Package 'ubair'

Description

Model agnostic function to calculate a number of common performance metrics on the reference time window. Uses the true data value and the predictions prediction for this calculation. The coverage is calculated from the columns value, prediction_lower and prediction_upper. Removes dates in the effect and buffer range as the model is not expected to be performing correctly for these times. The incorrectness is precisely what we are using for estimating the effect.

Usage

calc_performance_metrics(predictions, date_effect_start = NULL, buffer = 0)

Arguments

Value

Named vector with performance metrics of the model

Description

Helps with analyzing predictions by comparing them with the true values on a number of relevant summary statistics.

Usage

calc_summary_statistics(predictions, date_effect_start = NULL, buffer = 0)

Arguments

Value

data.frame of summary statistics with columns true and prediction

Description

Cleans a data table of environmental measurements by filtering for a specific station, removing duplicates, and optionally aggregating the data on a daily basis using the mean.

Usage

clean_data(env_data, station, aggregate_daily = FALSE)

Arguments

Details

Duplicate rows (by date, Komponente, and Station) are removed. A warning is issued if duplicates are found.

Value

A data.table:

• If aggregate_daily = TRUE: Contains columns for station, component, day, year, and the daily mean value of the measurements.

• If aggregate_daily = FALSE: Contains cleaned data with duplicates removed.

Examples

# Example data
env_data <- data.table::data.table(
  Station = c("DENW094", "DENW094", "DENW006", "DENW094"),
  Komponente = c("NO2", "O3", "NO2", "NO2"),
  Wert = c(45, 30, 50, 40),
  date = as.POSIXct(c(
    "2023-01-01 08:00:00", "2023-01-01 09:00:00",
    "2023-01-01 08:00:00", "2023-01-02 08:00:00"
  )),
  Komponente_txt = c(
    "Nitrogen Dioxide", "Ozone", "Nitrogen Dioxide", "Nitrogen Dioxide"
  )
)

# Clean data for StationA without aggregation
cleaned_data <- clean_data(env_data, station = "DENW094", aggregate_daily = FALSE)
print(cleaned_data)

Description

Copies the default params.yaml file, included with the package, to a specified destination directory. This is useful for initializing parameter files for custom edits.

Usage

copy_default_params(dest_dir)

Arguments

Details

The params.yaml file contains default model parameters for various configurations such as LightGBM, dynamic regression, and others. See the load_params()‘ documentation for an example of the file’s structure.

Value

Nothing is returned. A message is displayed upon successful copying.

Examples

copy_default_params(tempdir())

Description

Takes a list of train and application data as prepared by split_data_counterfactual() and removes a polynomial, exponential or cubic spline spline trend function. Trend is obtained only from train data. Use as part of preprocessing before training a model based on decision trees, i.e. random forest and lightgbm. For the other methods it may be helpful but they are generally able to deal with trends themselves. Therefore we recommend to try out different versions and guide decisisions using the model evaluation metrics from calc_performance_metrics().

Usage

detrend(split_data, mode = "linear", num_splines = 5, log_transform = FALSE)

Arguments

Details

Apply retrend_predictions() to predictions to return to the original data units.

Value

List of 3 elements. 2 dataframes: detrended train, apply and the trend function

Examples

data(mock_env_data)
split_data <- list(
  train = mock_env_data[1:80, ],
  apply = mock_env_data[81:100, ]
)
detrended_list <- detrend(split_data, mode = "linear")
detrended_train <- detrended_list$train
detrended_apply <- detrended_list$apply
trend <- detrended_list$model

Description

Calculates an estimate for the absolute and relative effect size of the external effect. The absolute effect is the difference between the model bias in the reference time and the effect time windows. The relative effect is the absolute effect divided by the mean true value in the reference window.

Usage

estimate_effect_size(df, date_effect_start, buffer = 0, verbose = FALSE)

Arguments

Details

Note: Since the bias is of the model is an average over predictions and true values, it is important, that the effect window is specified correctly. Imagine a scenario like a fire which strongly affects the outcome for one hour and is gone the next hour. If we use a two week effect window, the estimated effect will be 14*24=336 times smaller compared to using a 1-hour effect window. Generally, we advise against studying very short effects (single hour or single day). The variability of results will be too large to learn anything meaningful.

Value

A list with two numbers: Absolute and relative estimated effect size.

Description

Identifies unique meteorological components from the provided environmental data, filtering only those that match the predefined UBA naming conventions. These components include "GLO", "LDR", "RFE", "TMP", "WIG", "WIR", "WIND_U", and "WIND_V".

Usage

get_meteo_available(env_data)

Arguments

Value

A vector of available meteorological components.

Examples

# Example environmental data
env_data <- data.table::data.table(
  Komponente = c("TMP", "NO2", "GLO", "WIR"),
  Wert = c(25, 40, 300, 50),
  date = as.POSIXct(c(
    "2023-01-01 08:00:00", "2023-01-01 09:00:00",
    "2023-01-01 10:00:00", "2023-01-01 11:00:00"
  ))
)
# Get available meteorological components
meteo_components <- get_meteo_available(env_data)
print(meteo_components)

Description

Reads a YAML file containing model parameters, including station settings, variables, and configurations for various models. If no file path is provided, the function defaults to loading params.yaml from the package's extdata directory.

Usage

load_params(filepath = NULL)

Arguments

Details

The YAML file should define parameters in a structured format, such as:

target: 'NO2'

lightgbm:
  nrounds: 200
  eta: 0.03
  num_leaves: 32

dynamic_regression:
  ntrain: 8760

random_forest:
  num.trees: 300
  max.depth: 10

meteo_variables:
  - GLO
  - TMP

Value

A list containing the parameters loaded from the YAML file.

Examples

params <- load_params()

Description

This function loads data from CSV files in the specified directory. It supports two formats:

Usage

load_uba_data_from_dir(data_dir)

Arguments

Details

1. "inv": Files must contain the following columns:

   • Station, Komponente, Datum, Uhrzeit, Wert.

2. "24Spalten": Files must contain:

   • Station, Komponente, Datum, and columns Wert01, ..., Wert24.

File names should include "inv" or "24Spalten" to indicate their format. The function scans recursively for .csv files in subdirectories and combines the data into a single data.table in long format. Files that are not in the exected format will be ignored.

Value

A data.table containing the loaded data in long format. Returns an error if no valid files are found or the resulting dataset is empty.

Description

A small dataset of environmental variables created for testing and examples. This dataset includes hourly observations with random values for meteorological and temporal variables.

Usage

mock_env_data

Format

A data frame with 100 rows and 12 variables:

date

POSIXct. Date and time of the observation (hourly increments).

value

Numeric. Randomly generated target variable.

GLO

Numeric. Global radiation in W/m² (random values between 0 and 1000).

TMP

Numeric. Temperature in °C (random values between -10 and 35).

RFE

Numeric. Rainfall in mm (random values between 0 and 50).

WIG

Numeric. Wind speed in m/s (random values between 0 and 20).

WIR

Numeric. Wind direction in degrees (random values between 0 and 360).

LDR

Numeric. Longwave downward radiation in W/m² (random values between 0 and 500).

day_julian

Integer. Julian day of the year, ranging from 1 to 10.

weekday

Integer. Day of the week, ranging from 1 (Monday) to 7 (Sunday).

hour

Integer. Hour of the day, ranging from 0 to 23.

date_unix

Numeric. UNIX timestamp (seconds since 1970-01-01 00:00:00 UTC).

Source

Generated within the package for example purposes.

Examples

data(mock_env_data)
head(mock_env_data)

Description

Smooths the predictions using a rolling mean, prepares the data for plotting, and generates the counterfactual plot for the application window. Data before the red box are reference window, red box is buffer and values after black, dotted line are effect window.

Usage

plot_counterfactual(
  predictions,
  params,
  window_size = 14,
  date_effect_start = NULL,
  buffer = 0,
  plot_pred_interval = TRUE
)

Arguments

Details

The optional grey ribbon is a prediction interval for the hourly values. The interpretation for a 90% prediction interval (to be defined in alpha parameter of run_counterfactual()) is that 90% of the true hourly values (not the rolled means) lie within the grey band. This might be helpful for getting an idea of the variance of the data and predictions.

Value

A ggplot object with the counterfactual plot. Can be adjusted further, e.g. set limits for the y-axis for better visualisation.

Description

This function produces descriptive time-series plots with smoothing for the meteorological and potential target variables that were measured at a station.

Usage

plot_station_measurements(
  env_data,
  variables,
  years = NULL,
  smoothing_factor = 1
)

Arguments

Value

A ggplot object. This object contains:

• A time-series line plot for each variable in variables.

• Smoothed lines, with smoothing defined by smoothing_factor.

Examples

library(data.table)
env_data <- data.table(
  Station = "Station_1",
  Komponente = rep(c("TMP", "NO2"), length.out = 100),
  Wert = rnorm(100, mean = 20, sd = 5),
  date = rep(seq.POSIXt(as.POSIXct("2022-01-01"), , "hour", 50), each = 2),
  year = 2022,
  Komponente_txt = rep(c("Temperature", "NO2"), length.out = 100)
)
plot <- plot_station_measurements(env_data, variables = c("TMP", "NO2"))

Description

Prepares environmental data by filtering for relevant components, converting the data to a wide format, and adding temporal features. Should be called before split_data_counterfactual()

Usage

prepare_data_for_modelling(env_data, params)

Arguments

Value

A data.table in wide format, with columns: date, one column per component, and temporal features like date_unix, day_julian, weekday, and hour.

Examples

env_data <- data.table::data.table(
  Station = c("StationA", "StationA", "StationA"),
  Komponente = c("NO2", "TMP", "NO2"),
  Wert = c(50, 20, 40),
  date = as.POSIXct(c("2023-01-01 10:00:00", "2023-01-01 11:00:00", "2023-01-02 12:00:00"))
)
params <- list(meteo_variables = c("TMP"), target = "NO2")
prepared_data <- prepare_data_for_modelling(env_data, params)
print(prepared_data)

Description

This function rescales the predicted values (prediction, prediction_lower, prediction_upper). The scaling is reversed using the means and standard deviations that were saved from the training data. It is the inverse function to scale_data() and should be used only in combination.

Usage

rescale_predictions(scale_result, dt_predictions)

Arguments

Value

A data frame with the predictions and numeric columns rescaled back to their original scale.

Examples

data(mock_env_data)
scale_res <- scale_data(
  train_data = mock_env_data[1:80, ],
  apply_data = mock_env_data[81:100, ]
)
params <- load_params()
res <- run_lightgbm(
  train = scale_res$train, test = scale_res$apply,
  params$lightgbm, alpha = 0.9, calc_shaps = FALSE
)
dt_predictions <- res$dt_predictions
rescaled_predictions <- rescale_predictions(scale_res, dt_predictions)

Description

Takes a dataframe of predictions as returned by any of the 'run_model' functions and restores a trend which was previously removed via detrend(). This is necessary for the predictions and the true values to have the same units. The function is basically the inverse function to detrend() and should only be used in combination with it.

Usage

retrend_predictions(dt_predictions, trend, log_transform = FALSE)

Arguments

Value

Retrended dataframe with same structure as dt_predictions which is returned by any of the run_model() functions.

Examples

data(mock_env_data)
split_data <- list(
  train = mock_env_data[1:80, ],
  apply = mock_env_data[81:100, ]
)
params <- load_params()
detrended_list <- detrend(split_data,
  mode = "linear"
)
trend <- detrended_list$model
detrended_train <- detrended_list$train
detrended_apply <- detrended_list$apply
result <- run_lightgbm(
  train = detrended_train,
  test = detrended_apply,
  model_params = params$lightgbm,
  alpha = 0.9,
  calc_shaps = FALSE
)
retrended_predictions <- retrend_predictions(result$dt_predictions, trend)

Description

Chains detrending, training of a selected model, prediction and retrending together for ease of use. See documentation of individual functions for details.

Usage

run_counterfactual(
  split_data,
  params,
  detrending_function = "none",
  model_type = "rf",
  alpha = 0.9,
  log_transform = FALSE,
  calc_shaps = FALSE
)

Arguments

Value

Data frame of predictions, model and importance

Examples

data(mock_env_data)
split_data <- list(
  train = mock_env_data[1:80, ],
  apply = mock_env_data[81:100, ]
)
params <- load_params()
res <- run_counterfactual(split_data, params, detrending_function = "linear")
prediction <- res$retrended_predictions
random_forest_model <- res$model

Description

This function trains a dynamic regression model with fourier transformed temporal features and meteorological variables as external regressors on the specified training dataset and makes predictions on the test dataset in a counterfactual scenario. This is referred to as a dynamic regression model in Forecasting: Principles and Practise, Chapter 10 - Dynamic regression models

Usage

run_dynamic_regression(train, test, params, alpha, calc_shaps)

Arguments

Details

Note: Runs the dynamic regression model for individualised use with own data pipeline. Otherwise use run_counterfactual() to call this function.

Value

Data frame of predictions and model

Description

Trains a feedforward neural network (FNN) model on the specified training dataset and makes predictions on the test dataset in a counterfactual scenario. The model uses meteorological variables and sin/cosine-transformed features. Scales the data before training and rescales predictions, as the model does not converge with unscaled data.

Usage

run_fnn(train, test, params, calc_shaps)

Arguments

Details

This function provides flexibility for users with their own data pipelines or workflows. For a simplified pipeline, consider using run_counterfactual().

Experiment with hyperparameters such as learning_rate, batchsize, hidden_layers, and num_epochs to improve performance.

Warning: Using many or large hidden layers in combination with a high number of epochs can lead to long training times.

Value

A list with three elements:

dt_predictions

A data frame containing the test data along with the predicted values:

prediction

The predicted values from the FNN model.

prediction_lower

The same predicted values, as no quantile model is available yet for FNN.

prediction_upper

The same predicted values, as no quantile model is available yet for FNN.

model

The trained FNN model object from the deepnet::nn.train() function.

importance

SHAP importance values (if calc_shaps = TRUE). Otherwise, NULL.

Examples

data(mock_env_data)
params <- load_params()
res <- run_fnn(
  train = mock_env_data[1:80, ],
  test = mock_env_data[81:100, ], params,
  calc_shaps = FALSE
)

Description

This function trains a gradient boosting model (lightgbm) on the specified training dataset and makes predictions on the test dataset in a counterfactual scenario. The model uses meteorological variables and temporal features.

Usage

run_lightgbm(train, test, model_params, alpha, calc_shaps)

Arguments

Details

Note: Runs the gradient boosting model for individualised use with own data pipeline. Otherwise use run_counterfactual() to call this function.

Value

List with data frame of predictions and model

Examples

data(mock_env_data)
split_data <- list(
  train = mock_env_data[1:80, ],
  apply = mock_env_data[81:100, ]
)
params <- load_params()
variables <- c("day_julian", "weekday", "hour", params$meteo_variables)
res <- run_lightgbm(
  train = mock_env_data[1:80, ],
  test = mock_env_data[81:100, ], params$lightgbm, alpha = 0.9,
  calc_shaps = FALSE
)
prediction <- res$dt_predictions
model <- res$model

Description

This function trains a random forest model (ranger) on the specified training dataset and makes predictions on the test dataset in a counterfactual scenario. The model uses meteorological variables and temporal features.

Usage

run_rf(train, test, model_params, alpha, calc_shaps)

Arguments

Details

Note: Runs the random forest model for individualised use with own data pipeline. Otherwise use run_counterfactual() to call this function.

Value

List with data frame of predictions and model

Description

This dataset contains environmental measurements from the Leipzig Mitte station provided by the Sächsisches Landesamt für Umwelt, Landwirtschaft und Geologie (LfULG). Alterations in the data: Codes for incorrect values have been removed.

Usage

sample_data_DESN025

Format

A data table with the following columns:

Station

Station identifier where the data was collected.

Komponente

The environmental component being measured (e.g., temperature, NO2).

Wert

The measured value of the component.

date

The timestamp for the observation, formatted as a Date-Time object in the format "YYYY-MM-DD HH:MM:SS" (e.g., "2010-01-01 07:00:00").

Komponente_txt

A textual description or label for the component.

The dataset is structured in a long format and is prepared for further transformation into a wide format for modelling.

Details

The dataset is licensed under the "Data Licence Germany – attribution – version 2.0 (DL-DE-BY-2.0)". (1) Any use will be permitted provided it fulfils the requirements of this "Data licence Germany – attribution – Version 2.0".

The data and meta-data provided may, for commercial and non-commercial use, in particular

• be copied, printed, presented, altered, processed and transmitted to third parties;

• be merged with own data and with the data of others and be combined to form new and independent datasets;

• be integrated in internal and external business processes, products and applications in public and non-public electronic networks.

(2) The user must ensure that the source note contains the following information:

• the name of the provider,

• the annotation "Data licence Germany – attribution – Version 2.0" or "dl-de/by-2-0" referring to the licence text available at www.govdata.de/dl-de/by-2-0, and

• a reference to the dataset (URI).

This applies only if the entity keeping the data provides the pieces of information 1-3 for the source note.

(3) Changes, editing, new designs or other amendments must be marked as such in the source note.

For more information on the license, visit https://www.govdata.de/dl-de/by-2-0.

Source

Sächsisches Landesamt für Umwelt, Landwirtschaft und Geologie (LfULG).

Examples

data(sample_data_DESN025)
params <- load_params()
dt_prepared <- prepare_data_for_modelling(sample_data_DESN025, params)

Description

This function standardizes numeric columns of the train_data and applies the same scaling (mean and standard deviation) to the corresponding columns in apply_data. It returns the standardized data along with the scaling parameters (means and standard deviations). This is particularly important for neural network approaches as they tend to be numerically unstable and deteriorate otherwise.

Usage

scale_data(train_data, apply_data)

Arguments

Value

A list containing the following elements:

Examples

data(mock_env_data)
detrended_list <- list(
  train = mock_env_data[1:80, ],
  apply = mock_env_data[81:100, ]
)
scale_result <- scale_data(
  train_data = detrended_list$train,
  apply_data = detrended_list$apply
)
scaled_train <- scale_result$train
scaled_apply <- scale_result$apply

Description

Splits prepared data into training and application datasets based on specified date ranges for a business-as-usual scenario. Data before application_start and after application_end is used as training data, while data within the date range is used for application.

Usage

split_data_counterfactual(dt_prepared, application_start, application_end)

Arguments

Value

A list with two elements:

train

Data outside the application period.

apply

Data within the application period.

Examples

dt_prepared <- data.table::data.table(
  date = as.Date(c("2023-01-01", "2023-01-05", "2023-01-10")),
  value = c(50, 60, 70)
)
result <- split_data_counterfactual(
  dt_prepared,
  application_start = as.Date("2023-01-03"),
  application_end = as.Date("2023-01-08")
)
print(result$train)
print(result$apply)