[ACG50-P01] A precipitation prediction using an artificial neural network model based on observation time series data
Keywords:Artificial Neural Network, precipitation prediction, observation time series data
In weather prediction, precipitation prediction is one of most difficult but strongly required for society. This work investigated performance of an artificial neural network (ANN) model for the precipitation prediction. In particular, this study examined how training of the ANN model improved the performance of precipitation prediction.
The datasets used in this study are observation time series data compiled by the Japan Meteorological Agency (JMA). The stations are five, i.e., Kumamoto, Fukuoka, Nagasaki, Kagoshima, and Oita located in Kyushu of Japan. The used observation variables are six, that is, hourly precipitation, temperature, pressure, vapor pressure, wind speed, and wind direction. Using these station data, this study diagnosed the performance of the ANN model's few-hour precipitation prediction at Kumamoto. In this study, the ANN prediction model was constructed by a multi-layer perceptron, in which the number of intermediate layers was set to be two for easy experiment and interpretation. Then, a variety of experiments were performed about training of the ANN prediction model. In particular, this work examined how stations and physical variables improved the precipitation prediction. As the evaluation of prediction outputs, this work diagnosed 1) the difference in maximum value, 2) the time lag of appearance of the maximum, and 3) the root mean squared error between and the prediction and real precipitation time series. In addition, a skill score, which was the ratio between the predicted precipitation and the real value larger than 10 mm/h, was used. This skill score has been designed to evaluate the performance of the JMA's physical prediction model for heavy precipitation events.
The results show that 1) Kumamoto's precipitation itself is most crucial for the prediction, 2) the data at western station effectively improve the prediction, and of course, 3) the most recent precipitation data is most valid. In order to improve the prediction at the maximum value in a precipitation event, this study removed the data with no precipitation from the training data as a preprocessing. It was found that this preprocessing effectively improved the maximum precipitation value from 30% to 80% of the real maximum value. The skill score of this ANN prediction model in precipitation, which was estimated for three-hour prediction, was that of JMA's physical prediction model in about 10 years ago. This study is still on going, and the further results will be shown at the conference, e.g., the results of the experiments increasing stations, improving the ANN structure, using other algorithms such as recurrent NN, and so on.
The datasets used in this study are observation time series data compiled by the Japan Meteorological Agency (JMA). The stations are five, i.e., Kumamoto, Fukuoka, Nagasaki, Kagoshima, and Oita located in Kyushu of Japan. The used observation variables are six, that is, hourly precipitation, temperature, pressure, vapor pressure, wind speed, and wind direction. Using these station data, this study diagnosed the performance of the ANN model's few-hour precipitation prediction at Kumamoto. In this study, the ANN prediction model was constructed by a multi-layer perceptron, in which the number of intermediate layers was set to be two for easy experiment and interpretation. Then, a variety of experiments were performed about training of the ANN prediction model. In particular, this work examined how stations and physical variables improved the precipitation prediction. As the evaluation of prediction outputs, this work diagnosed 1) the difference in maximum value, 2) the time lag of appearance of the maximum, and 3) the root mean squared error between and the prediction and real precipitation time series. In addition, a skill score, which was the ratio between the predicted precipitation and the real value larger than 10 mm/h, was used. This skill score has been designed to evaluate the performance of the JMA's physical prediction model for heavy precipitation events.
The results show that 1) Kumamoto's precipitation itself is most crucial for the prediction, 2) the data at western station effectively improve the prediction, and of course, 3) the most recent precipitation data is most valid. In order to improve the prediction at the maximum value in a precipitation event, this study removed the data with no precipitation from the training data as a preprocessing. It was found that this preprocessing effectively improved the maximum precipitation value from 30% to 80% of the real maximum value. The skill score of this ANN prediction model in precipitation, which was estimated for three-hour prediction, was that of JMA's physical prediction model in about 10 years ago. This study is still on going, and the further results will be shown at the conference, e.g., the results of the experiments increasing stations, improving the ANN structure, using other algorithms such as recurrent NN, and so on.