Inventory management has long been regarded as one of the most significant catalysts for efficiency, profitability, and supply chain sustainability. For production firms in particular, inventory policy directly affects manufacturing planning, customer service level, and cost. Stockholding costs rise with overstocking, as does obsolescence and wastage [6]. Stockouts also generate order fulfilment delays, frequency of reordering, and customer dissatisfaction. With heightened global competition, decreasing product life cycles, and unpredictable patterns of demand, effective inventory management has become a fundamental dilemma for both researchers and practitioners [10].

Early inventory modelling schemes such as the Economic Order Quantity (EOQ) and Economic Production Quantity (EPQ) models laid a foundation for analysing ordering vs. holding cost trade-offs [19]. Later extensions added stochastic demand, shortage costs, and multi-period planning, often using linear or dynamic programming. While these models were beneficial, their deterministic demand assumption, restricted product range, and absence of capacity constraints make them less applicable in today's data-intensive supply chains. In addition, static models will not be able to cope with rapidly evolving markets or spontaneous peaks in demand caused by campaigns, promotions, or external shocks [11].

Artificial Intelligence (AI) and Machine Learning (ML) advancements in recent history have created new avenues for inventory forecasting and decision-making. Techniques such as Long Short-Term Memory (LSTM) networks, Gradient Boosting Machines (GBM), and Support Vector Machines (SVM) have been found to be very accurate in demand forecasting, particularly with time-series data that is seasonal and nonlinear [2]. Reinforcement Learning (RL) has been found helpful in adaptive decision-making, enabling dynamic reordering policy adaptation. But even with such developments, AI-based approaches are typically applied separately and are not integrated into mathematical models of optimization that manage cost reduction and inventory allocation [5].

In parallel, Internet of Things (IoT) and sensor-fitted supply chains enable real-time monitoring of warehouse operations and stock levels. With math-based optimization and AI-driven forecasting, IoT data streams are able to improve responsiveness by synchronizing replenishment and procurement to real-time stock levels [4]. All of this integration can serve to reduce stock-outs and overstocking, improve customer satisfaction, and reduce operational costs [8].

This paper addresses the aforementioned challenges by proposing a hybrid method that integrates AI-based demand forecasting, linear programming-based inventory optimization, and real-time data accumulation through IoT. The AI section uses LSTM and GBM for accurate demand forecasting, with incorporation of external factors such as campaign and promotional data. Forecast outputs are subsequently fed into a linear programming (LP) model that calculates supply quantities within the cost and capacity constraints [3]. Finally, real-time IoT-based data streams are utilized for dynamically configuring inventory policies in order to make them responsive under fluctuating conditions. The main contributions of this work are threefold:

Towards the creation of a hybrid AI–mathematical optimization framework that merges forecasting and inventory allocation to fill the gap between them.

• Towards incorporating real-time IoT data streams into the optimization process so that adaptive and responsive decisions can be made.
• Empirical illustration of the suggested methodology via a case study for a manufacturing warehouse with increases in forecasting accuracy, cost savings, and service levels over conventional models.

The motivation and gap in research for the proposed framework are encapsulated that the existing inventory models and AI-exclusive methodologies have their own drawbacks, whereas the proposed hybrid framework combines AI forecasting, mathematical optimization, and IoT-driven updates to provide a more adaptive and feasible solution.

Optimal inventory management has been widely researched across operations research, supply chain management, and computational intelligence. Initially, most of the research was concerned with the traditional models like the Economic Order Quantity (EOQ) and Economic Production Quantity (EPQ). These models offered a theoretical base by optimizing ordering and holding costs assuming deterministic conditions. Though helpful in revealing underlying trade-offs, their use of static parameters and disregard of real-world complications, including capacity constraints and uncertainty in demand, constrained their utility [7].

Later research added to these models by using stochastic formulations, dynamic programming, and linear programming methodologies. These studies tried to include uncertainty, shortage penalties, and multi-period planning. Their focus was still limited, though, as they tended to focus on cost minimization without the full context of changing market dynamics or adaptive decision-making [13].

Emergence of Artificial Intelligence (AI) and Machine Learning (ML) brought new demand forecasting techniques and inventory decision-making. Techniques like Long Short-Term Memory (LSTM) networks and Gradient Boosting Machines (GBM) have proven to be more accurate in capturing nonlinear and seasonal patterns in demand [1]. Reinforcement Learning (RL) has also been used in adaptive inventory control, where the policies change with shifting environments over time [17]. Whereas these AI techniques are more accurate in their forecasts compared to the conventional models, they are usually applied separately and fail to consider optimization constraints like warehouse space or costs of procurement [12].

A further new stream of research highlights the potential of Internet of Things (IoT) technologies and real-time data in supply chain visibility. IoT-based systems enable real-time tracking of inventories to inform reactive decision-making [10]. Yet, the majority of research considers IoT as a monitoring platform instead of incorporating it into decision-optimization models. Consequently, its full potential for adaptive, data-driven inventory optimization is yet to be explored [13].

Proposed Hybrid Framework

The proposed framework consolidates three complementary modules: (i) AI-based demand forecasting, (ii) mathematical optimization of inventory decisions, and (iii) IoT-enabled real-time monitoring. The three modules combined offer an adaptive system that reduce costs with high service levels in dynamic situations.

Forecasting Module (AI-based)

The forecasting module leverages advanced Machine Learning models such as Long Short-Term Memory (LSTM) networks and Gradient Boosting Machines (GBM).

• LSTM captures sequential and seasonal demand patterns.
• GBM incorporates promotional and campaign-driven demand spikes.

The general form of forecasted demand is expressed as:

where:

• : Forecasted demand at period ,
• : Historical sales data,
• : Campaign/promotion indicators,
• : Seasonal or trend-related features,
• : AI forecasting model (LSTM/GBM).

Optimization Module (Linear Programming)

The outputs of the AI forecasting model are fed into a Linear Programming (LP) model that optimizes procurement and inventory decisions under cost and capacity constraints.

Decision Variables:

• : Supply quantity in period ,

:  Ending inventory in period .

Objective Function:

where:

• : Procurement cost per unit in period ,
• : Unit holding cost per period,
• : Total inventory-related cost.

Constraints:

1. Inventory Balance
2. Capacity Constraint
3. Non-negativity

Real-Time IoT Integration

IoT-enabled RFID tags and sensors track live inventory levels. The observed data stream is continuously compared against forecasted inventory . If a deviation occurs (e.g., unexpected stock-out), the system triggers an adjustment in the optimization module.

where represents corrective procurement or redistribution, recalculated in real time.

Conceptual Workflow

• AI models generate multi-period demand forecasts.
• Forecasts serve as inputs to LP-based optimization for supply and storage allocation.
• IoT real-time data validates inventory levels, triggering adjustments if deviations are detected.
• The integrated system ensures cost minimization, demand fulfilment and adaptive responsiveness.

Workflow of the proposed hybrid system: AI-based demand forecasting generates forecasts that feed into an LP optimization module for procurement decisions; warehouse inventory is monitored by IoT sensors, which provide real-time feedback to both optimization and forecasting modules for dynamic adjustment.

The research methodology adheres to a systematic framework for the incorporation of AI-based forecasting, mathematical optimization, and real time feedback IoT-based inventory management. The process consists of four stages: data preparation, demand forecasting, optimization modelling, and real-time adjustment.

Data Collection and Preprocessing

• Historical demand data: 24–36 months of monthly demand data were utilized to identify seasonal and trend fluctuations.
• External factors: Promotional events, campaign indicators, and seasonal characteristics were incorporated as input features in order to capture spikes in demand.

Preprocessing: Missing values were filled in and normalization was used to scale features for machine learning algorithms. Data was divided into training (70%), validation (15%), and testing (15%) sets.

AI-based Demand Forecasting

Two AI models were employed for forecasting:

• Long Short-Term Memory (LSTM): Captures sequential and temporal dependencies in demand data.
• Gradient Boosting Machine (GBM): Handles nonlinearities and incorporates external features such as promotions and campaigns.

The models were trained using historical sales. seasonal attributes and campaign indicators. Evaluation metrics included:

• Mean Absolute Percentage Error (MAPE),
• Root Mean Square Error (RMSE),
• Forecast Accuracy (%).

The best-performing forecast model was selected as input for the optimization module.

Optimization Model

The forecasted demand values served as input to a Linear Programming (LP) model that determines procurement and inventory allocation. The LP model minimizes total cost while satisfying demand and capacity constraints, as outlined in Section 3.2.

The optimization model was solved using Python’s PuLP package and cross-validated with Excel Solver for smaller instances. Decision outputs included:

• Optimal procurement quantities per period,
• Inventory levels at each time step,
• Shortages (if any) and associated penalty costs.

IoT-enabled Real-Time Adjustment

IoT sensors installed on warehouse shelves offered up-to-date inventory visibility. Differences between actual and predicted inventory were continuously tracked.

• If stock-out risk was identified → immediate corrective procurement was initiated.
• If there was excess inventory → supply was decreased in future periods.
• This feedback mechanism ensured adaptive decision-making, so the system could better manage demand fluctuations and outliers.

To validate the suggested hybrid framework, a case study was conducted on a manufacturing warehouse that produces cement products. The warehouse has a maximum capacity of 400 units with an initial inventory of 150 units. Six months of historical demand data along with promotional event indicators were utilized to cover seasonal fluctuations. The objective was to compare the performance between traditional models, AI-only forecasting, and the suggested hybrid AI + LP + IoT system.

Input Data

Forecasting Performance

Observation: LSTM outperformed other models due to its ability to capture seasonality and sudden spikes from promotional events.

Optimization Results

(Hybrid Framework)

Observation:

The results of the case study validate that the suggested hybrid model is superior to both conventional and AI-based methods for inventory optimization. Conventional EOQ-based models could not handle demand fluctuations and promotional activity, so either there would be excess inventory or shortage. For instance, during months of high demand (Month 3), EOQ underestimates the demand, resulting in unsatisfied demand and lower profitability.

The AI-solo models (GBM, LSTM) showed much greater forecasting accuracy than EOQ. Specifically, LSTM picked up seasonal patterns and campaign-induced spikes better and achieved accuracy of almost 90%. This is highlighted in figure 2, which shows the comparison between actual and forecasted demand. The graph makes it clear that the AI models, especially LSTM, are capable of tracking fluctuations and spikes much more effectively than EOQ.

• Inventory remained within the warehouse capacity threshold (400 units).
• A shortage of just one (20 units) occurred in Month 3 during an event promotion was rectified by IoT feedback in subsequent or future orders.
• Profitability improved by 14% compared with traditional supply policies.

Fig 1: Actual vs Forecasted Demand

Nevertheless, while showing robust predictive capability, AI-solo approaches did not optimize procurement and storage actions under capacity and cost limitations. This shortcoming implies that forecasting accuracy is not enough to address practical inventory management. The hybrid AI + LP + IoT system overcame these limitations by combining predictive and prescriptive elements. As shown in figure 3, optimal supply decisions closely align with actual demand, demonstrating how the hybrid model successfully converts forecasts into practical procurement plans within capacity constraints.

Fig 2: Optimal Supply vs Actual Demand

In addition, real-time IoT-fed feedback loops updated deviances and lowered the risk of stock-outs, making adaptive decision-making possible. This is illustrated in figure 4, which highlights how ending inventory levels remained stable across the months and how shortages were minimized to just one instance during Month 3. The figure emphasizes how the hybrid model enabled capacity-sensitive planning and adaptive corrections through IoT integration.

Fig 3: Ending Inventory and Shortage Levels

From a management standpoint, the model offers three key advantages:

1. Capacity-sensitive planning: Procurement is maximized within warehouse capacities, avoiding both overstocking and understocking of warehouse space.
2. Cost savings: By matching supply with projected demand and adjusting according to real-time inputs, the model minimizes excess holding costs and shortage penalties.
3. Operations flexibility: IoT connectivity allows managers to dynamically react to sudden changes in demand or disruptions, better enabling resilience in uncertain markets.

In reality, this methodology can be realized by production companies looking to move from traditional inventory models to intelligent, data-based systems. The model demands average computational resources, bringing it into reach for not just large-scale corporations but even SMEs. Notably, the integration of AI forecasting, optimization, and IoT-based feedback ensures that the system is both precise and manageable, fulfilling the gap between industrial application and academic research.

This research presented a hybrid model that marries AI-driven forecasting, linear programming optimization, and IoT-based real-time monitoring for inventory control in manufacturing. The model effectively overcame the weaknesses of classical EOQ-based models and purely AI-based methods by putting together predictive precision, cost minimization, and responsive adaptability. A case study proved that the model enhanced demand satisfaction to 95% and raised profitability by 14%, providing effective use of warehouse capacity and minimizing the risks of overstocking and stockouts.

Future research will involve expanding the framework into multi-echelon supply chains, with coordination among multiple nodes providing further efficiency gains. The integration of stochastic demand modelling, sustainability indicators (e.g., carbon footprint, waste avoidance), and more advanced reinforcement learning algorithms can potentially offer even greater flexibility. Such expansions will allow the construction of intelligent, self-improving systems that meet Industry 4.0 goals, providing robust and sustainable supply chain operations.

1. Nweje, U., & Taiwo, M. (2025). Leveraging Artificial Intelligence for predictive supply chain management, focus on how AI-driven tools are revolutionizing demand forecasting and inventory optimization. International Journal of Science and Research Archive, 14(1), 230-250.
2. Kagalwala, H., Radhakrishnan, G. V., Mohammed, I. A., Kothinti, R. R., & Kulkarni, N. (2025). Predictive analytics in supply chain management: The role of AI and machine learning in demand forecasting. Advances in Consumer Research, 2, 142-149.
3. Sajja, G. S., Addula, S. R., Meesala, M. K., & Ravipati, P. (2025, June). Optimizing inventory management through AI-driven demand forecasting for improved supply chain responsiveness and accuracy. In AIP Conference Proceedings (Vol. 3306, No. 1, p. 050003). AIP Publishing LLC.
4. Sandra, M., Narayanamoorthy, S., Suvitha, K., Pamucar, D., & Kang, D. (2025). A Novel AI Analytics Driven Forecasting Paradigm for Strategic Inventory Management Through Fuzzy Multi-criteria Decision Model in Supply Chains. International Journal of Fuzzy Systems, 1-21.
5. Bhavikatta, N. B. (2025). AI-Driven Inventory Optimization in Supply Chains: A Comprehensive Review on Reducing Stockouts and Mitigating Overstock Risks. Journal of Computer Science and Technology Studies, 7(7), 01-13.
6. Verma, P. (2024). Transforming Supply Chains Through AI: Demand Forecasting, Inventory Management, and Dynamic Optimization. Integrated Journal of Science and Technology, 1(3).
7. Choudhuri, S. S. (2024). AI-Driven Supply Chain Optimization: Enhancing Inventory Management, Demand Forecasting, and Logistics within ERP Systems. International Journal of Science and Research (IJSR), 13(3), 927-933.
8. Islam, M. K., Ahmed, H., Al Bashar, M., & Taher, M. A. (2024). Role of artificial intelligence and machine learning in optimizing inventory management across global industrial manufacturing & supply chain: A multi-country review. International Journal of Management Information Systems and Data Science, 1(2), 1-14.
9. Patil, D. (2024). Artificial Intelligence-Driven Supply Chain Optimization: Enhancing Demand Forecasting And Cost Reduction. Available at SSRN 5057408.
10. Amosu, O. R., Kumar, P., Ogunsuji, Y. M., Oni, S., & Faworaja, O. (2024). AI-driven demand forecasting: Enhancing inventory management and customer satisfaction. World Journal of Advanced Research and Reviews, 23(2), 100-110.
11. Mitta, N. R. (2023). AI-Driven Optimization of Supply Chain Networks in Manufacturing: Utilizing Machine Learning for Demand Forecasting, Inventory Management, and Logistics Efficiency. Los Angeles Journal of Intelligent Systems and Pattern Recognition, 3, 404-446.
12. Onotole, E. F., Ogunyankinnu, T., Osunkanmibi, A. A., Adeoye, Y., Ukatu, C. E., & Ajayi, O. A. (2023). AI-Driven Optimization for Vendor-Managed Inventory in Dynamic Supply Chains.
13. Onukwulu, E. C., Agho, M. O., & Eyo-Udo, N. L. (2023). Developing a framework for AI-driven optimization of supply chains in energy sector. Global Journal of Advanced Research and Reviews, 1(2), 82-101.
14. Olagunju, E. (2022). Integrating AI-driven demand forecasting with cost-efficiency models in biopharmaceutical distribution systems. Int J Eng Technol Res Manag.
15. Hassan, N. A. B. (2022). Integrating Advanced Artificial Intelligence Models for Predictive Analytics and Real-Time Optimization in Global Supply Chain Networks. International Journal of Applied Business Intelligence, 2(12), 1-6.
16. Nazeer, A. (2021). AI-powered predictive analytics for supply chain optimization: A risk-resilient framework. International Journal of Emerging Trends in Computer Science and Information Technology, 2(1), 12-18.
17. Kalusivalingam, A. K., Sharma, A., Patel, N., & Singh, V. (2020). Optimizing inventory management with AI: Leveraging deep reinforcement learning and neural networks for enhanced demand forecasting and stock replenishment. International Journal of AI and ML, 1(3).
18. Kalusivalingam, A. K., Sharma, A., Patel, N., & Singh, V. (2020). Enhancing Supply Chain Visibility through AI: Implementing Neural Networks and Reinforcement Learning Algorithms. International Journal of AI and ML, 1(2).
19. Dash, R., McMurtrey, M., Rebman, C., & Kar, U. K. (2019). Application of artificial intelligence in automation of supply chain management. Journal of Strategic Innovation and Sustainability, 14(3), 43-53.