85 Integrating Mechanistic Models with AI for Precision Feeding of Sows

Abstract

Abstract Conventional feeding for sows is usually based on the average herd’s nutrient requirements. Thus, sows can be under- or over- fed leading to extra feed costs and environmental losses. New technologies, as sensors and AI, bring opportunities to measure and integrate individual variability into nutrient requirements estimations. The objective is therefore to go towards precision feeding (PF) combining on-farm data as input for a dynamic nutritional model with smart feeders to provide individual and daily-adjusted rations. As a first step, a mechanistic model (InraPorc) was upgraded and applied to databases to calculate daily nutrient requirements at the individual scale for sows. For lactating sows, it highlighted that milk production and appetite influenced the amount and composition of the optimal ration to be fed to each sow. For gestating sows, it showed that parity, gestation stage, and activity level influenced nutrient requirements. The second step was to develop algorithms to predict the parameters of interest defined in the first step and not measured daily on-farm. For lactating sows, feed intake and litter weight at weaning (as proxy for milk production) were accurately predicted using supervised methods: respectively, clustering k-shape and a linear regression. For gestating sows, an algorithm is being developed to identify individual activities via video recordings. The third step is to test on farm the decision support systems (DSS) composed of the models and algorithms. An interface allows the link between the DSS and the feeders, and another allows the farmers to enter observational data. During on-farm trials, nitrogen and phosphorus excretions as well as feed costs were reduced for sows fed with PF compared to sows fed a conventional diet. To conclude, AI allows mechanistic models and algorithms to be integrated on farm for sows for an on real-time individual adjustment of the nutrient supply.