Moji Weather, which has made a lot of money by betting on advertising, has opened up a new way to make money

[51CTO.com original article] A friend’s travel request was rejected because the PM2.5 index of the real-time air quality warning of Moji Weather (hereinafter referred to as Moji) was off the charts. Like most tool apps, Moji uses simple and crude advertisements. According to the prospectus submitted by Moji for listing on the Growth Enterprise Market at the end of 2016, advertising revenue accounted for more than 95%, with an annual profit of more than 20 million yuan, which can be said to be a huge profit from advertising.

Can the monetization method of betting on advertising really go far? From the perspective of Moji's usage scenarios, its user stickiness and product matrix need to be improved, which will also directly affect the future advertising conversion rate. Recently, it was learned that Moji began to lay out B-side business from the end of 2015 to the beginning of 2016, devoting itself to the development of the company's accumulated technical talents, massive meteorological data, research in the meteorological field, and the country's openness to meteorological data, providing enterprise-level meteorological services for industries with special meteorological needs, which will be a new way to attract money.

Whether it is traffic monetization on the C-end or customized weather service solutions on the B-end, meteorological data is the cornerstone for commercial success. Wang Lei, senior technical director of Moji, introduced that Moji APP has been accumulating data since its founding in the Saipan era until today, and all the original data has been retained. In the past four or five years, big data in niche areas has become increasingly important, and Moji has also begun to make efforts to build big data platforms and invest manpower, trying to find models and rules in users and weather forecasts from massive meteorological data. Based on these, personalized recommendations and refined services are made, of course, including commercial operations.

Sources and Analysis of Massive Meteorological Data

Data Source

Observation data is a necessary and sufficient condition for weather forecasting, which directly affects the accuracy of the forecast. The ideal state of observation data is to have enough observation points, so that the weather forecast will be more accurate. For example, if 100 observation points are set up in Beijing, but Beijing is vast, what are the temperatures and air pressures between points?

MoJi's data mainly comes from three aspects: third-party meteorological organizations, various device sensors and Shijing community.

The third aspect is to cooperate with other meteorological companies, such as the China Meteorological Administration, the GFS of NOAA (National Oceanic and Atmospheric Administration), the European Center for Medium-Range Weather Forecasts (EC), Japan's meteorological satellite data, and the data of the China Meteorological Administration, etc. These organizations have global observation data, which are remotely sensed by satellites, so the quality is relatively high, about 500G per day.

The second aspect is various sensors. For example, the mobile phones of C-end users of Moji are basically equipped with barometers and temperature sensors. There is also cooperation with General Motors, and there are residual meters on the car, and the number of sensors is very large. Meizu and Huawei mobile phones are also pre-installed with the Moji App, from which relevant data can be obtained. The amount of this part of data is about 80 million per day, but due to different forms, the data on each device has deviations and fluctuations, resulting in uneven quality, and unified processing is required before it can be put into use.

The third aspect is Shijing Community, which is a real-time weather community. About 100,000 weather photos are uploaded to Shijing Community every day. The total meteorological picture resources reach 100 million. It is currently the largest real-time weather picture community in China.

Data analysis

Moji's data analysis is divided into two parts: one is the short-term forecast of 0 to 2 hours and the short-term forecast of 2 to 8 hours. The other part is the medium- and long-term forecast of 8 hours to 15 days. Machine learning is mainly used in short-term forecasts, mainly using the industry's more advanced neural learning networks, such as FCN networks (full convolutional networks) and Conv-LSTM networks (long short-term memory networks) and other specific branches to eliminate noise from data from different sources, and then integrate them to learn the historical change trends of these meteorological data.

In addition to allowing C-end users to check the real-time weather before going out, B-end users can also combine weather forecasts to save costs and improve efficiency. Zhang Mingming, VP of Moji Commercialization, said that industries such as transportation, terminal logistics, and agriculture have relatively high demands for short-term forecasts. For example, in the case of Sinopec logistics and distribution, the two parties integrated historical weather data and Sinopec’s sales data to conduct model training, helping Sinopec solve the problem of long-term and cross-provincial resource allocation.

The technology used in short-term forecasting is machine learning, while medium- and long-term forecasting is another system. The following content will focus on the specific development history and technical details of short-term forecasting.

Application of machine learning in Moji Weather

Development History and System Architecture of Short-term Forecast

The Moji short-term forecast system started to develop from the internal test of the patrol team leader in 2015, and has gradually matured. The following figure shows the development process of Moji short-term forecast:

Throughout its development, Moji has three milestone events :

​​

• In August 2016, a deep learning CNN network was used for the first time to process noise in radar echoes, which is effective for both single radar noise and partially mixed radar noise.
• In March 2017, by combining artificial intelligence technology with traditional model forecasting technology, GFS forecast data was added as a supplementary data source in areas without radar coverage, and the accuracy was further improved.
• In May 2017, deep learning algorithms were applied in the short-term core extrapolation link.

The following figure shows the top-level design of Moji short-term forecast :

​​

The top-level design of the Moji short-term forecast consists of three parts: input (data source, feedback, WRF), middle layer (denoising, extrapolation) and output (forecast map).

Main technologies involved in short-term forecasting systems

There are many technologies used in the short-term forecast system. Here we mainly share two parts: algorithm model and algorithm implementation.

The algorithm models include deep learning image denoising algorithm (CNN network), deep learning image extrapolation algorithm (RNN recurrent network), machine learning rain and snow discrimination model (SVM support vector machine classification), pattern forecast data fusion, etc.

The algorithm implementations include Google Tensorflow deep learning framework, Caffe deep learning framework, Opencv image processing library, Sklearn machine learning library, etc.

Denoising technology

The denoising problem in deep learning can be attributed to a task called "image semantic segmentation". Compared with ordinary image classification tasks, this type of task requires marking the category of each point in the image. For radar image denoising, it is necessary to mark each pixel point whether it is a normal echo or noise. Moji has accumulated a large number of noise data sets through manual collection, and has manually marked and trained them. At present, the denoising process has iterated three versions.

Extrapolation Technology

The extrapolation technology uses the RNN (recurrent neural network) algorithm.

Here is a simplified diagram of RNN (Recurrent Neural Network) :

​​

The forecasting problem is a type of time series problem, and recurrent neural networks are very suitable for dealing with time series related problems.

Below is a simplified diagram of the LSTM (Long Short-Term Memory) network :

​​

Short-term forecast extrapolation requires multiple cycles of calculation. Traditional recurrent neural networks will obviously lose the initial input data features of the network during so many cycles, resulting in low accuracy of the extrapolated images. Long-term and short-term networks can significantly improve this problem.

Moji uses recent historical images to input into the recurrent network in chronological order, and then the network generates future prediction graphs based on the changing trends of the images and the laws of historical changes.

Rain and snow discrimination technology

The following figure shows the rain and snow forecast based on GFS forecast data :

The following figure shows the classification of rain and snow based on model forecast data :

​​

Since the short-term forecast results can only predict the precipitation conditions but cannot distinguish the type of precipitation, an additional model is needed to judge the precipitation type. The rain and snow discrimination model uses machine learning methods to determine whether the precipitation type is rain or snow, so as to provide the final actual weather type.

Numerical forecasting technology

Below is the national radar station coverage map:

The following are the model forecast data :

At present, MoJi is also doing some research and application in traditional numerical weather forecasting.

People have been using numerical weather forecasting methods to solve weather forecasting problems since the last century. By solving numerical solutions to a series of dynamic and thermodynamic partial differential equations that describe atmospheric motion, they calculate the future state of the atmosphere and thus predict future weather.

However, the development of this forecasting method depends on the development of computer technology, because its computational workload is quite astonishing. Each time step has 500 million spatial grid points, the spatial scale extends from hundreds of meters to thousands of kilometers, and the time scale ranges from a few seconds to a few weeks.

Currently, Moji's meteorological research team is also engaged in related research and applications. The finest model resolution is 3-5km, and each time step requires the calculation of millions (1.4 million) spatial grid points. The amount of calculation is also quite astonishing, and it requires the use of supercomputer or cluster resources to be applied.

About the future

At present, market competition in the meteorological field is becoming increasingly fierce. I don’t know how far Moji can go in making money by providing enterprise-level meteorological services to enterprise-level users based on its leading technology of massive meteorological data analysis. But at least Moji’s transformation from a functional App operator to an Internet comprehensive meteorological service provider has made Moji’s path wider and wider.

Top 3 hot articles recommended this month

​​Things about hybrid cloud: how to make public cloud and private cloud achieve 1+1>2​​

​​AR/VR experts tell you: What will the future scene of human-computer interaction look like?​​

How to balance security and performance? Exploration and practice of HTTPS optimization for e-commerce websites

[51CTO original article, please indicate the original author and source as 51CTO.com when reprinting on partner sites]