The repository motivates LSTM-based univariate multistep forecasting as an imputation technique for missing sequential data. The temporal and between-feature correlation in the absence, which often characterizes the missing data of sensors, undermines the ability of traditional imputation techniques.

For this purpose, the imputation accuracy of two LSTM-based architectures is compared with linear and spline imputation upon a dataset of meteorological data from Berlin in 2023. The data is provided by the Deutscher Wetterdienst (German Weather Service).

Structure of the repository:

1. datasets: hourly measurements of X meteorological characteristic for Berlin 2023, w/ and without imputation
2. utilities: required functions
3. EMDA.ipynb: Exploratory missing data analysis
4. imputation.ipynb: imputation based upon EDA results
5. LSTM_results.ipynb: comparison of different imputation techniques

Datasource:

The usage of the meteorological data is regulated by the "Creative Commons BY 4.0" (CC BY 4.0) and covers the following data sets:

• Hourly station observations of 2 m air temperature and humidity for Germany, Version v24.03
• Hourly station observations of precipitation for Germany, Version v24.03
• Hourly station observations of pressure for Germany, Version v24.03
• Hourly station observations of solar incoming (total/diffuse) and longwave downward radiation for Germany, Version v24.03
• Hourly mean value from station observations of wind speed and wind direction for Germany, Version v24.03

Ackknowledgement:

The architecture for multi-step forecast, applied for missing data imputation, has been altered from the tutorial by Brownlee (2020).

The original owner of the data and code used in this thesis retains ownership of the data and code.

Brownlee, J. (2020). Multi-step time series forecasting with machine learning for electricity usage.