Data-Driven Simulation and Optimization Approaches To Incorporate Production Variability in Sales and Operations Planning

Bruno A. Calfa, Anshul Agarwal, Scott J. Bury, John M. Wassick, Ignacio E. GrossmannIndustrial 2015 · DOI

TL;DR

This paper introduces two novel data-driven frameworks to better plan production in the face of variability, a common challenge in manufacturing. By using historical data to model production deviations and then employing simulation-optimization or bi-objective optimization, companies can generate more robust production plans that balance profit with risk.

What Problem Does This Paper Address?

Sales and Operations Planning (S&OP) is a critical process for aligning a company's strategic goals with its daily operations, ensuring demand is met by supply. A significant challenge in S&OP, particularly in the operations planning stage, is production variability. This refers to the difference between planned production rates and actual achieved rates, often due to unforeseen events. Ignoring this variability can lead to suboptimal plans, unmet demand, and financial losses. While existing literature has addressed demand uncertainty, incorporating production variability into tactical production planning within S&OP has received less attention. This paper aims to fill that gap by providing data-driven methods to explicitly account for this variability.

Methods

The core of the proposed methodology involves two main steps:

  1. Modeling Production Variability: The authors use quantile regression to model the distribution of deviations between planned and actual production rates. This statistical technique is powerful because it can capture how the variability (i.e., the spread of deviations) changes depending on the planned production rate. Historical data of planned vs. actual production is used to train these models.
  2. Incorporating Variability into Planning: Two optimization-based frameworks are proposed:
  3. Simulation-Optimization (Sim-Opt): This approach alternates between a Derivative-Free Optimization (DFO) solver that proposes production targets and a simulator that evaluates these targets under various production variability scenarios. The scenarios are generated by sampling from the quantile regression models. This method is flexible but can be computationally expensive due to the number of simulations required.
  4. Bi-Objective Optimization (Bi-Opt): This framework directly incorporates production variability into a two-stage stochastic programming model. It simultaneously optimizes for two objectives: maximizing expected profit and minimizing a chosen risk measure (e.g., variance, semivariance, or CVaR). This method offers explicit handling of constraints and risk but may lead to larger and more complex optimization models.

The paper also discusses how to handle dependencies between plants in highly integrated networks, proposing methods like linear/kernel regression, joint distribution estimation, and bootstrapping.

Datasets

The paper uses two types of data for evaluation:

  • Motivating Example: A small, hypothetical process network is used to illustrate the concepts.
  • Industrial Case Study: A real-world production planning problem from a chemical company is used to demonstrate the practical applicability of the proposed methods.

The data used for modeling production variability (deviations between planned and actual production rates) is historical production data.

Key Insights and Ideas

  • Quantile Regression is Key: Using quantile regression to model production variability is a significant contribution, as it allows for a more nuanced understanding of how variability changes with planned production levels, rather than assuming a constant variance.
  • Trade-off Between Profit and Risk: The study clearly demonstrates that accounting for production variability leads to a trade-off between average profit and risk. Plans that are optimal under deterministic assumptions can be risky when variability is considered.
  • Bi-Objective Optimization Offers Better Performance: The Bi-Opt framework generally yields higher profits and better service levels compared to the Sim-Opt approach, especially when risk is explicitly managed.
  • Integrated Networks Require Special Handling: In highly integrated chemical plants, production variability in one plant can directly impact others. The paper highlights the need for modeling these interdependencies, moving beyond simple independence assumptions.
  • Reliability vs. Margin: The motivating example cleverly illustrates how companies might allocate resources to less profitable but more reliable plants when risk is a significant concern, a counter-intuitive but rational decision under uncertainty.

Key Results

  • The Bi-Opt framework consistently produced more profitable and less risky production plans compared to deterministic or risk-neutral stochastic approaches. For instance, in the motivating example, Bi-Opt achieved a higher average profit (320.60 m.u.) with a lower standard deviation (25.82 m.u.) than a risk-neutral approach with fixed targets (272.84 m.u. profit, 28.30 m.u. std dev).
  • The Sim-Opt framework was found to be more computationally intensive than Bi-Opt, requiring significantly more time and function evaluations to achieve comparable or sometimes inferior results.
  • In the industrial case study, the proposed methods for handling plant interdependencies were shown to be crucial for avoiding infeasibilities in highly integrated systems.

Limitations and Future Work

  • Computational Cost: While the Bi-Opt framework is more efficient than Sim-Opt, both can still be computationally demanding, especially for large-scale problems with many scenarios.
  • Data Availability: The effectiveness of the data-driven approach relies on the availability of sufficient historical production data to accurately model variability.
  • Future Work: The authors suggest exploring more advanced risk measures and extending the methodology to account for other sources of uncertainty beyond production variability. Further research into more efficient algorithms for handling complex, integrated networks is also implied.

Summary updated

Abstract

We propose two data-driven, optimization-based frameworks (simulation-optimization and bi-objective optimization) to account for production variability in the operations planning stage of the sales and operations planning (S&OP) of an enterprise. Production variability is measured as the deviation between historical planned (target) and actual (achieved) production rates. A statistical technique, namely, quantile regression, is used to model the distribution of deviation values given planned production rates. Scenarios are constructed by sampling from the distribution of deviation values and used as inputs to the proposed optimization-based frameworks. Advantages and disadvantages of the two proposed frameworks are discussed. The applicability of the proposed methodology is illustrated with a detailed analysis of the results of a motivating example and a real-world production planning problem from a chemical company.
BibTeX
@inproceedings{calfa2015data,
  author = {Bruno A. Calfa and Anshul Agarwal and Scott J. Bury and John M. Wassick and Ignacio E. Grossmann},
  title = {{Data-Driven Simulation and Optimization Approaches To Incorporate Production Variability in Sales and Operations Planning}},
  booktitle = {Industrial},
  year = {2015},
  doi = {10.1021/acs.iecr.5b01273},
}

Loading related content...

Figure