Deep Learning with Long Short-Term Memory Networks for Air Temperature Predictions

Abstract

Temperature is a commonly used meteorological variable that plays an important role in society, agricultural production and the economy. In this paper, a stacked long short-term memory network (Stacked LSTM) is used to process temperature time series data and to provide temperature prediction results every half hour. Through the comparison of training and prediction with the two benchmark algorithms of deep neural network (DNN) and random forest (RF) on data generated under different sliding windows, it is found that the network built by Stacked LSTM is superior in the MSE, RMSE and MAE and more stable in R Squared. The prediction accuracy of the Stacked LSTM network is further improved by fusing various models in a linear fusion mode. The comparison shows that the result of combining the random forest and Stacked LSTM is not greatly different from the Stacked LSTM, while the result of combining the Stacked LSTM and deep neural network is the optimal. Specifically, the prediction results of combining the DNN with Stacked-LSTM are 1.7%, 3.8%, and 8.7% higher than the Stacked LSTM, DNN and random forest alone, respectively.