Neural networks & Weather prediction Peter Polák Faculty of Electrical Engineering and Informatics Department of Cybernetics and Artificial Intelligence peter.polak.6@student.tuke.sk Abstract - The need of weather prediction with sufficient accuracy can be seen when considering the benefits that it would bring to whether the control flow of rivers in agricultural interests or only as an little help in selecting the holiday period. While the necessary data for the needs of such a prediction have been available for a long time, the complex relations between data and their effect on the prediction is proving to be difficult to use classical analysis, and here by the integration of neural network we can achieve learning these relationships instead of their complex analysis, thus using less effort we can obtain sufficiently accurate prediction. Key words - artificial intelligence, neural networks, weather prediction, weather forecast I. INTRODUCTION Weather forecast is the most wanted product of meteorological services, where user requirements are, that prediction accuracy should be the highest and the predicted interval longest possible. The weather forecast is actually about prediction of atmospheric events in a given area taking into consideration that the atmosphere is a nonlinear dynamic system whose state at some future point in time is not adequately predictable in the deterministic sense. Nowadays, there are these 3 methods mainly used for weather prediction: Numerical Statistical Synoptic Current predictions are the result of using a combination of these methods, where each predictive method captures atmospheric phenomena only in a particular approximation, thus predictions have rather probabilistic than deterministic nature. Numerical methods are based on numerical integration of differential equations, which in certain model approaches describe dynamics and thermodynamics of the atmosphere. Statistical forecasting of meteorological elements or meteorological fields is based on the statistical characteristics of files meteorological elements. In the synoptic forecasts meteorologist based on his subjective experience and by empirical rules extrapolates future development of atmospheric processes, which can be much more efficient realizable by an appropriate neural network. II. HOW TO PREDICTION WITH NN Neural network is a massively parallel processor, which tends to preservation of experimental knowledge and their further use. Neural networks are very convenient in prediction because they assume the role of so-called universal function approximator, which in this case actually means that thanks to main property of the neural networks, which I mentioned above, we can learn neural network to simulate the course of the weather, and thus to predict its future short-term development with insufficient accuracy. Regarding the model of neural network topology for weather prediction, the most common form is the use of multilayer feed forward neural networks, also referred as MLP. This topology is composed of several layers of neurons, where each layer is fully connected with the next. To be able to learn neural network something, it is first necessary to obtain enough samples of training data, in this case it means the meteorological data from the certain area of the atmosphere. It is also necessary to decide which meteorological variables select for appropriate prediction and these selected variables are then ours neural network inputs. Problem occurs when variables are supplied to the input of the network, where could values like "overcast" be a problem, therefore normalization of inputs is required, where the values of the variables are modified appropriately to the actual inputs values of the neural network, where eg. value "overcasts" will be assigned the value 0.3 in the interval (-1,1), where -1 represents the state of "storm" and 1 is "sun". Of course for values of the inputs, we need the values of outputs and here we come to another problem, which is to determine the length of the forecast period, as I mentioned above, it is one of the parameters that we want to be the longest, but the larger he gets, so more we decrease even more important parameter accuracy. After selecting the forecast period and obtaining input and output data, we can move on to learning phase of the neural network. In the learning phase we bring to the inputs of the neural network normalized variables of meteorological data, and calculate the net output according to weights values. If the output corresponds to the desired value, then we continue with the next element of the training set, however, if the output does not respond, there is a readjusting of weight values, which consists in calculating the weights change by back propagation of errors method. The whole learning process is repeated a number of cycles or until the error function reaches the desired value. After completion of learning, the testing phase takes place. Test data have the same format as the training and the difference is that weight of neural network do not change anymore, it is only a determination of error, and thus the accuracy of the network. III. REALIZATION A. Prediction of temperature and state of clouds The most common objective of weather prediction either using neural network or not remains forecast temperature and clouds condition. Nowadays almost every Meteorological Institute takes into account of these parameters also the output of a dedicated neural network for that purpose. B. Rainfall prediction Hydro meteorologist Tony Hall from the National Weather Service in Texas using the software BrainMaker, which is used to implement the application using the neural network, created model consisting of four interconnected neural network for predicting rainfall in the area. Two neural networks are used to determine the amount of precipitation, one for summer and one for winter season, where nineteen input variables are used for both. The other two neural networks determine the probability of these deductions. The overall accuracy of the system reached 85%. C. Solar activity prediction Dr. Henrik Lundstedt from Lund Observatory in Sweden created a system to predict solar harmful activities. This solar activity resulting in a loss of signal, power outages, pipeline corrosion, failure of geological measurements equipment and other adverse effects, which could be prevented by their prediction. Designed neural network has 37 inputs, describing the behavior of the solar wind, the result of which are mentioned malicious activities. IV. CONCLUSION By integration of a neural network to the weather prediction has been simplified approach to achieving the predictions with maintaining a sufficiently high accuracy, which in my opinion will lead to even more frequent deployment of neural network in such systems. REFERENCES [1] ADAMCAK, Jan – Weather prediction using neural networks http://neuron.tuke.sk/adamcak/theses/Adamcak-Jaksa-BC-thesis-2010.pdf [2] SINCAK, Peter – Neural networks: Master approach 1 http://www.student.ui.sk/ZNS?action=AttachFile&do=view&target=NS1.pdf [3] Rainfall prediction http://www.calsci.com/Weather.html [4] Solar activity prediction http://www.calsci.com/SolarFlares.html