News & Updates

This article highlights the work carried out in TWINVEST Task 5.3 by LUT University, which focuses on developing a hybrid, physics-aware Energy Production Forecasting Model. Accurate short-term power forecasting is important as it helps the digital twin compare future operating scenarios, support planning, and improve confidence in production estimates.

Three main components within the forecasting model

The forecasting model combines three main components: a comprehensive feature preprocessing pipeline, a Seq2Seq LSTM forecasting architecture with attention mechanisms, and a physics-based wake post-processing layer. The framework was applied to two TWINVEST physical use cases: the PPC Kouromantri Wind Power Plant in Greece and the Zorlu Gökçedağ Wind Power Plant in Türkiye. It was evaluated for both day-ahead and hour-ahead forecasting horizons. 

The feature pre-processing pipeline improves data quality by transforming raw SCADA data into model-ready features using physics-based feature engineering, anomaly detection, physics-constrained imputation, variational mode decomposition of wind speed, temporal lag features, rolling statistics, and multivariate feature selection. The forecasting model uses a Seq2Seq LSTM structure with attention to capture time-based production patterns between historical power output, meteorological variables, and future operating conditions. Finally, a Gaussian wake deficit model is used to apply physics-based correction and aggregate turbine-level predictions into farm-level forecasts. 

Strong forecasting performance across both use cases and time horizons

The developed EPFM showed strong forecasting performance across the two physical use cases and across both hour-ahead and day-ahead horizons. On cleaned test data, the model achieved R² values close to 1.0, with WAPE values ranging from 0.84% to 6.57% depending on the use case and forecast horizon. The results confirm that combining deep learning with physics-aware preprocessing and wake post-processing can provide accurate and operationally useful wind-power forecasts for digital twin applications. 

One of the main insights from LUT’s work is that data quality and physical consistency are as important as the forecasting architecture itself. By combining anomaly detection, physics-constrained imputation, and wake-aware post-processing with Seq2Seq LSTM forecasting, the EPFM provides a more robust basis for digital twin deployment than a purely data-driven model alone.