The Weather Sensory Data Prediction based on LSTM

2011-13330 SNU CLS Seokmo-Yoo, August 2020.

Based on Long Short Term Memory(Sepp Hochreiter et al. 1997), which is effective model to predict time series data, I developed a simple prediction model of temperature and precipitation in weather sensory data. After training the model by observed data, I compared the result with the real data and analyzed the error of the result by numerical methods to discuss about accuracy.

Methodology

Check out snu2020-graduation-thesis.ipynb for detail.

• using Tensorflow Keras in Python 3

Data

Check out data/OBS_ASOS_DD_19071001-20200609.csv for detail.

• Source: https://data.kma.go.kr/cmmn/main.do, Open Weather Data Portal, accessed on 2020-06-10.
• Record: ASOS/Daily/Average, Minimum, Maximum Temperature(degree Celcius) and Daily Precipitation(mm/day)
• Range: 1907-07-01 to 2020-06-09 except 1950-01-01 to 1953-12-31 due to Korean War
• Observation Point: 37.57142°N 126.9658°E 86m, Seoul 108

Model

• Normalize to [-1, 1]
• LSTM Layer(Unit: 100)
• Dense Layer(Unit: 1)
• Denormalize

Result

Average Temperature

Accuracy: MSE 3.1986, R^2 0.9706

• Total test set

• Lastest 120 days test set

• Training loss by epochs

• Scatter plot of ground truths verus predicted values

Minimum Temperature

Accuracy: MSE 3.7829, R^2 0.9656

• Total test set

• Lastest 120 days test set

• Training loss by epochs

• Scatter plot of ground truths verus predicted values

Maximum Temperature

Accuracy: MSE 5.3571, R^2 0.9530

• Total test set

• Lastest 120 days test set

• Training loss by epochs

• Scatter plot of ground truths verus predicted values

Daily Precipitation

Accuracy: MSE 181.0255, R^2 0.1642

• Total test set

• Lastest 120 days test set

• Training loss by epochs

• Scatter plot of ground truths verus predicted values