Comparative Study on Soft Computing Approach in Weather Forecasting: Analytical Essay

NB: We do not resell your papers. Upon ordering, we do an original paper exclusively for you.

NB: All your data is kept safe from the public.

Abstract– In a developing country, like India where the agriculture & industries are base for the national economy, the weather conditions play leading role for their proper development and smooth running. Therefore having accurate weather forecasting information may allow farmers or industry managers to make better decisions on managing their farms. Soft computing using ANN is an innovative approach to construct a computationally intelligent system that is able to process nonlinear weather conditions within a specific domain, and make prediction. A number of researches have been done or being done using Soft Computing Approach for forecasting. In this paper the presentation is all about to present the comparative study of several researches and some key findings that are initials for better start any soft computing model for prediction.

Keywords– Soft Computing, Artificial Neural Network (ANN), Back Propagation Algorithms, Multilayer Feed Forward Neural network (MLFFNN), Mean Square Error (MSE).

I.Introduction

Soft Computing is an efficient approach for forecasting, whether it is weather forecasting or any other else. A number of researches have been done focusing on the usefulness of soft computing approach in forecasting area. Here the presentation is utilizing some of researches to make some conclusions that are initials to be taken in account when planning forecasting using ANN technique. In traditional Weather forecasting approaches like:

(a) The empirical approach and
(b) The dynamical approach.

The first approach is based upon the occurrence of analogues and is often referred to by meteorologists as analogue forecasting. This approach is useful for predicting local-scale weather if recorded cases are plentiful. The second approach is based upon the equations and forward simulations of the atmosphere, and is often referred to as computer modeling. Because of the grid Coarseness, the dynamical approach is only useful for modeling large-scale weather phenomena and may not predict short-term weather efficiently. But for local scale & short term weather forecasting the approach of artificial neural networks (ANNs) is so efficient and a little bit easy.

ANNs provide a methodology for solving many types of non-linear problems that are difficult to solve by traditional techniques. Most meteorological processes often exhibit temporal and spatial variability, and are further plagued by issues of non-linearity of physical processes, conflicting spatial and temporal scale and uncertainty in parameter estimates. With ANNs, there exists the capability to extract the relationship between the inputs and outputs of a process, without the physics being explicitly provided. Thus, these properties of ANNs are well suited to the problem of weather forecasting under consideration. The popular soft computing techniques is ANN which performs nonlinear mapping between inputs and outputs, has lately provided alternative approaches to weather forecasting and so many researchers have taken in their research and come into the conclusion that ANN is best suited for forecasting. In the following paper the main objective is to find out some basic fundamentals and initials to make conventions about the ANN & forecasting.

Figure 1: Schematic diagram of data acquisition system

II. Implementation

To summarize the working of our system we can categorize the major components in following manner:

Data recording scheme
Parameter selection and user requirement definition
Neuro-fuzzy training and prediction scheme
Generating weather forecasting system

1. Data recording scheme:

The experimental setup consists of highly sensitive atmospheric pressure sensor and atmospheric temperature sensor . We have linked the sensors to the data logging computer using an interface cable to the parallel port LPT1. Data acquisition software records data in real time at a fixed time intervals of 4 seconds. Data precision for pressure sensor is 1 mbar and for temperature sensor is 0.1℃. System has capability to record the data continuously for several days without break in data recording.

2. Parameter selection and user’s requirement definition

We considered atmospheric pressure as a primary parameter and atmospheric temperature and relative humidity secondary type. Other parameters are also considered like wind direction and wind velocity.

Atmospheric pressure changes at any given place on earth are minute, hence, a very high sensitivity of pressure sensor is essential. Atmospheric pressure also changes with altitude and hence base atmospheric pressure differs from place to place depending on the altitude of the place from mean sea level. User requirement involves the collection of different input values and variables; including the selection of one or more forecasting measurements, the forecasting range, parameters such as the local forecast and distance of altitude from the sea level, and no. of samples per recording and interval of recording.

3. Neuro-fuzzy training and prediction scheme

Having collected and preprocessed all of the relevant weather information, our system starts the appropriate network training and forecasting, which is based on the back propagation in a neuro-fuzzy network. Table I shows the different categories defined for the fuzzification of the weather conditions.

The fuzzy data for predicting the occurrence of either rain or not rain ,bad weather ,good weather are based on the above given table I. While local pressure is measured in absolute scale , for simplicity of understanding of weather conditions pressure at sea level is computer with altitude information as correction factor to predict weather .

The following conditions are to be analyzed —

Constant low pressure (Bad weather)
Constant high pressure (Good weather)
Negative slope(Good to bad weather)
Positive slope (Bad to good weather)
Dual Slope (Variable condition of weather)

4. Generating Weather prediction

From recorded pressure data the generated weather forecast under stable low-pressure condition was rainy weather. As only two hours record was used for testing, we did not expect great results from it. Pressure changes recorded at fixed pressure rate of -0.564 mbar/h. It is very small change and is considered to be constant.

III. Results and Findings

After going through the above detailed study we find the following key results that play a major role in any forecasting model building. Apart from the traditional forecasting systems, ANN based forecasting is much feasible & best suited. Applying soft computing could be one of the best alternatives for local and short scale weather forecasting. The Study says any forecasting system using Artificial Neural Network & Back propagation Algorithms depends on following:

The data: That we are going to acquire should be valid, Authentic and in proper format.
The Variables: Means how many different kinds of data variables we use for input training set. As Temperature, Pressure, Relative Humidity, Due Point, wind speed, cloud status etc. The training results are dependent on the inter-relationship of these variables so it should be chosen so carefully and according to need. Many the variables, better the result.
Data Analysis: These variables are interrelated & inter-dependent. So the interrelationships of these variables are a big factor in training set preparation & training of ANN. So the normalization should be done certainly & so carefully before making the training set.
Dataset: The data that we acquire for training of our model plays a vital role in forecasting accuracy. This describes how much data we acquire for the training of proposed model.
Training set: The training set is one the most considerable entity of our research work. Even it could be said the backbone of the ANN based forecasting system. It contains the input matrix & target matrix which contains the collection of unit input & unit output for the ANN correspondingly. The better the training set, better the result.
Architecture of ANN: Once the training set is prepared the next most important thing to discuss is Architecture. The architecture of any artificial neural network is defined by the layers, numbers of neurons etc. Different forecasting models requires different forecasting architecture. The best suited ANN architecture for any forecasting model is the subject of research for a researcher. following are the key points to be taken in account when developing a forecasting model:
Types of Network: The training and forecasting of any model is dependent upon the types of the network i.e. Multilayer Perception (MLP), Multilayer Feed Forward network (MLFFN) etc. in some cases MLP may suit best or in some cases it may be MLFFN. The appropriate type of network may converse fast for prediction.
No. Of Hidden Layers: Our forecasting result‟s accuracy is highly dependent on the numbers of hidden layers. Some problem may converse in single layer ANN or some may converse in multiple layers ANN. Although single layer network is appropriate for solving any problem but it may be so less accurate. For better result multi layer network may be used that may converse slow but produce much better result. One drawback is that it goes complex. Multi layer Feed Forward neural Network (MLFFNN) is found the best for forecasting the weather condition.
Algorithms: There are number of training algorithms are available but appropriate selection of training algorithms may leads fast and accurate forecasting results. Back propagation algorithm is found best suited with MLFFNN for forecasting the weather prediction.
Activation function: A number of activation functions are available for training the network but selection the appropriate activation function may leads to better conversion. We can select any one of following TANSIG, LOGSIG & PURELIN.
Weights/Bias: Apart from all above, another important factor is the initialization of weights and bias. Proper initialization of weights and bias may leads the network to converse fast and in proper direction.
Learning Rate: Learning rate is another factor that leads the training of network with a given constant factor. High the learning rate fast the conversion & low the learning rate slow the conversion. A small learning rate may leads smooth conversion better results but slow and a high learning leads fast but less accurate result. Initially we may take any randomly value for learning rate.
Threshold/Momentum: If we want an output by any particular condition then we can set threshold value .when the threshold value is achieved the output is generated else not. Another is the momentum that could be set for smooth conversion of network with the provided momentum factor.

Figure 2: Architecture of an ANN based model

In the Above shown figure 1 Inputs can be Temperature, Rain, Wind, Humidity, etc.

From conducted experiments we find the following key changes in the atmospheric pressure signature that can be related to dynamic states of atmospheric conditions and for meaningful short duration weather prediction. We are noticed the following key features in atmospheric pressure patterns that were related to the weather conditions and indicated trend of the future weather of the place. These conditions were:

Stable day-night pressure gradient – indicator of stable weather -sun shine.
Sudden pressure fall – indication of likely thunderstorm.
Sudden pressure rise – indicator for windy day.
Change in pressure slope – change of weather state in either way.

These trends were actually found in the recorded data and the selection of August month for the experiments was benefiting, as there were almost all types of signatures present in the atmospheric pressure indicating weather changes.

IV. Model Performance

Experiments were carried out on different locations of the different city. The performance of the experiments from stations having short difference from the sea level in the form of altitude is the best. These observations were carried out on the time series basis. The results for a single station or multiple stations do not produce a large difference in performance. The best result that is achieved by our model is due to the availability of a large amount of input data for the model to select the right variables. Thus the model has a greater chance of producing better prediction results .It is found and can be deduced that the correlation between the data at time t and t+1 is high; therefore, it is easier to build a successful model.

V. Conclusion

After going through all the above study & discussion we see that applying soft computing model for forecasting the weather conditions is most feasible rather than any other short term & local based weather forecasting approach. Through the implementation of this system, we illustrate how an intelligent system can be efficiently integrated with a neuro-Fuzzy prediction model to implement an online weather information retrieval, analysis, and prediction system by using electronic sensors. Increasing parameters for weather modeling may help in predicting weather changes to greater extend in comparison with than simple model used in present case.

References

https://www.sciencedaily.com/terms/weather_forecasting.htm
https://study.com/academy/lesson/methods-principles-of-weather-forecasting.html
Arvind Sharma and Prof. Manish Manoria, “A Weather Forecasting System using concept of Soft Computing: A new approach”, 2006; https://ieeexplore.ieee.org/document/4289915
Govind Kumar Rahul, Madhu Khurana, “A Comparative Study Review of Soft Computing Approach in Weather Forecasting”, 2012; http://www.ijsce.org/wp-content/uploads/papers/v2i5/E1053102512.pdf

Click Here To Order Now!