Probabilistic Finite Automata and Randomness in Nature: a New Approach in the Modelling and Prediction of Climatic Parameters1 L.Mora-Lópeza, R.Morales-Buenoa, M.Sidrach-de-Cardonab, F.Trigueroa a

Dpto.Lenguajes y C.Computación ([email protected], [email protected], [email protected]) b Dpto. Física Aplicada II. ([email protected]) a,b E.T.S.I.Informática.Universidad de Málaga. Campus Teatinos. 29071 Málaga

Abstract: A model to characterize and predict continuous time series from machine learning techniques is proposed. This model includes the following three steps: dynamic discretization of continuous values, construction of probabilistic finite automata and prediction of new series with randomness. The first problem in most models from machine learning is that they are developed for discrete values; however, most phenomena in nature are continuous. To convert these continuous values into discrete values a dynamic discretization method has been used. With the obtained discrete series, we have built probabilistic finite automata which include all the representative information which the series contain. The use of probabilistic finite automata allows us to consider, in an easy way, the different relationships between the values in the series for different environmental conditions. The learning algorithm to build these automata is polynomial in the sample size. An algorithm to predict new series has been proposed. This algorithm incorporates the randomness in nature: values are generated using the cumulative probability distribution function -included in the automata- and a random number to select the new value. After finishing the three steps of the model, the similarity between the predicted series and the real ones has been checked. For this, a new adaptable test based on the classical Kolmogorov-Smirnov two-sample test has been developed; this test takes into account the continuous nature of climatic data. The cumulative distribution function of observed and generated series has been compared using the concept of indistinguishable values. Finally, the proposed model has been applied in a practical cases: the study of hourly global solar radiation series. Keywords: Machine Learning, Modelling Climatic Data, Time Series

1.

INTRODUCTION

The fundamental idea in this paper is the use of probabilistic finite automata (PFA) as a means of representing the relationships observed in climatic data series. PFAs are mathematical models used in the machine learning field. Traditionally, the analysis of time series has been carried out using stochastic process theory. One of the most detailed analyses of statistical methods for time series research was done by Box et al [1976].

1

The goal of data analysis by time series is to find models which are able to reproduce the statistical characteristics of the series. Moreover, these models allow us to predict the next value of the series from its predecessors. Probabilistic finite automata have been used to model several types of natural sequences. Examples of such applications are: universal data compression, Rissman [1983], analysis of biological sequences, for DNA and proteins, Krogh et al. [1993], analysis of natural language, for handwriting and speech, Nadas [1984], Rabiner

This work has been partially supported by FACA Project number PB98-0937-C04-01 of the CICYT, Spain. FACA is a part of the FRESCO project

78

[1994] and Ron et al. [1998], etc. Different classes of automata have been developed. For instance, acyclic probabilistic finite automata have been used for modeling distributions on short sequences, Ron et al. [1998]; probabilistic suffix automata, based on variable order Markov models, have been used to construct a model of the English language, Ron et al. [1994]. All these automata allow us to take into account the temporal relationships in a series. These machine learning models are very useful to study systems in which the concept to learn presents probabilistic behaviour. The prediction of climatic variables is an example of these types of concepts. In these systems the recorded variables are insufficient to exactly determine the future values, due to the random nature of these variables. The systems in which these models can be used must have the following properties: −= Present probabilistic behaviour or uncertainty. This uncertainty can be due to several factors. For example, for the prediction of climatic variables the number of parameters which affect them is very high. −= Although there is uncertainty in these systems, there is always some structure within this uncertainty. This paper describes how to use certain models from the machine learning field in the analysis and prediction of climatic parameters. The model we propose is based on the Probabilistic Finite Automata Theory. Our goal is to use PFAs to represent all the relationships observed in natural time series and to use these PFAs to predict new values of the series. Moreover, an adaptable test based on the classic Kolmogorov-Smirnov twosample test has been used to check the proposed model. Finally, preliminary results of the model obtained for a climatic parameter are shown.

2.

period t1 to tm we will use the symbols y1y2...ym. So, in the series x5x3...x3, the symbol y1corresponds to the value x5, the symbol y2 to x3 and so on. - Q is a finite collection of states. Each state corresponds to a subsequence of the discretized time series. The maximum size of a state -number of symbols- is bounded by a value N fixed in advance. This value is related to the number of previous values which will be considered to determine the next value in the series and depends on “memory” of the series. - τ: Q × Σ → Q is the transition function - γ: Q × Σ → [0,1] is the next symbol probability function - q0 ∈ Q, is the initial state The function γ satisfies the following requirement: For every q ∈ Q and for every xi∈Σ, Σxi∈Σ γ(q, xi)=1. Moreover, the following conditions are required: - The transition function τ can be undefined only on states q∈Q and symbols x∈Σ, for which γ(q,x)=0; - The function τ can be extended to be defined on Q × Σ* in the following recursive manner: τ(q,y1,y2,...yt)= τ(τ(q,y1,y2,...,yt-1),yt). where yi∈Σ. Graphically, each state is represented by a node and the edges going out of each state are labeled by symbols drawn from the alphabet. Moreover, each state has an associated probability vector which is composed of the probability of the next symbol for each of the symbols of the alphabet. For instance, in figure 1 a simple PFA is shown. i (0.5,0.5) (0.4,0.6) 0

00

PROBABILISTIC FINITE AUTOMATA

1 (0.5,0.5)

01

10

11 (0.5,0.5)

(0.25,0.75) (0.5,0.5) (0.25,0.75)

2. 1 Introduction We propose using a mathematical model called probabilistic finite automata (PFA). We propose the use of this mathematical model to represent a univariate time series. Formally, a PFA is a 5-tuple (Σ,Q,τ,γ,q0) where: - Σ is a finite alphabet; that is, a set of discrete symbols corresponding to the different continuous values of the analyzed parameter. The different symbols of Σ will be represented by xi. For a series, the values observed can be x5x3,...x3 To represent the different observable series for a

79

Figure 1. Example of probabilistic finite automata In this PFA, the alphabet, Σ, is composed of the symbols 0 and 1. The states of the system, Q, are described in each node of the automata: initial (i), 0, 1, 00, 01, 10 and 11. For instance, the state labeled 01 corresponds to the following sequence of values in the series: 1 as the last value and 0 as the previous. The associated vectors at each state (node) are the probabilities which each symbol of the alphabet has to appear in the next moment, after the sequence of symbol that label the node has appeared. For instance, the node labeled with

10, has the associated vector (0.25,0.75); this means that if the current state is 10, then the next symbol can be 0, with a probability of 0.25 and 1 with a probability of 0.75. The continuous and discontinuous arrows represent the transition function between states (discontinuous for 0, and continuous for 1). For instance, if the current state is 10, and the next symbol is 0, then the following state will be labeled with 00; but if the next symbol is 1, then the following state will be labeled with 01.

3.

In the PFA shown in Figure 1, the states 01 and 11 have the same probability vector as state 1. That is, when the symbol 1 appears, it is not necessary to know the preceding value to determine the probabilities of the next symbol, since in both cases, (0 or 1), the probabilities vector of the next symbol is (0.5,0.5). Therefore, the PFA of Figure 1 can be converted into the PFA shown in Figure 2.

3. The set PSS -Possible Subsequence Set- is initialized with all sequences of order 1. Each element in this set corresponds to a sequence of discrete values. Take o=1 as the initial value of the order –that is, size of subsequences to consider.

i (0.5,0.5) (0.4,0.6) 0

00 (0.25,0.75)

BUILDING PROBABILISTIC AUTOMATA

FINITE

3.1 Algorithm to build probabilistic finite automata. The following algorithm is used to construct the PFA: 1. Compute the series of discrete values. 2. Initialize the PFA with a node, with label null sequence.

4. If there are elements of order o in PSS, pick any of these elements, Y. Using all discrete sequences in the series, compute the frequency of Y. If 4.a and 4.b are true, then go to 5, else go to 6. 4.a The frequency of this sequence is greater than the threshold frequency.

1 (0.5,0.5)

4.b For some xp∈Σ, the probability of occurrence of the subsequence Yxp is not equal to the probability of the subsequence final(Y)xp, s, that is:

10 (0.25,0.75)

P(xp|Y)≠P(xp|final(Y)).

Figure 2. Simplified probabilistic finite automata This class of PFA is used to represent variable order Markov models. These simplified automata are the automata proposed in this paper. They capture the same information with fewer states than the original automata. Moreover, they allow us to take into account, for each state, a different number of previous values in the series. Let us define some concepts that we will use to build the PFAs for climatic data series. Let Σ={x1,x2,...,xn} be the set of discrete values of the analized variable and Σ* denotes the set of all possible sequences which can be obtained with these values. For any integer N, ΣN denotes the set of all possible sequences of length N and Σ≤N is the set of all possible sequences with length less than or equal to N. For any subsequence, Y, represented by y1...ym, where yi∈Σ, the following notations will be used: −=

The longest final subsequence of Y, different from Y, will be final(Y)=y2...ym −= The set of all final subsequences of Y will be, last(Y)={yi...ym|1≤i≤m} In the next section we explain how to build a PFA for a time series.

80

(not equal: when the ratio between the probabilities is significantly greater than one) 5.Do 5.a Add to the PFA a node, labeled with Y, and compute its corresponding probabilities vector. 5.b For each amplified sequence, Yxp: if the probability of this augmented sequence is greater than the threshold probability, then include it in PSS. 6. Remove the analyzed subsequence, Y, from PSS. 7. If there are no elements of order o in PSS, add 1 to the value of o. If o ≤N and there are elements of length o in PSS, then go to 4, else Stop.

3.2

Predicting new values

A PFA can be used as a mechanism for generating finite sequences of values in the following manner. Start from an initial value selected from the alphabet, called the initial state. If qt is the current state, labeled by the sequence Y=y1...yt, then the next symbol is chosen (probabilistically) according to γ(qt,⋅). If x∈Σ is the chosen symbol, then the next state, qt+1, is τ(qt,x). The label of this new

state, Y’, will be the longest final subsequence of Yx in the PFA, that is:

must be used in the test, we propose using a bootstrap procedure.

Y’=Max{last(Yx)}∈PFA. The process continues until the length of the required sequence is reached. Moreover, if Pt(Y) denotes the probability that a PFA generates a sequence Y=y1...yt-1yt, then: t −1

P t (Y ) = ∏ γ (q i , y i +1 ). i =0

This definition implies that Pt(⋅) is in fact a probability distribution over the symbols of sequence, i.e.:

P t (Y ) = 1. Y ∈Σ*

4. HOW VALIDATED

THE

MODEL

CAN

BE

For a recorded time series, the following steps must be followed to use the proposed model. First, if the time series has continuous values, then these values must be discretized. After this, the PFA is built using the discrete series. With the PFA and the generation method described above, new values for the time series can be generated. In order to compare the simulated series to the real ones, several statistical tests can be used. The hypothesis that both series have the same mean and variance will be checked. The frequency histograms of the recorded and simulated series are also analyzed. To make this comparison, we propose the use of an adaptable goodness-of-fit test, which is based on the twosample Kolmogorov-Smirnov test, described in Rohatgi [1976]. The objective of this adaptable test is to determine if two distribution functions FY(.) and FZ(.) are the same, except for possible changes in location and scale. Specifically, we have checked the null hypothesis that there exist two unknow values µ and σ such that Zi and µ+σYj have the same distribution. Using distribution functions it is possible to express our null hypotheses as follows:

H 0 : ∃µ ∈ ℜ and σ ∈ (0,+∞ ) / ∀u ∈ ℜ

F X (u ) = FY (

u−µ ) σ

Replacing unknow parameters µ and σ by estimates introduces additional random terms in the statistic. Therefore, to obtain the critical values that

81

5. A PRACTICAL CASE: USING PROBABILISTIC FINITE AUTOMATA FOR CLIMATIC DATA The probabilistic finite automata presented in the previous sections have been used to characterize and predict a climatic variable; the hourly global radiation received on a surface on the ground. For this variable, time series are recorded by meteorological stations at regular time intervals. We need a stationary time series. From the original series we have calculated the series of the hourly clearness index, which are stationary. The following question - which we have solved- is the discretization of these series. The recorded values are continuous whereas the proposed mathematical model uses discrete values. The discretization method used is explained later. The PFAs have been built using the discrete series obtained and new values of the series generated. Finally, we have checked these values using several tests.

5.1

Data set

The data of the hourly exposure series of global radiation, {Gh(t)}, which are used in this work were recorded over several years in nine Spanish meteorological stations. In total, 745 months were accounted for. The pertinent latitudes range from 36ºN to 44ºN. The annual average values of daily global solar radiation for these locations range from 11MJm-2 to 18 MJm-2.

5.2

Discretization of the series

The goal is to use an effective and efficient method to transform continuous values into discrete ones using the overall information included in the series and, when possible, feedback with the learning system. To do this, the discrete value which corresponds to a continuous value has been calculated using qualitative reasoning, taking into account the evolution of the series. Qualitative models have been used in different areas in order to obtain a representation of the domain based on properties (qualities) of the systems; see, for instance, Forbus[1984], Keller et al. [1984] and Kuipers[1984]. We have used a qualitative dynamic discrete conversion method, described in Mora et al.[2000]. It is dynamic because the discrete value associated to a particular continuous value can change over

time: that is, the same continuous value can be discretized into different values, depending on the previous values observed in the series. It is qualitative because only those changes which are qualitatively significant appear in the discretized series.

only component of probabilities vector -for the current state, qt- which satisfies:

5.3 series

5.5

PFA for global hourly solar radiation

The parameter used to build the PFA is the hourly clearness index, defined as: K h = Gh / Gh , 0

where Gh is the hourly global radiation and Gh,0 is the extraterrestrial hourly global radiation. The alphabet of the PFA is: Σ={0,1,...,7} The relationship between the values of the clearness index and the symbols of the alphabet is the following. For the first symbol of a series, the discrete value of the series will be calculated using:

ì0 0 ≤ K h < 0.35 ê K − 0.35 ú + 1 0.35 ≤ K h < 0.65 Yh = í ê h ë 0.05 K h≥0.65 î7 where A means the integer value of A. For following values of the continuous series discretes values will be obtained using algorithm described in Mora et al.[2000] and intervals aforementioned.

j i =1γ ( qt , xi )

≥ r AND x j −1 |

j −1 i =1 γ ( qt , xi )

Dpto.Lenguajes y C.Computación ([email protected], [email protected], [email protected]) b Dpto. Física Aplicada II. ([email protected]) a,b E.T.S.I.Informática.Universidad de Málaga. Campus Teatinos. 29071 Málaga

Abstract: A model to characterize and predict continuous time series from machine learning techniques is proposed. This model includes the following three steps: dynamic discretization of continuous values, construction of probabilistic finite automata and prediction of new series with randomness. The first problem in most models from machine learning is that they are developed for discrete values; however, most phenomena in nature are continuous. To convert these continuous values into discrete values a dynamic discretization method has been used. With the obtained discrete series, we have built probabilistic finite automata which include all the representative information which the series contain. The use of probabilistic finite automata allows us to consider, in an easy way, the different relationships between the values in the series for different environmental conditions. The learning algorithm to build these automata is polynomial in the sample size. An algorithm to predict new series has been proposed. This algorithm incorporates the randomness in nature: values are generated using the cumulative probability distribution function -included in the automata- and a random number to select the new value. After finishing the three steps of the model, the similarity between the predicted series and the real ones has been checked. For this, a new adaptable test based on the classical Kolmogorov-Smirnov two-sample test has been developed; this test takes into account the continuous nature of climatic data. The cumulative distribution function of observed and generated series has been compared using the concept of indistinguishable values. Finally, the proposed model has been applied in a practical cases: the study of hourly global solar radiation series. Keywords: Machine Learning, Modelling Climatic Data, Time Series

1.

INTRODUCTION

The fundamental idea in this paper is the use of probabilistic finite automata (PFA) as a means of representing the relationships observed in climatic data series. PFAs are mathematical models used in the machine learning field. Traditionally, the analysis of time series has been carried out using stochastic process theory. One of the most detailed analyses of statistical methods for time series research was done by Box et al [1976].

1

The goal of data analysis by time series is to find models which are able to reproduce the statistical characteristics of the series. Moreover, these models allow us to predict the next value of the series from its predecessors. Probabilistic finite automata have been used to model several types of natural sequences. Examples of such applications are: universal data compression, Rissman [1983], analysis of biological sequences, for DNA and proteins, Krogh et al. [1993], analysis of natural language, for handwriting and speech, Nadas [1984], Rabiner

This work has been partially supported by FACA Project number PB98-0937-C04-01 of the CICYT, Spain. FACA is a part of the FRESCO project

78

[1994] and Ron et al. [1998], etc. Different classes of automata have been developed. For instance, acyclic probabilistic finite automata have been used for modeling distributions on short sequences, Ron et al. [1998]; probabilistic suffix automata, based on variable order Markov models, have been used to construct a model of the English language, Ron et al. [1994]. All these automata allow us to take into account the temporal relationships in a series. These machine learning models are very useful to study systems in which the concept to learn presents probabilistic behaviour. The prediction of climatic variables is an example of these types of concepts. In these systems the recorded variables are insufficient to exactly determine the future values, due to the random nature of these variables. The systems in which these models can be used must have the following properties: −= Present probabilistic behaviour or uncertainty. This uncertainty can be due to several factors. For example, for the prediction of climatic variables the number of parameters which affect them is very high. −= Although there is uncertainty in these systems, there is always some structure within this uncertainty. This paper describes how to use certain models from the machine learning field in the analysis and prediction of climatic parameters. The model we propose is based on the Probabilistic Finite Automata Theory. Our goal is to use PFAs to represent all the relationships observed in natural time series and to use these PFAs to predict new values of the series. Moreover, an adaptable test based on the classic Kolmogorov-Smirnov twosample test has been used to check the proposed model. Finally, preliminary results of the model obtained for a climatic parameter are shown.

2.

period t1 to tm we will use the symbols y1y2...ym. So, in the series x5x3...x3, the symbol y1corresponds to the value x5, the symbol y2 to x3 and so on. - Q is a finite collection of states. Each state corresponds to a subsequence of the discretized time series. The maximum size of a state -number of symbols- is bounded by a value N fixed in advance. This value is related to the number of previous values which will be considered to determine the next value in the series and depends on “memory” of the series. - τ: Q × Σ → Q is the transition function - γ: Q × Σ → [0,1] is the next symbol probability function - q0 ∈ Q, is the initial state The function γ satisfies the following requirement: For every q ∈ Q and for every xi∈Σ, Σxi∈Σ γ(q, xi)=1. Moreover, the following conditions are required: - The transition function τ can be undefined only on states q∈Q and symbols x∈Σ, for which γ(q,x)=0; - The function τ can be extended to be defined on Q × Σ* in the following recursive manner: τ(q,y1,y2,...yt)= τ(τ(q,y1,y2,...,yt-1),yt). where yi∈Σ. Graphically, each state is represented by a node and the edges going out of each state are labeled by symbols drawn from the alphabet. Moreover, each state has an associated probability vector which is composed of the probability of the next symbol for each of the symbols of the alphabet. For instance, in figure 1 a simple PFA is shown. i (0.5,0.5) (0.4,0.6) 0

00

PROBABILISTIC FINITE AUTOMATA

1 (0.5,0.5)

01

10

11 (0.5,0.5)

(0.25,0.75) (0.5,0.5) (0.25,0.75)

2. 1 Introduction We propose using a mathematical model called probabilistic finite automata (PFA). We propose the use of this mathematical model to represent a univariate time series. Formally, a PFA is a 5-tuple (Σ,Q,τ,γ,q0) where: - Σ is a finite alphabet; that is, a set of discrete symbols corresponding to the different continuous values of the analyzed parameter. The different symbols of Σ will be represented by xi. For a series, the values observed can be x5x3,...x3 To represent the different observable series for a

79

Figure 1. Example of probabilistic finite automata In this PFA, the alphabet, Σ, is composed of the symbols 0 and 1. The states of the system, Q, are described in each node of the automata: initial (i), 0, 1, 00, 01, 10 and 11. For instance, the state labeled 01 corresponds to the following sequence of values in the series: 1 as the last value and 0 as the previous. The associated vectors at each state (node) are the probabilities which each symbol of the alphabet has to appear in the next moment, after the sequence of symbol that label the node has appeared. For instance, the node labeled with

10, has the associated vector (0.25,0.75); this means that if the current state is 10, then the next symbol can be 0, with a probability of 0.25 and 1 with a probability of 0.75. The continuous and discontinuous arrows represent the transition function between states (discontinuous for 0, and continuous for 1). For instance, if the current state is 10, and the next symbol is 0, then the following state will be labeled with 00; but if the next symbol is 1, then the following state will be labeled with 01.

3.

In the PFA shown in Figure 1, the states 01 and 11 have the same probability vector as state 1. That is, when the symbol 1 appears, it is not necessary to know the preceding value to determine the probabilities of the next symbol, since in both cases, (0 or 1), the probabilities vector of the next symbol is (0.5,0.5). Therefore, the PFA of Figure 1 can be converted into the PFA shown in Figure 2.

3. The set PSS -Possible Subsequence Set- is initialized with all sequences of order 1. Each element in this set corresponds to a sequence of discrete values. Take o=1 as the initial value of the order –that is, size of subsequences to consider.

i (0.5,0.5) (0.4,0.6) 0

00 (0.25,0.75)

BUILDING PROBABILISTIC AUTOMATA

FINITE

3.1 Algorithm to build probabilistic finite automata. The following algorithm is used to construct the PFA: 1. Compute the series of discrete values. 2. Initialize the PFA with a node, with label null sequence.

4. If there are elements of order o in PSS, pick any of these elements, Y. Using all discrete sequences in the series, compute the frequency of Y. If 4.a and 4.b are true, then go to 5, else go to 6. 4.a The frequency of this sequence is greater than the threshold frequency.

1 (0.5,0.5)

4.b For some xp∈Σ, the probability of occurrence of the subsequence Yxp is not equal to the probability of the subsequence final(Y)xp, s, that is:

10 (0.25,0.75)

P(xp|Y)≠P(xp|final(Y)).

Figure 2. Simplified probabilistic finite automata This class of PFA is used to represent variable order Markov models. These simplified automata are the automata proposed in this paper. They capture the same information with fewer states than the original automata. Moreover, they allow us to take into account, for each state, a different number of previous values in the series. Let us define some concepts that we will use to build the PFAs for climatic data series. Let Σ={x1,x2,...,xn} be the set of discrete values of the analized variable and Σ* denotes the set of all possible sequences which can be obtained with these values. For any integer N, ΣN denotes the set of all possible sequences of length N and Σ≤N is the set of all possible sequences with length less than or equal to N. For any subsequence, Y, represented by y1...ym, where yi∈Σ, the following notations will be used: −=

The longest final subsequence of Y, different from Y, will be final(Y)=y2...ym −= The set of all final subsequences of Y will be, last(Y)={yi...ym|1≤i≤m} In the next section we explain how to build a PFA for a time series.

80

(not equal: when the ratio between the probabilities is significantly greater than one) 5.Do 5.a Add to the PFA a node, labeled with Y, and compute its corresponding probabilities vector. 5.b For each amplified sequence, Yxp: if the probability of this augmented sequence is greater than the threshold probability, then include it in PSS. 6. Remove the analyzed subsequence, Y, from PSS. 7. If there are no elements of order o in PSS, add 1 to the value of o. If o ≤N and there are elements of length o in PSS, then go to 4, else Stop.

3.2

Predicting new values

A PFA can be used as a mechanism for generating finite sequences of values in the following manner. Start from an initial value selected from the alphabet, called the initial state. If qt is the current state, labeled by the sequence Y=y1...yt, then the next symbol is chosen (probabilistically) according to γ(qt,⋅). If x∈Σ is the chosen symbol, then the next state, qt+1, is τ(qt,x). The label of this new

state, Y’, will be the longest final subsequence of Yx in the PFA, that is:

must be used in the test, we propose using a bootstrap procedure.

Y’=Max{last(Yx)}∈PFA. The process continues until the length of the required sequence is reached. Moreover, if Pt(Y) denotes the probability that a PFA generates a sequence Y=y1...yt-1yt, then: t −1

P t (Y ) = ∏ γ (q i , y i +1 ). i =0

This definition implies that Pt(⋅) is in fact a probability distribution over the symbols of sequence, i.e.:

P t (Y ) = 1. Y ∈Σ*

4. HOW VALIDATED

THE

MODEL

CAN

BE

For a recorded time series, the following steps must be followed to use the proposed model. First, if the time series has continuous values, then these values must be discretized. After this, the PFA is built using the discrete series. With the PFA and the generation method described above, new values for the time series can be generated. In order to compare the simulated series to the real ones, several statistical tests can be used. The hypothesis that both series have the same mean and variance will be checked. The frequency histograms of the recorded and simulated series are also analyzed. To make this comparison, we propose the use of an adaptable goodness-of-fit test, which is based on the twosample Kolmogorov-Smirnov test, described in Rohatgi [1976]. The objective of this adaptable test is to determine if two distribution functions FY(.) and FZ(.) are the same, except for possible changes in location and scale. Specifically, we have checked the null hypothesis that there exist two unknow values µ and σ such that Zi and µ+σYj have the same distribution. Using distribution functions it is possible to express our null hypotheses as follows:

H 0 : ∃µ ∈ ℜ and σ ∈ (0,+∞ ) / ∀u ∈ ℜ

F X (u ) = FY (

u−µ ) σ

Replacing unknow parameters µ and σ by estimates introduces additional random terms in the statistic. Therefore, to obtain the critical values that

81

5. A PRACTICAL CASE: USING PROBABILISTIC FINITE AUTOMATA FOR CLIMATIC DATA The probabilistic finite automata presented in the previous sections have been used to characterize and predict a climatic variable; the hourly global radiation received on a surface on the ground. For this variable, time series are recorded by meteorological stations at regular time intervals. We need a stationary time series. From the original series we have calculated the series of the hourly clearness index, which are stationary. The following question - which we have solved- is the discretization of these series. The recorded values are continuous whereas the proposed mathematical model uses discrete values. The discretization method used is explained later. The PFAs have been built using the discrete series obtained and new values of the series generated. Finally, we have checked these values using several tests.

5.1

Data set

The data of the hourly exposure series of global radiation, {Gh(t)}, which are used in this work were recorded over several years in nine Spanish meteorological stations. In total, 745 months were accounted for. The pertinent latitudes range from 36ºN to 44ºN. The annual average values of daily global solar radiation for these locations range from 11MJm-2 to 18 MJm-2.

5.2

Discretization of the series

The goal is to use an effective and efficient method to transform continuous values into discrete ones using the overall information included in the series and, when possible, feedback with the learning system. To do this, the discrete value which corresponds to a continuous value has been calculated using qualitative reasoning, taking into account the evolution of the series. Qualitative models have been used in different areas in order to obtain a representation of the domain based on properties (qualities) of the systems; see, for instance, Forbus[1984], Keller et al. [1984] and Kuipers[1984]. We have used a qualitative dynamic discrete conversion method, described in Mora et al.[2000]. It is dynamic because the discrete value associated to a particular continuous value can change over

time: that is, the same continuous value can be discretized into different values, depending on the previous values observed in the series. It is qualitative because only those changes which are qualitatively significant appear in the discretized series.

only component of probabilities vector -for the current state, qt- which satisfies:

5.3 series

5.5

PFA for global hourly solar radiation

The parameter used to build the PFA is the hourly clearness index, defined as: K h = Gh / Gh , 0

where Gh is the hourly global radiation and Gh,0 is the extraterrestrial hourly global radiation. The alphabet of the PFA is: Σ={0,1,...,7} The relationship between the values of the clearness index and the symbols of the alphabet is the following. For the first symbol of a series, the discrete value of the series will be calculated using:

ì0 0 ≤ K h < 0.35 ê K − 0.35 ú + 1 0.35 ≤ K h < 0.65 Yh = í ê h ë 0.05 K h≥0.65 î7 where A means the integer value of A. For following values of the continuous series discretes values will be obtained using algorithm described in Mora et al.[2000] and intervals aforementioned.

j i =1γ ( qt , xi )

≥ r AND x j −1 |

j −1 i =1 γ ( qt , xi )