Mining Goes Digital – Mueller et al. (Eds)
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
463
Comparison between regression models and neural networks
applied to forecast geometallurgical variables
F.G.F. Niquini & J.F.C.L. Costa
Universidade Federal do Rio Grande do Sul, Porto Alegre, Rio Grande do Sul, Brazil
ABSTRACT: Geometallurgical variables are normally forecasted by regression models,
which usually shows good results. When non-linear variables need to be predicted, as the
metallurgical recovery, accurate and precise forecasts are not always achieved by this technique
application. Once neural networks use non-linear activation functions, the predictions for
non-linear variables tends to be very accurate. Aiming at comparing the forecasts achieved
when regression and neural networks are applied, data from a Brazilian zinc mine was
used, where tailings, zinc concentrate and lead concentrate mass recoveries are the variables
which compose mass balance. The zinc metallurgical recovery was calculated for zinc and
lead concentrates and tailings, aiming at knowing where and how much of the zinc content
is going during mineral processing. The results obtained showed the forecast superiority
achieved when neural networks were used, in addition to illustrate the possibility to generate
a unique model to predict simultaneously six dependent variables.
1 INTRODUCTION
Geometallurgy can be defined as the discipline which analyzes geological, mineralogical,
chemical, physical and environmental data collected from a mineral deposit, aiming at
explaining the factors affecting ore recovery at a processing plant. It can consider economical
parameters, such as ore recoveries and concentrate grades or environmental variables, as
water consumption, energy and tailings generation. More often feasibility studies embed
geometallurgical variables, having in mind the benefits achieved with this practice: a better
financial adherence in a project or in a mine under operation; a better use of the ore; and a
better management of environmental impacts caused by the mine or the plant.
Two of the most studied geometallurgical variables are the mass and the metallurgical
recovery in the concentrate. Forecast the mass and the metallurgical recovery of non-saleable
products, in other words, of tailings, is not a common practice in geometallurgical studies,
but it is relevant when the environmental aspects are taken under consideration. Determining tailings mass recovery bring information about how much materials is sent to the tailings
dam when a block is mined and processed, making possible to better study the tailings deposition characteristics. Tailings metallurgical recovery reports how much of the economical
metal was not sent to the concentrate by the beneficiation process and because of that was
discarded, leading to financial losses for the operation. Note that mapping masses of all
products (inputs and outputs) and their metallic content gives a better knowledge about the
ore that will be processed. These prediction models require a closed sum (mass input = mass
output). Fulfilling this premise is not a trivial and is considered as one of the challenges of
creating robust geometallurgical models.
Another challenge related to the treatment of geometallurgical data is that most are
non-additive (Richmond & Shaw, 2009). Carrasco et al. (2008) proved that mass recovery
is an additive variable but metallurgical recovery is not and because of that it needs to
be studied with prudence. According to Coward et al. (2009), additivity is the “property
that allows the mean of some variables to be calculated by a simple linear average”, or
© 2019 Taylor & Francis Group, London, ISBN 978-0-367-33604-2
463
Comparison between regression models and neural networks
applied to forecast geometallurgical variables
F.G.F. Niquini & J.F.C.L. Costa
Universidade Federal do Rio Grande do Sul, Porto Alegre, Rio Grande do Sul, Brazil
ABSTRACT: Geometallurgical variables are normally forecasted by regression models,
which usually shows good results. When non-linear variables need to be predicted, as the
metallurgical recovery, accurate and precise forecasts are not always achieved by this technique
application. Once neural networks use non-linear activation functions, the predictions for
non-linear variables tends to be very accurate. Aiming at comparing the forecasts achieved
when regression and neural networks are applied, data from a Brazilian zinc mine was
used, where tailings, zinc concentrate and lead concentrate mass recoveries are the variables
which compose mass balance. The zinc metallurgical recovery was calculated for zinc and
lead concentrates and tailings, aiming at knowing where and how much of the zinc content
is going during mineral processing. The results obtained showed the forecast superiority
achieved when neural networks were used, in addition to illustrate the possibility to generate
a unique model to predict simultaneously six dependent variables.
1 INTRODUCTION
Geometallurgy can be defined as the discipline which analyzes geological, mineralogical,
chemical, physical and environmental data collected from a mineral deposit, aiming at
explaining the factors affecting ore recovery at a processing plant. It can consider economical
parameters, such as ore recoveries and concentrate grades or environmental variables, as
water consumption, energy and tailings generation. More often feasibility studies embed
geometallurgical variables, having in mind the benefits achieved with this practice: a better
financial adherence in a project or in a mine under operation; a better use of the ore; and a
better management of environmental impacts caused by the mine or the plant.
Two of the most studied geometallurgical variables are the mass and the metallurgical
recovery in the concentrate. Forecast the mass and the metallurgical recovery of non-saleable
products, in other words, of tailings, is not a common practice in geometallurgical studies,
but it is relevant when the environmental aspects are taken under consideration. Determining tailings mass recovery bring information about how much materials is sent to the tailings
dam when a block is mined and processed, making possible to better study the tailings deposition characteristics. Tailings metallurgical recovery reports how much of the economical
metal was not sent to the concentrate by the beneficiation process and because of that was
discarded, leading to financial losses for the operation. Note that mapping masses of all
products (inputs and outputs) and their metallic content gives a better knowledge about the
ore that will be processed. These prediction models require a closed sum (mass input = mass
output). Fulfilling this premise is not a trivial and is considered as one of the challenges of
creating robust geometallurgical models.
Another challenge related to the treatment of geometallurgical data is that most are
non-additive (Richmond & Shaw, 2009). Carrasco et al. (2008) proved that mass recovery
is an additive variable but metallurgical recovery is not and because of that it needs to
be studied with prudence. According to Coward et al. (2009), additivity is the “property
that allows the mean of some variables to be calculated by a simple linear average”, or