pyWLFRFD

Water Level prediction from RainFall Data

Version: 0.0.3    
Author : Md. Nazmuddoha Ansary    

Version and Requirements

• Python == 3.6.8
• Keras==2.3.1
• numpy==1.17.3
• pandas==0.25.2

1. Create a Virtualenv
2. pip3 install -r requirements.txt

NOTE: tensorflow is used as backend

DataSet

The dataset comprises of three parts:

1. RainFall data:

• Recorded Daily
• Starting Date: 01-01-07
• End Date : 31-12-15
• Lat Info : 22.875 to 25.875 with increase of 0.25
• Lon Info : 88.875 to 94.875 with increase of 0.25
• Total Data Points: 325 locations
• Basin : Meghna
• Zone : Borak

2. Water Level data:

• Recorded Daily
• Starting Date: 01-01-10
• End Date : 31-12-14
• Outlet: Bhairab Bazar

3. Shape File for ROI

        ROI- {lat}_{lon}  
        ['25.125_93.375', '25.125_93.625', '23.625_92.875', '23.875_92.875', '23.875_93.125',  
         '23.875_93.375', '24.125_93.125', '24.375_93.125', '24.625_93.125', '24.875_93.875',   
         '23.375_92.625', '23.625_92.625', '23.875_92.375', '23.875_92.625', '24.125_92.625',   
         '24.125_92.875', '24.375_92.875', '25.375_93.875', '24.625_93.625', '24.875_93.625',   
         '24.375_92.625', '24.625_92.625', '24.625_92.875', '24.875_92.625', '24.875_92.875',  
         '24.875_93.125', '25.125_93.125', '24.625_93.375', '24.875_93.375', '24.625_93.875',  
         '25.375_93.625', '25.375_94.125', '24.375_93.375', '24.125_93.375', '24.375_93.625',   
         '25.125_92.875', '25.125_93.875']  

NOTES:

• Prediction period in the future is tuned by: pred_len in config.json (default=1 day)
• The input sequence comprises of both water level data and rainfall data with seq_len in config.json (default=7 days)
• From the 5 Years of mergeable data available (2010,2011,2012,2013,2014): 2014 is taken as test data and 2010-2013 is taken as train data
• Maximum Rainfall is recorded to be : ~ 585 mm (Taken as int for normalization ease)
• Maximum WaterLevel is recorded to be : ~ 20 m (Taken as int for normalization ease)

Execution

• Change The following Values in config.json

    "ARGS":  
    {  
        "SHP_FILE"                : "/home/ansary/RESEARCH/RainFall/Data/Borak/Borak.shp",  
        "WATER_LEVEL_CSV"         : "/home/ansary/RESEARCH/RainFall/Data/Raw/waterlevel.csv",  
        "RAINFALL_CSV"            : "/home/ansary/RESEARCH/RainFall/Data/Raw/rainfall.csv",  
        "OUTPUT_DIR"              : "/home/ansary/RESEARCH/RainFall/Data/"  
    }  
  

• Run main.py

version: 0.0.2

    usage: main.py [-h] model_type

    WaterLevel Data Prediction From RainFall Data
    
    positional arguments:
    
    model_type  name of the LSTM model to be USED. Available: StackedLSTM 
    
    optional arguments:
    
    -h, --help  show this help message and exit

Results

• If execution is successful a folder called h5s should be created with the following folder tree:

        h5s              
        ├── X_Eval.h5
        ├── X_Train.h5
        ├── Y_Eval.h5
        └── Y_Train.h5
  

• After Complete Training and Testing a folder named {MODELNAME}_MAX_FEATs:{maximum_features_to_be_extracted}_EPOCHS:{num_of_training_epochs} (which is considered as identifier for a certain setting) will be created in the OUTPUT_DIR which will contain the following folder tree:

        identifier
        ├── arch.png
        ├── loss.png
        ├── model.h5
        └── test.png
  

NOTE

• An Example idenditier looks like: StackedLSTM_MAX_FEATs:64_EPOCHS:500

Model Name= StackedLSTM, Max Feats = 64, Epochs = 500

• arch.png plots the model structre
• loss.png plots training history (overfit/ underfit identification)
• test.png plots the test data and prediction
• model.h5 is the model_weight file for deployment

RNN Models

• Stacked LSTM
• Bidirectional LSTM

To Be Added in version 0.0.5

• CNN LSTM

To Be Added in version 0.0.6

• ConvLSTM

To Be Added in version 0.0.7

Stacked LSTM

The implemented Model Structre is as follows:

• NUMBER OF TRAINING EPOCHS: 500
• BATCH SIZE: 128
• MAXIMUM FEATURE TAKEN AT ANY LAYER: 64
• R2 SCORE: 0.9654419438558642
• Mean Absolute Error : 0.1374554411601445

The training loss history and prediction with target plot is as follows:

ENVIRONMENT

OS          : Ubuntu 18.04.3 LTS (64-bit) Bionic Beaver        
Memory      : 7.7 GiB  
Processor   : Intel® Core™ i5-8250U CPU @ 1.60GHz × 8    
Graphics    : Intel® UHD Graphics 620 (Kabylake GT2)  
Gnome       : 3.28.2