Enhancing Wind Power Forecasting with Adaptive Wind Speed Calibration (C-LSTM) and a Hybrid (LTC+XGBoost) Model

Merve Nur Akyer, Michael Nickel, Cagil Karakas

Keywords: Time Series Analysis, Calibrated LSTM (C-LSTM), Forecasting, Wind Power, Liquid Time-Constant (LTC)

Abstract: Accurate forecasting of wind power is essential for grid stability and integration of renewable energy. This work presents a hybrid framework for the prediction of short-term wind power in four Norwegian bidding zones. Models are trained at the wind park level and aggregated to zone-level forecasts. We combine physics-informed feature engineering with Liquid Time-Constant Networks (LTC), XGBoost, and a Calibrated LSTM (C-LSTM) module. LTC captures nonlinear temporal dynamics, XGBoost handles structured inputs, and C-LSTM adaptively corrects wind speed forecasts during inference. The model achieves a Mean Absolute Error of 10.9 MW and a RMSE of 10.92 MW, corresponding to less than 9% error relative to average zone-level production ($\sim$130 MW), demonstrating robust performance under various wind conditions.

Serve As Reviewer: ~Merve_Nur_Akyer1

Submission Number: 15