e-ISSN: 2582-5208

International Research Journal of Modernization in Engineering Technology and Science

( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:04/Issue:03/March-2022 Impact Factor- 6.752 [Link]
RAINFALL PREDICTION SYSTEM
Shravani Kadam*1, Dhanesh Salgaonkar*2, Bhishmesh Chaudhari*3,
Santosh Tamboli*4
*1,2,3Student, Department Of Information Technology, Vidyalankar Institute Of Technology,
Mumbai, Maharashtra, India.
*4Asst. Prof., Department Of Information Technology, Vidyalankar Institute Of Technology,
Mumbai, Maharashtra, India.
ABSTRACT
In modern times, global warming is affecting the entire world, having a major impact on humankind and
accelerating climate change. As a result, the atmosphere and ocean are warming, sea levels are rising, and
floods and droughts are occurring. One of the major consequences of this is uneven rainfall / precipitation.
Precipitation forecasting now-a-days is a tedious task that is being considered by most of the major world
authorities. Precipitation is a climatic factor that affects the various human activities on which they depend.
Like Agricultural production, construction, energy production and tourism. This makes rainfall a serious
problem and requires better rainfall forecasts. For these various reasons, accurate forecasting of rainfall is
paramount. There are many ways to predict it, but the one that is chosen for this project is to observe and
collect previous year's rainfall data which is collected over 10 years of rainfall measurements. Next, we forecast
rainfall for the next day. Therefore, this project seeks to optimize results and find suitable machine learning
models for predicting rainfall. We will also compare the machine learning algorithms and methods used in
machine learning. Methodologies include logistic regression, support vector classification, random forest,
catboost, and xgboost. Compare each of these algorithms individually, taking into account the parameters that
apply to machine learning, such as accuracy.
Keywords: Accuracy, Forecasting, Machine Learning Algorithms, Rainfall.
I. INTRODUCTION
Rainfall forecasting is one of the challenging and uncertain tasks that has a major impact on human society.
Timely and accurate forecasts help you proactively reduce human and financial losses. Heavy rain forecasts are
a big problem for the Meteorological Bureau because they are closely related to the economy and people's lives.
This is the cause of natural disasters such as floods and droughts that people all around the world face each
year. As global warming progresses, rainfall detection and prediction has become a major issue in countries
where appropriate technology is not available, and if done correctly, it can serve a variety of purposes such as
agriculture, health and drinking. The accuracy of precipitation forecasts is very important where the economy
is heavily dependent on agriculture. Due to the dynamic nature of the atmosphere, statistical methods do not
provide excellent accuracy in predicting precipitation. The non-linearity of precipitation data makes machine
learning and AI better techniques. Predictions help people take precautions, and predictions must be accurate.
The purpose of this project is to provide non-experts with easy access to the techniques and approaches used in
the field of precipitation prediction, to provide comparative studies between various machine learning
techniques, and to provide early precipitation prediction and to identify the best machine learning algorithm
for you. The objective of this system is to understand & analyze the data collected for rainfall for all the states
and derive conclusive information & statistics regarding the rainfall pattern across the country as well as to
design an interactive and user-friendly web interface for the system for user’s ease of understanding the results
obtained from the input data provided.
II. LITERATURE SURVEY
A. Kala and Dr. S. Ganesh Vaidyanathan in [1] have used Artificial Neural Network for implementing their idea.
Their procedure for rainfall prediction includes gathering the weather data, then preprocessing it, building the
Feed Forward Neural Network (FFNN) model with training data and then validating it with testing data and in
the end evaluating the model by comparing desired and actual output. There was 0.935 accuracy using Artificial
Neural Network. R. Kingsy Grace and B. Suganya in their paper has said that they have implemented their
[Link] @International Research Journal of Modernization in Engineering, Technology and Science
[2194]
 e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:04/Issue:03/March-2022 Impact Factor- 6.752 [Link]
model using machine learning [2]. They have compared various models such as Deep Convolutional Neural
Network, Genetic Programming, ANN, Linear Regression, Hybrid Neural Network, Likelihood, LSTM and
ConvNet. There was 99% accuracy of the Multiple Linear Regression. Hiyam Abobaker Yousif Ahmed and
Sondos W. A. Mohamed in [3] have implemented their model using linear regression. Their whole idea is done
in 2 major parts i.e., first is data collection and selection and second is data cleaning and transformation. There
was 85% accuracy given by the Multiple Linear Regression model. CMAK Zeelan Basha, Nagulla Bhavana,
Ponduru Bhavya, Sowmya V. has implemented a model on the same topic using machine learning and deep
learning techniques [4]. They have taken the help of Auto-Regressive Integrated Moving Average (AREMA
Model), Artificial Neural Network, Support Vector Machine and Self Organizing Map. The outcomes intend that
in terms of MSE and RMSE, their architecture outruns other approaches. B. Vasantha, R. Tamilkodi, L.
Venkateswara Kiran in [5] have forecasted the rainfall applying actual time global climate parameters. They
have mainly used convolution neural systems to anticipate weather parameters which in result will give
meaningful designs to understand the forecasting. They believe that the results precision is expected to be
extended by 70%. Anjali Samad, Bhagyanidhi, Vaibhav Gautam, Piyush Jain, Sangeeta, Kanishka Sarkar has done
precipitation forecasting applying Long Short Term Memory Neural Network [6]. This model is based on
Australian dataset which covers seasonal decomposition methods. They claim that their model has been
forecasting accurately for check cases but there was some increase in error rate while handling the outliers.
Rose Ellen N. Macabiog, Jennifer C. Dela Cruz have done the model building on rainfall predictive approach
applying classification [7]. They have done data collection, data preprocessing, building the predictive model,
model evaluation and selection and finally testing and evaluation. They observed that we can get most accurate
results if we use all 5 attributes for the model. Course KNN, Fine Gaussian SVM, Neural Network given the best
0.811 accuracy Arief Bramanto Wicaksono Putra, Rheo Malani, Bedi Suprapty, Achamad Fanany Onnilita Gaffar
have implemented Deep Auto Encoder using Semi CNN [8]. Their method of working includes Auto-Encoder
Neural Network (AENN), Convolution Neural Network (CNN), proposed method, model data time series
autoregressive (AR model). This approach is one of the recent ones for rainfall prediction with 99%
performance. Yuana Ratna Sari, Esmeralda Contessa Djamal, Fikri Nugraha in [9] have done the precipitation
forecasting using 1-D CNN. They have done preprocessing and convolutional neural network (CNN). They
observed that there was 0.9463 accuracy of training data and 0.8146 accuracy of testing data. Also, they
configured that more the layers you use in the model, more is the accuracy. Eslam Hussein, Mehrdad
Ghaziasgar, Christopher Thron in [10] have applied Support Vector Machine Classification. The methodology
used consists data gathering, preprocessing, classification, comparison between different SVM inputs,
comparison between regional predictions. Nikhil Oswal in [11] has used machine learning techniques to
implement his model. His methodology includes data exploration and analysis, data preprocessing, modelling
and evaluating. At the end he has concluded that Australian rainfall is not certain there is not a particular
relation between time and shower. Still, he is able to find certain pattern and develop high performance model.
III. PROPOSED SYSTEM
The proposed solution to this is to predict precipitation in short-term forecasts. Designing and developing a
precipitation forecasting system with a web interface (GUI). To predict precipitation a few days ago at a
particular location. Accurate and accurate forecasts help develop better strategies for agriculture and water
storage, and are also informed about floods to implement precautionary measures. Prediction systems are
implemented comparing various machine learning algorithms such as SVR, ANN, and multiple regression. The
data is analyzed and visualized using histograms, graphs, etc. to derive meaningful information from the
patterns of precipitation data obtained. The results of the various ML algorithms implemented are compared
with accuracy. The aim of the proposed study is to be effective and efficient in predicting rainfall with maximum
accuracy and precision.
IV. METHODOLOGY
The first and most important step in the process is data collection. We have collected our data from the official
website of the Indian Government from the year 1951 to 2015 from all the districts and for all months. Next
step is Converting the data into the correct format to conduct experiments i.e. doing preprocessing tasks on our
dataset like replacing missing and null values with the mean of the column and detecting and removing outliers
[Link] @International Research Journal of Modernization in Engineering, Technology and Science
[2195]
 e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:04/Issue:03/March-2022 Impact Factor- 6.752 [Link]
and visualizing using boxplots. Next step involves analysis of the data and observing variations in the patterns
of rainfall to derive conclusions and also to determine the correlation between the different parameters. After
that, we visualize our analyzed data using bar graphs, heat maps, histograms, etc. for a better and easy
understanding of the information obtained. After that, we try to predict the average rainfall by separating data
into training and testing datasets. We apply various statistical and machine learning approaches like SVR, KNN,
logistic regression, etc. in prediction and make analysis over various approaches. At last, we intend to create a
simple, user-friendly, and interactive webpage so that users can input different parameters and can get
predictions accordingly. Figure 1 depicts the methodology.

Figure 1: Methodolgy
Figure 2 explains the basic data flow of the entire methodology and explains each of the individual components
of our approach. The different components are data collection, data cleaning and analysis, data visualization,
splitting into training and testing set, implementing all the mentioned ML Models, input state, month and other
values through a web interface to predict the rainfall.

Figure 2: Flow of the system

[Link] @International Research Journal of Modernization in Engineering, Technology and Science
[2196]
 e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:04/Issue:03/March-2022 Impact Factor- 6.752 [Link]
V. IMPLEMENTATION AND RESULTS
Dataset selection and acquisition:
Authors together decided that the dataset on which they will be working should have ample number of rows or
tuples which helps in good and accurate prediction also it should have necessary attributes so that they can
determine how many can be effectively used for the prediction considering their co-relation. After comparing
different datasets, they came to conclusion that the one that is available on Kaggle with title “Rain in Australia –
Predict next-day rain in Australia” owned by Joe Young would be ideal for their prediction model. It has 23
distinct attributes and 145460 tuples of data. Out of 23 attributes 16 attributes contain numerical values and
remaining 7 columns have categorical values. Furthermore, that 16 numerical values can be subdivided into 2
discrete and 14 continuous values.
Data pre-processing:
After successfully loading the dataset, we checked if this dataset has null values or not. It was the same case
which happens with most of the dataset meaning there were null values present. 21 out of 23 attributes has
some or more null values. After checking the percentage null values, there were 4 such attributes of which more
than 35% of rows were not having any data. Replacing that amount of data with only one value, that could be
mean or median, affects the variance because of its central tendency, also it alters visualization considerably.
So, for those 4 attributes, we have defined a function which will randomly fit a value at the place of null or NaN.
Then for the remaining columns which have continuous numerical values we used the method of replacing null
values with median value of that column. After that, attributes having categorical value with null values are also
replaced with distinct but dummy values. At this point there were zero value in the entire dataset.
Our next aim was to convert categorical values into numerical values which will later will be used for
prediction. Categorical values such as yes and no are replaced with one, zero respectively; location, wind
direction these attributes are assigned distinct representative number. After making the dataset completely
numerical we checked for outliers in the attributes using boxplot. Outliers provide useful insights into the data
you are investigating and can affect statistical results. This can help you find inconsistencies in the statistical
process and find errors. For outlier removal we first define a range using inter quartile range in which if a value
resides it isn’t considered as an outlier. But if not, then it needs to get changed so that it won’t bother the
prediction. So, for that we looked for lowest and highest value of that range. Attribute value lower than that of
lowest range value is replaced with the lowest value and higher attribute value than highest range value is
replaced by highest value. After the process of pre-processing, we also drew some graphs and plots along with
boxplot such as histogram, distplot, countplot, subplot for the visualization.
Model validation:
A validation process involves comparing a trained model to a test set of similar data. The testing set is different
from the training set from which the model was trained. Model validation follows model training in order to
find a perfect model with the most ideal performance. The testing data set is utilized principally to check the
speculation capacity of a prepared model. Model validation aims to find an ideal model that has the best
performance. For this first thing that we have done is train test split. Because a model can't be evaluated by
using the same data it was trained on. It needs to get evaluated it with fresh data. So, we did 80-20 percent split
of our total dataset. 80 percent is used for training and remaining for testing.
Another problem we are considering here is handling imbalanced data. An imbalanced dataset is one where one
class has a very high number of observations while the other has many fewer, i.e., one class has very high
observations while the other has very few observations. A standard approach to deal with an imbalanced
dataset is to resample the data. There are two kinds of resampling: oversampling and undersampling.
Oversampling is more commonly used than undersampling. The reason is that when we undersample, we tend
to remove instances from the data that may contain valuable information. A technique called Synthetic Minority
Oversampling Technique (SMOTE) generates synthetic samples for the minority class, thereby reducing the
overfitting problem posed by random oversampling. In this method, interpolation between positive instances
that lie together is used to create new instances based on the feature space. That’s why we did perform this
technique and gathered resampled data. Then we applied this data to seven different algorithms to check which
gives us the best results among the seven. The seven algorithms include CatBoost Classifier, Random Forest
[Link] @International Research Journal of Modernization in Engineering, Technology and Science
[2197]
 e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:04/Issue:03/March-2022 Impact Factor- 6.752 [Link]
Classifier, Logistic Regression, Gaussian Naïve Bayes, K Neighbours Classifier, XGBoost Classifier and Support
Vector Classifier. After considering every algorithm’s accuracy, RoC curve and AuC score, we decided that
CatBoost will be most appropriate algorithm for prediction. Below is the table having accuracy and AuC score
for each of the algorithms.
Table 1. Accuracy and AuC score of the algorithms
Sr. No. Algorithm Accuracy AuC Score
1. Catboost Classifier 0.9112 0.89
Random Forest
2. 0.8614 0.87
Classifier
3. Logistic Regression 0.7686 0.85
4. Gaussian Naïve Bayes 0.7501 0.82
5. K Neighbours Classifier 0.7491 0.79
6. XGboost Classifier 0.9081 0.88
Support Vector
7. 0.7770 0.85
Classifier
Feature selection:
As the crucial features are determined during this step of Data cleaning, feature selection is vital to the process.
By selecting the relevant features, we not only get rid of the unimportant ones but also increase the
performance of our model. We are using mutual_info_classif method in our project. This is an interesting
method of selecting features based on mutual information(entropy) gain that is applicable to classification
problems. Using this method, the model is more accurate due to its univariate filtering. In univariate methods,
because features are calculated separately instead of in groups, the top ten performing variables seem to
perform poorly when they are grouped. This results in the selection of suboptimal features. Nevertheless,
univariate filtering methods are relatively quick and can be used for screening, which leads to better
performance and less training. That’s why we figured out mutual information values of every attribute,
arranged it decreasing order (Higher the value, more the dependency) and selected top 10 attributes which
were well suited for prediction. Taking these 10 attributes and performing resampling we used CatBoost
Classifier for the prediction model. Figure 3 illustrates the value on which we got the best 10 attributes.

Figure 3: Mutual Information Score

Interface building:
A web interface has been built to help users understand the importance of building prediction models. In order
to create the web interface, we used flask. It is a web framework which is a Python module that lets you create
web applications easily. One of the important parts of this interface is analysis of the dataset that we have done
[Link] @International Research Journal of Modernization in Engineering, Technology and Science
[2198]
 e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:04/Issue:03/March-2022 Impact Factor- 6.752 [Link]
using PowerBI. Power BI is a business intelligence tool that allows users to visualize raw data and analyze it as
well as present it in the form of actionable insights. We have made use of line chart, heatmap, bar graph,
decomposition tree, area chart, radial chart to analyze our data. We have also used slicer for periodic or time-
oriented analysis.
Figure 4 and 5 holds the view of our powerbi analysis dashboard on which you can see all the charts and plots.
There are two dedicated predictor pages designed for making prediction using user’s data. On one page we are
predicting using limited attributes that we have got using feature extraction. And on the other we can enter
data of all 22 entities for prediction.

Figure 4: PowerBi Analysis Dashboard 1

Figure 5: PowerBi Analysis Dashboard 2

VI. CONCLUSION
In this paper, we have explored and applied several preprocessing steps and implemented various machine
learning algorithms. Further we have carried out a comparative study to understand the overall performance of
the algorithms and select the most accurate and effective algorithm to implement. Out of the seven algorithms
implemented, Catboost classifier proved to be the most efficient ml model with an accuracy of 91.12% and AuC
score of 0.89. We have also developed an user-friendly front end system to accept different input parameters to
observe how the input data affects the model prediction and how accurately it will predict the rainfall for the
next day. A well-developed front end and interactive dashboard provides easy and better understanding of the
data and the built system. We concluded that Australia's rainfall pattern is irregular and uncertain. Using
Feature extraction and data visualization through PowerBI dashboard we were able to recognize and figure out
relationship among the data and identify the factors which have significant effect on the rainfall received in the
Australian Regions.

[Link] @International Research Journal of Modernization in Engineering, Technology and Science

[2199]
 e-ISSN: 2582-5208
International Research Journal of Modernization in Engineering Technology and Science
( Peer-Reviewed, Open Access, Fully Refereed International Journal )
Volume:04/Issue:03/March-2022 Impact Factor- 6.752 [Link]
VII. FUTURE SCOPE
More changes & ideas can be adapted such as integrating the system with a flood prediction and alert model
pertaining to the rivers. The system can also include another prediction model to predict whether it will rain
tomorrow or not based on weather data for that particular day in major cities. Additional Features such as real-
time updates & major alerts around the country can be displayed on the website for precautionary measures.
VIII. REFERENCES
[1] Kala, A., & Vaidyanathan, S. G. (2018, July). Prediction of rainfall using artificial neural network. 2018
International Conference on Inventive Research in Computing Applications (ICIRCA).
[2] Grace, R. K., & Suganya, B. (2020, March). Machine Learning based Rainfall Prediction. 2020 6th
International Conference on Advanced Computing and Communication Systems (ICACCS).
[3] Ahmed, H. A. Y., & Mohamed, S. W. A. (2021, February 26). Rainfall Prediction using Multiple Linear
Regressions Model. 2020 International Conference on Computer, Control, Electrical, and Electronics
Engineering (ICCCEEE).
[4] Basha, C. Z., Bhavana, N., Bhavya, P., & V, S. (2020, July). Rainfall Prediction using Machine Learning &
Deep Learning Techniques. 2020 International Conference on Electronics and Sustainable
Communication Systems (ICESC).
[5] Vasantha, B., Tamilkodi, R., & kiran, L. V. (2019, March). Rainfall pattern prediction using real time
global climate parameters through machine learning. 2019 International Conference on Vision
Towards Emerging Trends in Communication and Networking (ViTECoN).
[6] Samad, A., Bhagyanidhi, Gautam, V., Jain, P., Sangeeta, & Sarkar, K. (2020, October 30). An approach for
rainfall prediction using long short term memory neural network. 2020 IEEE 5th International
Conference on Computing Communication and Automation (ICCCA).
[7] Macabiog, R. E. N., & Dela Cruz, J. C. (2019, November). Rainfall Predictive Approach for La Trinidad,
Benguet using Machine Learning Classification. 2019 IEEE 11th International Conference on Humanoid,
Nanotechnology, Information Technology, Communication and Control, Environment, and Management
( HNICEM ).
[8] Wicaksono Putra, A. B., Malani, R., Suprapty, B., & Onnilita Gaffar, A. F. (2020, July). A deep auto encoder
semi convolution neural network for yearly rainfall prediction. 2020 International Seminar on
Intelligent Technology and Its Applications (ISITIA).
[9] Sari, Y. R., Djamal, E. C., & Nugraha, F. (2020, September 15). Daily rainfall prediction using one
dimensional convolutional neural networks. 2020 3rd International Conference on Computer and
Informatics Engineering (IC2IE).
[10] Hussein, E., Ghaziasgar, M., & Thron, C. (2020, July). Regional rainfall prediction using support vector
machine classification of large-scale precipitation maps. 2020 IEEE 23rd International Conference on
Information Fusion (FUSION).
[11] Oswal, N. (2021). Predicting Rainfall using Machine Learning Techniques. Institute of Electrical and
Electronics Engineers (IEEE).

[Link] @International Research Journal of Modernization in Engineering, Technology and Science

[2200]