This project involves the development of automatic dynamic control of fire modelling simulations.
Automatic dynamic control of the CFD solution process is being pursued in fire modelling applications for two primary reasons. Firstly, with such a system it may be possible to automate the control of CFD simulations and dynamically adjust the control parameters in order to ensure the stability and convergence of fire field model simulations. In this way, the technique would remove some of the “black art” associated with fire field modelling and make it accessible to a wider audience. Secondly, it is hoped that through the optimal setting of control parameters, these techniques will also lead to a reduction in simulation times. The development and use of these techniques is being explored through the use of the SMARTFIRE fire field model.
At the heart of fire field models is a control-volume based CFD code. This is used to simulate the fluid dynamics of fire development by numerically solving a set of differential equations that describe the laws governing the physical processes inside the domain. The simulated domain is meshed and the solution is found by performing a series of consecutive approximations, whose quality is measured by the residual error defined as the magnitude of change between two adjacent approximations. The residual error is calculated separately for each solved variable and the time step converges if the residuals of all the variables fall below the specified tolerance.
Early attempts by FSEG at achieving automatic dynamic control were based on a rule-driven system applied to two-dimensional fire scenarios. This system was able to reduce the total number of sweeps by 50% for relatively simple cases. However, the control techniques failed to provide any tangible benefits in more complex three-dimensional cases.
Following further analysis of the problems that plagued the earlier system, a new architecture based on a heuristic search was proposed and subsequently implemented. The new system (called ICS – Intelligent Control System) controls both the relaxation and the time step size while also adjusting the maximum number of iteration in order to guarantee full convergence of every time step. All the changes to the simulation control parameters are applied between the time steps and are preceded by a comprehensive heuristic search in order to obtain the best control parameters. During the search the simulation does not advance since a single time step is repeated several times using different control parameters. The relevant results (i.e. residual error graphs) are stored and then analysed using a sophisticated assessment algorithm. The algorithm determines which of the experimental time steps provided the best results and then applies the control parameters from that experiment in the subsequent time steps.
There are two alternating stages in ICS-controlled simulations:
During testing the ICS proved to be successful in guaranteeing full convergence of all time steps and in improving the overall stability of the simulation. However, the cost of the search proved to be very significant and therefore the initial system was impractical for performing full-scale CFD simulations. Nevertheless, the results were very encouraging and, with the new knowledge that emerged as a result of the tests, the next version of the ICS was developed with the goal of providing the required speed up while still guaranteeing the full convergence.
Consequently, the control system was refined in order to minimise the search cost. For example, one set of modifications provides speed up when the simulation is progressing smoothly while another adjustments help recover from divergence. In order to use this technique to limit the search, the control system must be able to perform accurate state recognition. Four significantly different states within a simulation were identified (Divergence, Slow Convergence, Oscillations, Normal Simulation), each having a specific set of modifications associated with it. The implementation of this procedure reduced the search cost by 70%.
The results from a non-trivial fire modelling case are included here as an example of the benefits that ICS provides. The scenario consists of a single compartment with a single door. A large fire is situated in the centre of the room, directly on the floor. The fire has a peak output of 2.1MW at t=450s and is defined by a growing heat release function. While a combustion model was not used in this example, there is no fundamental reason way one could not be used with this technique. The mesh contains approximately 20,000 cells and the initial time step size is 5s. Since the conventional i.e. non-controlled simulations experienced major problems with convergence, two different set-ups were used (dt = 5s, and dt = 1s). For the time step size of 5s, over 50% of the time steps failed to converge. When the time step size was reduced to 1s, the number of non-converged steps dropped to around 10%. The ICS-controlled simulation had the same initial set-up (initial time step size 1s) with the heuristic search performed after every 25s of the simulated time.
Figure 1 shows the number of sweeps required to complete every 5 seconds of the simulated time for the three differently controlled simulations. One of the graphs (no ICS, dt=5s) has a clear cut-off line at 500 sweeps. This was the maximum number of sweeps allowed for each time step (the number was fixed since the simulation was non-controlled) and therefore all the time steps that reached that number did not fully converge. More than 50% of the time steps failed to converge with a time step of 5s. In the second uncontrolled case (no ICS, dt=1s), 10% of the time steps diverged. In this case the graph shows no clear cut-off line as the scale in the graph required consolidating sweeps from 5 consecutive time steps (since the time step size in the simulation was 1s). The third curve in this graph shows the results from the ICS-controlled run (without the search cost included). It can be clearly seen that at each point the ICS-controlled simulation required substantially fewer sweeps to simulate any given 5-second period. With a 1s time step, the conventionally controlled system required 43,844 sweeps to complete the simulation while the ICS-controlled simulation required 17,641 sweeps (including experimental sweeps). The reduction in sweep number corresponds to the reduction in execution time since the overhead introduced by the ICS is minimal (the processing overhead without the search cost which is substantial and is included in improvement analysis).
Figure 1. Number of sweeps per 5s of the simulated time (for three different runs)
In our example, using a time step of 1s, there is a time saving of 60% in using ICS. It must be noted that the search cost consists of the time spent performing the experiment and the time spent assessing them but the cost of the assessment is negligible and therefore not used in improvement calculation.
The main improvement gained from the use of the ICS is the fact that all the time steps converged therefore ensuring the accuracy of the results. The control system was always able to recover from any problems, especially divergence. Consequently, the results are believed to be more accurate than the ones from standard simulation although a formal study is still to be conducted. In addition there were substantial savings in execution time as compared with non-controlled run. There was a 60% real reduction in simulation time.
The SMARTFIRE ICS system has been tested on a number of examples. In each case, using the ICS system leads to a substantial reduction in simulation time and more importantly, has produced complete convergence (for each time step) and can effectively recover from any solution excursions/faults.
For information relating to the SMARTFIRE fire field model visit our
web pages. You will find a range of publications relating to
SMARTFIRE and our other research on the FSEG Publications
pages both CMS press and external publications.