A hybrid framework for optimal design and control strategies of retractable roof stadium based on CNN-LSTM and NSGA-II

Abstract

This study addresses the dual challenges of time-varying environmental responses and multi-objective optimization in retractable roof stadiums. We propose a hybrid framework integrating Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) and Non-dominated Sorting Genetic Algorithm II (NSGA-II) to resolve competing environmental performance objectives, specifically addressing daylight and thermal comfort. Validated through a case study of a real retractable roof tennis court, the CNN-LSTM surrogate model achieved notable predictive accuracy with mean R² values of 0.948 for Period Spatial Useful Daylight Illuminance (psUDI) and 0.914 for Thermal Comfort Measure Points Percentage (TCMP). The retractable roof control strategy demonstrated significant performance enhancements, yielding an approximate 6.25-fold increase in the average psUDI and a 41.06 % improvement in the average TCMP compared to the fully open roof baseline. The framework and empirical findings provide theoretical foundation and a performance assessment tool for large-scale retractable roof stadiums, advancing the development of data-driven control strategies for dynamic building envelopes.