Non-Gaussian wind pressure characteristics and peak factor estimation for heliostats: A wind tunnel and convolutional neural network–attention mechanism–long short-term memory network study

Heliostats are sensitive to wind load due to their characteristics of light weight and high flexibility, and prominent blunt body effect, leading to a significant effect on the safety of structural operation and concentrating performance. This study examines the non-Gaussian characteristics of wind pressure and their impact on peak factors for heliostats through wind tunnel experiments and particle image velocimetry. The probability density distribution of the standardized wind pressure coefficient and the maximum mean wind pressure of the heliostat among all wind directions and mirror inclinations are statistically analyzed. Spatial and statistical distribution of skewness and kurtosis of fluctuating wind pressure is analyzed with 2310 samples to illustrate the non-Gaussian properties of wind pressure on the mirror. The skewness values at the edge and corner of the mirror are observed to be less than −1, while the corresponding kurtosis values exceed 7. This suggests that the wind pressure in these regions exhibits pronounced non-Gaussian characteristics. Subsequently, the Hermite model method (HM), the modified HM method, and the improved HM method were employed to calculate the peak factors for the heliostat. The comparative analysis demonstrates that the modified HM method is better suited for calculating the peak factor of the heliostat structure. And then the maximum peak factors across were calculated by this method. The results indicate that the peak factor reaches a maximum value exceeding 5.5, with the peak factor at the edges and corners of the mirror reaching values larger than 7.5. The Convolutional Neural Network–Attention Mechanism–Long Short-Term Memory Network model is employed to predict the peak factor. The peak factor predicted by the model is largely in agreement with the observed peak factor. The maximum absolute error between the predicted and actual peak factor at local measuring points is only 0.32, which demonstrates that the artificial neural network model can accurately predict the peak factor of wind pressure.