Software Requirements Specification for the Solar Storm Forecasting¶
|Author:||Simar Preet Singh|
2013.06.23 - Document created
This document describes the software requirements specification for the Solar Storm forecasting.
Describes the scope of this requirements specification.
- European Mars Analog Station
- Italian Mars Society
- Geostationary Operational Environmental Satellite
- National Geophysical Data Center
- Solar Dynamics Observatory
- Daily Solar Data
- Space Environment Services Center
The magnetosphere around the Earth protects us to certain extent, from the constant bombardment by charged particles from the sun. But, Much of Mars’ atmosphere, on the other hand, is exposed directly to these fast-moving particles from the sun and the effects of solar flares. These storms of solar radiation can disrupt satellite communication, resource information, electrical power, and radar. Energy in the form of hard x-rays can also damage space craft electronics. The module aims to forecst such events and issue relevant warning.
The goal of the module is to provide a Neural Network implementation trained using local database to be constructed. This trained Neural Net can then be interfaced with the Tango to issue a warning based on the type of solar flare.
Describe all the users and their expectations for this package
Although, the archive data for the training is readily available in . The module is constrained by the continued availability of functional data from the respective satellites.
A validation set of data will be maintained for the diagnostic requirements.
Development and Test Factors¶
The Software Engineering Practices Guidelines for the ERAS Project in  to be followed.
The planned steps for the design and implementation of the model :
- Variable selection
- Data collection
- Data preprocessing
- Training and validation sets
- Neural network paradigms
- Evaluation criteria
- Neural network training
This procedure is not a single-pass one, and may require the revisiting of previous steps especially between training and variable selection. Although, the implementation step is listed as last one, it is being given careful consideration prior to collecting data.
Use Case: Data collection and integration¶
The main focus is Data colection and preprocessing.
Raw data, local database
The raw data (txt files) must be downloaded on lacal machine.
The raw data from the warehouse in  is to be parsed and the data to be stored on local database (preferably using Mysql ). The data collected from the txt files will be integrated in database using the date as key. An example of the DSD file is in . Using this:
The following feature sets will be extarcted
- Radio flux
- SESC Sunspot number
- Sunspot area
- New regions
- X-ray background flux
The database will then be seperated into training and validation sets to be used for the neural network training.
Although, it was initially thought of using image data from SDO in . But, it is presently generating about 1.5TB of data daily and even the downsampled images would require immense processing power and bandwidth (SDO is receiving about 700Mb every 36 secs). Such processing power is not currently available for this implementation.Still, attempts wil be made to find any source of processed data access points or APIs which may provide us the preprocessed data.
The database split into training and validation sets.
Use Case: Neural network training¶
The focus will be to train the neural network to classify the Solar flares.
Neural network, local database
The training database must be available.
Using the training database, four different Neural networks will be trained where each neural net will be trained to classify the features into a different class ( classes to be trained for X, M, C, A&B ). Each of these neural nets will be trained for one class only. The extracted feature set in the database will be used to identify the class as the output.
The following Neural Network paradigms will be considered :
- number of hidden layers
- number of hidden neurons
- transfer functions
Additional factors considered for the training :
- number of training iterations
- learning rate
After the training, the validation set will be used to verify the performance for the neural network.
As an alternate course, a neural network consisting of multiple outputs to classify the features into the respective classes can be trained. Based on the input feature set, the output will be the corresponding class. The performance of both the implementations can be analysed to identify the most suitable solution.
A trained neural network implementation.
Use Case: Solar storm warning¶
Trained neural network as server, client that responds to warning
The neural network has access to input data feed.
The input features would be fed to the trained neural network. As the network has already been trained offline, the implemented neural network should be able to provide fast response. In case of a warning, the relevant warning will be issued specifying the type of forecast.