Date of Award
Doctor of Philosophy
Process, Energy & Transport Engineering
Dr. Michael D. Murphy
Milk production forecasting in the dairy industry has been an independent research topic since the early 20th century. The accurate prediction of milk yield can benefit both the processor (creameries) and the producer (dairy farmer) through developing short-term production schedules, planning long-term road maps, facilitating trade and investment in the dairy industry, improving business operations, optimising the existing infrastructure of the dairy industry, and reducing operating costs. Additionally, due to the innate characteristics of the milk production process, the accurate prediction of milk yield has been a challenging issue in the dairy industry. With the abolishment of EU milk quotas in 2015, the business requirements of milk production forecasting from the dairy industry has become increasingly important. However, to date, most of the existing modelling techniques are data dependent and each case study utilises specific data based on unique conditions. Consequently, it is difficult to compare the prediction performance of each candidate model for forecasting milk as both the data types and origins are independent from study to study. This body of work proposes an integrated forecasting framework XIX concentrating on milk production forecasting using heterogeneous input data combinations based on animal data, milk production, weather variables and other possible records that can be applied to milk yield forecasting on either the herd level or the individual cow level. The first objective of this study concerned the development of the Milk Production Forecast Optimisation System (MPFOS). The MPFOS focused on data processing, automated model configuration and optimisation, and multiple model comparisons at a global level. Multiple categories of milk yield prediction models were chosen in the model library of the MPFOS. Separated databases existed for functionality and scalability in the MPFOS, including the milk yield database, the cow description database and the weather database. With the built-in filter in MPFOS, appropriate sample herds and individual cows were filtered and processed as input datasets for different customised model simulation scenarios. The MPFOS was designed for the purpose of comparing the effectiveness of multiple milk yield prediction models and for assessing the suitability of multiple data input configurations and sources. For forecasting milk yield at the herd level, the MPFOS automatically generated the optimal configuration for each of the tested milk production forecast models and benchmarked their performance over a short (10-day), medium (30-day) and long (365-day) term prediction horizon. The MPFOS found the most accurate model for the short (the NARX model), medium and long (the surface fitting model) terms with R2 values equalling 0.98, 0.97 and 0.97 for the short, medium and long term, respectively. The statistical analysis demonstrated the effectiveness of the MPFOS as a model configuration and comparison tool. For forecasting milk yield at the individual cow level, the MPFOS was utilised to conduct two exploratory analyses on the effectiveness of adding exogenous (parity and meteorological) data to the milk production modelling XX procedure. The MPFOS evaluated the most accurate model based on the prediction horizon length and on the number of input parameters such as 1) historical parity weighting trends and 2) the utilisation of meteorological parameters. As the exploratory analysis into utilising parity data in the modelling process showed, despite varying results between two cow groups, cow parity weighting profiles had a substantial effect on the success rate of the treatments. Removal of the first lactation and applying static parity weight were shown to be the two most successful input treatments. These results highlight the importance of examining the accuracy of milk prediction models and model training strategies across multiple time horizons. While the exploratory analysis into meteorological data in the modelling process demonstrated that based on statistical analysis results, 1) the introduction of sunshine hours, precipitation and soil temperature data resulted in a minor improvement in the prediction accuracy of the models over the short, medium and long-term forecast horizons. 2) Sunshine hours was shown to have the largest impact on milk production forecast accuracy with an improvement observed in 60% and 70% of all predictions (for all test cows from both groups). However, the overall improvement in accuracy was small with a maximum forecast error reduction of 4.3%. Thus, the utilisation of meteorological parameters in milk production forecasting did not have a substantial impact on the overall forecast accuracy. One possible reason for this may be due to modern management techniques employed on dairy farms, reducing the impact of weather variation on feed intake and lessening the direct effect on milk production yield. The MPFOS architecture developed in this study showed to be an efficient and capable system for automatic milk production data pre-processing, model configuration and comparison of model categories over varying prediction horizons. The MPFOS has proven to be a XXI comprehensive and convenient architecture, which can perform calculations for milk yield prediction at either herd level or individual cow level, and automatically generate the output results and analysis. The MPFOS may be a useful tool for conducting exploratory analyses of incorporating other exogenous data types. In addition, the MPFOS can be extended (addition or removal of models in the model library) and modularised. Therefore the MPFOS will be a useful benchmark platform and integrated solution for future model comparisons.
Zhang, Fan, "An Optimal Milk Production Model Selection and Configuration System for Dairy Cows" (2017). PhDs [online].
Available at: https://sword.cit.ie/engdiss/3