High spatiotemporal resolution atmospheric wind field detection has important applications in pollution transport and diffusion, extreme weather monitoring, numerical weather forecasting, wind resource assessment, and other areas. Coherent Doppler lidar, as an active laser remote sensing device, acquires high spatiotemporal resolution vector wind field vertical-structure information. However, in practical applications, factors such as platform or power supply stability, and weather conditions can lead to missing wind profiles, limiting the application scope of wind-sensing lidar. Deep learning methods based on historical data modeling have been widely used in wind field prediction. The long short-term memory (LSTM) network shows good performance in wind field prediction. However, most studies mainly focus on one-dimensional temporal or spatial wind fields, while atmospheric wind fields exhibit both temporal and vertical spatial characteristics. Doppler lidar, as a high spatiotemporal resolution atmospheric wind field detection tool, obtains spatiotemporal two-dimensional wind field information. Therefore, we propose a method using a bidirectional long short-term memory (Bi-LSTM) model applied to wind field detection with lidar for wind profile prediction. The aim is to fully utilize the spatiotemporal two-dimensional wind field data observed by the lidar, train a temporal Bi-LSTM model to capture the temporal variation characteristics of wind profiles, predict future wind profiles, interpolate missing wind profiles, and acquire more continuous wind field information.

Our study focuses on Doppler lidar atmospheric wind field detection experiments in Juehua Island, Liaoning Province, China. We utilize complete wind profile data for modeling and validation to predict and interpolate deficient wind profiles detected by the lidar. Previous complete wind profile data segments serve as the training and validation sets to establish wind profile prediction models based on a time-series Bi-LSTM model and a non-time series convolutional neural network (CNN) model for the zonal component u and meridional component v of the wind profiles. We train the models using the same parameter settings, including step size, number of iterations, loss function, and optimization algorithm. We evaluate the wind profile prediction performance of the Bi-LSTM and CNN models using various metrics such as coefficient of determination (R2), root mean square error (RMSE), and mean absolute error (MAE). The Bi-LSTM model with superior validation wind profile prediction performance is then used for deficient wind profile prediction and interpolation to obtain more continuous wind field information.

Based on the evaluation results of wind profile prediction (Fig. 4), the Bi-LSTM model shows similar trends and ranges in performance evaluation metrics R², RMSE, and MAE for different look-back steps. As a temporal network, the Bi-LSTM model exhibits consistent performance across different look-back steps, indicating that wind profiles have short-term or long-term temporal dependencies that allow prediction based on past wind profiles at various time steps. With an increase in prediction time steps, errors accumulate gradually. After the 16th time step iteration, the model’s predictive capability rapidly declines, with R² values for predicting u and v components falling below 0.5, indicating an inability to accurately forecast wind profiles beyond that point. This suggests that the Bi-LSTM model demonstrates good short-term predictive ability for the next 15 wind profiles (within the next 2.5 h). Comparing the wind profile prediction performance of the temporal Bi-LSTM model with the non-temporal CNN model, the box plot analysis (Fig. 5) reveals that the CNN model shows greater variability in R² values for wind profile prediction across different look-backs, indicating a more pronounced influence of the look-back parameter on wind profile prediction and greater uncertainty introduced by the choice of look-backs. The Bi-LSTM model outperforms the CNN model in predicting u and v profiles, likely due to its ability to capture temporal features of wind profiles. In short-term wind prediction, the Bi-LSTM model exhibits lower variability in R² values across different look-backs, demonstrating greater robustness in wind profile prediction. Compared to the CNN model, the Bi-LSTM model achieves higher R² values and lower errors in prediction. The differences in predictive performance may stem from the CNN model’s proficiency in extracting local features using convolutional kernels, while wind profiles, as time series data, exhibit features closely related to preceding and subsequent time steps, potentially limiting the CNN model’s performance in handling such time-dependent wind profile data. In contrast, the Bi-LSTM network, with bidirectional LSTM layers, considers features of wind profiles from multiple time steps in both directions, enabling it to better capture dependencies in time series data and make more accurate wind profile predictions. Future work involves incorporating time-series data such as boundary layer height, temperature, humidity, and pressure as input features to further explore the Bi-LSTM model’s wind profile prediction performance (Fig. 8). Additionally, we find it necessary to increase the number of training time steps to achieve better wind profile prediction results.

In the present study, we propose a method for wind profile prediction using a Bi-LSTM model applied to wind field detection with lidar. The aim is to fully utilize the spatiotemporal two-dimensional wind field data observed by the lidar. By training a temporal Bi-LSTM model to extract the temporal variations of wind profiles, we predict future wind profiles and interpolate missing wind profiles. We conduct a comparison between the temporal Bi-LSTM model and the non-temporal CNN model in wind profile prediction. Our study reveals that the temporal Bi-LSTM model exhibits higher robustness in short-term wind field prediction compared to the non-temporal CNN model.