The objective of this project is to systematically investigate how different modelling decisions influence prediction performance of a forecasting model for predicting the 2 m air temperature (t2m) and the 10 m wind components (u10 and v10) for the city of Cologne using ERA5-Land hourly time-series data. The dataset covers a five-year period from 2021 to 2026, providing hourly observations of the target variables.
For this purpose, a three-part ablation study was conducted to analyse the effects of:
- Single-task versus multi-task learning setups,
- the length of the historical input window, and
- the addition of input features.
The forecasting task consists of predicting all target variables with a fixed forecast horizon of 24 hours (T+24). As the primary forecasting model, a Long Short-Term Memory (LSTM) network was chosen. To provide reference, a separate seasonal persistence baselines were used for each target variable.
The directory "results" contains all final results of the ablation study as well as the baseline as parquet files that have been used to write the report. They can be read in to check what the final state of the project was.
Experiments were conducted on the following hardware:
- CPU: AMD Ryzen 5 3600 6-Core Processor, 3600 MHz
- GPU: NVIDIA GeForce RTX 5060 Ti, 16 GB VRAM
- RAM: 16 GB
- OS: Windows 10