The application of AI to fa?ade optimisation with instantaneous dynamic simulation.

Introduction and Background

Passive design is the art of moulding a building’s fabric and form to create spaces that have a reduced energy demand and an enhanced environmental quality.

As the interaction of buildings with their content and environment is complex, dynamic simulation is often critical to the process of approaching an optimum passive design.

The process is used most frequently to determine if naturally ventilated, or mixed mode spaces will overheat, to improve space quality, and to predict energy demand for heating and cooling systems.

Dynamic simulation software is the gold standard in the detailed study of heat and mass transfer in buildings but there are challenges to its implementation:

·???????? Expensive software is required

·???????? The resulting accuracy is very dependent on the operator expertise

·???????? It is time consuming to build models and run simulations

As a result of these constraints, simulation is often used to test the perfomance of designs rather than forming an integrated part of the passive design optimisation process.?

Multiple simulations are required to identify optimum designs, and limitations of time, cost, and expertise are regular barriers to perfection.

Dynamic simulation has been in use since 1963 but only became widely used within the construction industry from the late 90’s.

The maths used in dynamic simulation software is intriguing but complex.? A model of the building and its systems is created in the form of a very large collection of simultaneous differential equations.? It is close to impossible to solve these equations explicitly, so the finite difference method is used which slowly steps through time iteratively, solving a series of partial differential equations. This makes the process relatively slow to complete even with modern computers.

Due to the above constraints, there have been many attempts made to simplify the calculation process.? These include the ISO 13790 method, which is used in SBEM, and the Passivhaus standard. It may come as a shock to some that the Passivhaus method uses the same calculation method as SBEM but there are some key enhancements in the Passivhaus method which result in an improved accuracy, particularly for energy estimation within high performance buildings.

Several other methods such as the CIBSE admittance method and the ASHRAE Radiant time series method have been used to produce rapid results with less computing power, but with a relatively poor accuracy, particularly for overheating calculations.

These methods are all less accurate than full dynamic simulations, particularly for overheating prediction but they are also relatively time consuming to complete, especially where multiple options must be tested as part of an optimisation process.

Instant dynamic simulation

If it were achievable, instant dynamic simulation would be invaluable for fa?ade optimisation as it would allow all possible configurations to be simulated and the energy and overheating effects to be analysed to determine the optimum arrangement.

Achieving instant dynamic simulation has however remained a 61-year-old mathematical challenge, as it has not been practical to achieve an exact solution to the equations involved.

Approaching the Challenge with Neural Networks

If we step back from the problem and consider it not as a process of cycling through a year’s weather data, but as a relationship function, then there is theoretically a direct relationship between input variables and outputs.? If a series of spot test simulations are produced, then a function that relates them could theoretically be developed.

While two dimensional functions can be generated for spot conditions, by traditional regression models such as polynomial regression, a dynamic simulation presents as a multi-dimensional function.? Multi-dimensional regression is very challenging when using traditional mathematical methods.

Neural networks are, however, particularly adept at tackling this type of challenge and it was hypothesised that they could potentially offer a solution to the provision of instantaneous dynamic simulations within a reasonable number of boundary conditions.

Attempting to Solve the Network

A neural network was programmed and? trained on data from thousands of dynamic simulations.? The simulations were generated using a python script that rapidly cycled through each scenario to build up the training data, as producing the quantity of data required would not be practical through traditional methods.

The network was tested and tuned until the results achieved were effectively identical to the dynamic simulations.

A typical classroom was used as a proof-of-concept model for the exercise, allowing a number of boundary conditions to be fixed.? While the neural network is capable of handling a large number of input variables there are some limitations on the number of variables that can be adjusted when preparing training data.? Five variables are however more than enough for a classroom optimisation exercise.? The constants that are fixed are typical for all new build classrooms.

Results

The result was an equation that can instantly and accurately predict the energy usage and overheating? results for a typical classroom for the five inputs in the following image:

The results have been shown to have a similar accuracy to the use of a full dynamic simulation process within the range required for optimisation.

The number of variables was then extended further using techniques such as the equating of variables with those used in the model and the introduction? of variables that have a linear relationship with the result.

Software

Once the method was validated, an intuitive software tool was developed that allows architects and engineers to enter the key details for a new school classroom and to achieve instant results.

The software saves a considerable amount of time as multiple fa?ade configurations can be tested with instant results produced.

The software also contains an advanced optimisation feature that allows the user to select any two inputs and view the outputs of energy, overheating, or daylight with all input variations, identifying the optimum point for each configuration.

Insights

Experimentation with the software resulted in a series of insights into the optimisation of modern classroom facades.

The software demonstrated that some historic assumptions in the optimisation of school designs were no longer valid since the introduction of heat recovery ventilation and heat pumps in schools.

A set of leaflets have been produced to share these insights and provide a quick guide to selecting optimum passive design conditions for both primary and secondary schools.

These leaflets will be circulated as part of a following post.

Conclusions

The use of neural networks allows architects and engineers to produce passive designs that are optimised to a level of detail that has not been practical with existing tools.

The number of conditions that can be checked and the speed of application allows insights and variable relationships to be gathered that have remained elusive when traditional spot checks simulation processes are performed.

This technique is not a complete replacement for dynamic simulation, particularly for existing buildings, but it offers an important additional tool for optimisation.

The tool produced is particularly useful for allowing architects and engineers to focus on the key variables that affect fa?ade optimisation without the complexities and distractions of using dynamic simulation.

The technique detailed could also be integrated into dynamic simulation software, allowing an optimisation process to take place once an initial simulation has been developed.

For further details please contact [email protected]

BDP Lab

This project was funded by BDP lab.

BDP lab, is a project that has been set up to support and encourage our staff in completing research and innovation projects that require investments beyond our normal project work.

Research projects and the exploration of new ideas is very important in the development of designs that improve quality and environmental performance within buildings.

This project is designed to foster innovation within the industry and assist in the development of better, more sustainable buildings.

?

?

?

?

?

?

?

?