TL;DR
This paper introduces a novel approach to Master Production Scheduling (MPS) by combining advanced Machine Learning (ML)-based probabilistic demand forecasting with stochastic optimization. This method significantly improves upon traditional techniques by more accurately capturing demand uncertainty, leading to more robust and profitable production plans, especially in volatile supply chain environments.
What Problem Does This Paper Address?
Master Production Scheduling (MPS) is a critical process in supply chain management, responsible for coordinating production to meet demand while respecting capacity constraints. However, traditional MPS methods often rely on simplified forecasts (like point forecasts) and basic safety stock calculations, failing to adequately account for the inherent uncertainty and complex patterns in demand, especially for seasonal or non-stationary time series. This oversight can lead to inefficiencies, stockouts, excess inventory, and ultimately, lost profits. The paper aims to bridge this gap by developing a more sophisticated approach that leverages the power of ML for probabilistic forecasting and integrates it into a stochastic optimization framework.
Methods
The core of this research is a novel "Separated Estimation and Optimisation" (SEO) framework. This framework consists of two main stages:
- Probabilistic Demand Forecasting: Instead of simple point forecasts, the paper uses Machine Learning (ML) models, specifically Neural Networks (NNs), to generate probabilistic forecasts. This means the models don't just predict a single demand value but rather a distribution of possible demand values for each product in each planning period. This is crucial for capturing uncertainty. The paper explores different probabilistic distributions, including Gaussian, Student's T, and Negative Binomial, and finds that more flexible distributions like Negative Binomial often perform better.
- Stochastic Optimisation: The probabilistic demand forecasts are then fed into a Mixed-Integer Linear Program (MILP) formulation of the MPS problem. This stochastic optimization approach uses the demand distributions to find the best production plan that minimizes expected costs and maximizes profit across all possible demand scenarios, rather than just a single predicted outcome.
The paper compares this novel approach against traditional methods, including:
- DG-DET (Degenerate Distribution - Deterministic Optimization): Uses point forecasts and deterministic optimization, essentially treating demand as certain.
- RB-STO (Residual-based Distribution - Stochastic Optimization): Uses point forecasts and then estimates demand distributions based on forecast errors.
Datasets
The study utilizes a real-world weekly demand dataset from a leading global manufacturer of household appliances. This dataset is comprehensive, encompassing:
- 17 distinct assembly groups.
- 303 unique location-product combinations.
- A planning horizon of the latest 26 weeks for testing, with all preceding data used for training in a rolling manner.
This real-world dataset is a significant aspect of the study, allowing for evaluation under practical conditions and extending findings that have often been limited to synthetic data in previous research.
Key Insights and Ideas
- Probabilistic Forecasting is Key: The most significant insight is that accurately modeling demand distributions (not just point estimates) is critical for effective MPS, especially with volatile demand.
- ML Outperforms Traditional Forecasting: ML models, particularly Neural Networks, demonstrate superior ability to capture complex, non-stationary, and seasonal demand patterns compared to traditional methods like Exponential Smoothing (ETS).
- Beyond Gaussian: The assumption of a Gaussian distribution for demand errors is often too simplistic. More flexible distributions, like the Negative Binomial, can significantly improve performance by better approximating real-world demand patterns.
- Stochastic Optimization is Robust: The study confirms that stochastic optimization approaches are more robust and yield better overall outcomes than deterministic methods, even when faced with significant demand uncertainty.
- No One-Size-Fits-All: While NN-based probabilistic forecasting combined with stochastic optimization generally performs best, the optimal approach can vary by product or assembly group, suggesting the need for tailored solutions.
Key Results
- Objective Function Improvement: The proposed ML-based probabilistic forecasting with stochastic optimization (specifically NN-Neg.-Bin. and NN-RB) achieved the highest objective function values, indicating superior profitability. For instance, NN-Neg.-Bin. achieved the best result, with NN-RB closely following at 99% of the best performance.
- Robustness: The Interquartile Range (IQR) of the relative objective function value showed that NN-Neg.-Bin. (IQR of 40.1) was more robust across different products than NN-RB (IQR of 48.9).
- Forecast Accuracy: While no single method was best across all error metrics, LGBM and NN-based methods generally showed better forecast accuracy.
- Computational Effort: While stochastic optimization took longer (average of 8 min 39 s) than deterministic optimization (average of 2.4 s), the time was deemed practical for real-world applications, especially given the performance gains.
- Challenging Traditional Assumptions: The NN-Gaussian approach performed poorly, highlighting the limitations of relying on this distribution.
Limitations and Future Work
- Focus on Demand Uncertainty: The current study primarily addresses demand uncertainty. Future work could incorporate other sources of uncertainty, such as supply disruptions and production lead time variability, for a more holistic approach.
- Rolling Planning Horizon: The paper suggests exploring the proposed approaches with a rolling planning horizon, which is more representative of dynamic operational environments.
- Capacity Constraints: While capacity constraints were considered, further research could explore alternating capacity constraints to simulate under- and overage situations more dynamically.
- Feature Engineering: More sophisticated feature engineering for ML models could potentially further enhance forecasting accuracy.
- Alternative Distributions: Investigating a wider range of probabilistic distributions for the PM-STO group could yield further improvements.
- Tailoring Approaches: The paper notes that while general guidelines can be provided, experimenting with different approaches tailored to specific assembly groups or individual products is crucial for achieving optimal results.
Summary updated