A bi-directional cross-channel RNN model for time-series forecasting of dairy production

Predicting milk production in dairy cattle is essential for precision livestock management, a goal that can be achieved by analyzing historical cow data, including health status, milk quality, and seasonal effects. This challenge is framed as a multivariate time-series forecasting problem, requiring the effective capture of temporal dependencies and interrelationships among dairy-related variables (features). Existing time-series models often struggle to adequately represent these intricate dynamics. To address this, we propose a recurrent neural network (RNN) architecture that leverages Gated Recurrent Units (GRUs) applied bidirectionally along the channel dimension to model these interactions efficiently. The proposed model is further enhanced with feed-forward layers to implicitly capture temporal dependencies. Our approach delivers superior or competitive performance, particularly in terms of various error metrics, compared to state-of-the-art methods when predicting cumulative milk income across different lactation periods. Additionally, it exhibits relatively lower time complexity than recently proposed Transformer and convolution-based models.