Statistical heterogeneity among clients constitutes a central challenge in federated learning: when local datasets follow non-identically distributed draws, the sample-size-weighted aggregation of FedAvg compromises convergence of the global model towards every client's local optimum. In time-series settings, this heterogeneity manifests as divergences in trend, seasonality, and volatility regimes across participants. The present dissertation proposes MARACATU, a metadata-driven clustering heuristic for federated learning in temporally dependent forecasting scenarios. At each communication round, every client extracts a compact vector of temporal descriptors from its local window, comprising statistical moments, autocorrelation, spectral profile, and peak statistics, and the server employs these vectors to partition clients into groups of compatible temporal regimes prior to applying FedAvg within each cluster. The heuristic is evaluated on the UK Power Networks Low Carbon London dataset, spanning four forecasting architectures (SFF, LSTM, GRU, TFT), three clustering algorithms (Agglomerative, SimClust, RepClust), and three temporal window sizes (6h, 16h, 24h). The results demonstrate that the benefit of metadata-driven clustering is contingent on the choice of clustering algorithm: globally, the heuristic does not outperform weight-based clustering, an effect attributable almost entirely to the Agglomerative method; nevertheless, RepClust paired with the 16h window produces statistically significant and consistent RMSE improvements across all four architectures, an advantage that is preserved irrespective of the federated aggregation method employed. The best-performing configuration, LSTM + FedNova + Metadata + 16h + RepClust, attains RMSE $0.2837 ± 0.0004, corresponding to a 3% reduction relative to LSTM + FedAvg + Metadata without any increase in the number of communication rounds.
.png)
.png)
