Inductive transfer applied to the prediction of ground water levels
LE3 .A278 2010
2010
Silver, Danny
Acadia University
Bachelor of Computer Science
Honours
Computer Science
This research uses Artificial Neural Networks with inductive transfer to predict the ground water level of the Greenwood monitoring well using weather data as input, as a proxy for the aquifer recharge rate. The flow rate of nearby streams (Stream Flow and Base Flow) are used as additional inputs or as secondary tasks. A basic single task learning model is constructed using only weather data for predictions. These results are compared to other single task and multiple task models through the use of mean absolute error to see if any improvement is provided over the basic single task model. The models use two days worth of weather data to predict the ground water level five days into the future. It is shown that models developed using five years worth of training data and inductive transfer perform as well as the base models constructed using ten years worth of training data. The multiple task models are also shown to perform as well as single task models with additional inputs. The neural networks show the greatest error when predicting the extremes of the ground water level over the test set. The use of additional sources of transfer, as well as the use of Stream Flow and Base Flow in combination to improve the accuracy of the models is suggested for future work.
The author retains copyright in this thesis. Any substantial copying or any other actions that exceed fair dealing or other exceptions in the Copyright Act require the permission of the author.
https://scholar.acadiau.ca/islandora/object/theses:731