Design and Simulation of Smart Food Supply Chains for Island Regions Based on AI-LM Using Long Short-Term Memory (LSTM) Approach

2.1 Food distribution challenges in island regions

Island regions are characterized by a group of small islands with limited accessibility, presenting unique logistical constraints that differ significantly from those in continental areas. Research conducted by Yudiyta et al. [11] revealed that the general characteristics of the island regions in Indonesia include limited agricultural land, high dependence on the fisheries sector, and poor food distribution infrastructure. The study used a post-positivist approach to understand the social and institutional dynamics that influence the success of policy implementation. The results showed that food security in island areas is highly vulnerable, due to not only difficulties in local production but also challenges in food distribution and accessibility between islands.

2.2 Blockchain and IoT in food distribution

Blockchain has been widely used to improve traceability and transparency in food distribution [12]. This system records every transaction and movement of goods in an immutable digital ledger, enabling rapid auditing and tracking of the source of problems. Duan et al. [13] explained that blockchain can reduce cases of food counterfeiting and increase consumer trust.

Meanwhile, the IoT plays a significant role in monitoring the physical condition of food during the distribution process [14]. IoT sensors can record temperature, humidity, and vibration data, which helps maintain food quality, especially fresh produce [15]. Complementary features between blockchain and IoT have been widely discussed in recent studies as a potential intelligent architecture for supply chains.

2.3 The concept of food supply chain and its efficiency

Food Supply Chain (FSC) is an integrated system that connects producers, distributors, and final consumers. Some common factors that greatly influence FSC efficiency are coordination, distribution speed, and adaptability to changes in demand and environmental disturbances [16]. Research conducted by Ganbold et al. [17] revealed that information technology capabilities play a significant role in driving supply chain integration (SCI), which ultimately has a positive impact on a company’s operational performance. IT capability dimensions, such as cross-functional applications and supply chain applications, significantly strengthen internal and external integration. In relation to island regions, logistics efficiency requires an intelligent systems approach capable of adapting to geographic and environmental uncertainties.

2.4 The application of artificial intelligence in the food sector

Artificial intelligence (AI) offers significant potential across all aspects of the food system, including agricultural production processes, distribution management, and food waste reduction. Soori et al. [18] stated that AI is capable of learning patterns and adapting to operational changes, making it highly suitable for dynamic and complex distribution systems.

Javaid et al. [19] highlighted various AI applications in the agricultural sector, including weather prediction, crop health monitoring, and early detection of pests and diseases based on digital images. This approach aligns with the concept of precision agriculture, aiming to increase efficiency, crop yields, and environmental sustainability.

Meanwhile, Onyeaka et al. [20] discovered the roles of AI in reducing food waste and strengthening the circular economy system in the food sector. AI can accurately predict demand, optimize inventory and distribution logistics, and classify food waste for efficient reuse or distribution. Their findings suggest that integrating AI can support resource efficiency, reduce carbon emissions, and contribute to achieving global food sustainability goals.

However, most of these studies have been conducted in the context of land infrastructure or metropolitan areas with high data access and connectivity. Studies on AI in geographically remote or island regions are still very limited, and even those that exist tend to focus on weather prediction or precision agriculture, rather than food logistics.

This section presents the design results of the AI-LM system, utilizing the LSTM approach, with a focus on the technical aspects of implementation. The design process involved developing a model architecture that integrated historical food demand data, the FSI, and environmental variables, including rainfall and ocean waves. The data were processed through an LSTM network to generate predictions of food demand and distribution in the island regions.

In addition to LSTM modeling, the system design was also equipped with a workflow automation scheme using n8n as a dynamic logistics distribution simulation framework. This stage demonstrated how data flows, from input to preprocessing, modeling, and prediction output, were integrated into a system architecture that can be tested and further developed.

4.1 AI-LM architecture design

The AI-LM system architecture was designed to handle the complexity of food distribution in island regions by utilizing dynamic multi-source data.

This design consisted of three main layers: a data input layer, an LSTM modeling layer, and an n8n workflow integration layer, which are described below.

1.      Data Input Layer

The data consisted of:

FSI for each island province.

Historical data on rainfall and ocean wave height.

Data on demand for basic food (rice, eggs, vegetables, and fresh fish).

The data were then formatted into a time series to suit the needs of LSTM predictive modeling.

2.      LSTM Modeling Layer

The LSTM network was chosen due to its ability to overcome the problem of vanishing gradients in long sequential data.

The architecture consisted of an input layer (multivariate time series), several hidden LSTM layers with tanh and sigmoid activation functions, and a dense layer to generate food demand prediction values.

This model was trained using historical data to predict demand patterns based on weather factors, ocean conditions, and consumption trends.

3.      Workflow Integration Layer (n8n)

The prediction results from LSTM were integrated with the n8n automation workflow.

The n8n managed the data flow from the input source (Google Sheets/statistics) to the prediction module, and then linked the output to the monitoring dashboard.

Through this integration, the system can provide dynamic notifications regarding potential spikes in demand and distribution risks due to weather/ocean disruptions.

Figure 2 shows the results of the AI-LM architecture design.

Figure 2. The AI-LM architecture design

4.2 n8n workflow design

The n8n workflow design for the AI-LM model focused on integrating input data, predictive processing with LSTM, and output distribution in the form of notifications and interactive dashboards. This modular workflow was easily expandable to meet the needs of the food supply chain in the island regions.

1. Node Input (Ingest)

• Connecting external data sources: Google Sheets, weather APIs (rainfall, wind, waves), and historical demand databases.
• Automatic triggers use Cron Scheduler, so the dataset is updated periodically without manual intervention (e.g., every 1–3 hours).

2. Node Pre-processing

• Cleaning (missing/outliers), scale normalization, category encoding, and time-series transformation (windowing/lag/rolling).
• Unifying heterogeneous data—rainfall, wave height, demand volume—into a uniform feature scheme for LSTM input.

3. Processing Node (LSTM)

• Executing a Python script or containerized application to perform time-series-based demand prediction using an LSTM model.
• Considering climate factors and seasonal patterns. Output: Estimated demand volume and alert threshold.

4. Output Distribution Node

• Saving results to Google Sheets/DB (historical log), displaying on dashboard, and sending automatic notifications (Email/Telegram bot) to stakeholders (distributors, department of agriculture/maritime, fishermen/farmers).

5. Monitoring & Evaluation Node

• Performance validation: compare predictions vs realizations (accuracy, MAE/RMSE, MAPE).
• Feedback loop for hyperparameter/feature updates when deviation increases to allow the model to adapt to geographic and consumption dynamics.

Figure 3 shows the n8n workflow design.

Next, Figure 4 presents the step-by-step flow of the process carried out on n8n.

The following is a detailed explanation of the process flow diagram of n8n:

1. Start / Trigger Workflow

• The workflow starts when the Execute Workflow button is pressed.
• This serves as the initial trigger for running the entire series of nodes.

Figure 3. The n8n workflow design

Figure 4. The n8n flowchart

                    2. Get Row(s) in Sheet

• This node retrieves row data from a specified Google Sheet.
• The retrieved data is usually raw input (e.g., transaction data, logistics requests, or respondent information).
  3. Code (Data Transformation 1)

• Data from the Google Sheet is processed using a script (JavaScript/Function Node) T.
• his stage is typically used to perform initial transformations, such as parsing text, filtering specific rows, or simple calculations.
  4. Code (Data Transformation 2)

• The data is processed again in the second code node.
• This stage is typically used for data cleaning and formatting, such as ensuring correct date formats, removing blank values, or standardizing variables.
  5. Edit Fields (Mapping)

• This node restructures the data to suit the output requirements.
• For example, renaming fields, adding new columns, or adjusting the JSON structure for ease of use in subsequent steps.
  6. Branching (Flow Branching)

• From here, the flow branches into two parallel outputs:
• Create a Row: Saves the processed data to a Google Sheet or other database.
• Send a Text Message: Sends a notification (e.g., WhatsApp, Telegram, or SMS) based on the newly entered data.

4.3 Simulation and testing

The simulation and testing procedures were carried out by building an LSTM-based workflow involving several stages. First, the quality of multi-source data, such as Google Sheets, weather APIs, and historical databases, was preprocessed through several stages, including normalization, cleaning, and the formation of time windows (sequences) as model input. Next, the LSTM model was trained on the training data using an early stopping mechanism to prevent overfitting, and data validation was employed to assess the model’s performance. The evaluation was conducted using accuracy metrics, including MSE, MAE, RMSE, and MAPE, to compare actual and predicted values. Furthermore, performance testing was performed by recording the average training time per epoch, inference time per sample, and processing throughput. Table 4 shows the results of the comparison between the actual values (FISC Actual) and predicted values (FISC Predicted) from the LSTM model in several provinces and certain months, and calculates the Absolute Error as the absolute difference between the two values.

Table 4. Comparison of actual results (FISC Actual) and predicted values (FISC Predicted)

The LSTM model’s prediction results demonstrate clear differences in accuracy between provinces as well as seasonal variations. In Papua, the actual FISC values exhibit distinct seasonal spikes, particularly in April–May (Ramadan and Idul Fitri), July (Idul Adha), and December (Christmas and New Year). For instance, the FISC Actual rose to 42.0 in April and 45.0 in May, while the predicted values remained lower at 53.30 and 51.50, producing errors of 11.58 and 6.3, respectively. Similarly, in December, the actual FISC increased sharply to 48.0, but the prediction dropped to 35.51, generating an error of 8.5. These deviations suggest that the model cannot yet adequately capture sudden surges linked to religious and cultural events, which are typically accompanied by rising household demand and supply chain disruptions.

In contrast, Southeast Sulawesi presents a more stable seasonal pattern. Actual FISC values show moderate increases during Ramadan–Idul Fitri, Idul Adha, and year-end holidays, while the LSTM model demonstrates better adaptability in this region. The prediction error declined from 15.38 in January to just 3.1 in June, and although fluctuations remain (e.g., May with an error of 13.3), the model generally follows the observed trend more closely.

These findings indicate that LSTM models are relatively more capable of learning patterns in areas with more stable data conditions [23, 24]. In areas with extreme conditions, longer historical data and additional features are required for more accurate predictions.

Table 5 shows a summary of the LSTM model evaluation results.

Table 5. Summary of LSTM model evaluation results

In Table 5, the evaluation results demonstrate that the LSTM model achieves satisfactory performance on the training dataset, with relatively low errors (MSE = 60.77; MAE = 6.93; RMSE = 7.80; MAPE = 9.53%). This suggests that the model is able to effectively capture temporal dependencies within the training period. However, when applied to the validation dataset, the errors increase substantially (MSE = 185.33; MAE = 13.44; RMSE = 13.61; MAPE = 29.07%), indicating a marked decline in predictive accuracy on unseen data. This discrepancy shows that, although the use of a 12-month dataset improves learning compared to shorter horizons, the model still struggles with generalization. The limitations are likely related to the relatively short one-year observation window, which does not fully capture inter-annual seasonal variations, and the inherent complexity of the LSTM architecture. Future research should therefore focus on extending the dataset to cover multiple years, introducing regularization methods (such as Dropout and L2 regularization), and employing province-specific modeling to better capture localized dynamics. Figure 5 presents the training history curve of the LSTM model for predicting FISC (FSI).

Figure 5. Training history curve of the LSTM model for predicting FISC (FSI)

4.4 Discussion

Simulation results indicate that the LSTM model is capable of accurately learning patterns in the training data. This is reflected in the Mean Squared Error (MSE) value of 0.0339 and Mean Absolute Error (MAE) of 0.1272 on the training dataset. This low error value indicates that the model successfully adapted to the characteristics of the training data. However, on the validation dataset, model performance decreases significantly, with an MSE value of 286.80 and an MAE of 16.05. A significant difference between the results on training and validation data indicates overfitting, where the model effectively memorizes patterns in the training data but struggles to generalize to new data [25].

This is a common phenomenon in the application of deep learning-based models, especially on datasets with a limited size. LSTM, as a variant of the Recurrent Neural Network (RNN), has a complex structure with a relatively large number of parameters, so it requires a sufficient amount of historical data to achieve optimal generalization ability [26]. In this study, only three months of data (January–March) were available for each province, which limited the model's temporal variation. Consequently, the model was unable to capture seasonal patterns and long-term fluctuations, which are necessary for food security analysis.

Furthermore, the prediction results showed variations in accuracy across provinces. In Papua, the model produced predictions significantly higher than the actual values, with an absolute error of 15–24 points. Papua has extreme conditions, such as high rainfall and ocean waves, which are very different from those in other provinces. In contrast to Southeast Sulawesi, the model showed more stable performance, with the error decreasing from 15.38 in January to 8.91 in March. These findings indicate that the model can learn patterns more easily in areas with stable conditions. Meanwhile, in areas with extreme conditions, additional data and more adaptive modeling techniques are required.

The training history graph also aligns with these findings. The training loss value decreases significantly until it stabilizes at a low level. Meanwhile, the validation loss decreases but plateaus at a higher level and remains constant after the 20^th epoch. This pattern confirms the indication of overfitting, where the model focuses too much on the training data and fails to improve accuracy on the validation data. There are several strategies to overcome this problem, such as extending the historical data horizon (at least 12–24 months), adding exogenous features like food prices and logistics indicators, using regularization techniques like dropout, and applying evaluation methods that are more suitable for time series data, for example, walk-forward validation [27].

The LSTM model in this study demonstrates excellent performance on the training data but does not perform optimally on the validation data. This indicates the need for improvement strategies in terms of data quality and quantity, as well as model design. By incorporating historical data and adapting the architecture, LSTM is expected to deliver more accurate and reliable predictions, thereby supporting food security analysis in island regions.

The authors gratefully acknowledge the financial support of the Ministry of Higher Education, Science, and Technology of the Republic of Indonesia for Fiscal Year 2025 under Contract No. 074/C3/DT.05.00/PL/2025.