Hybrid Machine Learning Framework for Herd-Level Cattle Growth Pattern and Weight Gain Forecasting in Grazing-Based Production Systems

This paper addresses the challenge of cattle growth forecasting in commercial grazing systems, where commercial grazing systems yield irregular livestock observations, which challenge cattle growth forecasting. The authors note that most livestock forecasting studies apply sequential forecasting models using regularly sampled observations collected under controlled production systems, while commercial grazing systems yield irregular observations because cattle voluntarily interact with weighing platforms. This irregularity—including uneven temporal spacing and inconsistent visitation frequency—introduces behavioural variability into animal-level measurements. The study notes that existing recurrent forecasting architectures often require stable temporal sequences and high observation density, assumptions that remain difficult to maintain in the context of voluntary livestock visitation behaviour associated with automated SOO weighing systems. The paper states that forecasting instability under sparse sensing conditions remains under-examined in commercial grazing environments.

The stated objectives were to: "(i) develop an operational herd-level forecasting framework using structured datasets derived from irregular livestock observations; (ii) compare hybrid ML architecture families in terms of predictive accuracy and computational efficiency; (iii) evaluate forecasting robustness across multiple operational prediction horizons and observation-density conditions; and (iv) identify dominant biological and environmental predictors influencing herd-level cattle growth dynamics."

The study was conducted at the Charles Sturt University (CSU) Farm, Wagga Wagga, Australia, between 2022 and 2024 under temperate grazing conditions typical of southeastern Australian beef systems. Three data streams were integrated: animal demographic records, automated live-weight observations, and climatic measurements. Live-weight observations were collected using the Optiweigh automated SOO weighing system, and environmental variables included daily rainfall and temperature measurements obtained from the nearest Australian Bureau of Meteorology (BOM) station. The dataset included 960 Angus and Limousin cattle monitored under commercial grazing conditions.

Raw live-weight observations were filtered such that consecutive fluctuations exceeding ±15 kg within 24 hours were excluded to reduce transient sensor instability and weighing artefacts. Weekly mean live weight was calculated per animal. Missing values associated with short temporal gaps of seven days or less were imputed using linear interpolation, while longer temporal gaps were imputed using k-nearest-neighbour regression. After quality control, a total of 686 cattle passed quality control and filtering for modelling analysis, generating 21,406 weekly observations across 2022 to 2024.

The paper characterises the irregular sensing dataset with metrics including a mean observation interval of 3.8 days, median weekly observations per animal of 2.9, missing weekly intervals of 18.6%, and an observation interval coefficient of variation of 42.3%.

The study aggregated animal-level data into herd-level trajectories because operational grazing enterprises commonly implement decisions at the herd scale. Weekly herd-level live weight was calculated as the mean of individual animal weights for each week. The paper reports that herd-level aggregation reduced stochastic fluctuations associated with irregular weighing frequency and improved the representation of seasonal growth dynamics in grazing systems. Specifically, variance decreased by 18.7%, and the CV decreased from 11.4% to 7.9% following aggregation. A comparison of individual-animal versus herd-level forecasting showed RMSE improving from 24.882 kg to 21.319 kg, and MAE from 18.471 kg to 15.462 kg following aggregation.

The methodology transformed irregular livestock observations into supervised learning datasets. Each weekly livestock observation was transformed into a temporally indexed forecasting record containing biological, environmental, and historical predictors. The forecasting record for animal i at week t was represented as:

Xi,t = [Wi,t−1, Ai,t, Rt, Rt−1, Rt−2, Tt, Tt−1, Tt−2, Gi, Si]

where "Wi,t−1 represents historical live weight, Ai,t represents animal age, Rt represents weekly rainfall, Rt−1 and Rt−2 represent one-week and two-week rainfall lag variables, Tt represents weekly temperature, Tt−1 and Tt−2 represent lagged temperature variables, Gi represents the genotype category, and Si represents the sex category. The forecasting target was defined as Yi,t+h = Wi,t+h, with forecasting horizons ranged from 1 to 24 weeks. Lagged environmental variables represented delayed climatic responses associated with pasture growth and forage availability."

Three baseline models were implemented: ARIMA, LSTM, and GRU. ARIMA used order selection using Akaike Information Criterion (AIC) optimisation, while LSTM and GRU used two hidden recurrent layers with dropout regularisation and Adam optimisation algorithm.

Four hybrid architecture families were evaluated. Residual hybrid architectures generated an initial prediction using a base learner. NNs were subsequently used to model residual prediction errors to refine final outputs, with models including Gradient Boosting (GB), Random Forest (RF), Support Vector Regression (SVR), and Extreme Gradient Boosting (XGBoost) combined with residual NNs. Stacked hybrid architectures combined predictions from multiple base learners and supplied these outputs to a NN meta-learner, including combinations of RF, GB, SVR, K-Nearest Neighbour, Decision Tree (DT), and Linear Regression (LR) models integrated with NN meta-learners. Cascade hybrid architectures sequentially propagated predictions from upstream models as additional features to downstream learners, including GB, RF, and SVR integrated with downstream NNs. Ensemble-assisted neural architectures generated intermediate predictions by applying ensemble techniques to multiple base learners, which were subsequently incorporated into neural networks to refine final predictions.

The study used animal-level hold-out validation with internal cross-validation applied exclusively within the training subset. Training and testing partitions were constructed using unique electronic identification (EID) codes assigned to individual cattle to minimise information leakage from repeated biological measurements. The final partition contained 686 unique animals, including 480 allocated to training and 206 allocated to testing, with the training subset containing 14,979 weekly records and the testing subset containing 6,427 weekly records. The training and testing partitions had comparable distributions, with mean live weights of 297.90 kg and 296.04 kg respectively.

The comparative analysis showed that baseline sequential models demonstrated lower predictive performance under heterogeneous livestock observation conditions compared with hybrid architectures. ARIMA produced the lowest predictive agreement among the evaluated approaches with R2 of 0.701, RMSE of 31.844 kg, and MAE of 24.905 kg. LSTM achieved R2 of 0.781, RMSE of 27.563 kg, and MAE of 20.771 kg. GRU achieved R2 of 0.794, RMSE of 26.918 kg, and MAE of 20.115 kg. The paper notes that sequential DL models remained sensitive to incomplete temporal observations despite interpolation-based sequence reconstruction.

The cascade GB→RF→NN architecture achieved the highest testing performance with R2 of 0.889, RMSE of 21.319 kg, and MAE of 15.462 kg. The RMSE improvement versus ARIMA was 33.1%. Statistical testing identified significant performance differences between the lowest-performing stacked architecture and the cascade GB→RF→NN architecture (p = 0.012, Cohen's d = 0.61), while performance differences between cascade GB→RF→NN and residual RF→NN architectures were not statistically significant (p = 0.184, Cohen's d = 0.14).

Predictive performance declined progressively as the forecasting horizon increased. For the cascade GB→RF→NN architecture, the 1-week horizon achieved R2 of 0.889, RMSE of 20.440 kg, and MAE of 14.633 kg; the 2-week horizon achieved R2 of 0.872, RMSE of 21.119 kg, and MAE of 14.959 kg; the 4-week horizon achieved R2 of 0.854, RMSE of 21.773 kg, and MAE of 15.791 kg; the 8-week horizon achieved R2 of 0.826, RMSE of 22.148 kg, and MAE of 16.014 kg; the 12-week horizon achieved R2 of 0.793, RMSE of 23.184 kg, and MAE of 16.921 kg; and the 24-week horizon achieved R2 of 0.742, RMSE of 23.426 kg, and MAE of 16.560 kg. The paper notes that MAE at the 24-week horizon remained slightly lower than the 12-week horizon despite increased RMSE because larger prediction outliers had a stronger effect on RMSE than on MAE at extended forecasting intervals.

Controlled observation-density scenarios were constructed: high-frequency (>80% valid weekly observations), medium-frequency (50–80%), and sparse-frequency (<50%). Hybrid forecasting architectures maintained comparatively stable forecasting performance, with less deterioration than sequential recurrent models as observation sparsity increased. The RMSE degradation from high to sparse frequency was: ARIMA 30.7%, LSTM 43.0%, GRU 40.7%, RF→NN 23.0%, and GB→RF→NN 20.5%. The cascade GB→RF→NN framework maintained the lowest RMSE degradation at 20.5%.

Feature-importance analysis identified animal age, rainfall, and temperature as dominant predictors influencing cattle live-weight forecasting. The relative importance values were: animal age 0.848, rainfall (t) 0.057, temperature (t) 0.020, rainfall (t-1) 0.019, temperature (t-1) 0.015, animal genotype 0.012, and animal sex 0.009. The paper states that animal age contributed the largest proportion of predictive variance and that the dominant contribution of animal age indicates a strong association between the physiological development stage and forecasting performance. The authors caution that the one-week rainfall lag should be interpreted as a short-term climatic signal, not a complete forage biomass response and that rainfall effects on forage availability were interpreted cautiously because the present study did not directly measure pasture biomass.

The cascade GB→RF→NN architecture had 63,929 parameters, a training time of 9.053 seconds, and inference time of 0.054 ms/sample, with moderate relative computational demand. For comparison, the SVR+KNN+RF→NN stacked architecture had 302,021 parameters, 39.177 seconds training time, and 0.990 ms/sample inference time, rated high computational demand. The paper notes that residual and cascade architectures maintained favourable predictive performance together with moderate computational demand and that reduced inference latency may improve compatibility with edge-based livestock monitoring devices and automated grazing-management platforms operating under continuous sensing conditions.

The analysis of genotype and sex categories found that Angus male cattle exhibited the highest mean weekly live-weight gain among the evaluated groups at 5.214 kg/week, followed by Limousin males at 4.872 kg/week, Limousin females at 4.221 kg/week, and Angus females at 3.104 kg/week. The paper notes that estimated daily gain equivalents ranged from 0.44 to 0.75 kg day−1 across evaluated groups and that "a mixed-effects ANOVA identified significant differences in growth between genotype and sex categories (p < 0.05)."

The paper presents a simplified economic interpretation framework using representative supplementary feeding costs between AUD 0.35 and 0.50 per kg dry matter (DM). Illustrative scenarios suggested that a 5% forecast error reduction could yield estimated annual economic impacts of AUD 6,388–9,125, a 10% reduction AUD 12,775–18,250, and a 15% reduction AUD 19,163–27,375 for a 500-head enterprise. The authors emphasise that economic values represent illustrative management scenarios and the framework does not represent enterprise-scale economic optimisation modelling.

The paper concludes that the study developed a hybrid ML framework for herd-level cattle forecasting under irregular sensing conditions in grazing systems and that hybrid architectures outperformed conventional sequential forecasting approaches across independent testing datasets and observation-density scenarios. The authors state that the framework may support decisions on feed allocation, grazing management, and livestock marketing under variable climatic conditions and that a structured forecasting representation improved the compatibility between automated livestock sensing observations and operational forecasting analytics across heterogeneous grazing environments. Future work should focus on external validation across multiple enterprises and integrate pasture biomass and remote-sensing variables within operational forecasting systems.

Here are specific improvements to AI forecasting systems based on the paper:

1. Use cascade hybrid architectures (GB→RF→NN) instead of sequential models for irregular time-series forecasting.

Build the core model as a pipeline where Gradient Boosting generates initial predictions, Random Forest refines them, and a neural network learns the final residual correction. This improves RMSE by 33.1% over ARIMA (21.319 kg vs. 31.844 kg) and achieves R2 of 0.889, outperforming pure LSTM (R2=0.781) and GRU (R2=0.794).

Resulting capability: The AI can accurately forecast herd-level weight up to 24 weeks ahead even with sparse, irregular sensor data.

2. Aggregate animal-level irregular observations into herd-level trajectories before modeling.

Instead of feeding noisy individual-animal sequences into the AI, compute weekly herd-level mean live weight. This reduces variance by 18.7% and improves RMSE from 24.882 kg to 21.319 kg.

Resulting capability: The AI produces more stable and reliable forecasts for operational decisions made at the herd scale, which is how grazing enterprises actually manage cattle.

3. Convert irregular observations into structured feature vectors with biological and environmental predictors.

Replace sequence-reconstruction (e.g., interpolation to fixed timesteps) with a supervised record containing: previous weight, animal age, current and lagged rainfall (t, t−1, t−2), current and lagged temperature, genotype, and sex. This avoids the instability of LSTM/GRU when temporal spacing is uneven.

Resulting capability: The AI leverages domain-relevant drivers (age = 84.8% importance, rainfall, temperature) and handles missing weeks up to 7 days via linear interpolation and longer gaps via k-NN regression without catastrophic failure.

4. Design the AI to remain robust under low observation density.

Test and train the model explicitly on sparse-frequency conditions (<50% valid weekly observations). The cascade GB→RF→NN degraded only 20.5% in RMSE from high to sparse density, compared to 43.0% for LSTM and 40.7% for GRU.

Resulting capability: The AI can maintain practical forecasting accuracy in commercial grazing environments where cattle voluntarily visit weigh platforms infrequently and irregularly.

5. Implement animal-level hold-out validation with EID-based partitioning to prevent data leakage.

Split data by unique electronic identification codes, not randomly by rows. Train on 480 animals, test on 206 animals, ensuring the same animal never appears in both sets.

Resulting capability: The AI's performance metrics reflect true generalization to new animals, avoiding inflated accuracy and enabling trusted deployment across different herds.

6. Optimize for computational efficiency to enable edge-device deployment.

Use the cascade GB→RF→NN architecture with 63,929 parameters, 9.05 s training time, and 0.054 ms inference per sample—far lighter than stacked architectures (302,021 parameters, 39.18 s, 0.99 ms/sample).

Resulting capability: The AI can run in real-time on edge-based livestock monitoring devices and automated grazing-management platforms, enabling continuous, low-latency forecasting without cloud dependence.

7. Incorporate tiered forecasting horizons with known degradation patterns.

Structure the system to report confidence intervals or adjustments as horizon increases. In the paper, R2 drops from 0.889 (1-week) to 0.742 (24-week) while RMSE grows from 20.44 kg to 23.43 kg.

Resulting capability: The AI can alert users to reduced certainty for long-range decisions (e.g., 12–24 weeks) and recommend re-forecasting as new data arrives.

8. Use feature importance analysis to guide model inputs.

Rank predictors automatically; animal age dominates (0.848), then rainfall (0.057), temperature (0.020), with genotype and sex minor.

Resulting capability: The AI can prioritize data collection efforts—e.g., ensuring reliable age records and weather feeds—and avoid over-reliance on noisy or irrelevant variables, while flagging caution for unmeasured pasture biomass.

9. Support scenario-based economic decision-making.

Integrate forecast outputs with a supplementary feed cost model (AUD 0.35–0.50 per kg DM). For a 500-head enterprise, a 10% RMSE reduction translates to AUD 12,775–18,250 annual impact.

Resulting capability: The AI can quantify the financial value of improved forecasting accuracy and help managers optimize feed allocation, grazing rotation, and marketing timing under variable climate conditions.

Commercial grazing systems yield irregular livestock observations, which challenge cattle growth forecasting. This study developed a hybrid machine learning framework for herd level cattle weight forecasting using automated sensing observations collected between 2022 and 2024 in southeastern Australia. Weekly live weight observations, demographic variables, and lagged environmental predictors were integrated into structured forecasting datasets. Herd level forecasting trajectories were generated through temporal aggregation of animal level predictions. Four hybrid architecture families were evaluated, including residual, stacked, cascade, and ensemble assisted frameworks. ARIMA, LSTM, and GRU models were used as comparative baselines. Independent testing demonstrated strong predictive agreement across multiple forecasting horizons. The cascade GB to RF to NN architecture achieved the best performance, with a test R squared of 0.889, RMSE of 21.319 kg, and MAE of 15.462 kg. Hybrid architectures maintained greater robustness than recurrent sequential models under sparse observation conditions. Forecasting error increased progressively across extended prediction horizons. Feature importance analysis identified animal age, rainfall, and temperature as dominant predictors influencing herd level growth forecasting. The proposed framework may support feed allocation, grazing management, and livestock marketing decisions under heterogeneous sensing environments.