Project case study · AI & Data Engineering
Machine learning for weather sensor forecasting
A predictive modelling project built around custom weather sensor data. My work combined AI development with team leadership, connecting data preparation and transmission with iterative model improvement.
By Berke Katrancı
- My role
- AI development and team leadership
- Context
- Practical project at TU Braunschweig
- Status
- Finished
The problem
The project required a predictive model for custom weather sensor readings. Working with the model also meant handling the surrounding data workflow: reading, cleaning, preparing and transmitting the sensor data.
This made the task a combination of data engineering and machine learning. A useful outcome depended on preparing the inputs as well as evaluating the model's predictions.
My contribution and approach
I worked on AI development and team leadership during the practical project at TU Braunschweig. The implementation used a Python stack covering data processing, machine learning and visualization.
- Read and prepare custom weather sensor data for predictive modelling.
- Work with Pandas and NumPy for data processing, and Scikit-learn and TensorFlow/Keras in the modelling stack.
- Iterate on the model and evaluate the resulting predictions, using Matplotlib for visualization.
- Combine technical development with coordination of the team's work.
Results and how to interpret them
The project summary reports an improvement in model accuracy from 22% to 67%. It also reports R² above 95% and mean squared error below 1 for some evaluated targets.
These figures describe different aspects of the evaluation and should not be treated as interchangeable scores. Their interpretation depends on the prediction target, the definition of accuracy, the target's scale and the evaluation split. They are project-specific results rather than a general weather forecasting benchmark.
Scope and limitations
The outcome demonstrates experience in sensor-data preparation, predictive modelling, evaluation and team leadership. It does not establish performance across other sensors, locations or forecasting horizons.
Comparing this work with another forecasting system would require the same target, data conditions and evaluation protocol.
Discuss this project
Get in touch to discuss the approach, my contribution or related AI and software work.
Contact Berke