Federated versus Centralized PV Capacity-Factor Forecasting: A Non-Inferiority Study Across Seven Stations in North Macedonia
Organizations: Kadir Has University, Istanbul, Turkey · Department of Chemical and Biological Engineering, Friedrich Alexander University Erlangen-Nuremberg, 91058 Erlangen, Germany · Faculty of Computer Science and Engineering, Ss. Cyril and Methodius University, Rugjer Boshkovik 16, Skopje, North Macedonia · Magix.AI, Vasil Gjorgov 19/1-3, Skopje, North Macedonia
Abstract
Background: Photovoltaic (PV) forecasts inform energy commitments by PV operators and aggregators and guide charge-discharge scheduling in solar-plus-storage systems. Federated learning enables collaborative model training through model updates when local operators cannot pool raw operational data. This study evaluated whether federated training was non-inferior to centralized training for PV capacity-factor prediction across seven stations in North Macedonia. A common multilayer perceptron was used in an offline forecasting comparison with a 24-hour timestamp lag. Exploratory validation compared local, centralized, Federated Averaging, and Federated Proximal (FedProx) training. FedProx (E = 1, mu = 1) was selected for a confirmatory comparison under a protocol fixed before test evaluation. The comparison used past-only offline inputs, five paired training seeds, macro-averaged mean absolute error (MAE), a prespecified non-inferiority margin of 0.005, and a paired day bootstrap with 10,000 repetitions over 49 test dates. Results: FedProx achieved a macro-MAE of 0.128397, compared with 0.133405 for centralized training. The paired difference (FedProx minus centralized) was -0.005008 (95% confidence interval: [-0.009926, -0.000537]). The upper confidence bound was below the prespecified margin, satisfying the non-inferiority criterion. All five seed-specific differences favored FedProx, and its mean MAE was lower at six of the seven stations. Post hoc calendar sensitivity analyses did not change the conclusion. Conclusions: These findings support further evaluation of federated forecasting for collaboration under data-pooling constraints. The study demonstrates predictive non-inferiority within the evaluated setting. Benefits for energy commitments and battery scheduling remain potential applications requiring operational evaluation.
Figures & tables
| Client | Rated | Retained | Excluded | Unrated | All rated kW | Excluded kW | |
|---|---|---|---|---|---|---|---|
| S1 | 1 | 1 | 0 | 2 | 44 | 0 | 44 |
| S2 | 1 | 1 | 0 | 1 | 44 | 0 | 44 |
| S3 | 5 | 5 | 0 | 2 | 220 | 0 | 220 |
| S4 | 8 | 8 | 0 | 2 | 440 | 0 | 440 |
| S5 | 14 | 11 | 3 | 4 | 770 | 165 | 605 |
| S6 | 11 | 11 | 0 | 2 | 605 | 0 | 605 |
| Group | Count | Timestamp | Contents |
|---|---|---|---|
| history | 1 | Capacity factor at | |
| geometry | 6 | cos_zenith , is_daylight , hour_sin/cos , doy_sin/cos | |
| weather_past | 5 | temperature_2m , direct_radiation , diffuse_radiation , global_tilted_irradiance , | |
| weather_future | 5 | Same five variables from target-hour ERA5 reanalysis |
| Client | Raw | Daylight | Model matrix | Matrix/daylight |
|---|---|---|---|---|
| S1 | 2,470 | 1,453 | 1,392 | 0.958 |
| S2 | 1,627 | 1,453 | 1,392 | 0.958 |
| S3 | 1,789 | 1,604 | 1,549 | 0.966 |
| S4 | 1,769 | 1,601 | 1,548 | 0.967 |
| S5 | 1,770 | 1,601 | 1,548 | 0.967 |
| S6 | 1,759 | 1,588 | 1,536 | 0.967 |
| Raw rows | Model-eligible rows | |||||
|---|---|---|---|---|---|---|
| Client | Train | Val. | Test | Train | Val. | Test |
| S1 | 1,729 | 370 | 371 | 882 | 247 | 263 |
| S2 | 1,138 | 244 | 245 | 934 | 229 | 229 |
| S3 | 1,252 | 268 | 269 | 1,051 | 246 | 252 |
| S4 | 1,238 | 265 | 266 | 1,051 | 248 | 249 |
| S5 | 1,239 | 265 | 266 | 1,051 | 248 | 249 |
| Variant A | Variant B | ||||||
|---|---|---|---|---|---|---|---|
| Model | A configuration | MAE | RMSE | Skill | MAE | RMSE | Skill |
| Persistence | — | 0.121 | 0.189 | 0.000 | 0.121 | 0.189 | 0.000 |
| Ridge | Local | 0.121 | 0.163 | 0.006 | 0.095 | 0.132 | 0.222 |
| Ridge | Centralized | 0.115 | 0.161 | 0.047 | 0.097 | 0.139 | 0.196 |
| MLP | Local | 0.102 | 0.147 | 0.157 | 0.086 | 0.123 | 0.286 |
| MLP | Centralized | 0.112 | 0.158 | 0.066 | 0.090 | 0.135 | 0.260 |
| Design | Centralized MAE | FedProx MAE | 95% CI | NI | |
|---|---|---|---|---|---|
| Union calendar (primary) 49 dates; synchronized | 0.133405 | 0.128397 | Pass | ||
| Common overlap (post hoc) 16 shared dates; synchronized | 0.139345 | 0.132578 | Pass | ||
| Client stratified (post hoc) 16–49 dates/client; independent | 0.133405 | 0.128397 | Pass |
| Seed | Centralized MLP | FedProx | |
|---|---|---|---|
| 0 | 0.130968 | 0.127309 | -0.003660 |
| 1 | 0.130857 | 0.127171 | -0.003686 |
| 2 | 0.131592 | 0.129054 | -0.002538 |
| 3 | 0.136819 | 0.128470 | -0.008349 |
| 4 | 0.136788 | 0.129981 | -0.006807 |
| Mean | 0.133405 | 0.128397 | -0.005008 |
| Station | C MAE | FP MAE | C RMSE | FP RMSE | C skill | FP skill | ||
|---|---|---|---|---|---|---|---|---|
| S1 | 263 | 0.120 | 0.115 | -0.005 | 0.177 | 0.161 | 0.085 | 0.123 |
| S2 | 229 | 0.127 | 0.126 | -0.002 | 0.187 | 0.174 | 0.039 | 0.050 |
| S3 | 252 | 0.122 | 0.117 | -0.006 | 0.183 | 0.165 | 0.053 | 0.097 |
| S4 | 249 | 0.123 | 0.112 | -0.011 | 0.175 | 0.156 | -0.003 | 0.087 |
| S5 | 249 | 0.140 | 0.131 | -0.008 | 0.187 | 0.172 | 0.107 | 0.161 |
| S6 | 247 | 0.134 | 0.125 | -0.009 | 0.179 | 0.164 | 0.098 | 0.161 |
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
| Stage | Seeds | Variant | Models and role |
|---|---|---|---|
| Exploratory validation | 0 | A, B | Local/centralized Ridge and MLP; FedAvg/FedProx, ; configuration selection and A/B comparison. |
| Validation robustness | 0–4 | A | Centralized MLP, FedAvg , and FedProx ; initialization robustness, not test inference. |
| Confirmatory test | 0–4 paired | A | Centralized MLP versus selected FedProx , only; frozen non-inferiority analysis. |
| Client | Train from | Train to | Val. to | Test to | Rows (train/val./test) |
|---|---|---|---|---|---|
| S1 | 2024-12-31 | 2025-04-17 | 2025-05-14 | 2025-05-31 | 1,729 / 370 / 371 |
| S2 | 2025-01-01 | 2025-05-01 | 2025-05-16 | 2025-05-31 | 1,138 / 244 / 245 |
| S3 | 2024-12-31 | 2025-04-28 | 2025-05-15 | 2025-05-31 | 1,252 / 268 / 269 |
| S4 | 2024-12-31 | 2025-04-28 | 2025-05-15 | 2025-05-31 | 1,238 / 265 / 266 |
| S5 | 2024-12-31 | 2025-04-28 | 2025-05-15 | 2025-05-31 | 1,239 / 265 / 266 |
| S6 | 2024-12-31 | 2025-04-28 | 2025-05-15 | 2025-05-31 | 1,231 / 264 / 264 |
| Variant | Input width | Nonconstant columns | Parameters |
| A | 12 | 11 | 2,945 |
| B | 17 | 16 | 3,265 |
| Setting | Centralized MLP | FedAvg / FedProx |
|---|---|---|
| Objective | MSE | MSE; FedProx adds |
| Optimizer state | Retained across epochs | Fresh per client-round |
| Budget / patience | 300 epochs / 25 epochs | 300 rounds / 25 rounds |
| Aggregation | None | Retained training-row-count weighted weights |
| Initialization | torch.manual_seed(s) | torch.manual_seed(s) |
| Shuffle stream | default_rng(s) | Persistent default_rng(s+i) for zero-based ordered client |
| CF percentage points | Upper 95% CI | Non-inferiority | Role | |
|---|---|---|---|---|
| 0 | 0 | -0.000537 | Pass | Post-hoc sensitivity |
| 0.0025 | 0.25 | -0.000537 | Pass | Post-hoc sensitivity |
| 0.005 | 0.5 | -0.000537 | Pass | Pre-specified primary |
| 0.0075 | 0.75 | -0.000537 | Pass | Post-hoc sensitivity |
| 0.010 | 1.0 | -0.000537 | Pass | Post-hoc sensitivity |
Solar Power Forecasting
Solar power forecasting predicts photovoltaic generation and irradiance from weather, imagery, plant data, and recent output to support reliable grid and energy decisions.
See all 49 papers