5:15 PM - 6:30 PM
[ACG43-P02] Precipitation predictions using RNN and LSTM with simple time series observations and the comparison between them.
Keywords:precipitation forecast, evalution functions, RNN
Precipitation prediction is one of most difficult predictions in weather and climate but strongly required for society. It is well known that there are three problems when artificial neural network (ANN) is employed to predict precipitation. The first is that the predicted local maximum in precipitation is not large enough; the second is that the time of occurrence of the maximum is lagged one interval time; the third is that the state of zero precipitation cannot be accurately represented. In this study, RNN and LSTM were employed to replace the multi-layer Perception (MLP) of ANN for the precipitation prediction, and the outputs were compared so as to explore the degree to which the above three problems could be improved.
Using simple time series observations, three different ANN models, that is, MLP, RNN, LSTM, were tested, two experiments were carried out for these three models. The data sets used in this study are observational time series data compiled by the Japan Meteorological Agency. The stations are five, i.e., Kumamoto, Kikuchi, Mashiki, Nagasaki, and Yatsushiro located in Kyushu of Japan. The used variables are six, that is, hourly precipitation, temperature, pressure, vapor pressure, wind speed, and wind direction. The evaluation functions are SS and RMSE, where SS is the ratio of observed values to predicted values larger than 10 mm/h, which varies between 0 and 1, and the large value means the better prediction performance.
When MLP was used for prediction, the predicted local maximum was less than one third of the observed value, and the occurrence of the maximum was delayed one hour. When RNN was used for prediction, this problem was obviously improved. The predicted maximum can reach about 90% of the observed one, but the maximum's occurrence was still delayed one hour. Finally, when we employed LSTM, the maximum's occurrence time was adjusted with large predicted maximum, which means the time-lag problem was certainly improved. When the ANN model is transformed from RNN to LSTM, the structure of LSTM seemed to improve the short-term dependency bottleneck of RNN, which means that the time series data can be better processed. Some unnecessary data can be forgotten and discarded, and the key data can be retained, which greatly improve the accuracy of the prediction.
When only data of Kumamoto was used for prediction, the best result of MLP was SS=0.45, RMSE=1.36, and for using 5 sites' data, it came to SS=0.54, RMSE=1.04. By adjusting the number of nodes in the two hidden layers, the results were further improved to SS=0.45, RMSE=1.04 for one site's experiment and SS=0.60, RMSE=1.04 in 5 sites' experiment. The results of RNN came to SS=0.46, RMSE=0.96 in one site's experiment and SS=0.61, RMSE=0.96 in 5 sites’ experiment. Finally, when the LSTM model was employed, our prediction reached the best so far, which was SS=0.62, RMSE=1.01 in one site's experiment and SS=0.72, RMSE=0.88 in 5 sites' experiment.
Using simple time series observations, three different ANN models, that is, MLP, RNN, LSTM, were tested, two experiments were carried out for these three models. The data sets used in this study are observational time series data compiled by the Japan Meteorological Agency. The stations are five, i.e., Kumamoto, Kikuchi, Mashiki, Nagasaki, and Yatsushiro located in Kyushu of Japan. The used variables are six, that is, hourly precipitation, temperature, pressure, vapor pressure, wind speed, and wind direction. The evaluation functions are SS and RMSE, where SS is the ratio of observed values to predicted values larger than 10 mm/h, which varies between 0 and 1, and the large value means the better prediction performance.
When MLP was used for prediction, the predicted local maximum was less than one third of the observed value, and the occurrence of the maximum was delayed one hour. When RNN was used for prediction, this problem was obviously improved. The predicted maximum can reach about 90% of the observed one, but the maximum's occurrence was still delayed one hour. Finally, when we employed LSTM, the maximum's occurrence time was adjusted with large predicted maximum, which means the time-lag problem was certainly improved. When the ANN model is transformed from RNN to LSTM, the structure of LSTM seemed to improve the short-term dependency bottleneck of RNN, which means that the time series data can be better processed. Some unnecessary data can be forgotten and discarded, and the key data can be retained, which greatly improve the accuracy of the prediction.
When only data of Kumamoto was used for prediction, the best result of MLP was SS=0.45, RMSE=1.36, and for using 5 sites' data, it came to SS=0.54, RMSE=1.04. By adjusting the number of nodes in the two hidden layers, the results were further improved to SS=0.45, RMSE=1.04 for one site's experiment and SS=0.60, RMSE=1.04 in 5 sites' experiment. The results of RNN came to SS=0.46, RMSE=0.96 in one site's experiment and SS=0.61, RMSE=0.96 in 5 sites’ experiment. Finally, when the LSTM model was employed, our prediction reached the best so far, which was SS=0.62, RMSE=1.01 in one site's experiment and SS=0.72, RMSE=0.88 in 5 sites' experiment.