time-series prediction of the offshore and onshore wind profiles using the autoencoding models: lidar and meteorological measurements based

The development of the reliable forecast model plays a vital role in describing the variability of the time-series (very- to long-term) wind profiles of a particular climatic zone. in this paper, the time-series multivariate forecasts and analysis of the air temperature, cnr, offshore/onshore wind profiles from the lidar and meteorological measurements based on 2–autoencoding architectures are presented. the historical datasets (lidar measurements and meteorological masts) of the selected multivariables at 5– and 10–minute intervals are collected. two autoencoding architectures (conv2d and gru encoding-decoding networks) in an unsupervised predictive operation are used for the time-series multivariable forecasting (1-288 horizons) and analysis of the: wind speed and wind direction, sectorwise windrose, cnr and prevailing air temperature. at the period of 48 timesteps, the time-series wind speed and direction variations are analyzed in determining the measurement data height with the steadiest wind flows for optimal loading of the large-scale wind turbine. studied finding results of the offshore wind profiles at different heights revealed the existence of a steadiest wind flow at 128.8 m height but driven by the atmospheric effects. also, the experimental findings revealed that the dominant wind flows of the onshore heights (10-20m) are impacted by the local surface irregularities and atmospheric effects. finally, the autoencoders performance is reported for the experimental offshore and onshore wind flow for different heights with and without the feature noise removal. upon validation and evaluation of the autoencoders with actual model, the gru autoencoder produces better forecast of the time-series of the onshore station multivariables, while the conv2d and gru architectures are needful for the predictions of the offshore station multivariables. the proposed model architectures clearly shown to be an essential forecast tool in providing a more robust wind resource estimates from the time-series predictions of the multivariable input sequences at a given location. lastly, the combined station height dataset of 78.8–158.8mwith the conv2d autoencoder reported the generalized score errors (me = 0.105 m/s and rmse = 0.420 m/s; me = -3.10 and rmse = 6.20) while the gru score errors (me = 0.019 m/s and rmse = 0.396 m/s; me = -2.90 and rmse = 7.90) for the predictions without the feature noise removal are reported.