Projects

Optimization of Hybrid Powertrains of Trains

Summary

Hybrid solutions are needed to reduce both the manufacturing costs of trains and the CO2 emissions of rail transport. Only those who succeed in doing so will be able to win tenders from the federal and state governments in the long term. Deciding which components to use and how to dimension them is a major challenge. Prosim efficiently finds the best individual train designs for each tender.

Problem

A large proportion of the routes in European regional transport are still served by diesel vehicles. However, current technology can no longer meet the requirements of pollutant emission limits. Even today, pure diesel vehicles are not allowed to enter tunnels and indoor stations. Battery-powered vehicles appear to be the solution for regional traffic for the time being. On closer inspection, however, it becomes clear that batteries are simply too large and expensive for the high energy volumes required by a train weighing up to 100 tons. Until now, the only choice was between an expensive, pure electric and the existing "cheap" diesel engine version.

Previous Problem Solution

Hybrid vehicles are the solution between the two extremes mentioned above. They allow quiet and emission-free operation in sections, avoid expensive idling of the diesel engine and can recover up to 60 of the energy expended during braking. Overall, a hybrid system can consume up to a quarter less energy than its counterpart with a pure combustion engine. Developers are now faced with the following question: How should the battery and diesel engine be dimensioned so that they meet both targets in the best possible way? This question is not easy to answer because there are many interdependencies. For example, the power requirement is not immediately apparent, as it depends on the vehicle mass. Conversely, the vehicle mass is in turn determined by the size of the battery and engine. Experts use complex simulation models for such problems to quantify alternative solutions. However, each simulation takes time. Even with two optimization variables (battery size and diesel engine performance), there are already many alternatives. If more are added, the number increases exponentially. On each train connection, there is an individual conflict: What distance (range) must my multiple unit train be able to cover and what CO2 guidelines must be met? If, for example, you want a vehicle that emits as little CO2 as possible on average for 20 different routes and at the same time has moderate costs, you need to run about 500,000 simulations to identify the best trade-off solutions. With a simulation time of 2 minutes per alternative, the pure simulation effort is just under 2 years. Such a process would be resource-intensive and therefore not economical, since employees have to start and evaluate all simulations manually.

Prosims Efficient Problem Solving

Prosim's software is linked to the existing simulation model via an interface and generates Pareto-optimal solutions for your target variables automatically and efficiently: Costs and CO2 emissions. From these, the customer can choose the most suitable trade-off when purchasing a train. Now the customer can decide how his train should be dimensioned to meet the environmental requirements while keeping the expenses as low as possible, taking into account his requirements for the route, which would be range and CO2 guidelines in this case. The customer has the results on his dashboard after only 3 days!

Optimization of Neural Networks for Bee Tracking

Summary

In order to bring the selected performance indicators of a neural network into the desired range for the application, an optimal adjustment of the hyper parameters is required. Which of these parameters are set in which way and in which combination is used is a time-consuming and, on a multi-criteria level, non-trivial challenge. Prosim individually finds the best compositions of the hyperparameters for neural networks, efficiently and time-saving.

Problem

The hyperparameters of a neural network are preset variables that define the architecture and control of a neural network. The choice of these hyperparameters is crucial for the successful training and subsequent performance of the algorithm. Depending on the size of the data set and its complexity, the training of a net can take several hours or even days. Only after a training run has been completed can performance indicators such as precision and recall, but also the reaction speed of a net, be evaluated. There is an immense variety of possible hyper parameter combinations. With ONLY 8 hyperparameters alone, the number of possible combinations can amount to over 8 million possibilities. To evaluate all these serially would take about 4000 years with an average training duration of 5h. Therefore, the question arises: How can we efficiently find those combinations of the hyperparameters that achieve an acceptable trade-off of the target indicators?

Previous Problem Solution

The hyperparameter optimization currently prevailing in the technical literature is based on the one hand on empirical values and the resulting "trial and error" approaches and on the other hand on the possibility of a unicritical optimization of hyperparameters. These methods are time-consuming and cannot guarantee to find all possible pareto-optimal combinations of parameters for several target indicators. But only with such solutions, the trade-off desired for the individual application can be selected from the multitude of possibilities.

Prosims Efficient Problem Solving

Prosims software usually requires simulation models for optimization. However, the software is also perfectly suited for the optimization of neural networks. Thereby, the next nets are trained with automatically parameterized hyperparameters. From the performance of the nets, the Prosim software derives correlations between the hyperparameters and the performance. Based on these correlations the Prosim software can approach the Pareto optimal solutions with a relatively small number of trainings.