Nidhika Yadav

Abstract– Climate change is here; natural calamities have become a regular phenomenon in world. There are various artificial intelligence-based techniques that can be used to predict and detect natural calamities, some of which include deep learning-based techniques such as RNN, LSTM, Transformers, Hybrid CNN models by considering the atmospheric data in form of low pressures and high pressures, humidity, moisture, geographic locations. These are taken at regular interval and proposed to be fed in LSTM and models are learned and built. However, these require high computational capacities. Due to restricted computational power, here, this paper only proposes how to use deep learning techniques in predicting climate calamities. While this paper also provides the proposals of the use of Fuzzy Inference Engine of AI for climate change calamities predictions, based on weather conditions read live using sensors, such as high amount of humidity for an example.  Example of modelling process are explained for experts in fields to make the exact models.

Keywords: Deep Learning, Fuzzy Inference Systems, Fuzzy Logic, LSTM, RNN, Climate Change, Acute Weather Forecasting Application, AI Software

1. Introduction

Climate change is here and its effects are clearly seen all around in every place, and this is not restricted to a small region. It is a global issue. If no one can control the climate change things predominately, one for sure can prepare the human life for safety. Also, as much as other species as possible can be saved with the timely notification of such warnings.

The weather monitoring systems works well. Then why issues like heavy rains, cloud bursts, forest wildfire are still not predicted in advance. What shall happen if such a thing, viz., natural calamity is detected in advance? Well, rescue operation can be projected in planning by town and measure such as excess water disposal systems can be dug in right places at right times, apart from that in case of forests, right amount of water can be supplied to these places, well in advance.

It is to note such a sophisticated weather monitoring system is well developed, still, people are not informed of the amount of rains that may occur, in time, so that they can safe essential things within right frameworks of time. The reason is yes, the weather and climate monitoring systems are sophisticated. Then has not much have been done to predict right amount of information using Artificial Intelligence?

This paper focus on the AI based proposals of use of deep learning and a Fuzzy Systems Inferencing technique to raise alarms when needed. AI based techniques can help plan travel and especially saving lives and natural resources apart from valuables, houses and other living beings. Firstly, a deep learning-based proposal is laid out in this article. Further, a locality-based alarm system is proposed using Fuzzy Logic, to raise alarms in locality automatically, in case of natural disasters are predicted. Both the models-deep learning and Fuzzy logic can be used for predictions of heavy rains to other extremes. Conditions such as very hot conditions in one area can lead to wildfires in that area, and heavy rains in other areas to mention a few. Hence the neighboring areas weather has to be considered. The aim is to raise an alarm for people on duty that AI has detected an extreme condition. Other calamities can be predicted in similar manner.

The article is organized as follow. Section 2 describes how deep learning-based techniques can be used for predicting exact weather effect, not just rains, but how much of rain, not just heat, but what shall be impact of heat and how much of it. Section 3, illustrate the use of Fuzzy Logic for weather forecast predictions and alarm raiser in localities. Section 4, concludes the article and also present the future work in this area.

• Deep Learning based Weather Forecast Proposal

Weather forecast can be done on basis of normal neural networks too. But adding the advantageous of deep learning in this system, helps understand non linear relations that exists in the weather data. The climatic relation, which cannot be humanly understood, between various climatic conditions can be learned in a nonlinear fashion by a deep learning model. Firstly, data is all there with the metrological systems. The data can be considered in form of time series data. Each instant of time over a specific time interval, for instant, for every 15 days, per hour data. The data for every hour is represented as the measurements done per minute, for all 12 hours. The article propose that the following measurements have to be collected per hour:

1. The measure of atmospheric pressure as a vector of 12 hours.
2. The humidity present in atmosphere as a vector of 12 hours.
3. The temperature as a vector of 12 hours.
4. The amount of rainfall as a vector of 12 hours.
5. Location, given as a categorical data.  The location can be categorically assigned to a cluster of geographic entity, can be taken as towns to start with, later can be refined to much smaller locations to track even the tiniest weather calamity for tracing and saving as much as is possible.

Why 12 hours have been set? Is this being good amount of time to be set for computing natural calamities? Well, this paper is a proposal of use of AI to predict and raise an alarm respectively. These changes of how many hours to select can be computed through experiments or experts in the climatology field. For each hour the vector is given as input to the deep learning model shall be a 60-dimensional vector which represent the above parameters in each 60 minutes of the hour. Hence there is a sequence of 12 hours each of dimensions of 60 for each of the five measured parameters, hence a 300-dimension vector. Further, once this system is developed the input can be changed to minute-by-minute basis as well, where each minute can be represented on a minor time framework.

• Deep Learning with RNN/LSTM for Weather Forecast – Proposal

When data has been collected it is fed into a recurrent neural network or a Long Short Term Memory Model (LSTM) which works on many to one basis or many to many. The two are explained below:

1. Many to one LSTM. This kind of network takes the sequential time data and predicts the output as a signal entity. The following kinds of outputs can be trained in this RNN/LSTM.
   1. A vector which shall predict the next instant weather parameters in form of quadruple (low/high pressure, low/high humidity, temperature, rainfall in mm, alarm-yes/no). Here the alarm can be yes/no which means a deeper analysis need to be done to predict an exact climate calamity.

• The network can be trained it in form of classifiers, as well, this can work only as an alarm raiser. The advantage of this over the first is lesser energy consumption for predictions. And deeper analysis can be computed once the alarm is set on.
• Many to many networks RNN/LSTM. This takes a series and outputs a series. This means it takes as input a series data of 12 hours each measured in four parameters as mentioned before (atmospheric pressure, humidity, temperature, rainfall, location) and predicts the next 12 hours exact weather warnings in right places.

The model can be tested with a basic neural network too, but RNN has advantage to deal with sequential data of this form. And LSTM overcomes many disadvantageous of RNN, hence the model should be directly applied to an LSTM. Though beginners can start with neural network or RNN.

The model works on real time reinforcements learning. The predicted output is sent back to the network with the exact output and the weights of the network learned through stochastic gradient backpropagation algorithm. It shall be running a complete sequential series of the weather data in much greater details than just a quadruple and the weights of network shall keep updating with time.

When data has been collected it is fed into convolution neural network (CNN) followed by application of long short-term memory model (LSTM) which works on many to one basis or many to many, as explained above. The key advantage of a hybrid model is that dimensionality of the problem is reduced. But the drawback is these models once trained can be learned on newer inputs and are static, once weights are learned they can only be tested. As against a LSTM model described above which keeps on learning and changing with times as per change in weather conditions.

One disadvantage of using non-hybrid models using only LSTM is that the weights are constantly changing as whether system is changing. This can be overcome by using markers or timestamps, which can be used when a similar weather is reached or the yearly weather change occurs, this should help in faster learning and lesser error rates, when run on a live data.

Now this data just doesn’t take a 12-hour vector each having four parameters of pressure, humidity, rain and temperature, but also take location as a parameter. Why location has been added in the mode? Location has been added in the data to emphasize that fact that an extreme low pressure and high heat in a place can cause heavy rains in other parts of world. The location has to be hence included in the data. Once a basic model has been set up an extended version can be set for all key geographic locations.  And this becomes a time series problem over locations of around 6000-7000 clusters of earth. The training and learning process can take up time, but once a similar input is fed in the system, the learned model is taken out from database and predictions can be made.  While the local weather system is always running and updating live for continuous changed in the atmosphere in the locality. This can work well if stamping is applied for places where weather changed very abruptly.

The proposed model requires high computational powers of system. The intention of paper is to provide a proposal of use of AI in raising an alarm in case of natural calamities. Here we also present a Fuzzy logic-based sample AI system to predict weather calamities and provide a proposal for raising automatic alarm and assist humans to take further actions. This is described in Section 3.

• Fuzzy Logic based technique for raising alarm: A Proposed Technique

Here a Fuzzy Inference Engine based basic model to predict a calamity caused by weather changes is proposed. This system works per locality and has the following inputs.

1. Pressure in area with following linguistic variables.
   1. Very Low-VL: This means a very low pressure is in area.
   2. Low- L- For low pressure in area
   3. Normal -N, this means pressure is normal in area
   4. High- H, this means a high pressure is in area
   5. Very High- VH, this means a very high pressure is in area
   6. Extreme. For an extreme climatic condition of this parameter

• Weather in neighboring area with following linguistic variables.
  • Low- L- For low impact on weather in neighboring areas
  • Normal -N, this means it is normal impact on neighboring areas
  • High- H, this means a bad weather in neighboring areas
  • Very High- VH, this means a very a bad weather in neighboring areas
  • Extreme. For an extreme climatic condition of this parameter
• Humidity with following linguistic variables.
  • Very Low-VL: This means a very low pressure is in area.
  • Low- L- For low pressure in area
  • Normal -N, this means pressure is normal in area
  • High- H, this means a high pressure is in area
  • Very High- VH, this means a very high pressure is in area
  • Extreme. For an extreme climatic condition of this parameter
• Temperature with following linguistic variables.
  • Very Low-VL: This means a very low temperature is in area.
  • Low- L- For low temperature in area
  • Normal -N, this means temperature is normal in area
  • High- H, this means a high temperature is in area
  • Very High- VH, this means a very high temperature is in area
  • Extreme. For an extreme climatic condition of this parameter
• Rainfall with following linguistic variables.
  • Very Low-VL: This means a very low rainfall is in area, with 0 being no rainfall.
  • Low- L- For low rainfall in area
  • Normal -N, this means rainfall is normal in area
  • High- H, this means a high rainfall is in area
  • Very High- VH, this means a very high rainfall is in area
  • Extreme. For an extreme climatic condition of this parameter

The following are the output of the proposed Fuzzy Inference System, alarm. The linguistic variables for weather alarm are as follows.

1. Alarm
   1. Low- L- For low harmful predictions of weather, a low alarm, sound of alarm can depend on the defuzzied value
   2. Normal -N, this means normal alarm, and needs human intervention to detect the harm that can be caused by the given conditions
   3. High- H, this means a high risk is involved, a high alarm is made, and necessary precautionary measures need to be taken in this region.

This fuzzy system needs to be run per locality, the measurements can be taken based on sensors and digital thermometers. These measurements need to be converted to or scaled down to range of 1 to 10. The experts can change the inputs of Fuzzy systems, if they like to work with exact values of humidity, pressures and other measurements.  

The proposed Fuzzy Inference System can be modeled are given in Fig 1, Fig 2 and Fig 3 below. However, the exact modelling requires experts in the respective fields to sit and finalize the values. The number of rules that the experts have to lay out here are 9375. Here, we give in Fig 4, some rules for producing alarm as per requirements, which was as required.

   Fig 1. Fuzzy Linguistic inputs for pressure and neighboring areas weather scaled in range 0 to 10

Fig 2. Fuzzy Linguistic inputs for humidity and temperature scaled in range 0 to 10

      Fig 3. Fuzzy Linguistic inputs for rainfall and alarm scaled in range 0 to 10 and 0 to 1 respectively

     Fig 4. Sample Fuzzy Rules for the proposed Alarm System using Fuzzy logic

This model works on the process of defuzzification of the inputs, once real valued data is obtained one can run the model so proposed. It is not necessary to scale the values, this has been providing for illustration of the proposed system.  The fuzzy rules can be learned once data is presented to the system in the format required by the Fuzzy Inference Engine.

Both these systems can work in tandem and have benefits, the second one can be switched on all time, while the first can be put to use once an alarm of weather forecast is raised by the AI-Software system. This once done can be send to humans or Fuzzy inference engine as described in Section 3, before human intervention.

• Conclusions and Future works

The models have been presented as a proposal. The deep learning model should perform better than Fuzzy Logic based model due to ability to learn non-linear data models in high dimensions. While Fuzzy Inferencing can help local climatologist get automated alarms for weather abnormalities, and they can take further steps henceforth. Further, deep learning-based models can learn the characteristic relationships that exits in weather and which cannot be explicitly constructed in a rule based Fuzzy Logic system. This article is for illustrating AI based climate calamities’ predictions. In future work, precise data is to be fed in the proposed models with exact measurements and optimizing on parameters for forecasting any calamities. Further, an AI based application can be made to make monthly predictions once this system gets up, this can help people plan travel and help them in saving their resources as well. Once such predictions start matching with the output of the environment, safety measures, treatment of local areas with other natural resources such as water can be put in measure. Water content in air plays a vital role in weather of an area, hence this is the future work that needs to worked once weather predictions using AI are accomplished.