FUSION: framework for urban solar intelligence and optimisation

: Prediction and Optimisation of Solar Photovoltaic Planning Using Deep Learning and Deep Reinforcement Learning Across Macro and Micro Spatial Scales

Abstract

Rapid urbanisation has intensified challenges such as population pressure, environmental degradation, and traffic congestion, many of which stem from inadequate or poorly adapted urban planning practices. To address these issues, this study proposes and validates an innovative framework—FUSION (Framework for Urban Solar Intelligence and Optimisation)—which integrates deep learning (DL) and deep reinforcement learning (DRL) for multi-scale urban land-use assessment and optimisation, with a specific focus on solar photovoltaic (PV) site suitability and layout planning.
The framework combines macro-scale top-down prediction using DL with micro-scale bottom-up optimisation using DRL, thereby overcoming the shortcomings of traditional planning approaches that either neglect local interests or rely excessively on subjective intuition. The research employs multi-source geospatial and socio-economic datasets, including digital elevation models, climate variables, population and road networks, land-use maps, and existing PV station records, across five major Australian cities (Sydney, Melbourne, Brisbane, Perth, and Adelaide).
Key findings demonstrate that DL models can accurately replicate and predict urban land-use patterns while providing interpretability through explainable AI (XAI), revealing the dominant spatial drivers of PV suitability such as slope, population density, and road accessibility. Meanwhile, DRL agents effectively optimise PV layouts by balancing population distribution and land availability, achieving up to 15–20% improvements in layout efficiency compared to baseline heuristics. The dual-scale framework further exhibits strong generalisation across different urban contexts, confirming its robustness and adaptability.
The contributions of this thesis are threefold: (1) it introduces the FUSION framework as a novel integration of DL and DRL for urban planning; (2) it establishes a comprehensive multi-source dataset and cross-city validation for PV suitability analysis; and (3) it provides interpretable and actionable insights for both policymakers and local stakeholders, bridging the gap between top-down planning logic and bottom-up optimisation strategies. Beyond these technical contributions, the novelty of this research lies in reframing urban modelling through the explicit integration of macro–micro and top-down–bottom-up perspectives. By addressing a long-standing divide in urban studies between strategic planning logics and localised optimisation needs, FUSION functions not only as a modelling tool but also as a conceptual framework for bridging scales and decision rationalities. This methodological contribution ensures the framework remains relevant even as AI techniques evolve, highlighting the enduring value of integrative and interpretable approaches for sustainable urban planning.

Date of Award	15 Nov 2025
Original language	English
Awarding Institution	University of Nottingham
Supervisor	Wu Deng (Supervisor), Anthony Graham Bellotti (Supervisor), Fiseha Berhanu Tesema (Supervisor) & Ali Cheshmehzangi (Supervisor)