Thermal modeling and Machine learning for optimizing heat transfer in smart city infrastructure balancing energy efficiency and Climate Impact

Abstract

The paper proposes a framework based on deep learning, transfer learning, and multi-objective optimisation to model and optimise heat transfer in smart city infrastructure to make them energy efficient and thermally comfortable. The framework in the paper contains a building thermal dynamics prediction model developed using hybrid CNN-LSTM on an extensive dataset (12.56 metric tonnes) of Indian buildings covering various characteristics, which is then fine-tuned with data from five major Indian cities. This predictive framework has a high generalisation capability of energy consumption and predicting indoor temperature profiles with the mean absolute errors (MAE) of building energy consumption ranging from 8.7 to 12.3 kWh and indoor temperature as 0.6 to 1.1 °C, respectively. Transfer learning is considerably improving the performance of the proposed model in newly added cities, which improved the MAE in the training cities (New Delhi and Mumbai) by 3.6 % and reduced the R^2 to 10.7 %. The multi-objective optimisation involving decision-making processes resulted in energy savings of 15.7 % to 22.3 % and improved comfort levels by 21.8 % to 28.5 % in the evaluated cities. The paper significantly contributes to developing a data-driven, generalisable, and interpretable framework, which can usher how to optimise heat transfer using deep learning to make smart city infrastructure resilient and comfortable. It also provides a novel solution to addressing the problems posed by energy efficiency and climate change in Indian cities. Policymakers and urban planners can utilise these key policy recommendations suggested in the paper to design new, liveable and self-sustaining urban environments in India.