Online forecasting of daily feed intake in lactating sows supported by offline time-series clustering, for precision livestock farming

According to precision livestock farming principles, it is essential apply feed intake forecasting processes real time feeding strategies in order improve the overall efficiency of chain. Considering lack a mechanistic model that predicts daily lactating sows, novel approach combining an online procedure with offline learning proposed. A database 39,090 lactations, from 6 different farms and containing first 20 records after farrowing, was used (1) identify consistent sets clusters trajectory curves offline, (2) test 3 predictive functions online. The homogeneity resulting assessed according Silhouette Calinski-Harabasz scores. quality Mean Error (ME), Root Square (RMSE). Time-series clustering k-Shape makes possible extract are scale-, shift- translate-invariant. best number obtained either global or at farm scale two. curve cluster characterized by mostly continuous increase over course lactation, second plateau starting about 10th day lactation. These identified very few studies available literature. When computed function specific curves, ME lactation ?0.08 kg/d, corresponding RMSE 1.06 kg/d. Though variability among sows period high, can be improved use curves. may on regular basis data directly similar practices. requires computing resources, could easily embedded smart feeder control systems as practical application for sows.