Forecasting Model of Wind Speed and Direction by Convolutional Neural Network - Deep Convolutional Long Short Term Memory

This paper serves forecast wind speed and direction using a convolutional neural network - deep convolutional long short term memory (CNN-DConvLSTM). The forecasting model merges merit between CNN and ConvLSTM to enhance forecasts accuracy of wind speed and direction. The input of the forecasting model is image data on two dimensions (2D) - coordinate, which expresses the time sequential image of the wind vector. The forecasting model consists of encoding and forecasting networks that forecast one hour later. The actual observed data was taken from the Automated Meteorological Data Acquisition System (AMeDAS)in Japan. The forecasting model performance was evaluated to mean absolute error (MAE) between forecasted data and observed data. For confirming effectiveness, the forecasting model was compared to the persistent model and fully connected - long short term memory (FC-LSTM). The proposed forecasting model by CNN-DConvLSTM for one-year data can improve by 25% for wind speed and 19.36% for wind direction over the persistent model which is a more significant accuracy than the LSTM model. It indicates that the CNN-DConvLSTM model is the strongest forecast model.