Time-Series Machine Learning for Short-Term Biomass Prediction in Patin Fish Farming Based on Operational Farm Data

Abstract

Small-scale aquaculture increasingly adopts sensor-based monitoring systems; however, most existing implementations remain descriptive and provide limited predictive support for operational decision-making. This study proposes a rolling time-series machine learning framework for short-term biomass prediction in ongoing Patin (Pangasius hypophthalmus) farming operations using real operational farm data. Two complete six-month production cycles were analyzed, each comprising daily feeding records, water quality parameters, survival information, and observed total biomass measurements. The prediction task was formulated as a seven-day ahead (t+7) supervised regression problem to align with weekly farm management and feeding adjustment practices. An XGBoost regression model was developed using rolling-window feature construction and evaluated using a cross-cycle validation strategy, where the model was trained on one production cycle and tested on an independent cycle. Model performance was assessed using MAE, RMSE, MAPE, and R² metrics and compared against persistence and linear regression baselines. Experimental results show that the proposed model achieves stable and accurate seven-day ahead biomass predictions as total biomass increases toward approximately 1.7 tonnes, consistently outperforming baseline approaches. The results further indicate that short-term growth dynamics are strongly influenced by recent feeding behavior and biomass history, highlighting the effectiveness of data-driven modeling for capturing non-linear growth patterns. Overall, the findings demonstrate the feasibility of deploying rolling machine learning models as a practical predictive decision-support layer for short-term biomass forecasting in Patin aquaculture systems.