Predicting solar radiation with Artificial Neural Network based on urban geometrical classification

Lila, Anas

Abstract

This research introduces the adaptation and development of an open source Artificial Neural Network (ANN) with the aim of predicting solar radiation falling on buildings for newly generated neighbourhoods in Aswan, Egypt as an example of a hot arid zone. The outcomes are a result of training the ANN on a database of classified urban geometries and their solar radiation simulation results for local weather conditions. The classification of this database was first introduced and discussed in (Lila and Lannon 2019). This paper discusses the different stages of developing the ANN code and its final version capabilities. The ANN code is developed to differentiate the training process from the prediction code to allow for the reuse of the trained ANN in multiple tests. The ANN code was tested for different database sizes to predict individual buildings’ solar radiation and was also used to predict solar radiation for urban configurations that were not part of the training process. The results of these ANN predictions were compared to conventional solar radiation simulation results to establish the accuracy and time saved. SCIENTIFIC INNOVATION AND RELEVANCE The focus on solar radiation is a part of a multi-stage proof of concept framework that produces a novel method to optimize performance based neighbourhood geometry. This study highlights the potential of applying ANN methods to predict solar radiation at the urban scale and the role it will play in allowing this prediction to be conducted for other simulation aspects at the neighbourhood scale i.e, energy performance, daylight availability and outdoor thermal comfort. The paper also discusses the limitation of this approach and future improvements to achieve higher efficiency. This process also utilizes an urban geometry classification method to build the training datasets for solar radiation simulation results. This research saves computational time to allow for performance based design decisions to be included in the early stages of urban design. These time savings will also pave the way for generating optimized neighbourhood geometrical options with solar radiation as a fitness function for this optimization process. Building an ANN that is based on novel geometrical classification of urban solar radiation is a new method to save simulation time and reach acceptable accuracy. PRELIMINARY RESULTS AND CONCLUSIONS The study used the Python programming language to build and develop the ANN code and a visual programming language platform (Grasshopper) to host the process of generating parametric models then simulating and predicting their performance. The solar radiation simulation was conducted using the Grasshopper plug-in Ladybug Tools. The differentiation between the training and prediction codes used the Pickle library which allowed for reusing the trained ANN in further generation and optimization studies of urban geometry. The final version of the developed ANN has shown a time saving of more than 90% when compared to the time consumed by traditional simulation for the same number of models. It achieved a coefficient of determination of 0.8 R2 value for unseen urban configurations and 0.9 R2 individual buildings prediction results correlation to simulation results These results have shown the capabilities of this method which allows for optimization between different neighbourhood geometrical iterations. This led to finalizing a proof of concept framework that can optimize neighbourhood geometry based on its solar radiation performance in the early stages of design with significantly less time than consumed for conventional brute force simulation methods

Citation

Lila, A. (2021). Predicting solar radiation with Artificial Neural Network based on urban geometrical classification. In 17th International IBPSA Building Simulation conference. Book of Abstracts (335-336)

• Conference Name: 17th International IBPSA Building Simulation Conference
• Start Date: Sep 1, 2021
• End Date: Sep 3, 2021
• Acceptance Date: Aug 1, 2021
• Online Publication Date: Sep 2, 2021
• Publication Date: Aug 1, 2021
• Deposit Date: Apr 26, 2022
• Pages: 335-336
• Series Title: 17th International IBPSA Building Simulation conference
• Book Title: 17th International IBPSA Building Simulation conference. Book of Abstracts
• Public URL: https://uwe-repository.worktribe.com/output/9192423
• Publisher URL: https://bs2021.org/wp-content/uploads/2021/08/BS2021-online-abstract-book-1.pdf