Summary of the context and overall objectives of the project (For the final period, include the conclusions of the action)
The European Green Deal Communication aims to transform the EU into a fair and prosperous society, with a modern, resource-efficient and competitive economy. Especially for the European Steel Industry digitalisation is a key enabler in this global restructuring process to realize a more efficient and sustainable industrial production.
Indeed, climate neutrality by 2050 as well as skyrocketing market prices for raw materials and energy are becoming ever greater challenges for Europe’s resource-intensive industries. Just in these times an optimized and target-oriented use of resources is mandatory, and a major objective of advanced production planning.
Production-planning systems (PPS) are widespread in modern steel production and an integral part of modern Enterprise Resource Planning (ERP) and Manufacturing Execution Systems (MES). However, there are usually significant differences between the assumptions made in the production planning process about the performance parameters of production resources and the reality of day-to-day operations. Furthermore, existing PPS lack the flexibility to deal with unforeseeable events frequently appearing in real industrial production. Modern production planning systems are not able to consider the current production state in their planning and this insufficient connection with the current production state inevitably leads to false assumptions resulting in wrong planning results and ultimately in loss of money.
Consequently this project aimed to improve flexibility of production scheduling in flat steel production through embedded real-time analytics of all available information coming from each plant involved. To reach this aim a DynReAct framework was implemented generating optimized production plans for each individual coil considering real-time plant performances.
This concept enables immediate reactions to critical situations like insufficient plant performance or coils that are outside the quality specifications. The optimal routing and sequencing is estimated using real-time-capable plant performance models derived from machine learning on large historical data, which are incorporated in multi-objective optimization methods. This approach enables the associated complexity to be mastered and other factors, such as quality, to be taken into account in scheduling in addition to the classic logistical targets.
The applicability of the DynReAct platform has been demonstrated at the tin-plate production site of the thyssenkrupp Rasselstein GmbH where there are multiple production steps (route pickling, cold rolling, degreasing, annealing, temper rolling, finishing) with free choice of different plants available.
The selected concept follows a hybrid scheduling approach combining three planning levels with different planning horizons as well as planning accuracies to provide robust production plans on the one hand and flexible reaction strategies in case of unexpected events or decreasing plant performances on the other hand.
As conclusion of the action, it can be stated that the optimization of production using efficient scheduling strategies provides a big potential for steel industry to improve their processes and accordingly to optimise the performance in general. Following the vision of Industry 4.0, the DynReAct project has enabled the creation of an advanced production planning system that includes the full use of all available information to dynamically respond to the current production situation and be able to predict and avoid potential production problems or bottlenecks in advance.
With the scheduling procedure and the plant performance models, the two basic pillars for the implementation of multi-objective production scheduling in the dynamic environment of the flat steel industry were implemented using a practical example. On the one hand, the hybrid solution with the planning pyramid and on the other hand the prediction models for the two dimensions quality and energy could be realized.
Work performed from the beginning of the project to the end of the period covered by the report and main results achieved so far (For the final period please include an overview of the results and their exploitation and dissemination)
The first half of this project was dedicated to the requirement specification and method development for dynamic product routing to enable the implementation of the DynReAct framework in the second half of the project.
The project started with the analysis of the existing planning system at the site of the industrial partner. For the local sequencing individual planning constraints exist for each production line. This includes hard constraints that have to be fulfilled by the planning and soft constraints that should be fulfilled if possible. A scenario approach to define functional requirements through the industrial use cases, using precise Use Case Cockburn standard template, was specified.
The main requirements and selection criteria for the product routing and the dynamic re-routing were investigated to avoid waste of materials, minimize costs and maximize productivity. Production planning at RAS is supposed to feed from five different systems at different planning levels which gather and serve a variety of information and data to help DynReAct to create real time planning strategies.
To guarantee the transferability of the solution to a wide scope in steel industry the flat steel planning was generalized and described as a flow shop with different production stages and unrelated machines in parallel i.e. flexible flow shop or hybrid flow shop.
At RAS the planning starts with the coil-order allocation where individual hot strips are allocated to related orders based on a target input material structure. For this problem a multi objective optimization technique was exploited, and the allocation problem was formulated as a binary integer linear programming. The proposed method consists in two phases: the first phase is used to find all the possible feasible pairings between hot rolled coils and production orders according to the parameters orders requirements and the coils parameters. The result of this first phase is a feasibility matrix that will be the input for the second phase whose aim is to solve the binary integer linear programming problem finding optimal pairing according the selected KPIs.
For the production planning use case, three different promising scheduling approaches were compared with each other. The results demonstrate that each method has its own specific advantages and disadvantages. It turns out that they complement each other excellently instead of competing. Thus, preference for a hybrid solution was the result of this comparison and a planning pyramid with three different levels has been created (Figure 2). This hybrid approach is expected to improve robustness and transparency of the production planning results considering a global production strategy. In the final solution implemented in the DynReAct project the production planning levels were solved by means of the following approaches:
Long-term planning (decomposition approach based on continuous flow model): In this order-less approach for the long-term horizon, the planning problem is simplified to a reasonable level. Hence, a manufacturing system can be approximately represented by a time-continuous or time-discrete dynamic model. The development of a dynamic model of a manufacturing system requires the definition of state (inventory at each production stage) and control variables (production). Besides, the capacity constraints are also considered in this model. The formulated continuous scheduling problem is an optimal control problem with state constraints, and thus can then be calculated in a rendering horizontal scheme. Different production routes of the relevant product groups can be taken into account in this mass flow method. At this level, the user can use these options to plan capacities and coordinate the product portfolio to be produced.
Mid-term planning (heuristic approach combining tabu search and travelling salesman problem (TSP)): To solve the mid-term planning problem at first a MOMILP approach combining Multi-Objective Optimization (MOO) techniques with a Mixed-Integer Linear Programming (MILP) formulation was proposed providing optimal solutions for the mid-term planning stage. However, the techniques require high computation time, which is a disadvantage that makes dynamic scheduling for continuous production difficult. Therefore, in order to provide robust and producible solutions for the production planners a new heuristic approach based on the combination of a tabu search algorithm with a TSP solver was implemented providing good order-based solutions in reasonable time.
Short-term planning (auction-based multi-agent approach): The auction-based multi-agent scheduling system is predestined for uncertain scheduling environments regarding resource availability and is characterized by its responsiveness, flexibility and robustness. On a negotiation platform, different agents represent production facilities, the coils to be produced, and auxiliary agents. Thereby, coils can bid on different auction processes at each resource with an available virtual budget depending on their status. Concerning the lowest virtual costs, the utility maximization of each resource agent is balanced with the coil agent’s objective. This approach acts towards the user as an adviser and delivers robust reaction strategies based on real-time process data.
Finally, based on the developed methods the DynReAct framework was designed by the partners and prototypically implemented at RAS. All components were integrated in a common framework as shown in Figure 3. As main interface to the current plant status a snapshot component was integrated in the final DynReAct system that provides all information relevant for the planning and the compliance with scheduling rules of the tin-plate production at RAS.
For the real-time consideration of plant performances individual interfaces to the required real-time data had to be provided. Realized as web-service in this service-oriented architecture (SOA) each individual plant performance model provides a plant KPI to the specific plant agent of the short-term planning. Since the finishing lines represent the highest complexity of alternative actions in the context of scheduling and offer the largest input of available process data, this production stage was selected as the focus for the plant performance models.
In a production with several production stages, each plant pursues its own individual goals, which make their own contribution to maximizing performance. This is due, for example, to the different characteristics of the plants, their respective semi-finished products or the corresponding processing methods. While some of these characteristics are fixed (e.g., construction-related), others are highly dynamic, difficult to specify, or even dependent on the coil to be processed (e.g., selected quality characteristics). As a result of this implied complexity, a deep analysis of historical data was performed to build a robust plant performance model with real-time prediction using supervised as well as unsupervised learning in combination with expert domain knowledge.
In the final production planning system two real-time plant performance models for the finishing lines have been implemented and can be considered by the short-time planning system on a per-coil base:
Quality Performance Dimension: For the plant performance model regarding quality, the use case of very sensitive products was taken. These materials are packaging materials that are in direct contact with dry foods such as milk powder and require a very bright surface. Small defects in the surface or any contaminants will cause the filling material to stick to the cans and cause quality control to fail. One of the main reasons for this quality problem are spots on the material surface coming from the production process. These products are produced on all tinning lines. In the present use case, the aim was to apply data analysis techniques to automatically identify problematic distributions of these relevant spots at an early stage in order to reduce or avoid the associated quality defects via rescheduling. The necessary surface data are supplied by automatic surface inspection systems from the production process. With respect to the online identification of problematic defect distributions, a deep embedding clustering (DEC) developed for industry applications was selected to perform this clustering task and trigger the impulse for appropriate rescheduling strategies in given cases.
Energy Performance Dimension: In the plant performance model for energy, the focus was placed on electrical energy, as this represents the main part of energy consumption at the finishing plants. The individual energy consumptions are collected online across all relevant components of each finishing line and linked to the accompanying material-piece-related production data for exploratory data analysis. This ensures the allocation of energy consumption per coil and order. In a first analysis, the individual features of the production data were examined with respect to their relevance for the variability of energy consumption at the individual finishing plants. Energy consumption over strip length was found to be the basis for measuring this energy dimension. This allows the effects of rescheduling measures or quality problems on this target variable to be recorded. A model based on the random forest approach was developed to predict the energy consumption of a coil on a selected finishing line.
Overview of the results and their exploitation and dissemination
In the RFCS project DynReAct a DynReAct framework has been implemented generating optimized production plans for each individual coil considering real-time plant performances. The applicability of this platform was demonstrated at tin-plate production where there are multiple production steps (route pickling, cold rolling, degreasing, annealing, temper rolling, finishing) with free choice of different plants available. The final release of this software platform has been successfully installed as demonstrator at the tin-plate production site of Rasselstein in Andernach.
The DynReAct long-term planning component was not fully implemented in the DynReAct prototype but verified in simulations (TRL 4). The final version for the RAS use-case simulation implements two liquids for the production flow which have different flow rates (t/h) due to different material widths.
Concluding, the TRLs reached for the single components of the DynReAct framework can be defined as follows:
Long-Term Planning (LTS) – TRL 4
Mid-Term Planning (MTS) – TRL 5
Short-Term Planning (STS) – TRL 5
Plant Performance models – TRL 4
DynReAct prototype with hierarchical planning system – TRL 4
The final DynReAct prototype installed at RAS exploits the MTS and STS components and is able to provide optimized plans for the finishing step in parallel with the production given the target weights for the single lines (result or the LTS) as input. In a successor pilot and demonstration project starting in July 2023 this prototype will be rolled-out to the full production and it is planned to publish a software platform for dynamic production planning in the steel industry and beyond under an open-source license.
Concerning further transferability of the developed solution the application of this technology in the metal industry is unquestionable. The DynReAct project suggests adopting the TIDE methodology for assessing the transferability of the solution to other industries manufacturing in a flexible flow shop production. The analysis considers key success factors, barriers, and potential mitigating actions.
Progress beyond the state of the art, expected results until the end of the project and potential impacts (including the socio-economic impact and the wider societal implications of the project so far)
Considering the main objective of this project to improve flexibility of production scheduling in flat steel production no approaches can be found in the international state-of-the-art covering all aspects of this concept adequately. Thus, this project was by far a significant step beyond the current state-of-the-art.
Referring to optimization algorithms, all described algorithms have been considered and implemented in different studies and projects to find an optimal schedule. The proposed problem decomposition by means of a hierarchical planning pyramid as described above which benefits from advantages of them in each optimization step is the strength and distinguished feature of this project.
European steel producers are facing increasing customer requirements in terms of product diversity, delivery capability, on-time delivery and constant quality whilst cost pressure continues to increase. Further, market conditions are characterized by increased variability, volatility, and complexity, which require from steel supply chains more agility than ever.
The industrial benefits of a commercial KPI driven planning and scheduling solution is stated by one PPS supplier as reaching better date fulfilment, delivery increase, higher productivity, less lead time and less planning time. This may give an idea of the potential of DynReAct taking into account that this standard PPS even does not consider any real-time plant performances. DynReAct provides a novel way of multi-criteria optimized and coil-oriented scheduling with an integrated knowledge-based approach in order to extract all required patterns relevant for the performance targets.
Another potential industrial impact when implementing dynamic production planning as developed in this research project may be an increased productivity and yield due to more homogeneous production plans, a minimization of setup-times and a reduced use of dummy and setup coils. Such increased productivity and yield again will increase resource-efficiency and competitiveness of European steel manufacturers and leads to a more sustainable production. Furthermore, an improved yield will lead to a decrease in energy consumption, waste, and CO2 emissions of the European Steel Industry. Additionally, a general improvement of the delivery performance will lead to a higher customer satisfaction, improving the image of the European Steel manufacturers and thus contribute to save costs as well.
Regarding the socio-economic impact it can be stated that DynReAct is able to transform the planning tasks from manual routine activities that rely on individual expert knowledge to the maintenance of planning constraints ensuring greater standardization and stability of the overall planning. Furthermore, trade-offs of relevant KPIs can be provided to the production team to understand correlations and gain production knowledge. This greater transparency regarding relevant KPIs again may lead to better overall planning results as well.
At the end of the project, the DynReAct solution was put into industrial practice by means of a prototype installation at RAS, which offers the industrial partner great potential for improving their processes and thus optimizing performance in general.