Towards positive energy districts with multi-vector storage and EVs: A multi-scale AI-driven UBEM framework

Recognizing the central role of urban energy districts in meeting European decarbonization targets, the present study introduces a hybrid, multi-scale modelling and optimization framework that bridges urban- and system-scale analysis for prosumer residential districts. The proposed architecture integrates the GIS-driven City Energy Analyst (CEA), dynamic energy system simulation, and machine-learning (ML) surrogate models within an NSGA-III optimization environment. CEA serves as the urban building energy modelling (UBEM) backbone, capturing the spatial and morphological characteristics of the building stock, while TRNSYS resolves the transient behavior of an integrated multi-vector energy system comprising PV, battery, H 2 system, and EVs. A Sobol low-discrepancy design of experiments is used to sample the high-dimensional design space and generate data to train three different ML surrogate regressors. These surrogates are then embedded in NSGA-III multi-criteria optimization, enabling efficient exploration of Pareto-optimal solutions and the identification of balanced knee-point designs. As a case study, a real residential district in Southern Italy is used, where district loads are characterized by generating stochastic user-centric and EV charging profiles via an agent-based LPG-CEA workflow, whose aggregated results are validated against official ARERA statistics. The results show that all ML regressors achieve strong forecasting accuracy across the three techno-enviro-economic targets, with a maximum R 2 exceeding 0.998, and that GPR consistently outperforms the tree-based ensemble models. The positive energy district scenario requires 23.5% higher upfront investment than the net-zero case but delivers a 35.7% reduction in life-cycle cost (LCC), the highest peak-shaving index (PSI = 0.450), and the highest round-trip efficiency (RTE) for the hydrogen storage subsystem. SHAP analysis further reveals that, after PV plant power, battery capacity is the second most influential driver of economic performance, whereas fuel-cell rated power is the dominant secondary contributor to environmental impact. • Toolchain for high-resolution urban multi-vector energy planning and optimization • Integrated AI-enabled GIS-UBEM framework for multi-scale prosumer district design • Data-driven NSGA-III algorithm yields Pareto-optimal PV–BESS–H 2 configurations • The GPR surrogate outperforms tree-based ensemble ML regressors with R 2 > 0.998 • The optimal PED design improves peak-shaving index, CO 2 mitigation and grid export