Lorenzo Bucchieri, Alessandro Marini, Mattia Olivero | EnginSoft

This case study details the design optimization of an axial steam turbine of 160 MW, focusing on maximizing the total-to-total isentropic efficiency of the last three low-pressure stages of the turbine. Specifically, the engineers considered the shape and angle of the blades. After more than one century of development, it is the advances in blade design that have contributed to improved steam turbine thermal efficiency. Since modern turbines already reach efficiency values beyond 90%, extracting any further improvement is a very challenging task. The engineers used a combination of three engineering approaches for this study to accomplish this: traditional trial-and-error, virtual optimization and direct optimization. The final verification of the steam turbine after the optimization, achieved an isentropic total-to-total efficiency gain of about 0.5% -- a small but vital improvement in today’s highly competitive, highly regulated market.

This article discusses the design optimization of an axial steam turbine (with a rated power of 160 MW), focusing on maximizing the total-to-total isentropic efficiency of the last three low-pressure stages. The turbine, which is designed and produced by Franco Tosi Meccanica SpA, was optimized in collaboration with EnginSoft, by using a modeFRONTIER work flow to explore different designs and completely manage the fluid dynamics simulations. This simulation involves geometry generation, model pre-processing, solving and post-processing, all of which was achieved with the tools provided by ANSYS. Thanks to a strategic selection of input parameters, output values, targets and project constraints, a thorough exploration of the possible parametric space was achieved by using efficient optimization algorithms.

Figure 2 – Multi-stage steam turbine design by Franco Tosi Meccanica

Details of the last three low-pressure stages – Sealings and cavities

The first device to be classified as a steam turbine was the aeolipile, proposed by the Greek mathematician Hero, of Alexandria, in the 1st century. Other steam-driven machines were described in the following centuries: in 1551, by Taqi al-Din in Egypt; in 1629, by Giovanni Branca in Italy; and in 1648, by John Wilkins in England. No significant developments occurred thereafter until the end of the 19th century when various inventors laid the groundwork for the modern steam turbine. In 1884, a British engineer, Sir Charles Algernon Parsons recognized the advantage of using a large number of stages in series toallow the extraction of the thermal energy from the steam in small steps. The invention of Parsons’ steam turbine made cheap and plentiful electricity possible and revolutionized marine transport and naval warfare.
After Parsons, a number of other variations of turbines that work effectively with steam were developed. During the 1880s, Gustaf de Laval of Sweden constructed small reaction turbines that rotated at about 40,000 RPM. From 1889 to 1897, de Laval built many turbines with capacities ranging from about 15 to several hundred horsepower. Auguste Rateau of France first developed multistage impulse turbines during the 1890s. At about the same time, Charles G. Curtis of the United States developed the velocity-compounded impulse stage. One of the founders of the modern theory of steam and gas turbines was Aurel Stodola, a Slovakian physicist, engineer and professor at the Swiss Polytechnical Institute in Zurich.

By 1900, the largest steam turbine-generator unit produced 1.2 MW; ten years later the capacity of such machines had increased to more than 30 MW. This far exceeded the output of even the largest steam engines, making steam turbines the principal movers in central power stations after the first decade of the 20th century. Steam turbines also gained pre-eminence in large-scale marine applications, first with vessels burning fossil fuels and then with those using nuclear power.

Despite the introduction of many alternative technologies during the intervening 120 years, currently it is estimated that more than 80% of the world’s electricity is generated using steam turbine systems driving rotary generators. The steam to drive these turbines is raised by burning fossil fuels-- mostly coal but also oil and gas (~65%), or by using nuclear power (~15%). Less common thermal sources for steam generation are solar power and geothermal energy. Because of its ability to develop tremendous power within a relatively small space, the steam turbine has superseded all other prime movers, except hydraulic turbines, for generating large amounts of electricity and for providing propulsive power for large, high-speed ships. Today, units capable of generating up to 2 GW of power can be mounted on a single shaft.

This article will focus on how Franco Tosi Meccanica and EnginSoft collaborated to find an answer to this last question.
Beginning with Parson’s concept and after more than a century of development, advances in blade design have contributed to improved steam turbine thermal efficiency. Considering that the efficiency of modern turbines can reach values beyond 90%, it is clear that any further improvement is very challenging to accomplish.
Theoretical methods and experimental tests are very useful in predicting and verifying the performance of every new design or redesign. However, the classical approach of “trial-and-error” using many experimental tests is very costly in terms of time and money and is also unable to identify exactly how to improve performance.
With the greater availability of substantial computer power and efficient numerical algorithms, Computational Fluid Dynamics (CFD) becomes an essential tool for engineers by enabling a wide variety of complex flow situations to be simulated, reducing the amount of testing required, increasing understanding, and accelerating development. As a result, CFD is now an established industrial design tool, helping to reduce design time-scales and improve processes throughout the engineering world.

Today, the designer has to cope with two key challenges to compete in the market: competitive sales targets and strict energy efficiency regulations. In this complex scenario, the designer must thus focus every day on “raising the bar”, knowing that an increase in performance of just a few percent often makes the difference. A design tool that is able to provide accurate, reliable, and automated predictions of the fluid flow behavior in steam turbines is essential to enable Franco Tosi Meccanica to gain a new competitive edge in the market. In this context, ANSYS proves to be a high-fidelity Computer-Assisted Engineering (CAE) tool to meet a turbine designers’ needs.

The last three stages of a low-pressure steam turbine are the focus of this study (Figure 1 and Figure 2). The objective was to optimize the three statoric rows by maximizing the total-to-total isentropic efficiency of the device for a given operating condition. The turbine blades’ shape is defined by the position of five airfoils arranged in a spanwise direction. A Bezier distribution for both the stacking angle (for flow incidence control) and the bowing angle (for flow separation control) was used to recreate the blade shape. Another variable was used for the number of blades in the statoric row. See Figure 3 and Table 1 for details. A parametric model is generated with these characteristics in ANSYS DesignModeler. From this point, using the ANSYS BladeEditor features, it was possible to extract the computational domain for a single passage. In fact, the first assumption made here was to consider just a single passage instead of the full wheel (Figure 4). The rotating speed compared to the mean stream velocity makes it well-suited to a mixing plane approach. Consequently, we obtained a huge advantage in terms of the reduction of the computational costs, both as a result of the size of the CFD model and of the steady state approach with the mixing plane.

Fig 3 | Blade geometry – Description of parameters

Figure 4 | Three stage, single passage – CFD model and Figure 5 | Three stage, single passage – Computational grid

Another assumption made was the simplification of the actual geometry of the steam turbine: the sealings, cavities, and rotor were left out of the computational domain. This was necessary to enable a completely automatic optimization workflow, in particular for the computational grid generation of ANSYS TurboGrid (Figure 5). A grid independency study was performed to apply the best compromise between speed and accuracy. The result was a very fast and robust procedure able to achieve a high-quality mesh and a well-defined boundary layer treatment for every configuration that was identified with a unique parameter set or design point.

The result of the optimization lay in the geometry for the three statoric blades. This generation depended on the simplified (“ideal”) layout of the flowpath considered here. After the optimization campaign, a final comparison between the “ideal” flow path and the actual one was performed to measure the losses created by the sealings and cavities.

Another assumption made was to consider each stage independently from the other in the optimization. In this way, the three stages, namely L-0 (the last one before the diffuser), L-1, and L-2, were treated separately in three different optimization stages. The boundary conditions for each CFD model were obtained by combining 1D data supplied by Franco Tosi Meccanica with a preliminary analysis of the single passage, three-stage turbine performed by EnginSoft.

An additional analysis of the single passage, three-stage turbine was performed after each optimization stage. The baseline geometry of the statoric blade was replaced with the geometry that resulted from the current optimization stage. In this way, we could verify if the new layout of the turbine performed better than the baseline configuration. Once this was assured, the optimization continued to the next stage.

Table 1 | Blade geometry – Description of parameters

Table 2 | Operating constraints (with respect to the baseline configuration)

Fig 6 | Reference for post-processing

ANSYS CFX was used to set up and solve the CFD analyses. The fluid flow was considered in steady, compressible, and turbulent conditions. The advection term was resolved with a “High Resolution” scheme (bounded second order accuracy). The RANS 2 equations for Shear Stress Transport (SST) was chosen for the turbulence model. The IAPWS library was used to characterize the steam as a real fluid. The liquid-vapor transition phase was considered under conditions of equilibrium. The Stage (or mixing plane) approach was considered for a multiple frame of reference (MFR). The numerical setup was optimized to achieve a good level of convergence in less than 100 iterations. In this way, each design point selected in the optimization campaign took just about 25 minutes to estimate (from the selection of the parameter set to the output of the post-processing procedure).

The results collected by the post-processing procedure were useful to understanding if the new design performed better compared to the baseline configuration. The objective was to maximize the total-to-total isentropic efficiency:

Several operating constraints had to be satisfied to guarantee the feasibility of a design point. For the particulars of these constraints, see Figure 6.

The choice of input parameters, output values, targets and project constraints greatly define an optimization process. Building on this point, several techniques can be selected to define which kind of optimization method is appropriate to apply to an engineering problem, especially in terms of time and cost. Traditional engineering based on “trial-and-error” was widely used in the past when automatic optimization tools were not available. Starting from a baseline configuration, the design was perturbed to get a new design point (hopefully with improved performance). The aspects of this perturbation were selected and applied by the designer and were fundamentally based on his experience. This process was repeated iteratively until the desired performance target was reached. From this description, it becomes clear that this approach cannot ever be completely automated because the decisions are made by the designer, and it is a time consuming and costly method. A schematic representation of this approach is shown in Figure 7.

Fig 7 | Optimization methods – Trial-and-error approach

Fig 8 | Optimization methods – Direct optimization approach

Fig 9 | Optimization methods – Virtual optimization approach

The evolution of this concept lead to the birth of modern optimization approaches, where algorithms replaced the designer in the selection of the new design points to be evaluated. Using efficient optimization algorithms, the parametric space can be explored broadly and intelligently in a completely automated way, in much less time compared to the traditional “trial-and-error” approach.

Two different modern approaches are available:

- Direct optimization: after an initial exploration of the parameter space through a Design of Experiment (DoE), each of the design points is explicitly evaluated, with the intention of reaching a desired target. See Figure 8;
- Virtual optimization: after an initial exploration of the parameter space through a DoE, a response surface is generated to provide a continuous representation of this space. In turn, this response surface is explored using a large number of virtual design points (not explicitly simulated) in order to find good candidates. These good candidates are then explicitly evaluated, enriching the database with which the response surface can be redefined. This process is repeated iteratively until the potential inaccuracy of the response surface decreases below a certain threshold. See Figure 9.

For both the direct or virtual optimization approaches, an optimization software that is able to control the workflow that automatically manages both the exploration of different design points and the management of the fluid dynamics simulations (such as geometry generation, model pre-processing, solution and post-processing by means of the CAE tools provided by ANSYS) is necessary. The software used for this purpose is ESTECO’s modeFRONTIER.

The modeFRONTIER’s optimization process could be split into 3 steps:

- The first step includes the creation of a logical workflow to graphically formulate the engineering design problem at hand, i.e. how the simulations must be performed and in which order. The “how” implies the choice of values/measures to be used and generated (inputs and outputs), the definition of the optimization objectives and the configuration of the most appropriate algorithms to explore and optimize the design space. A representation of the optimization workflow that was defined for this study is shown in Figure 10;
- The second step consists of the evaluation of the designs, as defined by the workflow. The evaluation, or “run”, can be monitored in real-time by means of charts and graphs, and with direct access to log and process files;
- The final step is the assessment and visualization of the results. The available tools allow an understanding of a problem’s important parameters on the basis of the exploration of the design space, reducing the number of significant parameters to be considered, making the optimization more efficient, re-arranging the data in a comprehensible manner, and extracting a clear meaning with which to make informed decisions. Specific analysis tools help convey relevant insights into the interaction effects and to visualize optimization trends. The Response Surface Models (RSM) tool allows for the training, comparison and validation of meta-models, speeding up the entire optimization process.

Table 3 | Optimization strategy

Table 4 | Results of the optimization campaign

In this study, all three different approaches were used, as summarized in Table 3.

In the classical “trial-and-error” approach, all the design points (almost 100) were explicitly evaluated by Franco Tosi Meccanica’s designers on their local workstations. This optimization stage took about one month to be accomplished.
For the optimization stages in which L-1 and L-2 were studied, EnginSoft adopted an automatic optimization procedure using modeFRONTIER™. The Multi-Objective Genetic Algorithm (MOGA-II) was the optimization algorithm selected for this campaign. The optimization stage for L-1 required the evaluation of about 1,000 design points, while for L-2 about 5,000 CFD calculations were performed. All these computations were performed on the EnginSoft cluster. A concurrent design-points strategy was adopted to reduce the total wall-clock time, so both optimization stages took about one month to complete.

The results of the optimization campaign are summarized in Table 4. The second column represents the improvement of the total-to-total isentropic efficiency evaluated for the respective turbine stage. In other words, these are the results of the three optimization stages. The last column represents the results of the additional analyses of the single passage, three-stage turbine that were performed after each optimization stage. As mentioned before, in these analyses the baseline geometry of the statoric blade was replaced by the geometry resulting from the respective optimization stage.

Even though the optimization stages were independent from each other, it is clear that a good overall trend was achieved. It is important to highlight that the efficiency of the baseline design was already quite high (above 90%). From this perspective, the results achieved are very remarkable.

The last phase of the study was the evaluation of the single passage, last three-stage steam turbine with its actual flowpath, i.e. including the sealings and cavities (see Figure 13 and Figure 14). A comparison between the baseline and optimized geometries of the statoric rows was performed. In real conditions, the improvement of the total-to-total isentropic efficiency achieved is about 0.5%.

Figure 10 – ESTECO modeFRONTIER – Optimization worklow

Figure 11 – Comparison between baseline (left) and optimized (right)

design of the statoric blades – Blade geometry

Figure 12 | Comparison between baseline (left) and optimized (right) design of the

statoric blades – Pressure field on blades

Figure 13 | Three stage, single passage – CFD model with actual sealings and cavities and

Figure 14 | Three stage, single passage – Computational grid with actual sealings and cavities

Figure 15 – Three stage, single passage – CFD model with actual sealings and

cavities – Static pressure field comparison @ 50% span – Baseline (above) and optimized (below)

Figure 16 –Three stage, single passage – CFD model with actual sealings

and cavities – Mach number field comparison @ 50% span – Baseline (above) and optimized (below)

The objective of the study was to optimize the last three low -pressure stages of a steam turbine. A global optimization was performed on the system. Using the geometry of the statoric blades, the purpose was to identify the optimal performance in terms of isentropic total-to-total efficiency. The optimization strategy adopted here was defined in three different stages with different approaches:

- Trial-and-error for L-0
- Virtual optimization (by means of a Response Surface Method) for L-1
- Direct optimization for L-2

Once the optimal design for each stage was selected, it was positioned in the three-stage steam turbine to confirm the improvement of the system. After the three optimization stages, a final verification of the steam turbine was performed, considering the real flow path with the sealings and cavities. In this complex scenario, the isentropic total-to-total efficiency gain is about 0.5%.

It is worth noting that, as summarized in Table 3, each optimization stage required the same amount of human engineering time (~1 month). This time also includes all the CPU time and the man-hours involved. In the final optimization stage, the engineer’s time is almost close to zero while the simulations are run automatically. Given that each design takes about 25 minutes to run on the EnginSoft cluster, and three concurrent design points were running simultaneously during the optimization stage, the total CPU time was about 1 month.

In a typical medium-sized enterprise where HPC computing clusters with over a hundred CPUs are now quite common, it is clear that such times could be dramatically reduced. For example, with a medium installation of 256 cores, the same job could be performed in just one week! This is a remarkable result, especially if we think about the time that can be saved in the development of a large machine like the one considered in this study.

CASE STUDY

The Arvedi Group approached the University of Trieste to find a solution to the uneven distribution of material inside the hopper of their blast furnace in Trieste, Italy.

optimization modefrontier rocky mechanics