Cliffs Natural Resources Pty Ltd (CNR) operates iron ore mines in the Koolyanobbing region of Western Australia. Ore is mined from three locations, separated by many kilometres. The ore is stockpiled at these locations, according to an in-house ore classification system (based on grade and source) and then trucked to the crushing and screening plant at Koolyanobbing. Lump and fines products are railed to Esperance for ship loading and export to Asian customers. CNR prides itself on the relatively low inter-shipment grade variability of the products._x000D_
The Koolyanobbing crusher is fed using a daily blend plan, generated to maintain lump and fines product grades within acceptable tolerance ranges around targets. Achieving low variability requires predicting lump and fines grades as accurately as practical from the estimated head grades of the ROM ores potentially going into the blend. The grade prediction model may be either a direct prediction of crushed lump and fines grades and lump percentage, or be split into two stages: the bias between blasthole estimated head grade and crusher head grade and then the lump-fines algorithm for splitting the head grade between lump and fines products. The lump-fines algorithm comprises the percentage of lump and the difference between the lump and fines grades._x000D_
We describe a weighted least squares (WLS) regression model for predicting crusher grades and lump proportion from the estimated head grade for Fe, P, SiO2, Al2O3, Mn and S. The method is applied to CNR production data, where the regression model explains approximately 60 per cent of the variance in the crushed ore grade, for both lump and fines. A further small but significant improvement in prediction can be achieved by including the ore classification data in the model._x000D_
The regression errors exhibit strong positive serial correlation, indicating trends in grade error across multiple blend records. To compensate for the longer-term error, an exponentially smoothed model was developed and applied to the daily grade blend errors. This gave an increase in the longer-term variance explained and therefore an improvement in grade prediction.