Authors: Yunhao Chen, Sicheng Mao, Yang Zhang, Zong Shang, Yushan Liu
Convolutional neural networks are a type of deep learning algorithm that have shown promising results in detecting features in images, and LSTM works pretty well on tasks involving the time dimension such as time series prediction. These networks have the ability to learn spatial and temporal dependencies in the data by integrating convolutional layers and LSTM layers, which makes them effective in capturing complex patterns in weather data. By training on historical radar reflectivity data, Convolutional LSTM neural networks can make accurate predictions of precipitation in the short-term, providing valuable information for weather forecasting and disaster preparedness.
The radar dataset which contains composite reflectivity are collected by KNMI, in a raw format of every 5 minutes. Each training data point consists of 36 consecutive radar raw files, the first 18 frames being used as features and the last 18 as the target to predict by neural network. Basically, we are trying to predict the precipitation in the next 1.5 hours given the data of the past 1.5 hours.
Training data: first 18 frames as X_train, last 18 frames as Y_train
Test data: first 18 frames as X_test, last 18 frames as Y_test
The training data consisted of radar images at 5-minute intervals from 20-23 pm for the first 10 days of January 2021, and the test data consisted of radar images at 5-minute intervals from 20-23 pm for the first 10 days of February 2021.
Due to the download restriction of a single api, we prepare the data for you so you don't need to run download_data.py with the trouble of reconfiguration your own api.
Open a terminal and
- install the
ramp-workflowlibrary (if not already done)
$ pip install ramp-workflow
- Follow the ramp-kits instructions from the wiki
Get started on this RAMP with the dedicated notebook.
To test the starting-kit, run
ramp-test --quick-test
Go to the ramp-workflow wiki for more help on the RAMP ecosystem.