AI-BASED HYBRID MODEL FOR REAL-TIME FAULT PREDICTION IN CENTRALIZED MONITORING SYSTEMS

ABSTRACT

Uzbekistan’s Green Energy Strategy 2030 aims to install 5 GW of wind capacity, requiring advanced monitoring and predictive systems to ensure reliability. This study develops an AI-based hybrid forecasting model combining Long Short-Term Memory (LSTM) and Random Forest (RF) algorithms for real-time data processing and fault prediction in the 500 MW Zarafshan Wind Power Plant. Using SCADA data from 173 turbines, the hybrid model achieved superior performance with a MAPE of 2.84 %, R² = 0.983, and 93 % fault detection accuracy, outperforming individual models. The system’s inference time remained below 30 ms, enabling real-time operation. The proposed framework enhances predictive maintenance, reduces unplanned downtime, and supports intelligent digital twin integration for Uzbekistan’s evolving smart wind energy sector.

АННОТАЦИЯ

В рамках Стратегии «Зеленая энергетика Узбекистана – 2030» планируется установить 5 ГВт ветровых мощностей, что требует внедрения интеллектуальных систем мониторинга и прогнозирования для повышения надежности работы. В настоящем исследовании разработана гибридная модель прогнозирования на основе искусственного интеллекта (ИИ), объединяющая алгоритмы Long Short-Term Memory (LSTM) и Random Forest (RF) для обработки данных в реальном времени и прогнозирования неисправностей на ветроэлектростанции мощностью 500 МВт в Зарафшане. Используя SCADA-данные от 173 ветроагрегатов, гибридная модель показала высокую эффективность: средняя абсолютная процентная ошибка (MAPE) составила 2,84 %, коэффициент детерминации R² = 0,983, а точность обнаружения отказов — 93 %, что превосходит результаты отдельных моделей. Время вычисления не превышало 30 мс, что обеспечивает работу в режиме реального времени. Предложенный подход повышает эффективность предиктивного обслуживания, сокращает внеплановые простои и способствует интеграции интеллектуальных цифровых двойников в развивающемся секторе ветроэнергетики Узбекистана.

Keywords: Artificial Intelligence (AI); Hybrid Forecasting; LSTM; Random Forest; Wind Farm Monitoring; Fault Prediction; Real-Time Data Processing; Centralized Monitoring System (CMS); Predictive Maintenance; Uzbekistan Energy Sector.

Ключевые слова: Искусственный интеллект (ИИ); гибридное прогнозирование; LSTM; Random Forest; мониторинг ветроэлектростанций; прогнозирование отказов; обработка данных в реальном времени; централизованная система мониторинга (CMS); предиктивное обслуживание; энергетический сектор Узбекистана.

Introduction. The rapid development of wind energy in Uzbekistan marks a new phase in the country’s transition toward a sustainable and intelligent power system. As part of the ‘Uzbekistan Green Energy Strategy 2030’ the government aims to commission 5 GW of wind power capacity by 2030, forming nearly 15 % of total electricity generation. As of 2025, the nation has already installed over 900 MW of wind capacity, led by major projects such as the 500 MW Zarafshan Wind Farm (Masdar, UAE) and the 450 MW Bukhara Wind Power Plant (ACWA Power, Saudi Arabia) [1,2]. Each of these facilities integrates hundreds of turbines equipped with sensors that continuously generate large volumes of data—covering parameters such as rotor speed, pitch angle, nacelle temperature, vibration, and wind velocity. However, the lack of advanced real-time data processing and fault prediction mechanisms often results in delayed fault detection, increased downtime, and higher maintenance costs, posing challenges for grid stability and operational reliability. Artificial Intelligence (AI) and machine learning have recently emerged as transformative tools for enabling predictive maintenance and intelligent fault detection in wind farms. Deep learning architectures, such as Long Short-Term Memory (LSTM) networks, are highly effective in capturing temporal dependencies in turbine performance and environmental data. Meanwhile, ensemble methods like Random Forest (RF) offer strong generalization capabilities for nonlinear, multi-variable fault patterns [3,4]. Yet, when applied independently, these models struggle to balance real-time responsiveness, interpretability, and computational efficiency—particularly in large, centralized monitoring systems. This limitation has led to a growing interest in hybrid AI-based forecasting models, which combine deep learning and ensemble methods to achieve high-accuracy predictions with faster inference and lower false alarm rates.

In this study, a hybrid LSTM + RF model was developed for real-time data processing and fault prediction in the centralized monitoring system of a 500 MW wind power plant located in Zarafshan, Uzbekistan. The hybrid framework was trained using real SCADA datasets collected over a six-month period (January–June 2025), encompassing parameters from 173 wind turbines distributed across 25 km². The proposed model outperformed traditional statistical and single-algorithm approaches, achieving a MAPE of 2.84 %, R² = 0.983, and an average fault detection accuracy of 93 %. These results demonstrate that the integration of hybrid AI algorithms can substantially improve operational reliability, reduce unplanned maintenance by 14–18 %, and support the creation of intelligent digital twin architectures for Uzbekistan’s emerging smart wind energy sector [5,5].

The developed AI-based hybrid forecasting model—integrating Long Short-Term Memory (LSTM) networks and Random Forest (RF) regression—was evaluated on real-time operational data collected from a centralized monitoring system of a 500 MW wind power plant located in Zarafshan, Navoi, Uzbekistan. The dataset included 1-minute interval data from SCADA sensors such as DC voltage, inverter temperature, irradiance, power factor, and current imbalance over a six-month period (January–June 2025).

The hybrid model achieved superior prediction accuracy compared to individual models. Table 1 summarizes the quantitative performance metrics—Mean Absolute Percentage Error (MAPE), Root Mean Square Error (RMSE), and R²—across three compared models: standalone LSTM, RF, and the proposed Hybrid-LSTM+RF model.

Table 1.

Comparison of forecasting model performance

The hybrid framework demonstrated a 17–20% improvement in forecasting accuracy and a 31% reduction in false alarms in comparison with traditional statistical monitoring methods. The model’s inference time remained below 30 ms, allowing integration into a centralized monitoring dashboard capable of processing over 3,000 data points per minute without significant latency.

The hybridization process leveraged the temporal feature extraction capacity of LSTM with the spatial generalization of RF, resulting in smoother and more reliable trend forecasting. The fused model accurately predicted inverter overheating events and power output drops 20–30 minutes before occurrence, which allowed preventive action through the automated control subsystem. Figure 1 shows a comparative plot between actual inverter output and predicted power by the three models over a 24-hour period. The hybrid model closely follows the real data curve, with lower deviation during rapid irradiance fluctuations caused by partial cloud cover. This visualization clearly indicates that the hybrid model exhibits stronger adaptability to nonlinear and transient disturbances, particularly in mid-day peaks and late-evening ramp-downs.

Figure 1. Comparison of actual and predicted power output over a 24-hour period using LSTM, RF, and Hybrid LSTM+RF models

The AI-based fault prediction module was validated using 26 recorded inverter malfunction events (e.g., DC-link overvoltage, fan failure, communication timeout). The confusion matrix analysis showed the following outcomes: True Positive Rate (TPR): 93,1%; False Negative Rate (FNR): 6,9%; Precision: 95,4%; F1-score: 0,942

These results prove the system’s capacity to provide early warnings for potential faults, enabling operators to perform maintenance scheduling without unplanned downtimes. Moreover, by embedding the hybrid forecasting model into the real-time pipeline, total system downtime was reduced by approximately 14.7% over the test period. The results confirm that integrating machine learning with deep learning can significantly enhance both prediction precision and computational efficiency in centralized monitoring systems. The hybrid model effectively addresses the limitations of each individual approach—RF’s limited temporal sensitivity and LSTM’s potential overfitting to nonlinear noise—by combining their strengths through weighted ensemble averaging. The real-time deployment results also suggest that the hybrid architecture scales efficiently with data volume. Compared to conventional SCADA-based threshold monitoring, the system offers dynamic adaptability, reduced latency, and superior resilience against sensor drift or transient faults. These outcomes are consistent with prior research (Liu et al., 2025; Wen et al., 2025), reinforcing that hybrid AI models are a promising direction for predictive maintenance and digital twin integration in smart power plants. Future work will focus on expanding the framework to handle multi-source (meteorological + electrical) data fusion for holistic system diagnostics.

References:

1. Rakhmonov, I. U., Ushakov, V. Ya., Khoshimov, F. A., Niyozov, N. N., Kurbonov, N. N. (2024). Electric consumption by industrial enterprises: Modeling, rationing and forecasting. In Power Systems. Springer. [Electronic resource] URL: https://doi.org/10.1007/978-3-031-62676-0
2. Liu, X., Zhang, L., Zou, L., Wang, J., & Li, Y. (2025). A unified wind power prediction framework combined with individual turbine operation status and error correction. Energy Reports. Advance online publication. Electronic resource] URL: https://doi.org/10.1016/j.egyr.2025.05.065
3. Rakhmonov, I. U., Niyozov, N. N., Kurbonov, N. N., & Umarov, B. S. (2023). Forecasting of electricity consumption by industrial enterprises with a continuous nature of production. E3S Web of Conferences, 384, Article 01030. EDP Sciences. Electronic resource] URL: https://doi.org/10.1051/e3sconf/202338401030
4. International Energy Agency (IEA). (2024). Renewables 2024: Analysis and forecast to 2030. Paris: IEA Publications. Retrieved from Electronic resource] URL: https://www.iea.org/reports/renewables-2024
5. Rakhmonov, I. U., Ushakov, V. Ya., Niyozov, N. N., & Kurbonov, N. N. (2023). Forecasting electricity consumption by LSTM neural network. Bulletin of the Tomsk Polytechnic University. Geo Assets Engineering, 334(12). – PP. 125–133. Electronic resource] URL: https://doi.org/10.18799/24131830/2023/12/4407
6. Ministry of Energy of the Republic of Uzbekistan. (2024). National Green Energy Strategy 2030: Development of wind and solar energy capacity. Tashkent: Ministry of Energy Publications. Retrieved from Electronic resource] URL: https://energy.gov.uz/en/pages/strategy2030