Data-Driven Multiple ARIMA through Neural Fusion for Enhanced Wind Power Prediction with Missing Data

Robust wind power forecasting in the presence of missing data remains a critical challenge for efficient grid management and energy trading. This paper proposes a novel methodology that integrates a data-driven approach for creating multiple Autoregressive Integrated Moving Average (ARIMA) models with with neural network-based fusion strategy. We first employ statistical properties to identify distinct operational regimes within the wind power data, training specialized ARIMA models. Then, we introduce neural networks for the robust fusion of these segmented ARIMA forecasts. These networks are engineered to: (1) learn and exploit intricate relationships between the predictions originating from different operational regimes, and (2) dynamically adapt its fusion strategy based on the availability of current data and the predicted reliability of each ARIMA model, effectively compensating for missing data points. Extensive experiments conducted on real-world wind power datasets characterized by varying missing data patterns demonstrate the superior accuracy and robustness of our proposed approach compared to conventional forecasting methods.