KINLI: Time Series Forecasting for Monitoring Poultry Health in Complex Pen Environments

TL;DR

This paper explores using machine learning to predict turkey health from noisy farm sensor data, finding that some models balance accuracy and ease of use for real-world farming.

Contribution

The paper introduces a novel real-world turkey farm dataset and evaluates diverse forecasting models for usability in low-tech agricultural settings.

Findings

Deep learning models like PatchTST achieve best forecasting accuracy on noisy poultry data.

Simpler models offer reliable predictions with minimal setup, suitable for low-tech environments.

Large language models show promise but face issues with computational inefficiency and pattern deterioration.

Abstract

The paper presents the KINLI project, which applies machine learning and deep learning techniques to time series forecasting for monitoring turkey health in poultry farms. Using a real-world dataset from turkey barns—characterized by noisy, incomplete, and irregular sensor data (e.g., food intake, water intake, environmental factors)—the study evaluates a wide range of forecasting models. These include statistical approaches (ARIMA and Prophet), classical machine learning models (XGBoost and LSTM), transformer-based architectures (Informer, Autoformer and FEDformer), and emerging time series foundation models (PatchTST, TimeLLM, and TimesFM). The authors compare models in terms of forecasting accuracy and practical usability, especially in settings with limited technical expertise. Results show that while deep learning models such as PatchTST perform best overall, simpler models can still offer reliable predictions with minimal setup. Large language models (LLMs) show potential but suffer from computational inefficiency and “pattern deterioration”. Ultimately, the study concludes that robust yet easy-to-use forecasting tools are essential for real-world agricultural applications, where automation and low maintenance are critical.

We analyze how to perform accurate time series forecasting for monitoring poultry health in a complex pen environment. To this end, we make use of a novel dataset consisting of a collection of real-world sensor data in the housing of turkeys. The dataset comprises features such as food intake, water intake, and various environmental values, which come with high variance, sensor defects, and unreliable timestamps. In this paper, we investigate different state-of-the-art forecasting algorithms to predict different features, as well as a variety of deep learning models such as different transformer models and time series foundational models. We evaluate both their forecasting accuracy as well as the efforts required to run the models in the first place. Our findings show that some of these aforementioned algorithms are able to produce satisfactory forecasting results on this highly challenging dataset while still remaining easy to use, which is key in a tech-distant industry such as poultry farming.

Figures