 'bottom'),
  plot(forecast, time_dimension = TRUE) + ggtitle('spatial and temporal prediction') +
    theme(legend.position = 'nonbottom')
  ), ncol = 2)