Forecasting silicon
in blast-furnace hot metal.

The problem

Silicon content is an important indicator in blast-furnace operation, but laboratory measurements arrive far less often than sensor data. That creates a difficult forecasting problem: dense time series, sparse labels, and a process whose behaviour changes over time.

My work in the project

My submitted project reports cover the data and modelling workflow: preparing sensor time series, investigating label availability, and comparing forecasting approaches. The reports document the experimental work and its limitations.

Making the comparison meaningful

The second phase compared 11 architectures at two temporal resolutions, giving 22 experiment configurations. The implementation used TensorFlow/Keras. Chronological 80/10/10 train, validation, and test splits preserved the ordering of the process data.

Standardization was fitted on training data. This matters because letting future information influence preprocessing can make a forecasting result appear stronger than it is.

A result, with its context

This is a validation result for a particular dataset and configuration. It is not a demonstrated production error rate, and it should not be compared directly with normalized-target metrics from the first phase.

What I took from it

In industrial machine learning, the data construction and evaluation protocol can matter as much as the network architecture. This project strengthened my approach to time alignment, sparse labels, temporal validation, and reporting the limits of a result.

The reports describe an offline research proof of concept. Online plant validation and operational deployment remain future work. Industrial source data is not published here.