Sie sind auf Seite 1von 6

Wavelet Analysis of Meteorological Data

Rudolf Rabenstein, Thorsten Bartosch Lehrstuhl fur Nachrichtentechnik, Cauerstrasse 7, D-91058 Erlangen, Germany Tel. *9131-857101, Fax *9131-303840, Email rabe@nt.e-technik.uni-erlangen.de

The Wavelet transformation is a modern signal analysis tool, which displays the information content of an observed signal simultaneously over time and frequency. Its application to meteorological data reveals the spectral components on di erent time scales (hours { days { years). Wavelet analysis provides an intuitive access to statistical properties of data series, that is not available with conventional stochastic analysis methods.

Introduction
The investigation of the properties of meteorological data is important for the analysis and classi cation of measured data, generation of synthetic data for simulation, and statistical description of solar systems. Since the temporal variations of meteorological data are nondeterministic, they are usually described by the conventional tools for stochastic functions like probability density function, auto- and cross-correlation functions, power density spectrum and correlation coe cients. An important feature of meteorological data that is not to be overlooked is that they contain information on di erent time scales. The most obvious examples are the seasonal and daily cycles with xed and well known periods. Depending on the location, also variations on the scale of several days to several weeks occur. They do not possess sharply de ned periods and may be present only at certain times of the year. As an example, temperature variations on this scale are important for solar thermal systems with large storage masses. On the other hand, uctuations in the cloud cover occur on a scale of minutes and hours and may exist today but not tomorrow. They play a role in the description of photovoltaic systems. The conventional tools of stochastic analysis named above are not very well suited for the description of such non-stationary, multi-scale functions. Either they contain only amplitude information (probability density function), or they show exclusively the time dependence (autoand cross-correlation functions) or the frequency dependence (power density spectrum) of the analysed data. In order to show which scales are present at which times, more sophisticated data analysis tools such as time-frequency representations are required. After a short introduction, two speci c approaches will be presented: short-time Fourier transformation and continuous wavelet transformation. Space does not permit an in-depth coverage of spectral analysis and wavelet theory (see 3, 6]). A comprehensible description of the basic ideas is given instead.

Time-Frequency Representations
There are two basic ways to represent the information content of a physical quantity: as a function of time or as a Fourier spectrum. However a function of time does not show the spectral content of the signal (see the plots at the bottom of gs. 1,3,4). Conversely does a Fourier spectrum not distinguish between characteristic time segments (e.g. summer { winter). Time-frequency representations provide an escape to this situation. They are a

convenient way to display the varying spectral content of non-stationary signals over time. They attempt to visualize the answer to the question: What is the energy of a signal at a given frequency and at a certain point in time? However, this answer can only be given with some uncertainty. One may obtain good time resolution at the cost of poor frequency resolution and vice versa. But even with this unavoidable limitations, time-frequency representations give a much clearer picture of the properties of a non-stationary signal than the time function, an auto-correlation function, or a power density spectrum alone.

Short-Time Fourier Transformation


The short-time Fourier transformation may be interpreted as a Fourier spectrum of data as seen through a sliding window of nite extension. The position of the window indicates the approximate time for which the spectrum is valid. For a rectangular window function of length T 8 < 1 T=2 < t < T=2 rT (t) = : ; (1) 0 else the short-time Fourier transformation of a function v (t) is given by
V (f; ) =

Z1
1

v (t)rT (t

)e

j 2 ftdt =

+T =2 Z v (t)e j 2 T =2
)ej 2 ft :

ftdt =

Z1
1

v (t)g(t; f; )dt

(2)

with the transformation kernel (g denotes complex conjugate of g )


g (t; f; ) = rT (t
(3)

It shows the Fourier spectrum depending on the frequency f for a segment of data seen through a window at position in time. Varying values of and f set up the desired time-frequency representation. High values for the window length T yield good frequency resolution but give poor temporal resolution and vice versa for short window length. The window length is xed and does not vary with scale. The rectangular window de ned by (1) is the simplest form of a data window. Various other forms of window functions (apertures) are frequently used. As an example, consider the short-time Fourier transformation of hourly normal solar rediation data with one year duration in g. 1 (Typical Measurement Year for Madison, USA, provided by the Solar Engineering Laboratory of the University of Wisconsin). Since the shorttime Fourier transformation V (f; ) is in general a complex quantity, its squared magnitude jV (f; )j2 is displayed (spectrogram). The original data set is shown on the bottom, while the image on the top shows the spectrogram. Dark colors repesent high signal energy at a given time (abscissa) and frequency f (ordinate). Of course the day-night cycle at a frequency of 1=(24h) = 0:0421=h is present throughout the whole year. Higher harmonics are less pronounced and are present only at certain times. The window length of T = 512 h was chosen for a reasonable compromise between temporal and spectral resolution on the daily scale. However, signal components with periods larger than the window length (e.g. seasonal variations) cannot be resolved well. Furthermore, they would hardly be distinguishable from the zero axis in the linear scale of g. 1. Adjusting the window length for proper resolution of seasonal variations would spoil the time resolution for the daily scale. This shows, that the short-time Fourier transformation produces a time-frequency representation which is not capable of resolving the di erent scales present in meteorological data. The reason for this limitation is the xed window length for all scales.

shorttime Fourier transformation, windowlength = 512h 0.25

0.2

frequency in 1/h

0.15

0.1

0.05

1000

2000

3000

4000

5000

6000

7000

8000

normal solar radiation in kJ/m2 3000 2000 1000 0

1000

2000

3000

4000 5000 time in h

6000

7000

8000

Figure 1: Short-time Fourier transformation spectrogram of solar radiation for hourly data of one year duration To illustrate this point further, the transformation kernel g (t; f; ) from (3) is shown in g. 2 (top row) for three values of f which correspond to a daily, weekly and monthly cycle (real part only). It is obvious that the xed window length (T = 512 h) is not equally well suited for all three scales. It should be reduced on the daily scale for better time resolution and enlarged on the monthly scale to improve frequency resolution.

Wavelet Analysis
Wavelet analysis is a di erent form of time-frequency representation which does not su er from the disadvantage of a xed window length. In simple terms, the continuous wavelet transformation di ers from the short-time Fourier transformation only in the choice of the transformation kernel. The starting point is a function (t) well localized in both time and frequency, the so called `wavelet'. The transformation kernel follows from the wavelet by time shift ( ) and scaling (a = b=(2 f ))
g (t; f; ) =
1 pa

2 (t) = ejbt e t =2 :

(4)

shorttime Fourier transformation 1 1 1

1 1200 0.2 0 0.2 1200 1600 time in h 1600

1 1 0 1000 2000 0 wavelet transformation 0.2 0 0.2 0 1000 2000 time in h 0.2 0 0.2 0

1000 2000

Figure 2: Transformation kernels for shorttime Fourier transformation (top) and continuous wavelet transformation (bottom)

1000 2000 time in h

wavelet transformation (morletwavelet, b = 11 )

10

frequency in 1/h

10

10

10

0 3000 2000 1000 0 0

1000

2000

3000

4000

5000

6000

7000

8000

normal solar radiation in kJ/m2

1000

2000

3000

4000 5000 time in h

6000

7000

8000

Figure 3: Wavelet-spectrogram of solar radiation for hourly data of one year duration The choice of the wavelet (t) is arbitrary within certain admissibility restrictions 3, 6]. A widely used function is the so called Morlet-wavelet as given in (4). Its Gaussian envelope ensures good time and frequency localization in the sense of a small time-bandwidth product.

The bottom row of g. 2 shows the real part of the transformation kernel for the wavelet transformation (Morlet-wavelet, b = 5). The self-similarity in (4) guarantees equal resolution properties on all scales. Fig. 3 shows the result of a wavelet analysis of the same data sets as in g. 1 (squared magnitude). It is now able to resolve all scales from seasonal 1=(8760h) = 1:2 10 4 =1h to daily 1=24h) = 4:2 10 2 =1h and even shorter variations.

Coherence
So far, wavelet analysis has been applied to a single data set only. But it is also possible to calculate cross wavelet spectra from two data sets, thus extending the idea of a cross correlation to the time-frequency representation. The normalised version of the cross wavelet spectrum is called coherence 5]. An example for coherence between normal solar radiation and ambient temperature is shown in g. 4. On the bottom are the two data sets, on the top is the coherence pattern computed with the wavelet transformation. Dark colors represent strong correlation. It is clear that solar radiation and air temperature share the same daily and yearly cycles. This high correlation shows up in the coherency pattern as dark horizontal lines at the frequencies 1=(24h) = 4:2 10 2 1=h and 1=(8760h) = 1:2 10 4 1=h. But there are also correlations on the half day scale (ca. 9 10 2 1=h), however not during the summer (hours 3000 { 5500). Furthermore, we see correlations on the monthly scale (1=(7 24h) = 6:0 10 3 1=h), but only for intervals of some hundred hours. (It should be noted that a Test Measurement Year is not an ideal test object for this analysis, since it does not contain the exact long term correlations.)

Applications
The examples considered so far showed the use of wavelet analysis for measured data. However, for the simulation of solar energy systems, synthetic data generators are preferred. There exist various models for generation of synthetic irradiance or ambient temperature data separately 2], but their joint generation with proper cross correlations is still a topic under research 4]. The rst problem is to identify these correlations over all relevant scales in order to incorporate them into a suitable model. The second problem is to test an implemented synthetic data generator against measured data to evaluate its performance. The coherence example presented above shows, that wavelet analysis is a suitable tool for this purpose. Other applications of wavelet theory to meteorological data can be found in 1, 5].

Conclusion
The wavelet transformation provides a very general approach to time-frequency representations and constitutes the state-of-the-art in modern signal analysis. It overcomes the limitations of the short-time Fourier transformation and shows the time-varying spectral content of a function within the bounds set by the uncertainty principle. Furthermore, it can be used to display the correlation between di erent but related functions (coherence). Wavelet-analysis is able to display information in an intuitive way, that would be hidden to the observer otherwise. This information can be used to explain observed facts or to construct and test synthetic data generators for multiple outputs with proper correlations.

wavelet coherence, (morletwavelet, b = 11 )


1

10

frequency in 1/h

10

10

10

1000

2000

3000

4000

5000

6000

7000

8000

normal solar radiation in KJ/m2 3000 2000 1000 0 0

1000

2000

3000 4000 5000 6000 ambient temperature in degree C

7000

8000

20 0 20 0

1000

2000

3000

4000 5000 time in h

6000

7000

8000

Figure 4: Coherency between solar radiation and air temperature

References
1] H.G. Beyer, C. Costanzo, and Ch. Reise. Multiresolution analysis of satellite-derived irradiance maps. Solar Energy, 55(1):9{20, 1995. 2] J. Boland. Time-series analysis of climatic variables. Solar Energy, 55(5):377{388, 1995. 3] A. Cohen and R.D. Ryan. Wavelets and Multiscale Signal Processing. Chapman & Hall, London, 1995. 4] Ch. Langer, D. Heinemann, and J. Schumacher. Synthesis of hourly ambient temperature time series correlated with solar radiation. In Proc. EuroSun'96, Freiburg, 1996. 5] P.C. Liu. Wavelet spectrum analysis and ocean wind waves. In E. Fousala-Georgiou and P. Kumar, editors, Wavelets in Geophysics, pages 151{166. Academic Press, 1994. 6] A. Mertins. Signaltheorie. Teubner, Stuttgart, 1996.

Das könnte Ihnen auch gefallen