Weather forecasting with data science approaches (2023)

In this article, I would like to show how we can use a data science algorithm for weather forecasting and compare some frameworks for classification and regression tasks.

Weather forecasting with data science approaches (1)

Weather forecasting is a quite difficult task. The Wiki said, “ Weather forecasting is the application of science and technology to predict the conditions of the atmosphere for a given location and time. Weather forecasts are made by collecting quantitative data about the current state of the atmosphere at a given place and using meteorology to project how the atmosphere will change.”

So in general, weather forecasting is about the data about the atmosphere. Of course, I would not model the atmosphere condition, but I can use a lot of data about it for weather forecasting. In this article, I will try to predict the precipitation for the next 3 hours. I will try to solve this task in two ways — as a binary classification (1- precipitation, 0- no precipitation) and regression, where I will try to predict the amount of precipitation in mm.

So, what is precipitation? Let’s ask it in Wiki.” In meteorology, precipitation is any product of the condensation of atmospheric water vapor that falls under gravity. The main forms of precipitation include drizzle, rain, sleet, snow, graupel, and hail. Precipitation occurs when a portion of the atmosphere becomes saturated with water vapor so that the water condenses and “precipitates”. ” Let look at the schema, which shows the process of the water cycle and precipitation occurrence.

Weather forecasting with data science approaches (2)

Of course, the precipitation occurrence is a very complicated process and those schema does not show the whole process. The Wiki said, “Precipitation is a major component of the water cycle and is responsible for depositing most of the freshwater on the planet. Mechanisms of producing precipitation include convective, stratiform, and orographic rainfall. Convective processes involve strong vertical motions that can cause the overturning of the atmosphere in that location within an hour and cause heavy precipitation, while stratiform processes involve weaker upward motions and less intense precipitation. Precipitation can be divided into three categories, based on whether it falls as liquid water, liquid water that freezes on contact with the surface, or ice. Mixtures of different types of precipitation, including types in different categories, can fall simultaneously. Liquid forms of precipitation include rain and drizzle. Rain or drizzle that freezes on contact within a subfreezing air mass is called “freezing rain” or “freezing drizzle”. Frozen forms of precipitation include snow, ice needles, ice pellets, hail, and graupel.”

So, let’s try to find some data and try to predict this natural phenomenon. In my experiment, I will use data from one meteorology station. Let’s examine the header of data.

So, I have got 29 variables and 22 064 records. Let me describe the column names:

  • T — Air temperature (degrees Celsius) at a 2-meter height above the earth\’s surface;
  • Po — Atmospheric pressure at weather station level (millimeters of mercury);
  • P — Atmospheric pressure reduced to mean sea level (millimeters of mercury);
  • Pa — Pressure tendency: changes in atmospheric pressure over the last three hours (millimeters of mercury);
  • U — Relative humidity (%) at a height of 2 meters above the earth\’s surface;
  • DD — Mean wind direction (compass points) at a height of 10–12 meters above the earth’s surface over the 10 minutes immediately preceding the observation;
  • Ff — Mean wind speed at a height of 10–12 meters above the earth’s surface over the 10 minutes immediately preceding the observation (meters per second);
  • ff10 — Maximum gust value at a height of 10–12 meters above the earth’s surface over the 10 minutes immediately preceding the observation (meters per second);
  • ff3 — Maximum gust value at a height of 10–12 meters above the earth’s surface between the periods of observations (meters per second);
  • N — Total cloud cover;
  • WW — Present weather reported from a weather station;
  • W1 — Past weather (weather between the periods of observation) 1;
  • W2 — Past weather (weather between the periods of observation) 2;
  • Tn — Minimum air temperature (degrees Celsius) during the past period (not exceeding 12 hours);
  • Tx — Maximum air temperature (degrees Celsius) during the past period (not exceeding 12 hours;
  • Cl — Clouds of the genera Stratocumulus, Stratus, Cumulus, and Cumulonimbus;
  • Nh — Amount of all the CL cloud present or, if no CL cloud is present, the amount of all the CM cloud present;
  • H — Height of the base of the lowest clouds (m);
  • Cm — Clouds of the genera Altocumulus, Altostratus, and Nimbostratus;
  • Ch — Clouds of the genera Cirrus, Cirrocumulus, and Cirrostratus;
  • VV — Horizontal visibility (km);
  • Td — Dewpoint temperature at a height of 2 meters above the earth\’s surface (degrees Celsius);
  • RRR — Amount of precipitation (millimeters);
  • tR — The period during which the specified amount of precipitation was accumulated;
  • E — State of the ground without snow or measurable ice cover;
  • Tg — The minimum soil surface temperature at night. (degrees Celsius);
  • E — State of the ground with snow or measurable ice cover;
  • sss — Snow depth (cm).

First of all, I would like to know the summary of this dataset and then select a useful column for further analysis.

I replace NULL numeric value with mean and then I can visualize this dataset. Let’s start with Temperature(T).

First of all, I would like to plot the time series of Temperature. On the plot, you can see how Temperature changes during the “2012–02–01” — “2019–07–13”.

Weather forecasting with data science approaches (3)

Here we can find seasonal of our data, but I would like to examine it more detail. For examining the changes year by year I create a box plot.

Weather forecasting with data science approaches (4)

The changes are very little on this chart, so let’s examine it by month and year.

Weather forecasting with data science approaches (5)

Wow, It was interesting to find that summer in most years is similar, but autumn and spring are different. Also, it would be interesting to plot some trends in maximum, minimum and median temperature overall period.

Weather forecasting with data science approaches (6)

(Video) Introduction to Forecasting in Machine Learning and Deep Learning

It is interesting, that the minimum temperature for the last years became higher. For more details, you can examine it in the next table.

The next stage of my data analysis is the calculation of the correlation between all numeric values.

Weather forecasting with data science approaches (7)

On this plot, we can see three interesting correlation:

  • The positive correlation (0.4) between Temperature and Horizontal visibility;
  • The negative correlation (-0.6) between Temperature and Relative humidity;
  • The negative correlation (-0.7) between Relative humidity and Horizontal visibility.

Also, I would like to analyze more detail Amounts of precipitation and to plot a similar to Temperature charts.

Let’s begin with the time series plot.

Weather forecasting with data science approaches (8)

The summary of precipitation in autumn 2013 was the highest for the last 5 years. Let’s calculate summary statistics for non zero precipitation values.

Weather forecasting with data science approaches (9)

And also I would like to know the summary value of precipitation overall period.

Weather forecasting with data science approaches (10)

Here we can see that in 2014 and 2015 were the lowest level of precipitation in this period.

What about categorical data and its influence on precipitation? Let’s start with wind direction (DD). We have got 16 different directions in the dataset and one event with no wind.

Weather forecasting with data science approaches (11)

We can see that most precipitation may come from north or north-west.

The next stage is the analysis of the cloud. First of all, I would like to analyze the count of a day with cloud and analysis of the influence of cloud characteristics on the precipitation.

Weather forecasting with data science approaches (12)

Cloud coverage is a very similar overall period and mostly it is not changed. Let’s also analyze cloud characteristics.

Weather forecasting with data science approaches (13)

(Video) Time Series Forecasting with Machine Learning

This chart shows, that most precipitation brings with cloud on high 300–600m. That’s all, that I want in exploratory data analysis.

Firstly, I calculate the date feature — quarter, month, week, weekday, month day, year day, the hour. All this feature I can calculate with the lubridate library.

The second stage of feature engineering is the calculation of lag value for some numerical features. For my analysis, I would like to calculate the lag value for T, Po, P, U, Ff, N, Nh, VV and my target RRR. It would be 3 lag on a particular day and 1 lag for the previous day. Final data set:

So, it is time to split the dataset for the train, validation, and test subset. Train data set in the range “2012–02–02 02:00:00”-”2018–12–31 23:00:00", validation data set in the range “2019–01–01 02:00:00”-”2019–03–31 21:00:00", and test data set in the range “2019–04–01 00:00:00”-”2019–07–13 15:00:00".

Also, it is important to check the balance of classes in the subset and distribution of the target value for the regression task.

Weather forecasting with data science approaches (14)

So, as a result, we have got unbalanced data sets for classification and a little bit different distribution for regression.

Classification

Ok, let’s start modeling. The framework, that I would like to test is H2O.

H2O is a fully open-source, distributed in-memory machine learning platform with linear scalability. H2O supports the most widely used statistical & machine learning algorithms including gradient boosted machines, generalized linear models, deep learning and more. H2O also has an industry-leading AutoML functionality that automatically runs through all the algorithms and their hyperparameters to produce a leaderboard of the best models.

In my experiment, I would like to use gradient boosted machines, generalized linear model, deep learning and AutoML. One of the H2O benefits is an integrated analysis of the importance of factors.

My first model would be — the generalized linear model for the classification task. As we see before our label is unbalanced, so we need to set balance_classes = TRUE. Also for classification, we need to set family = ‘binomial’. So, let’s build it and look at the importance of the features.

Weather forecasting with data science approaches (15)

As we can see, the most important factor is U humidity one hour before precipitation. And what about accuracy? Let’s calculate all main metrics:

Weather forecasting with data science approaches (16)

So in general, our model predicts good, but the precipitation we can predict only in 56%.

Let’s try another model. My second model is gradient boosted machines. I also to set balance_classes = TRUE. The importance of the features will look like:

Weather forecasting with data science approaches (17)

The most important factors are different lags of precipitation and the cloud cover. And what about the accuracy of this model:

Weather forecasting with data science approaches (18)

The result is better. We can predict precipitation more accurate. I would like to try a more complicated model. It would be a deep learning algorithm, which H2O also provided. I also to set balance_classes = TRUE. The importance of the features will look like:

Weather forecasting with data science approaches (19)

(Video) Big Data Assimilation Revolutionizing Weather Prediction

Wow, it is interesting the most important factors if all lag of precipitation and only then are humidity and cloud coverage. And what about the accuracy:

Weather forecasting with data science approaches (20)

The result is worst then in GBM algorithm it can be connected with my deep learning model overfitting.

H2o also has got Automated machine learning. Let’s build 50 models and look at the results. TOP 5 models that were created — GBM with different hyperparameters. As usual. let’s look at the most important factors:

Weather forecasting with data science approaches (21)

You can see factors importance are very similar to my own GBM model, and what about accuracy:

Weather forecasting with data science approaches (22)

AUC is better than in my own GBM model, but TNR is worst. Also, it took more time for calculation.

Regression

Let’s build the regression model and try to predict the amount of precipitation.

Our first model is linear regression. As in the classification task, we can look at the features importances:

Weather forecasting with data science approaches (23)

The humidity has the most significant value. And what about accuracy? In this task, I would like to use 3 metrics: MAPE, MAE, MSE.

Weather forecasting with data science approaches (24)

The result of the GLM model is pretty bad.

Let’s try another model — gradient boosted machines. Firstly, I would like to look at the features’ importance:

Weather forecasting with data science approaches (25)

It is different from the GLM model, here we can see that feature precipitation in the previous hour has the most value.

The metrics are pretty good:

Weather forecasting with data science approaches (26)

Our complicated model as in classification task would be deep learning. The features importance looks like this:

Weather forecasting with data science approaches (27)

(Video) Weather Forecast from sensor data , Azure IOT Hub, Stream analytics and Machine Learning Studio

The most important features are the amount of precipitation in the previous and current hours. What about accuracy:

Weather forecasting with data science approaches (28)

The accuracy is lover then in GBM model, I think it is because of DL is too complicated.

The last one is AML. TOP 5 models are:

Weather forecasting with data science approaches (29)

As usual. let’s look at the most important factors of GBM model from AML:

Weather forecasting with data science approaches (30)

The result is very similar to my GBM model.

Let’s took at the results of the best AML model — StackedEnsemble_AllModels_AutoML:

Weather forecasting with data science approaches (31)

It is the best result of all the models that I created.

The H2O framework is good and easy to use. It can give a good result with a minimum time to implementation. It also has got a lot of hyperparameters to tune.

Let’s try another one — KERAS.

Keras is a high-level neural networks API developed with a focus on enabling fast experimentation. Being able to go from idea to result with the least possible delay is key to doing good research. Keras has the following key features:

  • Allows the same code to run on CPU or on GPU, seamlessly.
  • User-friendly API which makes it easy to quickly prototype deep learning models.
  • Built-in support for convolutional networks (for computer vision), recurrent networks (for sequence processing), and any combination of both.
  • Supports arbitrary network architectures: multi-input or multi-output models, layer sharing, model sharing, etc. This means that Keras is appropriate for building essentially any deep learning model, from a memory network to a neural Turing machine.
  • Is capable of running on top of multiple back-ends including TensorFlow, CNTK, or Theano.

I would like to build my neural network with multi-input and multi-output.

For input, I would like to use the same features that I use for H2O models. For categorical features, I would use the Entity Embeddings approach. More detail about this approach you can read in this paper — link.

Multi-output helps me to solve 2 tasks (classification and regression) at the same time.

The Keras framework has not own features importance module, but you can use Lime.

Let’s look at the results:

  • Classification:

Weather forecasting with data science approaches (32)

  • Regression:

Weather forecasting with data science approaches (33)

Keras’ neural network brings cool results. This model performs mostly better than the previous ones.

Weather forecasting is a really difficult task. The machine learning algorithms can help in prediction for a short term period. To get good results we need to use lag features or use RNN architecture in the neural networks.

All code you can find in the Git repository — link.

(Video) Time Series Forecasting and Analysis of Delhi's Weather with Deep Learning | Part 1

FAQs

How data science is used in weather forecasting? ›

Weather forecasts are made by collecting the maximum amount of data possible about the current state of the atmosphere (particularly the temperature, humidity, and wind) and using the understanding of atmospheric processes to determine how the atmosphere evolves in the future.

Which algorithm is best for weather forecasting? ›

The work proposes to predict a day's weather conditions. For this the previous seven days weather is taken into consideration along with fortnight weather conditions of past years.

What type of data is involved in weather forecasting? ›

Weather forecasts are made by collecting quantitative data about the current state of the atmosphere at a given place and using meteorology to project how the atmosphere will change.” So in general, weather forecasting is about the data about the atmosphere.

Which weather prediction model is most accurate? ›

Global models with worldwide weather forecasts

The ECMWF is generally considered to be the most accurate global model, with the US's GFS slightly behind.

Is big data used for weather forecasting? ›

In weather forecasting issue, big data is capable of upgrading the process of decision-making [13]. The importance of precise prediction can not be neglected; thus, we can use big data analytics in weather forecasting to achieve this goal.

What is weather data analysis? ›

Definition of Weather Analysis

It is a process that breaks the object of the analysis into parts so that each part can be examined in detail. This process determines its nature, function, and other characteristics.

How is machine learning used in weather forecasting? ›

Using a convolutional neural network, the authors developed a machine learning weather prediction system called Deep Learning Weather Prediction (DLWP). The model is trained on past weather data, which differs from standard numerical weather prediction models that create mathematical representations of physical laws.

Which machine learning algorithm is used for weather forecasting? ›

We are using four different machine learning algorithms for predicting weather. Algorithms – K-Nearest Neighbor, Decision Tree, Random Forest Algorithm, Linear Regression.

Is weather prediction using machine learning? ›

Some studies have shown that machine learning-based forecast systems can predict general weather patterns as well as numerical weather prediction models while using only a fraction of the computing power the models require.

What are 4 ways to collect weather data? ›

Observational data collected by doppler radar, radiosondes, weather satellites, buoys and other instruments are fed into computerized NWS numerical forecast models. The models use equations, along with new and past weather data, to provide forecast guidance to our meteorologists.

What are the 3 types of weather forecasting? ›

Types of Weather Forecasting

Short-range forecasts are predictions made between one and seven days before they happen. Medium-range forecasts are usually given between one week and four weeks in advance. Long-range forecasts are given between one month and a year in advance.

What are 5 ways we collect data on the weather? ›

Thermometers measure temperature, radar measures rain or snow locations and movements, barometers measure air pressure, rain gauges measure amounts of rain, wind vanes measure wind direction, anemometers measure wind speed, transmissometers measure visibility, and hygrometers measure humidity.

How many weather forecast models are there? ›

The three primary used synoptic forecast models are the North American Mesoscale Model or NAM (formally ETA), the Global Forecast System or GFS (formally AVN and MRF), and the long standing Nested Grid Model or NGM. There are also other models such as the RUC, Canadian Model, European Model.

How are weather forecasts models created? ›

Ground radar, weather balloons, aircraft, satellites, ocean buoys and more can provide three-dimensional observations that a model can use. This allows meteorologists to simulate what the atmosphere is currently doing and predict what will happen in the next few days or, for some models, hours.

Which weather app is most accurate? ›

2022's Top 10 Most Accurate Weather Apps Forecasting:
  1. The Weather Channel – Best Free Weather App. ...
  2. AccuWeather – Most Accurate Weather App. ...
  3. WeatherBug – Best App for Free Weather Alerts. ...
  4. Dark Sky - Best Weather App for iPhone. ...
  5. Shadow Weather – Best Weather App for Android. ...
  6. FlowX – Most Interactive Weather App.

How does big data help in understanding weather patterns? ›

Forecasting weather requires analysis of a large set of data at an instantaneous rate. The size and complexity of weather patterns make it very difficult to predict the weather a few days in advance [7]. Researchers have to refine their instruments and techniques to analyze and make more accurate weather forecasts.

What is weather forecasting system? ›

Weather forecasting is the application of science and technology to predict the conditions of the atmosphere for a given location and time. People have attempted to predict the weather informally for millennia and formally since the 19th century.

Videos

1. Stephan Siemen - Using Python in Weather Forecasting
(PyData)
2. Deep Uncertainty Quantification: A Machine Learning Approach for Weather Forecasting
(KDD2019)
3. Purely Data-Driven Approaches to Weather Prediction: Promise and Perils | Suman Ravuri | DISCOVERY
(AI for Good)
4. 2015 Unidata Users Workshop: Severe weather forecasting with Python and data science tools
(Unidata)
5. S2S forecasting using large ensembles of data-driven global weather prediction models
(US CLIVAR)
6. Dr. Takemasa Miyoshi | Big Data Assimilation A New Science for Weather Prediction and Beyond
(ESSICUMD)
Top Articles
Latest Posts
Article information

Author: Kelle Weber

Last Updated: 09/10/2022

Views: 5531

Rating: 4.2 / 5 (73 voted)

Reviews: 88% of readers found this page helpful

Author information

Name: Kelle Weber

Birthday: 2000-08-05

Address: 6796 Juan Square, Markfort, MN 58988

Phone: +8215934114615

Job: Hospitality Director

Hobby: tabletop games, Foreign language learning, Leather crafting, Horseback riding, Swimming, Knapping, Handball

Introduction: My name is Kelle Weber, I am a magnificent, enchanting, fair, joyous, light, determined, joyous person who loves writing and wants to share my knowledge and understanding with you.