Scholarlink

This study conducts a systematic evaluation of convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) integrated within a physics-informed neural network (PINN) framework for wind turbine power prediction. Using SCADA data from a 2.3 MW turbine in Turkey, we compare four configurations: CNN–MSE, CNN–PINN, LSTM–MSE, and LSTM–PINN. The LSTM–MSE model achieved the best overall performance, significantly outperforming the PINN variants in key error metrics. The CNN–PINN configuration showed intermediate results, while the PINN models exhibited extremely high MAPE values, indicating substantial difficulties in balancing physical constraints with data fitting. These findings highlight three critical insights: (1) LSTMs temporal feature modeling remains dominant despite CNNs spatial pattern extraction capabilities; (2) current PINN implementations may inadequately capture turbine-specific physics under variable operational conditions; and (3) hybrid architectures require carefully tuned weighting between data-driven and physics-based loss components. Overall, the results challenge prevailing assumptions about PINN superiority in this context and suggest the need for refined constraint-integration strategies in renewable energy forecasting systems.

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

Integrating Physical Laws and Deep Learning: A Comparative Study of CNN and LSTM PINNs for Wind Energy

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Pablo-Andrés Buestán-Andrade,
  • Matilde Santos,
  • Jesús-Enrique Sierra-García,
  • Nathalia-Michelle Peralta-Vásconez

摘要

This study conducts a systematic evaluation of convolutional neural networks (CNNs) and long short-term memory networks (LSTMs) integrated within a physics-informed neural network (PINN) framework for wind turbine power prediction. Using SCADA data from a 2.3 MW turbine in Turkey, we compare four configurations: CNN–MSE, CNN–PINN, LSTM–MSE, and LSTM–PINN. The LSTM–MSE model achieved the best overall performance, significantly outperforming the PINN variants in key error metrics. The CNN–PINN configuration showed intermediate results, while the PINN models exhibited extremely high MAPE values, indicating substantial difficulties in balancing physical constraints with data fitting. These findings highlight three critical insights: (1) LSTMs temporal feature modeling remains dominant despite CNNs spatial pattern extraction capabilities; (2) current PINN implementations may inadequately capture turbine-specific physics under variable operational conditions; and (3) hybrid architectures require carefully tuned weighting between data-driven and physics-based loss components. Overall, the results challenge prevailing assumptions about PINN superiority in this context and suggest the need for refined constraint-integration strategies in renewable energy forecasting systems.