Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
WEATHER PREDICTION EXPERT SYSTEM APPROACHES (Ceng-568 Literature Survey) Bulent KISKAC and Harun YARDIMCI Computer Engineering Middle East Technical University Ankara,TURKEY June, 2004 ABSTRACT Weather prediction approaches are challenged by complex weather phenomena with limited observations and past data. Weather phenomena have many parameters that are impossible to enumerate and measure. Short-term local weather prediction accuracy and timeliness can save million dollars while a large-scale prediction can save billion dollars and lives. Fuzzy Case Based approach and Artificial Neural Network approach are practical and successful implementations for local area short-term weather forecast. Increasing development on communication systems enabled weather forecast expert systems to integrate and share resources and thus hybrid system has emerged. Even though these improvements on weather forecast, these expert systems can not be fully reliable since weather forecast is chaotic problem. 1. INTRODUCTION Weather is one of the most effective environmental constraints in every phase of our lives. We are subject to adjusting ourselves with respect to weather condition from our dressing habits to strategic organizational planning activities, since the adverse weather conditions may cause a considerable damage on our lives and properties. We need to be on alert to these adverse weather conditions by taking some precautions and using prediction mechanisms for early warning of hazardous weather phenomena. Weather prediction is an indispensable requirement for all of us. There is a general and increasing interest on weather information, since every day we habitually give an ear to weather forecast news for local and large-scale long-term or short-term weather predictions. Leading weather research institutions and companies have been developing weather prediction systems capable of detecting, predicting and forecasting weather phenomena and hazards by utilizing state-of-the-science technologies. Thus weather prediction utilizations fields and prediction accuracy increases monotonically by the time. In this survey we try to give readers an overview about weather prediction phenomena, expert systems approaches, main domain specific problems, and solution methodologies. We made this research in parallel with our KE project about a local airbase short-term weather prediction implementation with Case Base Reasoning KNN algorithm, Fuzzy logic and Artificial Neural Network implementation and their performance comparisons. That is why our survey mainly focuses on local and short-term predictions when covering all literatures about weather prediction expert systems. 1 1.1 Weather Prediction Problem Characteristics? Weather prediction is a complicated procedure that includes multiple specialized fields of expertise. “There are only two methods to predict weather: the empirical approach and the dynamical approach” (Lorenz 19) [1]. Lorenz thus separated weather forecasting methodologies into two main branches in terms of numerical modeling and scientific processing (AI) of meteorological data. Hansen[2] explains this classification as follows; The empirical approach is based upon the happenings of comparable cases (i.e., similar weather situations). The empirical approach is useful for predicting local-scale weather if recorded cases are plentiful (e.g., cloud ceiling and visibility in a few square kilometers around an airport). The dynamical approach is based upon equations of the atmosphere and is commonly referred to as computer modeling. The dynamical approach is only useful for modeling large-scale weather phenomena (e.g., general wind direction over a few thousand square kilometers). The dynamical approach is a NWP method that takes inputs readings from weather stations, weather buoys, satellite images, atmospheric probes and other sources and computes solar energy effects and global rotational effects, latitude effects, ocean currents effects, landmasses effects and produces global weather predictions. NWP models build these computation results as initial conditions to systems of equations that describe the atmosphere. The models are run statically and produce forecasts on a six hourly basis that serves as the foundation of all forecasts. Output from these models is useful for large scale and longer-term forecasts. However, the complete description of the atmosphere in this form is far beyond current capability that requires more improvements, more measurement, and more parameters. Since the accuracy of the models is inherently dependent upon the initial conditions that are inherently incomplete, these systems are not able to produces results that satisfies everybody in local and shortterm cases. The empirical approach mostly depends on past meteorological data assuming that initially similar conditioned weather conditions shall go in parallel for some time enough to make a short range weather prediction. Huschke[3] defines a “persistence forecast” as “a forecast that the future weather conditions will be the same as the present conditions.” This approach is an analogical method of weather forecasting that bases predictions for the present case on the outcomes of similar past cases. Utilizing past meteorological data is not as easy as it seems, since there are many problems to be handled, like deciding similar cases in continuous weather parameters. There are many different solution trials like using Case Base Reasoning, Artificial Neural Network, Decision Trees and Rule Production System etc. 1.2 Short term and long term forecast Lately expert systems using artificial intelligence (AI) techniques have been used to forecast visibility, marine fog, precipitation, severe weather and other climatological conditions [e.g., 4,5,6,7]. Expert systems on smaller scale model have many of the characteristic advantages that typically arise when AI techniques are applied. First they 2 are guided and taught by human domain experts so they become more adaptive with this heuristic guidance for using knowledge acquisition techniques. Secondly they can cope with uncertainty, incompleteness and recency. Thirdly they are not computationally intensive as much as NWP systems. Large Scale weather prediction information can easily be accessible from weather report services for a big region or a city. They report like tomorrow heavy rain expected in this city and visibility will be 10 km. and so on. However, large-scale static forecasts are not sufficiently stable on a small scale where local effects and recency become significant or predominant. 1.3 Forecast Types Weather forecast expert systems can be categorized as terminal aerodrome forecasts (TAFs), public forecasts and marine forecasts in terms of field of interest and scale. In addition to these, many specialized forecasts are produced for specific purposes (e.g. weather pollution forecasting, agricultural predictions) TAFs are required to be most precise both in terms of measurable weather conditions and in terms of timing. TAF forecasts of the height of low cloud ceiling are expected to be accurate to within 100 feet; forecasts of the horizontal visibility on the ground, when there is dense obstruction to visibility, such as fog or snow, are expected to be accurate to within 400 meters Vislocky [8]; and forecasts of the time of change from one flying category to another are expected to be accurate to within one hour Wilson [10]. On the other hand public and marine forecasts can be less precise. In public forecasts, for example, it may be sufficient to predict “cloudiness this morning,” and in marine forecasts, it may be sufficient to predict “fog patches forming this afternoon.” 1.4 Importance of TAF Airports organize air traffic mainly based on local weather condition that has agile parameters that may change in a short time. These parameters such as horizontal visibility and vertical visibility (ceiling), fogginess, precipitation level etc. can be very serious threat for flight safety and cost. When ceiling and visibility at a busy airport are low, in order to maximize safety, the rate of planes landing is reduced. When ceiling and visibility at a destination airport are forecast to be low at a flight’s scheduled arrival time, its departure may be delayed in order to minimize traffic congestion and related costs. For Airport weather forecast timeliness and correctness are the main issues. An examination of the causes and effects of flight delays at the three main airports serving New York City concluded that a correctly forecast timing of a ceiling and visibility event (i.e., a significant change) could be expected to result in a savings of approximately $480,000 per event at La Guardia Airport [10]. 3 2. WEATHER PREDICTION EXPERT SYSTEMS From now on, since NWP approaches are another research subject. In this survey, we will focus on expert systems with empirical approaches that use past meteorological data and AI algorithms. We will explain their relative success and their constraining factors. Our research sources are mainly related with experimental papers, our own flight alert system implementation (FA-1) has many clues, interesting results, and illuminating information for readers. 2.1 Case Base Reasoner Expert Systems Case-based reasoning has been used to solve problems in diverse areas including decision support, help desk support, product cataloguing and maintenance support [11, 12]. In order to solve a current problem a case-based reasoner requires a library of past problems and the solutions that were used to solve the past problems. The case-based reasoner works by using a measure of similarity to retrieve past problems that are most similar to the current problem. The reasoner then combines and adapts the solutions to the most similar past problems to generate a proposed best solution to the current problem. Clearly the designs of the similarity metric and the adaptation algorithm are crucial to the functionality of any case-based reasoner. There are many similarity metrics that can be used in case-based reasoners for retrieving similar past case. Euclidean norms and Hamming distances are the most common; rule-based metrics have been discussed [13]; some metrics use domain knowledge [14]; similarity measures have been adapted to use fuzzy set based formalisms [15] and others use hybrid fuzzy measures [16]. FA-1 system uses Euclidean norms to retrieve past similar case. First it calculates relative distances of past data rows with respect to our current case attributes, after finding relative distances we select K nearest rows. Past data is stored consecutively in a relational database and in order of hourly interval. Easiest way is actually to give K=1 so we get nearest neighbor quickly but this is not proper because we have to take probability distribution of cases into account, maybe this 1 nearest neighbor has noisy and incomplete data. For this reason we select K=16 and calculated every attribute value by taking a weighted average. You can clearly see how to KNN work for continuous values in Figure-1 This result most probably does not exist in past meteorology database but again we calculate 1 nearest neighbor this second process ensure us that we are not most probably affected by noisy data and abnormal cases. After finding most similar past record we retrieve next record as a solution for hourly weather prediction. We assume every attribute has equal weight with respect to each other. For example Temperature attribute has (min= –20, max= +50) dynamic range when Ceiling attribute has (min=0, max=100000feet) dynamic range. Euclidean distance equation naturally gives more weight to Ceiling attribute. We need to stretch attributes to same dynamic range in accordance with their probability distribution. This stretching method requires some statistical approaches. FA-1 is now limited to correctness of this stretching. So we 4 conclude that even such a simple implementation is not easy to handle that require domain expertise and guidance. Figure 1 One of surveyed papers claims that they used fuzzy set based formalisms and achieved great success [2]. Actually this second method seems very promising because they convert all database attributes from continuous values to discrete values and in every fuzzy attribute values set, each member has differing weight; for example rain condition and drizzle condition are very near to each other but snow has totally different event has totally different conditions. Fuzzy logic imparts to case-based reasoning the perceptiveness and case discriminating ability of a domain expert. The fuzzy KNN technique retrieves similar cases by emulating a domain expert who understands and interprets similar cases. The main contribution of fuzzy logic to case-based reasoning (CBR) is that it enables us to use common words to directly acquire domain knowledge about feature salience. This knowledge enables us to retrieve a few most similar cases from a large temporal database, which in turn helps us to avoid the problems of case adaptation and case authoring. Case Base Reasoning expert systems require very large historical weather observation records, these records must be noise-free and consistent and well organized. On the other hand local weather condition may change by the passing years because of other environmental effects. For example a newly constructed damn, or a destructed forest may change local effects; so yearly rainfall amount will change depending on these incidents. Now historical data will be misleading anymore. That is why there must be a forgetting mechanism for historical meteorological data. Last 30-40 years of meteorological data observation will be enough to produce satisfactory results [2]. Forgetting mechanisms [18] are used in many case based systems to help control the amount of memory used by the case library. Forgetting can be achieved by the use of 5 recency in the similarity algorithm and weighting each passing years decreasingly or removing older cases up to a point without adversely affecting the accuracy of the forecasts produced by the system. During the problem-solving process, a number of similar solutions are retrieved not one of which is completely valid. The retrieved solutions need to be adapted to arrive at a valid solution to the current problem. Adaptation mechanisms are often user-driven and can be rule-based or rely on generalization and refinement heuristics [19]. Adaptation mechanisms have been categorized as compositional or transformational [20]. Compositional adaptation occurs when knowledge is used to combine portions of approximate solutions to achieve the desired solution. Transformational adaptation occurs when more comprehensive knowledge needs to be used to add extra components to the retrieved solutions to achieve a satisfactory result. Hansen[2] has built an airport TAF Expert System by using fuzzy K Nearest Neighbor Case Based Approach, he performed fuzzy logic by setting attribute values to different sets with a different membership rates. Domain experts can easily change these weights of attribute values in fuzzy set memberships. He does not open his detailed work but he has broadcasted experimental results, he claims that his similarity measure approach has given the best results in comparison to other approaches. The quality of TAFs is determined by the accuracy of the forecast weather elements and the timeliness of issue. Hansen thought that combination of the attributes is so crucial to obtain high accuracy in short-term weather prediction. Even though, he tested varying attributes set in Figure-2. Hansen’s [2] test helps selecting related attributes pairs and combinations. We can find related and dependent attributes by different test over attributes pairs and combinations. Figure 2 6 Combinations of weather attributes considered are; • cig & vis – cloud ceiling height and visibility alone • pres + cig & vis – mean sea level pressure and cloud ceiling height and visibility • pcpn + cig & vis – precipitation type and cloud ceiling height and visibility • temps + cig & vis – dry bulb temperature and dew point temperature and cloud ceiling height and visibility • time + cig & vis – time of day and time of year and cloud ceiling height and visibility • wind + cig & vis – wind speed and direction and cloud ceiling height and visibility • All of the aforementioned attributes. He also tested effect of varying case base in Figure-3. As the history goes back there is a small increase for reliability but after 4 year increase slows down. This kind of test can help us to decide our database size. At the first year reliability start from %90 percents succeeded. So we should not much be concerned about our small database size. Figure 3 He tested varying k, the number of nearest neighbors that are used as the basis of predictions (K = 1,2, 4, 8, ... , 256) and finds that maximum accuracy is achieved with K = 16 (For that reason we got K=16 to achieve maximum reliability in our FA-1 system). Assuming that fuzzy KNN similarity metric is effective, if k is too small, prediction result accuracy should suffer from sample size being too small (i.e., not representative), and if k is too large, prediction result accuracy should taper off because of the inclusion of an increasing number of decreasingly similar cases. Accordingly, Figure-4 shows that the fuzzy KNN similarity metric is effective. 7 Figure 4 Hansen[2] start using fuzzy logic in large scale from beginning to end of implementation. He gets rid of continuous attributes by converting them into fuzzy membership set with different weights. Solution sets members are also categorized and classified. For example we have retrieved 10 nearest neighbors. We need to find visibility attribute of current case from 10 visibility attribute values of nearest neighbors. Neighbor 1 Neighbor 2 Neighbor 3 Neighbor 4 Neighbor 5 Neighbor 6 Neighbor 7 Neighbor 8 Neighbor 9 Neighbor 10 Table 1 Neighbor similarity based on Neighbor other attributes Visibilities slightly near Visibility=30 near Visibility=26 very near Visibility=24 slightly near Visibility=31 slightly near Visibility=14 near Visibility=29 very near Visibility=23 very near Visibility=22 slightly near Visibility=10 very near Visibility=25 Weight .25 .50 1 .25 .25 .50 1 1 .25 1 6 7.5 13 24 7.6 3.5 14.5 23 22 2.5 25 140/6~=23 Table 1 show a simple visibility calculation on 10 retrieved neighbors and we found 23 km which is neighbor 7. 8 An example attribute values similarity table is given for precipitation type on Figure-5. Figure 5 2.2 Weather Prediction Using Artificial Neural Network For an expert system implementation we surveyed a paper of Manan[23]. This project implemented an expert system that is capable to learn the pattern of rainfalls in order to produce a precise forecasting result. He used the multiplayer perceptron with error back-propagation algorithm to provide prediction of rainfall. The ANN technique is applied to the hourly local weather data for a specific area. His main motivation for selecting ANN algorithm comes from the analogy that climate and rainfall are highly non-linear phenomena and the multilayer perceptron neural network has been designed to function well in such non-linear phenomena. A feed-forward multilayer neural network consists of an input layer and output layer. There are also one or more hidden layers in between the input and the output layer with some number of neurons on each. 9 distinct weather features are selected time of the day, pressure, dry bulb temperature, wet bulb temperature, dew point, wind speed, wind direction, cloud cover and whether it rained in the past hour. Since our target is to forecast rainfall, initial condition is the rest of features. They are fed to each input neurons while rainfall attribute is fed to output neuron. The network learns the relationship between the input-output data in the training set via network training in which the weights are modified by presenting input-output patterns of the training set until a prescribed error criterion is fulfilled. This simple implementation experiments can give us practical results. Back propagation algorithm adjusts weights of links between neurons in order to decrease some of squared errors between expected outputs and observed output values until a specified termination condition is met. This termination condition can be an error threshold or a number of loop count (epoch), Learning rate and momentum constants are used when tuning weights in a step logic. All of these parameters and constants should be customized to produce best result from constructed neural network. 9 Actually this implementation seems very easy with respect to our FA-1 expert system. We need to forecast all multiple features at the same time. For a selected hourly weather observation, relevant parameters are fed to input neurons while next hour parameters are fed to output neurons. Topology of neural (number of hidden units and layers) network may seriously affect performance. In Figure-6 results of learning and testing in different architecture networks and parameters settings Manan [23]. This experiment enables us to make a generalization about ANN performance for weather prediction. These different network configurations produce approximately the same result. We can conclude that ANN is successful at most %80, on the other hand there may be some reasons affecting performance, maybe training data was not enough to represent all behavior of weather phenomena. Figure 6 2.3 Hybrid Weather Prediction For every special purpose we are forced to implement different projects and to utilize different methods. In this way it is impossible to produce this specific implementation for everywhere. Weather prediction systems must be integrated. Nowadays, hybrid systems swiftly spread all around the world. They use multiple methods and sensors to feed each other. In this way even our simple weather forecast implementations will be able to receive input readings from satellites to weather doppler radars or even other expert system forecasts will guide our expert systems. This newly developing approach will certainly increase weather forecast accuracy and timeliness Up to now we mentioned about empirical methods based on occurrence of analog or similar weather situations, whereas dynamical methods can guide our empirical methods and vice versa. In practice, hybrid methods use dynamic models and empirical observations. In addition to using statistical methods that infer estimated expected distributions under specified conditions. Hybrid systems cannot go further than being a decision support system because problem area is inherently complicated and not deterministic. 10 • • • • Here are the some of the Hybrid Forecast Decision Support Systems; Data fusion: intelligent integration of output of various models, observational data, and forecaster input using fuzzy logic [22] Data mining: C5.0 pattern recognition software for generating decision trees based on data mining, freeware by Ross Quinlan (http://www.rulequest.com), like CART Incorporate AutoNowcast of weather radar in 2004-2005 (http://www.rap.ucar.edu/projects/nowcast) Incorporate satellite image cloud-type classification algorithms [23] Hybrid systems designs must obey some standardization rules for easy integration; • Generic: no-name, conceptual design that could link and integrate the most useful elements of WIND, AVISA, MultiAlert, SCRIBE, FPA, URP, and so on in evolving WSP application. • Modular: shows where distinct sub-tools / agents can be developed. Working in this way, individual developers could work on isolated sub-problems and anticipate how to plug their results into a larger shared system. As technology inevitably improves, improved modules can be easily installed and quickly implemented. • User-centered: forecast decision support systems from forecaster's point of view, designed to increase situational awareness. • Hybrid: combines complementary sources of knowledge, forecasters and AI, to increase the quality of input data and output information. Intelligent integration of data, information, and model output, and use of adaptive forecasting strategies are intrinsic in this design. 3. COMPARISON of MENTIONED APPROACHES Case Base Reasoning systems can easily use fuzzy logic while other methods such as ANN and Decision Trees etc. it is hard to manage. Case Base Reasoning experiment results outperform ANN experiment results. Thus we can conclude that Case Base Reasoning is inherently associated with weather phenomena. But none of the advantages are free for Case Base Reasoning. As we said Case Base Reasoning uses stored past data, and retrieve the solution from this database with respect to present situation. It has to find all the Euclidian distance for every instance in the database. It needs much more computation power and large memory. From ANN point of view, the system trained at once, and the weight values are stored in the simple database table. When the new current instance comes, ANN easily computes the output values. ANN can be used to simple predictions like whether rain is expected and clear weather expected but not multiple parameter result. There are many other test results for decision trees and rule based productions system and statistical forecast systems. Our implementation and handling errors made these approaches less successful. This methods are hard to model and construct that is 11 why they are not much preferred when such a successful Case Base Reasoning approach exist. CONCLUSION Weather prediction is inherently complex process, so it impossible to wait %100 accurate forecast results since we cannot measure all factors that may be local and global scales. Weather Prediction systems are more likely decision support system than expert systems because they need guidance and weather predictions must be evaluated by human interference. Hybrid systems are very promising for integration of current expert systems on large scale. On Internet, there are many web weather report and forecast service systems that enable our implementations to fetch some weather information. Since there are many weather related data needs to be processed, hybrid system can reduce intensive forecast computations. In stead of computing some detailed information, we can easily pick up knowledge about a satellite post-processing reports and comments Fuzzy k nearest neighbor in Case Base Reasoning is the most promising approach with successful test results. All reviewed test results are limited to past meteorological data that will represent all possible weather condition distributions. Data collecting and acquisition are initial and one of the most critical parts of expert systems computations. Automated data collecting systems must be available rather than using human manual inputs. Local data collection is very important if we want to make very precise forecasting. Even though these fast developments, meteorologists go without artificial intelligence benefits, they still use classical methods. There will be many fields that use weather prediction for specific purposes. So we are always bound to meteorological reports and predictions. Abbreviations: AI : Artificial Intelligence CBR : case-based reasoning DBMS : database management systems DSS : decision support system FOH : Frequency of Hits ES : Expert Systems FA-1 : Flight Alert Systems [Bulent KISKAC , Harun YARDIMCI] FAR : False Alarm Ratio KE : Knowledge Engineering KNN : K nearest neighbor MOS : model output statistics (NWP + climatology + statistics) NWP : Numeric Weather Prediction PC : Persistence Forecast 12 POD TAF VFR : Probability of Detection : Terminal Aerodrome Forecast : Visual Flight Rules References 1. Lorenz, E. N. (1969a) Three approaches to atmospheric predictability, Bulletin of the American Meteorological Society, 50, 345–349. 2. Eng Int Syst (2002) 3: 139–146 © 2002 CRL Publishing Ltd “A fuzzy case-based system for weather prediction” Denis Riordan and Bjarne K Hansen Faculty of Computer Science, Dalhousie University, Scotia, Canada, 3. R. E. Huschke (editor), Glossary of Meteorology, American Meteorological Society, Boston, Massachusetts, USA, pp 106, 419, 1959. 4. B. J. Conway, Expert systems and weather forecasting, Meteorological Magazine, 118, pp23–30, 1989. 5. T. Hall, Precipitation forecasting using a neural network, Weather and Forecasting, Boston, 14, Iss. 3; 338–346, Jun 1999. 6. Marzban and G. J. Stumpf, A neural network for tornado prediction based on Doppler radar-derived attributes. J. Appl. Meteor., 35, pp617–626, 1996. 7. C. Marzban, A Bayesian neural network for severe-hail size prediction, Weather and Forecasting, Boston; Vol. 16, Iss. 5, pp600–611, Oct 2001. 8. R. L. Vislocky and J. M. Fritsch, An automated, observations based system for short-term prediction of ceiling and visibility, Weather Forecasting, 12, pp31–43, 1997. 9. L. J. Wilson and R. Sarrazin, A classical REEP short-range forecast procedure, Weather and Forecasting, 4, pp502–516, 1989. 10. S. S. Allan, S. G. Gaddy and J. E. Evans, Delay Causality and Reduction at the New York City Airports Using Terminal Weather Information Systems, Lincoln Laboratory, Massachusetts Institute of Technology, Lexington, Mass. 11. R. Bergman, S. Breen, M. Göker, M. Manago & S. Weiss, Developing Industrial Case-Based Reasoning Applications The INRECA-Methodology, Springer: Berlin, 1998. 12. Leake, D. B., CBR in context. The present and future; in Leake, D. B. (ed.), CaseBased Reasoning: Experiences, Lessons & Future Directions, American Association for Artificial 13 13. M. Sebag and M. Schoenauer, A rule-based similarity measure, In Proceedings of the First European Workshop(EWCBR-93), pp119–131, Nov. 1-5, Kaiserslautern, Germany,1993. 14. J. Surma and K. Vanhoof, Integrating rules and cases for the classification task, In Proceedings of the First International Conference (ICCBR-95), pp325–334, Oct.26-30, Sesimbra, Portugal, 1995. 15. D. Dubois, F. Esteva, P. Garcia, L. Godo, R. López de Mántaras & H. Prade Case-Based Reasoning: A Fuzzy Approach, Lecture Notes in Computer Science, Springer, vol. 1566, pp79–91, 1999. 16. T. W. Liao, Z. Zhang and C. R. Mount, Similarity measures for retrieval in casebased reasoning systems, Applied Artificial Intelligence, 12, pp267–288,1998. 17. J. M. Keller, M. R. Gray and J. A. Givens, JR, A Fuzzy Knearest Neighbour Algorithm, IEEE Trans. Syst., Man, Cybern., vol. SMC-15, 4, pp580–585, 1985. 18. R. López de Mántaras and E. Plaza, Case-Based Reasoning: an overview, AI Communications, 10, pp. 21–29, 1997. 19. D. B. Leake, CBR in Context: The Present and The Future,Case-Based Reasoning: Experiences, Lessons, and Future Directions, D. B. Leake, ed., Menlo Park, Calif., AAAI Press, 1996. 20. W. Wilke, B. Smythe & Pádraig Cunningham, Using Configuration Techniques for Aaptation, Case-Based ReasoningTechnology From Foundations to Applications, Springer: Berlin, 1998. 21. Abdul Manan Ahmad1 , Chia Su Chuan2 and Fatimah Mohamad Department of Software Engineering, Faculty of Computer Science & Information System Universiti Teknologi Malaysia, 81310 Skudai, Johore, Malaysia 22. Intelligent Weather Systems, RAP, NCAR, http://www.rap.ucar.edu/technology/iws Shel Gerding and William Myers, 2003: Adaptive data fusion of meteorological forecast modules, 3rd Conference on Artificial Intelligence Applications to Environmental Science, AMS. 23. Tag, Paul M., Bankert, Richard L., Brody, L. Robin. 2000: An AVHRR Multiple Cloud- Type Classification Package. Journal of Applied Meteorology: Vol. 39, No. 2, pp. 125-134. 14