The final project with small omissions is presented in notebook.ipynb file.
The project with more thorough explanations, justifications and frighteningly excessive figures is stored in notebook_extended.ipynb.
The task is provided by the Zyfra company. The company develops solutions for the efficient operation of industrial enterprises.
Task:
The provided datasets contain parameters of gold flotation process. The data is indexed by the date and time the information was received (date feature in datasets). Data is collected every hour and is already ordered chronologically. Parameters adjacent in time are often similar.
Prepare a prototype machine learning model.
The model should predict the recovery rate of gold from gold ore.
The model will help optimize production so as not to launch an enterprise with unprofitable characteristics.
Technological process:
When the mined ore undergoes primary processing, a crushed mixture is obtained. It is sent for flotation (enrichment) and two-stage purification:
Feature names look like this: [stage].[parameter_type].[parameter_name]
Example: rougher.input.feed_ag
Possible values for block [stage]:
rougher: flotation stage
primary_cleaner: primary cleaning stage
secondary_cleaner: secondary cleaning stage
final: final characteristics
Possible values for block [parameter_type]:
input: raw material parameters
output: product parameters
state: parameters characterizing stage
calculation: calculated characteristics
Possible values for block [parameter_name]:
feed_{component}: feedstock component concentration, %
tail_{component}: tail component concentration, %
concentrate_{component}: component concentration, %
feed_rate: input feed rate
feed_size: input granule size
floatbank_..._air: air volume at process stage
floatbank_..._level: water level at process stage
floatbank_..._{reagent}: amount of added flotation reagent at process stage
Flotation reagents:
- Xanthate: collector reagent;
- Sulphate (in this production, sodium sulfide);
- Depressant (sodium silicate).
- Recovery. Formula for recovery calculation:
where:
recovery is the recovery percentage(rougher.output.recovery);
C is the share of gold in the concentrate after flotation/cleaning (rougher.output.concentrate_au);
F is the share of gold in the raw material/concentrate before flotation/cleaning(rougher.input.feed_au);
T is the share of gold in final tailings after flotation/cleaning (rougher.output.tail_au).
- SMAPE (Symmetric Mean Absolute Percentage Error) is used as the main metric for models evaluation in this task:
where:
yi is the target recovery percentage;
ŷi is the predicted recovery percentage;
N is the number of observations in the dataset.
SMAPE is calculated separately for model predictions on rougher.output.recovery and final.output.recovery, and then two metrics are weighed to produce the final result:
