Examination of the Working Modes of the 240 MW Wind Power Plant Ai-Based Short-Term Wind Power Forecasting System Multi-Sensor Wind Speed Prediction and Hybrid Ensemble Approach
Date Issued
2025
Author(s)
Aliyev, Rasim
Abstract
The rapid growth of large-scale wind power in Azerbaijan requires accurate short-term forecasting tools to support secure and economical power system operation. This thesis develops and evaluates an AI-based short-term wind speed and power forecasting framework using high-resolution measurements from a 120-m meteorological mast located near a 240 MW wind power plant. The dataset consists of one year of 1-second observations from multiple anemometers, wind vanes, temperature, humidity, and pressure sensors, which are preprocessed, quality-controlled, and aggregated to operationally relevant time scales.
The methodology combines advanced feature engineering, data decomposition, and machine-learning models. Physics-informed features such as wind shear, veer, turbulence intensity, and stability indicators are derived from the multi-level measurements. Tree-based algorithms (XGBoost and Random Forest) are used as fast and robust baselines, while a deep learning architecture based on a CNN–BiLSTM network with an attention mechanism is implemented to capture complex temporal dependencies. A regime-based design is adopted in which XGBoost provides accurate forecasts over the full operating range, and a Random Forest specialist is activated for high-wind conditions that are critical for turbine loading and grid security. Forecasted wind speeds are converted into active power using the manufacturer’s power curve of the Envision 171/6.5 MW turbines.
The results show that the proposed framework achieves high accuracy and good generalization for 15-minute-resolution forecasts up to 24 hours ahead, with substantial error reductions in the high-wind regime compared to a single global model. The study demonstrates that combining multi-sensor measurements, physics-aware features, and hybrid AI models can deliver operationally meaningful wind power forecasts, supporting day-ahead scheduling, reserve allocation, and more reliable integration of large-scale wind farms into the Azerbaijani power system.
The methodology combines advanced feature engineering, data decomposition, and machine-learning models. Physics-informed features such as wind shear, veer, turbulence intensity, and stability indicators are derived from the multi-level measurements. Tree-based algorithms (XGBoost and Random Forest) are used as fast and robust baselines, while a deep learning architecture based on a CNN–BiLSTM network with an attention mechanism is implemented to capture complex temporal dependencies. A regime-based design is adopted in which XGBoost provides accurate forecasts over the full operating range, and a Random Forest specialist is activated for high-wind conditions that are critical for turbine loading and grid security. Forecasted wind speeds are converted into active power using the manufacturer’s power curve of the Envision 171/6.5 MW turbines.
The results show that the proposed framework achieves high accuracy and good generalization for 15-minute-resolution forecasts up to 24 hours ahead, with substantial error reductions in the high-wind regime compared to a single global model. The study demonstrates that combining multi-sensor measurements, physics-aware features, and hybrid AI models can deliver operationally meaningful wind power forecasts, supporting day-ahead scheduling, reserve allocation, and more reliable integration of large-scale wind farms into the Azerbaijani power system.