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