Integrating production planning using APC and other technologies
Advanced process control can deliver value chain optimization by integrating various types of production planning in the hydrocarbon process industry
By Simon Rogers
The hydrocarbon industry, like many other process industries, has typically optimized its value chain using a siloed approach for supply chain planning, production planning, production scheduling, process control, production accounting, and other business processes.
With a siloed approach, each of the individual business processes has a limited view of the complete value chain, and each optimizes decisions according to its view. This typically results in nonoptimal performance due to the lack of a holistic view and integration.
For example, those involved with operational planning often have poor knowledge of actual plant conditions, use linear models with a limited range of validity, and employ a time-consuming updating and correction process. In addition, the logistics are not generally considered.
Production scheduling also usually happens with a limited knowledge of actual plant conditions, and optimal plans are not implemented due to logistics constraints and limited optimization capability in the scheduling system. This leads to limited economic consideration in the creation of the production schedule.
Operators often do not update advanced control strategies based on schedule changes, and they have little or no knowledge of intermediate feedstock pricing.
Production accounting must often rely upon poor quality source data, and a largely heuristic and manual reconciliation process, with the resulting information not being used effectively in the other business processes.
The solution to these and other value chain optimization issues is to change the focus from siloed optimization to an integrated approach. This article describes how to implement such an approach in general, and then examines a number of opportunities to use the latest digital technologies to support this integration.
Leaders use an integrated approach
An integrated approach to planning should be used to automate and increase the speed of decisions. Figure 1 depicts some of the main components of such an approach, with data and analysis results continuously and rapidly transferred among different business processes and organizations.
Unfortunately, most organizations do not operate in this manner, and instead use manual optimization processes. They also rely too heavily on limited resources, particularly their skilled subject matter experts (SMEs), who are always in short supply.
The linear models used in production planning have a limited range of validity, restricting their accuracy. Updating and synchronizing each of the tools used in production planning, scheduling, accounting, and process control and optimization is time consuming and relies on SMEs.
There is a largely heuristic data reconciliation process, which is also time consuming and introduces many opportunities for errors. Data is siloed and difficult to manage, with manual processes slowing data exchange efforts. There is often delayed or no recognition by operations of opportunities to expand constraints.
Data-driven optimization addresses these and other issues in a predictive rather than a retrospective fashion. This creates opportunities to expand the optimization envelope by challenging perceived constraints. These opportunities can be addressed by:
- Improving data quantity and quality using first principle models to reconcile data and calculate unmeasured variables.
- Using machine learning (ML) to perform multivariate analysis of planning, simulation, and measured data.
- Automating and simplifying the use of first principle models.
- Diagnosing performance on the basis of past experience.
- Using knowledge graphs to manage and visualize information.
By combining first principle simulation and machine learning, it is possible to:
- Automate data reconciliation and production accounting.
- Automate the comparison of measured, simulated, planned, and optimal key performance indicators (KPIs).
- Analyze changes in the KPIs over time to reveal improvement opportunities and determine when to update models.
- Combine synthetic and measured data in the generation of ML models.
- Automate the calibration of simulation models.
- Automate the generation of planning and scheduling models.
- Use refinery-wide simulation models to evaluate improvement ideas.
- Investigate and implement worthwhile ideas via management of change.
Figure 1. An integrated approach uses simulation and other tools to closely coordinate actions among the business units responsible for process industry operations.
Use case 1: Value chain knowledge graph
Knowledge graphs are the underlying technology used to support web search and digital assistants. Enterprise knowledge graphs are increasingly being used to structure and integrate information to support natural language programming. In many organizations, value chain optimization is split among different business units, each with a limited view of the overall activities. Value chain data is in different databases, spreadsheets, and other unstructured documents—creating data silos with poor integration.
Nonlinear programming can be used to extract valuable knowledge and metadata from the various sources of value chain data. Knowledge graphs can then be employed to manage, integrate, and visualize value chain information.
The knowledge graph in figure 2 shows the connections between information, and it provides a comprehensive and holistic view of the value chain. It can be used to improve the optimization and management of the value chain in real time and improve risk management. The knowledge graph can also be used to structure historical improvement opportunities, and to assist in identifying opportunities for improving value chain optimization using cognitive analytics.
Figure 2. Knowledge graphs can be created to show the required relationships among various business units, along with expected results from improving interactions.
Use case 2: Demand forecasting
Feedstock, product prices, and product demand are increasingly volatile. Optimization of the value chain often relies upon poor price and demand forecasts and is therefore suboptimal.
ML can use more historical data and a wider set of data to improve the forecast of prices and demand. High-performance computing and AI allow the optimization of multiple scenarios to reduce risk and increase the robustness of the plans.
Historical data can be used to determine the inputs that have the most impact on the demand, such as the season, and overall economic conditions as indicated by GDP, prices, and competition. A predictive model can be created with minimal error, and these predictions can be used for multi-period optimization of multiple scenarios.
Figure 3. Production scheduling can be automated by linking the business units responsible for baseline updates, and for the scheduling of crude, refinery, blending, and shipping operations.
Use case 3: Retrospective analysis
Planners within the organization want to look at expected market conditions and propose production plans to maximize profit. An event-based schedule meets the day-to-day plan targets as closely as possible given the logistical constraints.
Due to changes in market conditions, model mismatches, or unexpected events, the final operation often differs significantly from the expected plan. Better models and more robust plans can reduce the impact of unexpected events and reduce the gap between the plan and schedule.
Machine learning can be used to perform multivariate analysis of historical plans, schedules, and actual operations to identify and reduce repeating patterns of discrepancies. Recurring active constraints can be tracked, and rigorous refinery-wide models can be used to identify which ones would lead to a significant economic impact.
The ML analysis of plan versus actual can also identify when it is necessary to improve the planning and scheduling models, which are a simplified and often linear approximation of a more rigorous model. These linear models result in artificial constraints and do not reveal all optimization opportunities.
The models require periodic regeneration to account for changes in plant performance. Maintenance and recalibration of these models is time consuming and requires a high level of involvement from process specialists.
Artificial intelligence can be used to automatically detect when the models need to be regenerated and to automate the calibration of the first-principle, physics-based models, along with the generation of simplified planning models.
Figure 4. Plantwide optimization requires coordination among various APC and other automated controllers.
Use case 4: Automated scheduling using AI
Generating a production schedule from a plan is an intractable mixed-integer nonlinear optimization problem. Current production scheduling practices use multiblend optimization for crude and product recipes, along with event-based simulation.
As a result, scheduling is an iterative trial and error, manual process. Schedulers have limited time to analyze discrepancies between actual and expected opening inventories, and to then optimize the schedule or investigate alternative schedules and scenarios, resulting in a suboptimal schedule significantly different from the plan.
Historical operation and scheduling data can be used with ML to identify the likely cause of inventory discrepancies, improve the daily production accounting process, and adjust the schedule baseline. AI can also be used to automate the scheduling process itself by analyzing the constraint violations (e.g., tank levels or product qualities) in the simulated schedule, and to then update the schedule tasks to alleviate the constraints while minimizing the deviation from the plan. Figure 3 illustrates this methodology in more detail, showing the progression from baseline updates and initial discrepancies to the assignment of shipments.
Use case 5: Plantwide optimization
In many cases, advanced process control (APC) objective functions and constraints are not routinely updated, and individual unit APC applications typically operate independently with no coordination between units. APC models are only optimized within the operating envelope used during the initial step testing.
It is possible to use dynamic real-time optimization to combine multiple APC applications, allowing the operation of upstream units to adjust to constraints in downstream units. Online rigorous models can provide additional inferential inputs to the APC and update the gains used in the APC models in real time (figure 4). In addition, it is possible to automate the download of the objective functions and targets from the scheduling system to the dynamic real-time optimizer.
As demonstrated by these use cases, value chain optimization can be significantly improved using a combination of a knowledge graph, rigorous models, and AI. The goal is integrated optimization of the entire value chain, leading to significant increases in profitability.