Abstract
Qualification Masters work
"Construction of Predective Complex and Looking at the Introduction of its in University Network"
Background
Adequate (true) weather forecast - is one of the important contemporary issues have a practical application. Very often, the weather conditions are for people miserable. In developed countries, weather and climate have long been categories economic. Each year the world's natural disasters kill nearly 250,000 human lives, the amount of damage to property, lies within 50-100 billion U.S. dollars. But the world statistics show if the trust hydrometeorological information and adequately react to it, it can prevent 30 to 40% of losses and to completely avoid human casualties.
Especially notable economic effect is the use of meteorological information in aviation, energy, construction, fisheries and shipping, agriculture.
Weather prediction from a scientific point of view - one of the greatest challenges Atmospheric Physics. Over the past decade, the development of ideas and methods prediction had moved forward, and this contributed to both development of mathematical approaches and improved methods of investigation atmosphere, and the use of modern computer technology. There Various methods for predicting weather phenomena and their values for example, synoptic, numerical, statistical methods, but in full none of them yet does not provide an accurate prediction [1]. This makes the topic and research in the field of weather forecasting is not only useful, but and relevant.
Aims
Purpose thesis - to develop a software system as a software product that combines a number of modern methods of forecasting meteorological data. Perform a sync with the weather station complex, integrated in an electronic network of university and adapt it to the laboratory studies (Work).
Object of the study are time series of basic meteoparametorov.
Subject of research - the methods of forecasting, quality and range of their results.
Hypothesis research - the process of the prediction will be successful and efficiently under the following conditions:
1) Direct interaction with the complex prognostic meteorological station;
2) The use of modern methods of working with time series, such as the method Eglaysa [4], the method of principal components, the method of informal (soft) logic;
3) Introduction to the practice of looking complex, its integration into electronic network of the University;
4) Adaptation of the complex under laboratory investigation.
Thus, the task of master's work are:
1. Sync software product developed by the meteorological station. Ensure the transfer of data from meteorological stations in the program;
2. Explore new methods of processing time series and put them in the existing product.
3. Carry tested prognostic complex;
4. Ensure data protection software;
5. Implement a complex electronic network of the University;
6. Adapting prognostic system for laboratory studies. Expand tools and provide high interactivity in its dialogue with user.
The practical significance is to develop a predictive complex that will enable an adequate short-term forecasts of the basic meteorological range of 12 hours. Kompeleks will be directed to use his laboratory and practical purposes.
Overview substantive research
Building predictive models based on processing time series obtained from the meteorological station Vantage Pro 2, established at the Faculty CST DonNTU. This weather station can take the following data:
- temperature;
- humidity;
- pressure;
- wind speed.
All data is stored on the server and the Department of KSM "AKIAM. Interval measurements 10 minutes. Thus, during its operation continuously formed and gradually accumulate a set of time series. The presence of this information makes the task of developing a realistic looking complex.
In recent decades, the nonlinear dynamics obtained a number of fundamental theoretical results and methods have been developed to justify the principle possibility to predict physical processes on the basis of their time series. The theoretical foundation for these developments and methods is the Takens theorem. One of his key ideas is that in constructing empirical models for time series as the missing variables can be use or successive values accessible observable, or its successive derivatives. It was proved that the reconstruction of the scalar time realization of a dynamical system and method of time delays, and the method of successive derivatives guarantee that the new variables will be obtained equivalent description of the original dynamic system with a sufficiently large reduced dimension vectors D . Namely, we must have condition D > 2d , where d - the dimension of M in phase space of the original system, which is modeled movement [2]. These statements constitute the contents of the famous theorems of Takens.
Analysis of recent research
Currently at the Department of KSM has developed a software product, the basis of which was to develop methods of short-term weather forecast for Time Series meteoparametorov [2]. It allows on the basis of the series, removed from weather station to put short-term forecasts of temperature, humidity, pressure, and wind speed. The algorithm of the complex is divided into several stages, which can submit a scheme (Fig. 1)
Figure 1 - Diagram of model for time series
Consider more detail each step:
- stage № 1. Series taken from weather stations are processed and systematized. Then conducted their analysis using different methods to identify pronounced patterns that could simplify the choice of model equations. This, for example, visual analysis in the form of graphs depending variable from time to time, the restoration phase trajectory, spectral and statistical analysis and others [3 ].;
- stage № 2. Forms the structure of the model. Originally selected type equations, then set the form included in their functions, then establish a connection between the dynamical variables (components of the vector x) with observable quantities a. The variables can serve themselves observed, but in a more general case, this relationship are given by a = h (x), where h called the measurement function. Often, introducing more random additive e: a = h (x) + E, to take into account measurement noise. To make the model more realistic, random additive injected frequently and in the equations themselves - the so-called dynamic noise.
Structure formation models - the most complex and creative stage modeling procedures. At this stage, select the type of equations, the type of incoming in their functions and their arguments.
Problem of determining the function's arguments is to determine the smallest dimension of the model provides an unambiguous prediction. For solve this problem using different assessment methods: the method of false neighbors The principal components method, the method of Grassberger - Procaccia, the method of well adjusted basis. The first of these is as follows.
It is based on verification of the properties that the phase trajectory, recovered in the space of sufficient dimension should not be self-intersections. When the test dimension D for each of the restored vectors xk seek out one (the closest) neighbor; increasing D 1, define some of the neighbors were wrong (highly dispersed), and which - true. Counts the ratio of the number of false neighbors to the total number of restored vectors. If an increase in D is the number becomes small for some value of D *, then the latter is the estimate for the dimension of space, which is achieved embedding of the trajectory of the simulated motion [1 ].;
- step number 3. The next stage is determining the structure of the model equations. For this purpose, various methods of approximation of functions of many variables: the method of generalized polynomial, the use of radial basis functions, artificial neural networks, local model search for the closest neighbors.
Most often to solve problems of approximation using two-layer ANN rarely - three layers. The increase in the number of layers does not lead to significant improvement. Improvements sought by increasing the number of neurons in layers K1, K2. The theoretical basis for the use of ANN - generalized approximation theorem (a special case is the theorem of Weierstrass), which states that any continuous function can be arbitrarily exactly evenly approximated by the ANN.
Procedure for calculating the parameters of the ANN by minimizing - her "training" - this challenge of a multidimensional nonlinear optimization, solution of which developed special "technology": an algorithm for error back-propagation, learning schedule, training with noise, stochastic optimization (genetic methods method of simulated annealing). ANN can contain a lot of unnecessary elements, and structure of this model (network architecture), it is desirable to make it more compact. For this neuron, the weights and thresholds which vary only slightly in the learning process excluded from the network.
If there are several alternative ANNs with different architectures, derived from training Exercise row, then the best of them is usually chosen by the lowest test error of approximation. For "Honest" indicator of the effectiveness of predictive models use another number (no training and not a test, because both are used to construct model), which is called "examination».
After selecting a structure to perform the "fit model". For this purpose, as generally carried out search of extreme values of some objective function For example, minimizing the sum of squared deviations of solution of model equations from the observed data. If necessary, at this stage the preliminary transformation of the observed number of: filtering of noise numerical differentiation or integration, etc. This is mainly Technical stage of numerical calculations, but here you need to make a choice principle calculation parameters and methods for its implementation.
The large number of existing configurations of neural networks, according to their orientation to the classes of tasks and the results of preliminary experiments in the work for the reconstruction of model equations were selected three network types: single-layer linear, two-layer nonlinear and generalized regression. To calculate the values are different prognostic prediction scheme based on the basic iteration, and ensemble variants [1].
- stage № 4. In the final phase checking the quality of the model. A checking the effectiveness of the model to achieve the required accuracy of prediction [1]. If the model is found to be satisfactory (good), the resulting model taken in the case, otherwise - back to revision at any stage presented in Fig. 1.
Testing
Results of the work reported at the I All-Ukrainian scientific-technical conference of students, graduate students and young scientists "Information management systems and computer monitors on 19-21 May 2010 and published in the appropriate Miscellany.
Current and planned results
As software development tools used among Matlab, which is a high-level technical computing language, interactive development environment of algorithms and advanced analysis tool data. Effectiveness is primarily due to its focus on matrix calculations with the program emulation of parallel computing and simplified means of assignment cycles. Successfully implemented a means to work with multidimensional arrays, large and sparse matrices and many types of data.
To date, implemented the first two tasks, namely:
- set up a database on the server department KSM, which is preservation of data from the meteorological station Vantage Pro 2, established at the Faculty CST DonNTU. Made setup and preparation of a server for its future synchronization software. Synchronized program server using the tools of Matlab;
- made an attempt to determine the principal components of time series meteorological data.
The queue implementation of the following tasks:
- introduction to the complex processing techniques of time series - the method of principal component, the method Eglaysa [4], methods of informal logic;
- extension tools software. Formation of its more interactive;
- approbation;
- the formation of information security software;
- implementing it in an electronic network of the University;
- adapted for laboratory research.
General scheme of the software system is shown in Figure 2.
Figure 2 - Scheme of work looking
complex
(animation: size - 67,3 КB; resolution - 600x300; shots quantity - 11; infinite number of repetition cycles; delay between shots - 1,3 мs; delay between last and first shot - 1,3 мs)
Conclusion
As the main task of building predictive of the complex was adopted by the task of improving and synchronization of the complex looking at the meteorological station, as well as integration it into an electronic network of the University As a result of an analysis of literature on methods of forecasting and optimization (compression) time series. Collection of meteorological parameters has been launched with local meteorological station, installed in Donetsk National Technical University, the possibility of forecasting weather parameters, using methods reconstruction of model equations, based on the analysis of time series, formulated the concept of forward-looking complex in an electronic network of the University.
When writing this Autosummary master work is not yet completed. Total willingness to work - in December 2010. Full text of the work and materials on the subject can be obtained from the author or his head after specified date.
References
1. Гриценко А.В. Реконструкция уравнений и прогнозирование метеопараметров по их временным рядам. – Донецк, ДонНТУ, 2010. – 149 с.
2. Takens F. Detecting strange attractors in turbulence // Lec. NotesMath., 1981. V. 898. 366-381.
3. Безручко Б.П., Смирнов Д.А. Математическое моделирование и хаотические временные ряды. Саратов: ГосУНЦ «Колледж», 2005. – 320 с.
4. Эглайс В.О. Аппроксимация табличных данных многомерным уравнением регрессии. – Вопросы динамики и прочности: Рига, 1981, Вып. 39. – с. 120-125.
5. Заенцев И.В. Нейронные сети: основные модели. Воронеж, 1999. 76 с.
6. Гриценко А.В. Прогнозирование метеопараметров по временным рядам. – Компьютерный мониторинг и информационные технологии 2008 / Материалы IV международной научной конференции студентов, аспирантов и молодых ученых. – Донецк, ДонНТУ – 2009.
7. Гриценко А.В. Использование локальных метеостанций для прогноза погоды. Метеокомплекс VANTAGE PRO 2. – Охрана окружающей среды и рациональное использование природных ресурсов 2009 / Материалы VIII международной научной конференции аспирантов и студентов. – Донецк, ДонНТУ – 2009.
8. Горбань А.Н. Функции многих переменных и нейронные сети // Соросовский образовательный журнал. 1998. № 12. С. 105-112.
9. Christopher R. Williams. Applying neural network models to prediction and data analysis in meteorology and oceanography // Journal of Atmospheric and Oceanic Technology, Volume 79, Issue 9 (September 1998) pp. 1855-1870
10. Will Dwinnell. Putting PCA to work. [Electronic resource] - Mode of access to the article: http://matlabdatamining.blogspot.com/20 10/02/putting-pca-to-work.html
