Search

Inaccuracies in spatial modelling of coal properties can impact coal resource and reserve estimates. Geostatistical methods which account for compositional data have previously been recommended as a preferred method for modelling of coal properties. These methods can quantify uncertainty, can use auxiliary information to improve estimates, and ensure that compositional data (such as coal proximate analysis) is modelled in a mathematically consistent way. However, geostatistical methods have drawbacks in that have rigid statistical assumptions about the distribution and stationarity of the target variable and require variogram interpretation which can be onerous when then number of domains and target variables are numerous. Due to these drawback, inverse distance weighting (a simple deterministic method) remains the most popular method for modelling coal properties in Australian resource estimates. To address the drawbacks of geostatistical methods, a machine learning approach based on the quantile regression forest algorithm is proposed as an alternative method to spatially model coal properties. This newly proposed method, accounts for spatial arrangement of data, requires minimal pre-processing steps, can quantify uncertainty and can easily incorporate auxiliary information of various types. To evaluate the performance of this method, the accuracy of predictions of coal proximate analysis properties, and coal density are compared to the predictions of inverse distance weighting and regression kriging. All methods incorporate isometric log ratio transform and back-transform of data in order to account for the compositional nature of coal proximate analysis. Data from an active coal mine in the Bowen Basin, Queensland Australia was used as basis of the results. A unique feature of the mine site in this basin is the presence of intrusion which impacts coal quality properties to various degrees. Using evaluation metrics from leave-one-out cross-validation, this paper demonstrates that quantile regression forest has higher accuracy, lower bias and higher precision than inverses distance weighting across all coal properties. The paper also shows that results of the new method are very similar or better than regression kriging. It is also demonstrated that the prediction maps produced by quantile regression forest could be used to more accurately map out coal impacted by intrusion compared to inverse distance weighting, and that inverse distance weighting overestimates the impact of intrusion and intrusion extent. A drawback of the method compared to inverse distance weighting, is that it is more computationally demanding, less intuitive and is currently not available in existing geological packages. [ABSTRACT FROM AUTHOR]

Identification of the extent of heat altered coals is important for coal mining and resource estimation because alteration directly affects key economic properties such as ash content, volatile matter, specific energy, sulphur, free swell index and beneficiation characteristics. The most reliable method to identify altered coal is through geochemical analysis of drill core samples; however this method is time consuming and costly. Cheaper alternative methods to identify altered coal is by macroscopic observation of non-core drill cuttings, but this is less reliable than core. Geophysical wireline log data is also used to identify altered coal, but with variable success. Over the past decades, Machine Learning methods have proven popular to automatically classify a variety of lithotypes from geophysical log data. However rarely has there been focus on predicting altered coal. In this paper we apply the most recent machine learning methods which include gradient boosted machines, random forests and artificial neural networks to automatically predict altered and non-altered lithotypes using geophysical log data. We use a massive data set comprising of 1230 samples from 263 boreholes from a highly intruded deposit in the Bowen Basin, Eastern Australia. To do this, we calculate each sample's distance from intrusion and predict their relative density from geophysical log inputs including gamma ray, caliper and compensated density. We then train our machine learning methods on 80% of the data to predict alteration class using the calculated distance to intrusion, predicted relative density, and geophysical log gamma ray, as primary inputs. We evaluated each of these machine learning methods on the remaining 20% of the data to determine the best performing model. Finally, using the best performing model we further split the altered and non-altered classification into lithotypes using a decision tree based on geological knowledge of the case study area. The results indicate that of the machine learning algorithms the random forest produced the best results with only 11 misclassifications across the entire data set of 1230 samples which represents <1% error. • Prediction of heat affected coals with a massive dataset and machine learning methods in the Bowen Basin. • Prediction of relative density and classification of altered and non-altered coals. • Lithology group definition based on a decision tree with <1% error. [ABSTRACT FROM AUTHOR]