Automated biofloc aquaculture farming system using machine learning algorithms / Abdelmoneim Mohamedosman Ahmed Bakhit

Faculty of Electrical and Electronics Engineering Technology

Thesis (Doctor of Philosophy) -- Universiti Malaysia Pahang – 2026

Includes bibliographical references

Biofloc Technology (BFT) is a sustainable approach used in intensive aquaculture systems to improve water quality by converting waste nutrients, particularly ammonia, into microbial biomass. This method reduces water exchange, minimizes environmental pollution, and supports sustainable fish production. However, the effectiveness of BFT depends on maintaining stable water-quality parameters such as dissolved oxygen (DO), pH, temperature, turbidity, and ammonia concentration. In practice, water quality is often monitored through manual sampling and laboratory analysis, which is time-consuming and reactive. Such delays hinder timely responses to rapid fluctuations in key parameters, leading to ammonia accumulation, pH imbalance, fish stress, and reduced productivity. Furthermore, the addition of beneficial bacteria and molasses as a carbon source may require up to 24 hours to take effect, while the microbial conversion process consumes significant oxygen. Therefore, relying solely on real-time monitoring without predictive capability is insufficient for effective water-quality management. To address these challenges, this study proposes a Smart Biofloc Monitoring and Control System (SBMCS) that integrates IoT-based water-quality sensors, machine learning (ML) models for 24-hour forecasting, and automated control mechanisms to enable proactive decision-making. Five ML models Autoregressive Integrated Moving Average (ARIMA), Exponential Smoothing (ETS), Support Vector Machine (SVM), Random Forest (RF), and Long Short-Term Memory (LSTM) were evaluated to determine the most accurate model for predicting water-quality parameters. The system was experimentally tested over a 90-day tilapia culture cycle in an indoor environment. Outlier analysis using the interquartile range (IQR) method identified anomalies in the dataset, including 41 pH values, 1,030 DO values, 91 turbidity values, and 498 temperature values. After removing outliers, missing data 10-50% were imputed using several techniques, where Expectation Maximization (EM) and Random Forest (RF) produced the lowest and most stable RMSE and MAE values compared with PMM, KNN, and missForest. Various smoothing methods were then applied to the cleaned dataset, with kernel smoothing demonstrating high accuracy across all water-quality parameters. For prediction, LSTM achieved the highest accuracy, with an R² of 0.99 for pH prediction, compared with 0.96 for SVM, 0.81 for RF, and 0.02 for both ARIMA and ETS. Similar performance trends were observed for DO, turbidity, electrical conductivity (EC), ammonia, and temperature, where LSTM and SVM consistently achieved R² values close to 0.99, while ARIMA and ETS showed poor predictive performance. Overall, LSTM outperformed the other models across all parameters. The effectiveness of SBMCS was further evaluated by comparing it with a traditional biofloc monitoring system (BMS). Over the 90-day culture period, SBMCS maintained more stable water-quality conditions, with an average pH of 7.0 ± 0.13 compared to 5.9 ± 0.88 in BMS, and an EC of 2.98 ± 1.63 compared to 0.85 ± 0.42. DO and turbidity levels were similar between the systems. In addition, SBMCS improved aquaculture performance, achieving a higher mean final fish weight 65.79 ± 8.31 g vs. 30.06 ± 8.22 g, greater mean fish length 12.5 ± 0.41 cm vs. 9.0 ± 0.82 cm, a lower feed conversion ratio 5.29 vs. 6.64, and a higher survival rate 95% vs. 86%. These results demonstrate that the proposed SBMCS can effectively enhance water-quality stability and improve fish growth performance in biofloc aquaculture systems.