Machine-learning algorithms for predicting on-farm direct water and electricity consumption on pasture based dairy farms

This study analysed the performance of a range of machine learning algorithms when applied to the prediction of electricity and on-farm direct water consumption on Irish dairy farms. Electricity and water consumption data were attained through the utilisation of a remote monitoring system installed on a study sample of 58 pasture-based, commercial Irish dairy farms between 2014 and 2016. In total, 15 and 20 dairy farm variables were analysed for their predictive power of monthly electricity and water consumption, respectively. These variables were related to milk production, stock numbers, infrastructural equipment, managerial procedures and environmental conditions. A CART decision tree algorithm, a random forest ensemble algorithm, an artificial neural network and a support vector machine algorithm were used to predict both water and electricity consumption. The methodology employed backward sequential variable selection to exclude variables, which added little predictive power. It also applied hyper-parameter tuning with nested cross-validation for calculating the prediction accuracy for each model on unseen data (data not utilised for model development). Electricity consumption was predicted to within 12% (relative prediction error (RPE)) using a support vector machine, while the random forest predicted water consumption to within 38%. Overall, the developed machine-learning models improved the RPE of electricity and water consumption by 54% and 23%, respectively, when compared to results previously obtained using a multiple linear regression approach. Further analysis found that during the January, February, November and December period, the support vector machine overpredicted electricity consumption by 4% (mean percentage error (MPE)) and water consumption by 21% (MPE), on average. However, overprediction was greatly reduced during the March - October period with overprediction of electricity consumption reduced to 1% while the overprediction of water consumption reduced to 8%. This was attributed to a phase shift between farms, where some farms produce milk all year round, some dry off earlier/later than others and some farms begin milking earlier/later resulting in an increased the coefficient of variance of milk production making it more difficult to model electricity and water accurately. Concurrently, large negative correlations were calculated between the number of dairy cows and absolute prediction error for electricity and water, respectively, suggesting improvements in electricity and water prediction accuracy may be achieved with increasing dairy cow numbers. The developed machine learning models may be utilised to provide key decision support information to both dairy farmers and policy makers or as a tool for conducting macro scale environmental analysis.