Quazi Md Yusuf

Mr. Quazi Md Yusuf


Research Focus & Metrics

Academic Career Length

1Year (2026-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
17 November 2026 in Preprints.org
Verified
Preprint
Open access
PDF file
Jahan AliJahan Ali
Jahan Ali
QYQuazi Md Yusuf
Quazi Md Yusuf
Nurul I. SarkarNurul I. Sarkar
Nurul I. Sarkar
NMNazmul Mridha
Nazmul Mridha
Accurate household-level electricity-demand forecasting underpins core smart-grid functions such as dynamic tariff design and balancing distributed generation, yet obtaining it without pooling residents’ raw consumption traces in one place remains an open operational challenge. Federated learning (FL) with homomorphic encryption (HE) meets it, but existing peer-to-peer encrypted pipelines treat every round the same way: all clients participate regardless of network condition, load or reliability, so encryption’s accuracy cost is paid uniformly rather than managed. This paper adds a Digital Twin-informed orchestration layer above a peer-to-peer FL pipeline protected with CKKS homomorphic encryption. A synchronized state store tracks each client’s real consumption statistics, simulated network and edge conditions, and measured training progress and trust score. A Proximal Policy Optimization (PPO) policy trained on a fast surrogate environment then decides which clients participate each round. We build and run the full pipeline on a coverage-filtered subset of 151 households from the Low Carbon London dataset (5,268,611 half-hourly readings). We evaluate it against seven baselines, a six-way ablation, a client-scale sweep and a robustness study under simulated unreliable clients. The orchestrated framework recovers most of the accuracy encryption costs, reaching a test MSE of 0.673 ± 0.005 against 0.694 ± 0.041 for unorchestrated encrypted peer-to-peer learning and 0.686 ± 0.007 for a Digital Twin-informed heuristic without a learned policy, while CKKS HE itself adds roughly 260 ms of aggregation time per round and multiplies transmitted payload by about 27× over plaintext. Across zero to thirty percent simulated unreliable clients, the orchestrated framework holds a consistently lower test error than naive selection at every level tested, though several comparisons fall short of conventional statistical significance at the two to three seeds this study could afford. We report this limitation alongside the results rather than around them.