Scholarlink

<p>This study presents a real-time intelligent Maximum Power Point Tracking (MPPT) framework integrating Transformer-based environmental forecasting with XGBoost (XGB) DC power prediction to enhance photovoltaic efficiency under dynamic environmental conditions. In Phase 1, the Transformer model achieved the lowest forecasting errors (Irradiation: MAE = 0.0590, RMSE = 0.0853 and module_temperature: MAE = 1.3029, RMSE = 1.9045), demonstrating superior capability in capturing nonlinear temporal dependencies of solar irradiance and module temperature. In Phase 2, the XGB regression model was employed to map forecasted parameters to DC power output, yielding MAE = 219.06 kW, RMSE = 477.78 kW, and R² = 0.9867. XGB achieved a strong balance between accuracy and computational cost, with the shortest training time (0.46–1.5&#xa0;s) among tested models, indicating an optimal relationship between MAE and training time. MATLAB validation showed that the Transformer-XGB MPPT achieved the highest steady-state efficiency (η = 0.9980) and minimized power oscillations compared to traditional incremental conductance and perturb and observe methods. SHAP analysis confirmed irradiance as the dominant predictive feature. Statistical tests (Friedman <i>p</i> = 0.00077) verified significant model differences. The proposed Transformer–XGB framework thus delivers accurate, explainable, and computationally efficient predictive MPPT control for real-time solar energy optimization.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Real-time ML-based MPPT system with integrated environmental forecasting and power prediction

  • 期刊论文

  • 发表日期：

    2026年5月19日

查看全文
  • Keerati Maneesai,
  • Nitipon Pongphaw

摘要

This study presents a real-time intelligent Maximum Power Point Tracking (MPPT) framework integrating Transformer-based environmental forecasting with XGBoost (XGB) DC power prediction to enhance photovoltaic efficiency under dynamic environmental conditions. In Phase 1, the Transformer model achieved the lowest forecasting errors (Irradiation: MAE = 0.0590, RMSE = 0.0853 and module_temperature: MAE = 1.3029, RMSE = 1.9045), demonstrating superior capability in capturing nonlinear temporal dependencies of solar irradiance and module temperature. In Phase 2, the XGB regression model was employed to map forecasted parameters to DC power output, yielding MAE = 219.06 kW, RMSE = 477.78 kW, and R² = 0.9867. XGB achieved a strong balance between accuracy and computational cost, with the shortest training time (0.46–1.5 s) among tested models, indicating an optimal relationship between MAE and training time. MATLAB validation showed that the Transformer-XGB MPPT achieved the highest steady-state efficiency (η = 0.9980) and minimized power oscillations compared to traditional incremental conductance and perturb and observe methods. SHAP analysis confirmed irradiance as the dominant predictive feature. Statistical tests (Friedman p = 0.00077) verified significant model differences. The proposed Transformer–XGB framework thus delivers accurate, explainable, and computationally efficient predictive MPPT control for real-time solar energy optimization.