Comparing multiple linear regression and support vector machine models for predicting electricity consumption on pasture based dairy farms

摘要

This study compared multiple linear regression (MLR) and support vector machine (SVM) models for predicting the annual electricity consumption of 20 Irish dairy farms, at a farm and catchment (combined) level. Model input variables were constrained tomilk production, stock numbers, infrastructural equipment and managerial procedures to allow predictions to take place on a large scale without the use of specialized equipment. The SVM model has previously been shown to reduce the prediction error of monthly electricity consumption by 54% compared to the MLR model. Results found both the MLR and SVM models predicted annual electricity consumption per farm to within 20%. However, the error of the SVM model reduced to 9% when two farms with the greatestmonthly prediction errors were removed. With herd sizes in excess of 190 dairy cows, these two farms were found to represent less than 3.3% of the Irish dairy farm demographic. Regarding the ability of each model to predict catchment level electricity consumption, the MLR model prediction resulted in an error of 4% while the SVM prediction resulted in an error of 9%. The improved accuracy of the MLR model when predicting electricity consumption at a catchment level was respective of a greater balance between the under and over prediction of electricity consumption across the 20 dairy farms. These models may be utilized to provide key decision support information to both dairy farmers and policy makers, or as a tool for conducting macro scale environmental analysis, for marketing Irish dairy products abroad.