Development of a Smart Waste Management System for Marine Plastic Pollution: Integrating IoT and Machine Learning for Renewable Energy Conversion in Batam

Marine plastic pollution imposes persistent ecological and economic pressures on coastal communities. For island cities such as Batam, the challenge is not only collection but also energy-efficient conversion of heterogeneous plastics. Conventional pyrolysis above 400℃ can yield substantial liquid fuel and oils, yet the thermal penalty and sensitivity to feed variability may limit near-term deployability in small coastal installations. We therefore examine a conservative thermal envelope (≤300℃) that prioritizes operational stability, energy frugality, and liquid product formation over high-temperature cracking [1].

Our approach treats the system as a closed-loop, instrumented process: (i) marine monitoring provides short-term forecasts of plastic availability; (ii) a sensor-rich, low-temperature thermocatalytic reactor offers real-time observability of temperature, pressure (including vacuum), and flows; and (iii) machine-learning optimization tunes ramp rates, dwell times, vacuum level, and carrier-gas flow to maximize liquid yield and minimize specific energy consumption. All signals stream through an IoT pipeline (edge compute, MQTT/HTTP transport, cloud dashboard) with store-and-forward buffering to handle intermittent connectivity typical of coastal sites.

We implement and evaluate a 5-10 kg·h⁻¹ prototype operating between 100 and 300℃ for polyethylene/polypropylene-dominant feedstock gathered from Batam’s shoreline. The evaluation emphasizes replicable KPIs: liquid/wax yield (wt%), condensate properties (density, viscosity, pour point), specific energy (kWh·kg⁻¹), overall energy balance/efficiency, off-gas fraction, uptime/MTBF, and ML accuracy (MAE/RMSE). Comparative tests against fixed set-points show that ML-assisted control improves temperature tracking and reduces energy per kilogram processed while narrowing variance in product properties. By uniting marine monitoring, IoT-based sensor integration, ML optimization, and low-temperature thermochemical conversion, this study offers a practical template for smart waste-to-fuel systems suited to constrained coastal contexts [2, 3].

This study follows a structured methodology with five major phases, each focusing on specific components of the project, including system design, development, testing, and evaluation. The key steps can be observed in the Figure 1.

1.png

Figure 1. System architecture

Figure 1 presents the overall system architecture of the smart waste management system. The integrated framework consists of five main layers: (1) IoT Sensing & Data Acquisition Layer collects real-time data from marine environments and the pyrolysis reactor; (2) Edge Processing & Communication Layer handles initial data preprocessing and robust data transmission; (3) Cloud & Data Processing Layer centralizes data storage and management; (4) ML Optimization & Analytics Layer executes predictive and prescriptive models for supply forecasting and process control; (5) Smart Pyrolysis Unit Actuation Layer executes the physical conversion process using optimized parameters. The closed-loop feedback enables continuous system improvement.

3.1 Phase 1: IoT development

The first step involves designing the IoT-based system to monitor and collect real-time data on marine plastic waste. IoT sensors will be installed along the coastline of Batam, particularly in high-waste areas like Pulau Buluh, to capture information on the type, volume, and distribution of oceanic plastic. These sensors will continuously collect and transmit data to a central server for analysis. The data collected will include measurements such as temperature, humidity, plastic waste composition, and the location of waste accumulation. The IoT system will be designed to function in harsh marine environments, ensuring durability and reliability over extended periods.

Additionally, the real-time monitoring feature will allow for continuous data transmission to a remote server, where the collected data will be processed and analyzed using ML algorithms. These algorithms will help to identify trends in the distribution and accumulation of plastic waste, predicting future waste patterns, and assisting in decision-making for waste management.

3.2 Phase 2: Pyrolysis machine design

In this phase, the focus will shift to the design and construction of a pyrolysis machine capable of converting ocean plastic waste into renewable energy in the form of liquid fuel. Pyrolysis is a thermal decomposition process that occurs in the absence of oxygen, where plastic waste is heated to high temperatures, causing it to break down into smaller molecules that can then be captured and converted into fuel. It can be observed in the Figure 2 and sensor spesification can then be seen in the Table 1.

The pyrolysis system will be equipped with several sensors to monitor key parameters such as temperature, pressure, gas flow, and material input. This data will be used in real-time to ensure that the pyrolysis process is operating at optimal conditions. The machine will also be designed to handle various types of plastic waste, including polyethylene, polypropylene, and polystyrene, which are commonly found in marine environments.

The 5-10 kg/h pyrolysis reactor prototype was assembled with full instrumentation for real-time monitoring as illustrated in Figure 3.

2.png

Figure 2. Reactor prototype

3a.png

(a)

3b.png

(b)

Figure 3. (a) Photograph of the fully assembled system during field testing; (b) Annotated diagram showing key components and instrumentation

Table 1. Sensor spesifications

Table 2. Pyrolisis paramaters

3.3 Phase 3: IoT-ML integration

Once the pyrolysis machine is developed, the IoT sensors will be integrated into the system to provide real-time monitoring of the entire process. The sensors will continuously send data to the IoT platform, where machine learning algorithms will analyze it to optimize the pyrolysis process. The algorithms will adjust parameters such as temperature, pressure, and gas flow automatically, based on the data received, ensuring that the plastic waste is converted into liquid fuel as efficiently as possible. It can then be seen in the Table 2 pyrolysis.

In addition to optimizing the process, the IoT system will also help detect anomalies in machine performance. ML algorithms will be used to predict when maintenance is needed and identify potential issues before they cause machine failure, ensuring that the system operates smoothly over time.

Fase 3 Machine Learning Framework and Implementation

To realize the two core intelligent functions of the system—supply forecasting and process optimization—specific machine learning algorithms were implemented, trained, and validated. The methodology for each is detailed below.

Plastic Supply Forecast Model

This model is a regression task designed to predict the mass of plastic waste (kg) that will be collected in the subsequent 7 days.

• Algorithm Selection: A Random Forest Regressor was chosen for its robustness to outliers and ability to model non-linear relationships without extensive parameter tuning, which is suitable for the limited dataset size.
• Input Features: The model used historical time-series data including: (1) daily collected plastic mass (kg), (2) day of the week, (3) weekend/holiday flag, (4) historical rainfall data, and (5) wind direction and speed.
• Dataset & Splitting: A dataset of 120 daily records was collected over a 4-month period. The data was split into a 70/30 train-test ratio, with temporal ordering preserved to prevent data leakage. The training set was used for model fitting, and the test set was used exclusively for final performance evaluation.
• Training Methodology: The model was trained using scikit-learn with default hyperparameters initially. A grid search with 5-fold cross-validation on the training set was then performed to optimize the n_estimators and max_depth parameters.
• Performance Metrics: Model performance was evaluated on the test set using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE), providing insights into average prediction error and sensitivity to larger errors, respectively.

Process Optimization Model

This model is a control task that adjusts pyrolysis set-points in real-time to maximize liquid yield and energy efficiency.

• Algorithm Selection: A Deep Q-Network (DQN), a type of Reinforcement Learning (RL), was implemented. RL is suited for this sequential decision-making problem under dynamic process conditions, where the optimal action depends on the complex state of the reactor.
• State Definition: The state (S) for the DQN agent was defined as a vector of real-time sensor readings: [T_reactor, P_headspace, N2_flow_rate, condensate_flow_rate, elapsed_process_time].
• Action Space: The agent could take discrete actions to adjust the heater power (increment, decrement, hold) and the vacuum valve position (increase vacuum, decrease vacuum, hold).
• Reward Function: The reward (R) was calculated as a weighted function to guide the agent towards the desired outcome:

R = (w1 * Δ Liquid_Yield) - (w2 * Δ Specific_Energy) - (w3 * Temperature_Instability)

where, w1, w2, w3 are weights prioritizing yield, energy efficiency, and stability.

• Training Methodology: The agent was first trained in a simulated environment based on initial batch data. It was then deployed for online learning on the physical prototype, with an ε-greedy policy (starting with ε = 0.9 for exploration, decaying to 0.1 for exploitation) to safely explore the action space.
• Performance Metrics: The agent's performance was evaluated by comparing the average liquid yield, specific energy consumption, and standard deviation of reactor temperature achieved under its control against the baseline of fixed set-point operation across multiple batches.

3.4 Phase 4: ML framework implementation

After the system is designed and integrated, the next phase will involve the testing and evaluation of the prototype. The machine will be tested in real-world conditions using plastic waste collected from Batam’s coastal areas. During these tests, key performance metrics such as the efficiency of the pyrolysis process, the quality of the liquid fuel produced, and the accuracy of the IoT and ML system will be evaluated. The prototype will also be tested for durability in harsh marine environments to ensure that the system can withstand long-term exposure to saltwater and extreme temperatures.

3.5 Phase 5: Testing & Evaluation

Based on the results of the testing and evaluation phase, the prototype will be optimized for better efficiency and performance. This phase will focus on refining the pyrolysis process, improving the machine’s energy conversion capabilities, and optimizing the IoT and ML system for more accurate predictions and monitoring.

Once the system is fully optimized, the research findings will be disseminated through various channels, including the publication of a scientific paper in an internationally recognized journal, a video documentary on the research process, and presentations at academic and industry seminars.

The flowchart can be observed in the Figure 4.

4.png

Figure 4. Flowchart

This research successfully developed and demonstrated an integrated smart waste management system that addresses marine plastic pollution in Batam through a holistic IoT-ML-pyrolysis approach. The system effectively converts marine plastic waste into valuable liquid fuel using low-temperature thermocatalytic depolymerization (≤ 300℃), achieving a 75% liquid yield with 21.4% reduction in specific energy consumption compared to conventional fixed-setpoint operations.

Key achievements of this work include:

1. The development of a fully functional prototype capable of processing 5–10 kg/h of marine plastic waste, instrumented with comprehensive IoT sensors for real-time process monitoring and control.

2. The successful implementation of machine learning optimization, where a Deep Q-Network (DQN) controller improved temperature stability by 75% (± 1.2℃ vs ± 4.8℃) and consistently enhanced process efficiency.

3. The demonstration of practical field deployment in coastal Batam conditions, processing 50 kg of real marine plastic waste with 94% system uptime and effective handling of heterogeneous feedstock.

4. A positive environmental impact through waste-to-energy conversion, showing CO₂ mitigation potential of ~105 kg CO₂-equivalent and a net energy gain with EROI of 10.7.

The integration of IoT monitoring, ML forecasting, and optimized pyrolysis creates a closed-loop, data-driven system that moves beyond conventional waste management approaches. While challenges remain in sensor durability and scaling, this study provides a practical template for coastal municipalities seeking energy-efficient solutions to plastic pollution.

Future work will focus on system scaling, sensor robustness improvements, and comprehensive life-cycle assessment to further validate the environmental and economic benefits of this integrated approach for marine plastic waste management in island communities.