Distributed uncertainty-aware multi-agent-inspired learning for wind power forecasting

Wind energy plays an increasingly important role in modern power systems, but its dependence on variable atmospheric conditions creates substantial operational uncertainty. This limits the effectiveness of traditional point forecasting for risk-aware decision making and motivates the use of distributed probabilistic forecasting methods that can better capture local variability and quantify uncertainty. This study presents a distributed probabilistic forecasting framework based on operational SCADA (Supervisory Control and Data Acquisition) measurements collected through dedicated sensing units. Each unit models the temporal dynamics of a specific SCADA feature using Long Short-Term Memory (LSTM) networks to estimate conditional quantiles. Measurement-level forecasts are then merged using three alternative strategies: Direct Quantile Concatenation, Quantile Range Representation (QRR), and Probability Weighted Quantile Aggregation. Results based on real-world SCADA data show that the distributed multi-agent-inspired framework outperforms a centralized baseline LSTM in both deterministic and probabilistic forecasting performance. Among the evaluated merging strategies, QRR achieved the best overall results, reducing RMSE by 7.7% (from 0.0815 to 0.0752) and improving CRPS by 7.9% (from 0.0355 to 0.0327) relative to the baseline. QRR also reduced the 90% Interval Score by 56.7% (from 0.6110 to 0.2643) while maintaining a reliable 90% prediction interval coverage of 89.4%. These findings indicate that modeling uncertainty at the individual sensor level provides a scalable and well-calibrated approach for probabilistic wind power forecasting.