입력 데이터 해상도에 따른 심층학습 알고리즘의 기상 변수 예측 정확도 평가

Abstract

We evaluated the performance of deep learning algorithms predicting air temperature in different time steps. Three different data sets were compiled at various time intervals covering days, hours, and minutes for three separate months (i.e., January, July, and November 2021) in two monitoring stations (i.e., one in Seoul 108 and the other in Jinju 192) from the Korea Meteorological Administration. Those data sets divided into 70 % for training and 30 % for testing were provided as inputs to two popular algorithms, the multi layer perceptron (MLP) and long short-term memory (LSTM). Our results showed that the MLP algorithm exhibited superior prediction performance for data recorded at one-minute intervals rather than those updated hourly or daily. In addition, the MLP algorithm was found to work best for data with seasonality. The predictive accuracy was, however, slightly lower for the MLP algorithm than for the LSTM algorithm which yielded error rates as low as 0.04 in terms of the mean absolute error. All these results implied that the use of high-frequency data played an important role in improving the performance of deep learning as well as the proposed methodology could be used to prioritize candidate algorithms with input data (resolution) for prediction of weather variables.