Abstract

To effectively manage pollution associated with mine waste rock, it is crucial to reliably predict the quantity and quality of drainage generated from waste rock piles. It is extremely difficult to make such predictions using traditional reactive transport models, because waste rock piles are highly heterogeneous in physical, chemical, and biological properties. Artificial intelligence (AI) approaches are reported to have advantages over traditional reactive transport models, because they are primarily algorithm-driven and tend to be easier and quicker to develop. However, the current AI-enabled models tend to oversimplify the contaminated drainage release process due to their reliance on limited inputs, such as weather monitoring data or a few measurable chemical indicators. This oversimplification is one of the reasons why AI approaches are often criticized for its black-box and hard-to-interpret nature.

This study aims to improve the previous AI-enabled models based on neural networks for prediction of drainage quantity, which refines local weather monitoring data to better estimate the net infiltration into waste rock piles. The weather monitoring data including total precipitation, temperature, relative humidity, and wind speed, obtained from Environment Canada, were transformed into rainfall, snowmelt, and evaporation that are directly correlated to net infiltration. The transformation was carried out in four steps: 1) combine weather data from different monitoring stations in one database representing the local weather conditions of the mine waste rock pile; 2) calculate the actual ambient temperature at the elevation of the waste rock pile; 3) simulate the rainfall and snowmelt using the climate model, which partitions total precipitation into rainfall and snowfall based on the calculated ambient temperature. The climate model also simulated the amount of snowmelt by considering snow accumulation-melting process; 4) Simulate evaporation by a simplified version of the standardized Penman equation. AI-enabled models based on neural networks were developed and trained with rainfall, snowmelt, evaporation, and waste rock volume as the inputs and daily drainage flow rate as the output. They performed better than the previously published model that used raw weather monitoring data as the input.