Reliable heating-load measurements are essential for building energy management, yet long-term missing data remain difficult to impute, especially across heterogeneous building portfolios. Existing generative imputation models can impute long missing sequences without recursively reusing previous estimates, but they are mainly designed for individual buildings and do not explicitly distinguish shared weather-dependent patterns from building-specific thermo-operational responses. Moreover, training separate models for different buildings increases model-management complexity in practical applications. To address these limitations, this study proposes a unified GAN-based framework for cross-building long-term heating-load imputation under both balanced and imbalanced data-volume conditions. The framework imputes missing heating-load data from conditioned weather and temporal variables in non-overlapping 24h windows. Its enhanced generative backbone introduces U-Net connections to preserve detailed load-profile information and PatchGAN discriminators to improve local temporal realism. Building on this backbone, the generator integrates a shared encoder, a Multi-gate Mixture-of-Experts (MMoE) module, and building-specific decoders, supporting joint weather-load representation learning with building-specific expert mixtures and decoders. We evaluate the framework on four primary case studies drawn from two real-world data sources and one simulated data source (single-family houses, multi-family houses, commercial buildings, and hotels), and we further validate generalisability using two supplementary case studies from an additional real-world data source (school and office buildings). Compared with separate single-building training under balanced data-volume settings, the proposed network reduces NRMSE by 9.70% on average, while substantially reducing model size and training time. Compared with transfer learning (TL) under imbalanced data-volume settings, it reduces NRMSE by up to 58.78%. These results demonstrate improved cross-building heating-load imputation and reduced model-storage requirements for the evaluated building portfolios.
Fault detection and localization in grid-connected photovoltaic (PV) systems is essential to limit short circuits, reduce safety risks, shorten system downtime and maintain high energy-conversion efficiency. Researchers have proposed many algorithms for this task, but most of them assume that the PV array topology is already known. They need the number of modules per string, the number of strings, the total number of modules and the condition of each module. Each algorithm is usually designed and validated on one specific PV configuration. As a result, it often performs poorly on systems with a different topology, operating conditions or components, and wrong topology assumptions lead to misclassified faults and incorrect fault locations. This paper proposes an artificial intelligence (AI) based method that identifies the PV array topology automatically from real-time electrical and environmental measurements. The method estimates the modules per string, number of strings, total modules, and healthy and faulty modules. It combines physics-informed constraints with a Random Forest regression model, trained and tested on 46,347 simulated PV samples using an 80:20 split. The model achieves a topology-level accuracy of 94% on unseen test data and outperforms XGBoost and support vector regression in both accuracy and computational cost. The method was further validated in real time on a hardware-in-the-loop (HIL) platform using two Typhoon HIL 404 units and a TI F28379D microcontroller, where it was integrated with a fault detection and localization algorithm. By removing the need for prior topology information, the proposed approach makes PV fault diagnosis more reliable across different installations.
This study evaluates how information access, architectural control and delayed adaptation affect hourly wind-power forecasting. Observed Bulgarian national generation for 2015–2024 supplies capacity-proportional regional targets, with a complete 366-day test in 2024. Independent measured production from ten Australian farms supplies a separate 184-day test. Statistical, recurrent, convolutional and attention models are compared with fixed and adaptive combinations under chronological selection and a declared observation delay. A harmonized Transformer–tokenwise pair matches initial parameters, sample order and numerical stopping. Its pooled normalized errors are 12.131% versus 12.108% in Bulgaria and 21.908% versus 21.594% on the farms. The Bulgarian paired attention contrast remains unresolved. A documented within-origin TCOAT-H48 adaptation gives 11.486% and 21.833%. Separate production-history controls assess richer inputs without treating them as architecture improvements. Three retrospective yearly refits of a Climatology/Ridge/HGB pool yield Hedge-minus-static intervals spanning zero in every primary comparison. Ten leave-one-farm-out fits increase HGB error from 23.828% to 24.257% relative to a matched no-identity included-farm control, a paired difference of +0.430 [+0.330, +0.526] percentage points. The original seven-expert experiment likewise establishes no general Hedge advantage over its calibration-selected expert, and similar errors can conceal different weight allocations. The contribution is empirical evidence about the conditions and transfer limits of additional model complexity. Exact regional accounting preserves national totals, reanalysis inputs and assumed delivery delays keep the evaluation retrospective.
