This project focuses on using an LSTM model to predict key chemical production outputs based on timeseries data from multiple reactors. The dataset consists of hourly resolution data collected over a year, totaling 37,728 samples from six reactors.
Each sample contains multiple sensor readings. For example, Reactor 2 includes:
- CB 2
- Erdgas 2
- Konst.Stufe 2
- Perlwasser 2
- Regelstufe 2
- Sorte 2
- V-Luft 2
- VL Temp 2
- Fuelöl 2
- Makeöl 2
- Makeöl Temperatur 2
- Makeöl Ventil 2
- CCT 2
- CTD 2
- FCC 2
- SCT 2
- C 2, H 2, N 2, O 2, S
- CO2 R2, SO2 R2
Additionally, some values are common across all six reactors:
- KD|Dampfmenge, KD|Restgasmenge, KD|NOx, KD|Rauchgasmenge, KD|SO2
- KE|Dampfmenge, KE|Restgasmenge, KE|NOx, KE|Rauchgasmenge, KE|SO2
The goal is to predict the following output variables:
- CB (the product)
- SO₂
- CO₂
- Dampfmenge (steam)
- Rauchgasmenge (tail gas, all reactors combined)
The input consists of all available sensor values except for the shared reactor-independent values:
- KD|Dampfmenge, KD|Restgasmenge, KD|NOx, KD|Rauchgasmenge, KD|SO2
- KE|Dampfmenge, KE|Restgasmenge, KE|NOx, KE|Rauchgasmenge, KE|SO2
Developed an LSTM model to predict key chemical production outputs from multivariate timeseries data (37,728 samples, six reactors), optimizing industrial process forecasting.