Local Maximum Ozone Concentration Prediction Using Neural Networks Dominik Wieland and Franz Wotawa Technische Universität Wien, Institut für Informationssysteme and Ludwig Wittgenstein Laboratory for Information Systems, Paniglgasse 16, A-1040 Wien, Email: wotawa@dbai.tuwien.ac.at From: AAAI Technical Report WS-99-07. Compilation copyright © 1999, AAAI (www.aaai.org). All rights reserved. Abstract This paper describes the use of Artificial Neural Networks (ANNs) for the short term prediction of maximum ozone concentrations in the East Austrian region. Various Multilayer Perceptron topologies (MLPs), Elman Networks (EN) and Modified Elman Networks (MEN) were tested. The individual models used ozone, temperature, cloud cover and wind data taken from the summer months from 1995 and 1996. The achieved results were satisfactory. Comparisons with alternative models showed that the neural approaches used in this study were superior. Introduction manifests itself in its effect on organisms as a poiOzone sonous gas. It results in irritation of the respiratory system and shows effects on the health especially of children. Empirically, the highest concentrations are found downwinds of conurbations such as the Viennese basin and mostly in the summer season. A reason for this is the emission of precessor substances and the available time for producing ozone. In order to provide adequate early warnings, it is important to have accurate and reliable forecasts of future high ozone levels. The ability to forecast ozone trends or exact values is not so important for the goverment than forecasting that the ozone value will reach a dangerous level where driving cars is prohibited and plants emitting to much precessor substances have to be shut down. In practise, there are many different models for ozone forecasting. Many of these use statistical approaches, such as correlation and regression analyses. These models are mostly simple and the accuracy of their results is not to be underestimated, however, peak ozone levels are not accurately predictable. As already suggested, it is just these peak levels which are most interesting. Therefore, we tested neural models on their ability for predicting the maximum ozone concentrations. Advantages and benefits of Artificial Neural Networks (ANNs) published so far include: Fault Tolerance: ANNs can use incomplete and corrupted data. The malfunction of part of the system causes no sudden failure of the whole system. This work was supported by the Austrian Science Fund Project N Z29-INF Parallelism: This feature encompasses not only the aforementioned fault tolerance but, also allows, through effective hardware implementation, a much quicker calculation of the system output. Adaptivity: ANNs are often capable of self organisation. This means that certain free system parameters do not have to be adjusted experimentally but are often set by the system itself. Non-linearity: The immanent non-linearity of most ANNs allows the calculation of complex and indisputable correlations. User friendly: ANNs are more user friendly than other models with similar capacities. Due to these characteristics of ANNs, it would appear that their ability to predict maximum ozone concentration is very promising. And in fact Acuna et al's paper (Acuna, Jorquera, & Perez 1996) entitled “Neural Network Model for Maximum Ozone Concentration Prediction” introduces the application of neural networks to ozone forecasting. They use a Multilayer Perceptron (MLP) model for predicting the maximum ozone concentrations in Santiago de Chile. Various MLP topologies were tested and according to the authors, provided satisfactory results. The Santiago model provided a good comparison for this work, the results of which compared favourably with those contained herein. In difference to (Acuna, Jorquera, & Perez 1996) we add new models, e.g., Elman Networks, and compare our results with a physical/chemical model that had been developed for the East Austrian Region around Vienna (Stohl, Wotawa, & Kromp-Kolb 1996). The work described in this paper is the first part of a general research program applying Artificial Intelligence concepts and techniques on issues of environmental research, using ozone forecasting as initial application area. In the next step we want to combine neural networks and qualitative reasoning as described in (Catala, Moreno, & Parra 1998) for the prediction of ozone values. In this context we are interested to use neural networks for directly predicting the (qualitative) severity levels of ozone concentration in an area instead of predicting the exact values (as done in this paper) and mapping it to the levels afterwards. As an advantage learning and forecasting should be speeded up. Another direction of research is the use of qualitative phys- ical/chemical models instead of using ordinary differential Finally, it is planned to compare the outcomes of equations. all models helping to find the appropriate AI technique in other domains. Basics of Neural Networks To be self contained we briefly recall the basics of neural networks. Neural networks can be seen in the sense of an abstract simulation of a biological neural system, such as the human brain. They are made up of many parallel working units, based upon single brain cells or neurons and the connections between them which simulate the axon interlaces of the brain. The most important part of an ANN is the neuron. Data is processed in the neurons insofar as they accept incoming data, transform it into an activation signal and switch it to the neuron output. Individual neurons are combined in a single network through directed and weighted connections. Such networks take on external input through their own input neurons and propagate these through the neuron connections to the output neurons. ANNs allow themselves to be placed in models with supervised and unsupervised training algorithms, whereby the data predictions mostly come into effect as supervised approaches. Some of the most important approaches are: Single and Multi-layered Feedforward Networks (such as MLPs and RBFNs, see (Haykin 1999)) Recurrent Neural Networks (RNNs) (e.g. Jordan and Elman networks, see (Pham & Liu 1995) and (Pham & Liu 1993)) Stochastic models for time series prediction (Markov Models, Hidden Markov Models, see (Kung 1993)) Unsupervised models (Art networks, self organizing feature maps, see (Cotrell, Girard, & Rouset 1997)) Other approaches e.g. alternative neuron models (see (Burg & Tschichold-Gürman 1997)) Evolutionary algorithms which can be used for training as well as for determining the topology of ANNs (see (Fang & Xi 1997)) Multilayer Perceptrons (MLPs) MLPs come about through the joining together of multiple non-linear perceptrons (see (Haykin 1999)) and are multilayered feedforward networks. Figure 1 shows the formal representation of a single neuron used in MLPs, consisting of an input, an activation, and an output function. Usually the input function computes the sum of all inputs using the given weights, i.e., where denotes the -th neuron in the -th layer, the weight between the neuron and ! , and " the number of neurons in layer $#&% . In most cases identity is used as the output function. Therefore this function is often ignored (as Artificial Neuron Input Activation Output j-1 y1 w1 i j xi j-1 yN j-1 wN j-1 j ai act j 1 yi i Figure 1: A single Neuron Output Layer Hidden Layer Input Layer Figure 2: Topology of an MLP it is in our case). The activation function ')(* takes the input value and computes the output value. The most popular activation function is the sigmoid function + % 2 %-,/.10 where denotes the output value of the -th neuron in the -th layer. Figure 2 shows the topology of a classic MLP. MLPs are normally trained through the Backpropogation algorithm by modifying the weights between the neurons. The Backpropagation Algorithm (BP) The BP algorithm tries to minimize the output error function of a network by adapting the weights of the network connections in the direction of its negative gradient. The error function is therefore half the square of the network output error compared to the desired target output: 3 6 4% 5 87 * # 9 :<; where index denotes the cells 9 of the output layer, * represents the target output, and the actual output of the ANN. The change in the weights runs parallel to the negative gradient of the error function: = 3 # ? +> @ ? The direct result is the back propagation learning rule distinguishing = BA)between D FAC layers: C 9 , =Edifferent with C '(*HG 7 : 7 * # : 1. Output layer 1 1 1 from the weather prediction model of the European Center for Medium Range Weather Forecasts (ECMWF 1995). Ozone and weather data were available for the periods 7.7.95 to 25.9.1995 and 1.5.1996 to 30.9.1996. In order to quantify the prediction ability of a certain forecast model, we use the Root Mean Square (RMS) error, defined as follows: Context Units Figure 3: Topology of the EN ')(*HG :JI LK C N M N M 7 where A is the learning rate or learning coefficient and regulates the speed of the convergence of the algorithm, C is also known as the error signal of a particular cell, and ')(* 7 : is the activation function of the respective neuron. ')(*HG denotes the derivation of ')(* . 2. Others C Elman Networks (ENs) ENs (Pham & Liu 1995) belong to the class of partially recursive networks. They can be seen as an extension of the MLPs whereby for each neuron of the hidden layer a state neuron is added to the input layer. At each stage the contents of the hidden neurons are copied to the state layer through fixed feedback links and fed back to the hidden layer in the next stage. In this way former information of time series is implicitly kept in the network and thus used to calculate the network output. Figure 3 shows the topology of the classic EN. The model series M01 used classic MLPs. The maximum ozone value for a day * was predicted for a day using the ozone levels for the day before and the temperature forecasts from a meteorological model for the same day * . Experiments were made with hidden layers of different sizes (3 to 10 hidden neurons), varying learning rates (1.0 to 0.2) and bias neurons. The first model M01a has 2 input, 5 hidden, and 1 output neuron. The MEN (see (Pham & Liu 1995)) differentiates itself from the classic EN by connecting state neurons with themselves. In this way each state neuron gets a certain inertia, which increases the capabilities of the network to dynamically memorize data. The following formula shows the output value of the jth neurons of the state layer. According to the authors, MENs are superior to the classic ENs in non linear time series problems. Their use is effective in solving complex problems such as the prediction of local maximum ozone concentrations. Ozone Prediction Models Time series of ozone measurement in the East Austrian region (as the average of five measurement points providing 3 hours fixed average values) was available for training and measuretesting of the developed ANNs. In addition to ments, model analyses and forecasts of temperature, cloud cover, and wind speed of the last, current, and the next 2 days were available. The meteorological forecasts originated where Z represents the actual measured and \ the predicted ozone values at the time respectively, and " denotes the number of observations. For the training of the developed ANNs, the raw data was divided into a training and a test set. In this way all data from 1996 were used for the training of the ANN and the data from 1995 was used for testing the prediction ability of each model. A validation set was not used due to lack of data. The optimum timing for the completion of the training process was determined experimentally. All models were trained with the BP algorithm. A data scale was used to feed the input pattern into the ANN. In this way the values of the individual measurements were limited by the minimum and maximum values of the respective measurement series and then transformed into the interval [0.1]. As part of the work described in this paper, MLPs and ENs were implemented and tested for their ability to perform predictions of ozone concentrations in the East Austrian region. The following model series were developed: The Modified Elman Network (MEN) PO Q* ,B% : FR PO * : , * : 7 7 7 : where 7 * represents the output of the j-th hidden neuron, and R denotes the neuron' s inertia. SUTWV YX I [7 Z #]\ : ; " The models in the series M02 were an extension of model M01. With the additional inputs of the forecasts for cloud cover and wind speed similar experiments as for series M01 were carried out. The models vary in the hidden layers, the lerning rates and the use of bias neurons. The first model M02a has 4 input, 5 hidden, and 1 output neuron. The models in the series M03 and M04 were dynamic models. Multiple measurements of one time series were presented to the MLP in parallel after a relevant precoding. The series M03 and M04 used data of only one time series (Ozone or temperature time series) and tried to find the optimum number of past values which were relevant to the target values. The models of series M03 used the ozone values from the past 4 days as inputs for forecasting the today' s ozone level. For the models M04 only the temperatures of the past 4 days were used. As model M04 scored surprisingly good results, further experiments with various hidden layers, learning rates and bias neurons were conducted. Model M01 M02 M03 M04 M05 M06 M07 M08 Version M01k M02f M03b M04e M05g M06f M07a M08c NN used MFFN MFFN MFFN MFFN MFFN EN MEN ( R =0,2) MEN ( R =0,4) #HU 5+1 5 2 2+1 5+1 8+1 5+1 5+1 LR 1 0,4 1 0,2 0,2 0,2 0,2 0,2 Bias Y N N Y Y Y Y Y Steps 1,000 100 3,000 1,000 100 5,000 3,000 5,000 RMS Trg. 9.4942 8.3786 10.6892 10.4792 8.2406 9.0740 8.9964 7.3741 RMS Test 11.2004 11.1768 15.0537 10.6651 10.8132 10.3186 10.5150 9.9579 Table 1: Comparison of the RMS errors of various tested models Model M08c - Training (1996) Ozone Concentration The models in the M05 series were bivariate time series models. Therefore they were a combination of models M03 and M04. The exact topology of the individual networks were originally developed from the Santiago models. Further experiments were carried out using wind and cloud cover data. The approaches of series M06, M07 and M08 used partially recursive ANNs: While M06 was testing classic ENs using bias neurons and hidden layers of varying sizes, in series M07 the MEN was used. Experiments with varying inertiae were carried out. Lastly wind and cloud cover data were again included, leading to model M08. The base model M06a has 2 input, 4 hidden, 1 output neurons and 4 context units. The ozone value from yesterday and the predicted temperature value for today is used as an input to the net. Model Ozon Day (a) Training Set Ozone Concentration Model M08c - Test (1995) More informations about the used models can be found in the appendix. Model Ozon Results and Discussion Table 1 shows the RMS error of the respective best models of all test series during the training and abstraction phases. #HU stands for the number of hidden units, LR for the learning rate, RMS Trg. and RMS test for the RMS error of the training and test sets respectively. The Bias column indicates whether bias neurons had been used or not, while #Steps give the necessary number of learning steps for the neural network in order to obtain the best result. The RMS error for all model series during the abstraction phase are presented in the appendix. The best results were scored by the model M08c. Figure 4 show both the actual ozone level variations and those predicted by model M08c during the learning phase (Summer 1996) and the abstraction phase (Summer 1995). In general, the results of the individual model series were stable and uniform. As a result of these facts, on the one hand, accurate forecast results were pretty much guaranteed. On the other hand, the limitations of the approach used in this work were to be seen. The results obtained by the different models are given in the appendix. A further step has been taken to compare the results shown herein with those of other models. These models are: The Persistence Model (PM): In this model, yesterday' s levels are used to predict today' s forecast. Day (b) Test Set Figure 4: Results of Neural Network Model08c Note: Ozone concentration is given in ppb (parts per billion). IMPO Model: This model uses a chemical/physical approach developed by the Institute for Meteorology and Physics of the Universität für Bodenkultur Wien (BOKU, University for Agricultural Science, Vienna) (see (Stohl, Wotawa, & Kromp-Kolb 1996)). Figure 5 shows the results of the IMPO model within the considered time period. A statistical model (Loibl 1996), which predicts today' s ozone maximum value using the regression function Z_^ B`ba cPced ,gf a h f h 5 ,F% a i9j l5 k Z_^ ^ where Z ^ denotes the predicted ozone value for day * , Z ^ denotes the ozone value from the day before, and k ^ is the temperature from day * . Model PM IMPO Stat. M. S3 S3 Version 1995 1995 1995 G1 G2 NN used Persistence physic./chem. Regression Fkt. MFFN MFFN #HU — — — 5+1 5+1 LR — — — — — Bias — — — Y Y #Steps — — — — — RMS Trg. 11.2245 12.5739 10.5856 — — RMS Test 14.4540 14.8043 12.0291 13.7000 15.4000 Table 2: RMS errors of various comparison models Note: PM, IMPO und Stat.M. stands for the persistence model, the IMPO model and the statistical model respectively. S3 denotes the best Santiago model (see (Acuna, Jorquera, & Perez 1996)), which was tested on the sets G1 und G2. The 2 rightmost columns show the RMS errors of the training and the test sets respectively. The Santiago Models (Acuna, Jorquera, & Perez 1996))1. In comparison to these reference models, the network architecture tested herein showed a more satisfactory performance. A comparison of the tables 1 and 2 shows that almost all the results of this project displayed a higher level of accuracy than those of the IMPO model, the PM and the statistical model. All models, with the exception of model M04, scored appreciably lower RMS errors during the abstraction phase than the two Santiago models described in (Acuna, Jorquera, & Perez 1996). In the comparison with the Santiago models, it must be noted, however, that the ozone level variations in Santiago were greater than those in the East Austrian test region. For most models it is true to say that the trends, as shown in Figure 4, of the short term ozone development could be accurately predicted. However, the tested models had problems in correctly forecasting extreme ozone peaks. The reason for this is that in all predicted measurement series the individual value variations were less extreme than in the actual measured ozone levels. This can be seen in the highly differential variances and standard deviations of the measured and the forecast time series. General Findings Through the experiments with the individual ANNs general findings on the topology and parameter settings of the ANN were gained. Amongst these are: In most models a hidden layer with five neurons was seen to be optimal. It was observed that networks with larger hidden layers could better reproduce once learned data. However, this has to be seen in relation to the more complex architecture and longer training periods of such networks. During the abstraction phase, networks with five hidden cells showed as a rule satisfactory results. Using lower learning rates (0.2 - 0.4) the forecasting ability of the relevant ANNs could be increased. The introduction of bias neurons in the input and hidden layers had a positive effect on the performance of the ANNs. In the MLP approaches lower RMS errors were observed. In EN and MEN models bias neurons had a stabilising effect on the prediction curves. 1 We compare the published results of the Santiago models with ours. A further positive effect on the ability of the ANNs to correctly predict local ozone values was brought about through the introduction of cloud cover and wind speed data. The best models of each series mostly used these supplementary values as was seen in models M02f and M08c. The ultimately tested RNNs appeared to have been a good choice in the forecast of ozone levels. The scored results were appreciably better than those achieved by MLFF networks. The altogether best result (model M08c) was achieved by a MEN with an inertia of 0.4. In general the optimum settings for the inertia parameter were between 0.2 and 0.4 and therefore surprisingly low. Large R -values (approaching 1) seemed to be inappropriate. Comparing the outcome of the IMPO model (see figure 5) with model M08c (figure 4) leads to the following conclusions. While the IMPO model tends to overestimate the maximum ozone concentration (in the extrem case with a factor nearby 2) this is not the case for the ANN. Both models follow the ozone curve although the predicted values are not equal to the real values. Conclusion In the scope of this work various ANN models for the short term forecast of local ozone maxima were developed. Extensive tests with varying topologies and parameter settings led to the results shown in table 1. The best result was achieved by model M08c which used a MEN and the standard BP algorithm. The RMS error during the abstraction phase was 9.958 (see also figure 4). The results were all in all satisfactory since the RMS errors of the best models of the respective series lay under the results of the comparison models (see table 2). It was possible to develop a forecast system for the test data from 1995 which, compared to the PM and IMPO models, showed a clear improvement and also seems to be superior to common statistical approaches. For more informations about the used networks, their results, and a discussion about possible improvements see (Wieland 1999). Future research include the use of qualitative neural networks and qualitative reasoning techniques for ozone prediction. This allows to compare several different approaches applied to the same problem and may help to select the appropriate techniques in other forecasting domains. Ozone Concentration IMPO Model (1996) Fang, J., and Xi, Y. 1997. Neural Network Design Based on Evolutionary Programming. Artificial Intelligence in Engineering 11:155–161. Model Ozon Haykin, S. 1999. Neural Networks – A Comprehensive Foundation. Prentice Hall. Kung, S. Y. 1993. Digital Neural Networks. PTR Prentice Hall. Loibl, W. 1996. Trendprognose regionaler Ozonmaxima unter Einbezug verschiedener meteorologischer Daten. Technical report, Report Nr. UBA-BE-058, Umweltbundesamt, Vienna, Austria. Day (a) Ozone Concentration IMPO Model (1995) Model Ozon Pham, D., and Liu, X. 1993. Identification of Linear and Nonlinear Dynamic Systems using Recurrent Neural Networks. Artificial Intelligence in Engineering 8:67–75. Pham, D. T., and Liu, X. 1995. Neural Networks for Identification, Prediction and Control. Springer Verlag. Day (b) Figure 5: Results of the IMPO Model Note: Ozone concentration is given in ppb (parts per billion). Acknowledgement The authors wish to thank Gerhard Wotawa from the Department of Meteorology and Physics, Universität für Bodenkultur, Vienna, Austria for his support and for his comments on earlier drafts of this paper. Ozone data were contributed by the Austrian Environmental Protection Agency (Umweltbundesamt, UBA) and by the government of Lower Austria. References Acuna, G.; Jorquera, H.; and Perez, R. 1996. Neural Network Model for Maximum Ozone Concentration Prediction. In Proceedings of the International Conference on Artificial Neural Networks ICANN-96, 263–268. Burg, T., and Tschichold-Gürman, N. 1997. An Extended Neuron Model for Efficient Time Series Generation and Prediction. In Proceedings of the International Conference on Artificial Neural Networks ICANN-97, 1005–1010. Catala, A.; Moreno, J. M.; and Parra, X. 1998. Neural Qualitative Systems. In Proceedings of the Workshop (W5) on Model-based Systems and Qualitative Reasoning of the 13th European Conference on Artificial Intelligence ECAI98, 12–20. Cotrell, M.; Girard, B.; and Rouset, P. 1997. Long Term forecasting by Combining Kohonen Algorithm and Standard Prevision. In Proceedings of the International Conference on Artificial Neural Networks ICANN-97, 993–998. European Centre for Medium Range Weather Forecasts (ECMWF), Reading, UK. 1995. User guide to ECMWF products Version 2.1. Stohl, A.; Wotawa, G.; and Kromp-Kolb, H. 1996. The IMPO modeling system description, sensitivity studies and applications. Technical report, Universität für Bodenkultur, Institut für Meteorologie und Physik, Türkenschanzstras̈e 18, A-1180 Wien. Wieland, D. 1999. Prognose lokaler Ozonmaxima unter Verwendung neuronaler Netze. Master's thesis, Technische Universität Wien, Vienna, Austria. Only available in German. Appendix A – Used Neuron Network Models In this section the specification of the considered network models are given. For all models the inputs, the number of hidden units #HU, the learning rate LR, and the use of bias neurons Bias, and the inertia Inertia for state neurons in modified Elman Networks are given. The ozone, temperature, cloud cover, and wind can be used as input. We 4o speed use an index for +mgn# #p% o f o % oq4)r to indicate whether the value is given for today (2), yesterday (1), the day before yesterday (0), another day before (-1), and so on (-2). All models predict the ozone value for today (Ozon ; ). Model M01a M01b M01c M01d M01e M01f M01g M01h M01i M01j M01k M01l M01m Inputs Ozone Ozone Ozone Ozone Ozone Ozone Ozone Ozone Ozone Ozone Ozone Ozone Ozone , Temp ; , Temp ; , Temp ; , Temp ; , Temp ; , Temp ; , Temp ; , Temp ; , Temp ; , Temp ; , Temp ; , Temp ; , Temp ; #HU 5 3 4 6 8 10 5 5 5 5 5+1 5+1 5+1 LR 1.0 1.0 1.0 1.0 1.0 1.0 0.8 0.6 0.4 0.2 1.0 0.4 0.2 Bias N N N N N N N N N N Y Y Y sModel M02a M02b M02c M02d M02e M02f M02g M02h M02i M02j Model M03a M03b M03c M03d Model M04a M04b M04c M04d M04e M04f M04g M04h Model M05a M05b M05c M05d M05e M05f M05g M05h M05i Inputs Ozone , Cloud ; Wind ; , Temp ; Ozone , Cloud ; Wind ; , Temp ; Ozone , Cloud ; Wind ; , Temp ; Ozone , Cloud ; Wind ; , Temp ; Ozone , Cloud ; Wind ; , Temp ; Ozone , Cloud ; Wind ; , Temp ; Ozone , Cloud ; Wind ; , Temp ; Ozone , Cloud ; Wind ; , Temp ; Ozone , Cloud ; Wind ; , Temp ; Ozone , Cloud ; Wind ; , Temp ; Inputs Ozone Ozonet Ozone Ozone Ozone Ozonet , #HU 5 LR 1.0 Bias N , 6 1.0 N , 8 1.0 N , 10 1.0 N , 4 1.0 N , 5 0.4 N , 5 0.2 N , 5+1 1.0 Y , 5+1 0.4 Y , 5+1 0.2 Y #HU 1 2 3 LR 1.0 1.0 1.0 Bias N N N 4 1.0 N ,Ozone , Ozonet , ; , Ozone , , Ozone Model M06a M06b M06c M06d M06e M06f Model M07a M07b M07c M07d M07e M07f Model M08a M08b M08c Inputs Temp ; Temp , Temp ; Temp t , Temp , Temp ; Temp , Tempt , Temp , Temp ; Temp ,Temp ; Temp ,Temp ; Temp ,Temp ; Temp ,Temp ; #HU 1 2 3 LR 1.0 1.0 1.0 Bias N N N 4 1.0 N 2+1 3+1 4+1 5+1 0.2 0.2 0.2 0.2 Y Y Y Y Inputs Ozone ,Temp ; Ozonet , Ozone , Temp ; Ozone ,Temp , Temp ; Ozonet , Ozone , Temp , Temp ; Ozone ,Temp ; , Cloud ; , Wind ; Ozonet , Ozone , Temp ; , Cloud ; , Wind ; Ozone , Temp , Temp ; , Cloud ; , Wind ; Ozonet , Ozone , Temp , Temp ; , Cloud ; , Wind ; Temp , Temp ; , Cloud ; , Wind ; #HU 6+1 5+1 LR 0.2 0.2 Bias Y Y 5+1 0.2 Y 4+1 0.2 Y 5+1 0.2 Y 5+1 0.2 Y 5+1 0.2 Y 5+1 0.2 Y 5+1 0.2 Y M08d Inputs Ozone ,Temp ; Ozone ,Temp ; Ozone ,Temp ; Ozone ,Temp ; Ozone ,Temp ; Ozone ,Temp ; Inputs Ozone ,Temp ; Ozone ,Temp ; Ozone ,Temp ; Ozone ,Temp ; Ozone ,Temp ; Ozone ,Temp ; Inputs Ozone ,Temp ; Cloud ; ,Wind ; Ozone ,Temp ; Cloud ; ,Wind ; Ozone ,Temp ; Cloud ; ,Wind ; Ozone ,Temp ; Cloud ; ,Wind ; #HU 4 3+1 4+1 5+1 6+1 8+1 #HU 5+1 5+1 5+1 5+1 8+1 8+1 #HU , 5+1 LR 1.0 0.2 0.2 0.2 0.2 0.2 LR 0.2 0.2 0.2 0.2 0.2 0.2 LR 0.2 Bias N Y Y Y Y Y Bias Y Y Y Y Y Y Bias Y Inertia 0.0 0.0 0.0 0.0 0.0 0.0 Inertia 0.2 0.4 0.6 0.8 0.2 0.4 Inertia 0.0 , 5+1 0.2 Y 0.2 , 5+1 0.2 Y 0.4 , 5+1 0.2 Y 0.6 Appendix B – RMS Errors The following tables shows the RMS error of all models of all test series during the abstraction phases (see (Wieland 1999)). #Steps give the necessary number of learning steps for the neural network in order to obtain the best result. #Steps 100 500 1000 3000 5000 M01a 12.7054 12.3203 12.1776 12.1714 12.1799 M01b M01c M01d M01e M01f 12.1960 12.2140 12.2118 12.2554 12.2460 12.2341 12.3053 12.1978 12.2186 12.2253 12.1165 12.1671 12.3947 12.2114 12.1483 #Steps 100 500 1000 3000 5000 M01g M01h M01i M01j M01k M01l 12.1442 12.1019 12.1799 12.1543 12.0105 12.0474 12.0999 11.9613 11.9500 12.0028 11.9310 11.9069 11.2004 11.4195 11.5968 11.9176 11.2626 11.2787 M02e 11.5549 12.2018 12.4597 11.9919 12.0421 M02f 11.1768 11.3759 11.7479 11.9430 11.8852 #Steps 100 500 1000 3000 5000 M01m 11.8898 11.7384 11.4646 #Steps 100 500 1000 3000 5000 M02a 11.3809 12.3983 12.0239 11.9833 12.0366 M02b 12.6112 11.8879 12.6211 12.1773 12.0859 M02c 12.5169 12.0402 12.9571 13.3754 12.9121 M02d 12.2704 12.3392 12.8433 12.4654 12.0287 #Steps 100 500 1000 3000 5000 M02g 11.2219 11.2256 12.0909 12.1283 11.7518 M02h 12.2785 12.0397 11.7547 11.8873 11.7883 M02i 11.3814 11.9295 12.0620 12.7961 13.1452 M02j 11.1829 11.6844 11.7748 12.0353 12.3596 #Steps 100 500 1000 3000 5000 M03a 17.9275 17.8842 17.8841 17.8841 17.8841 M03b 15.2594 15.1088 15.0786 15.0537 15.0939 M03c 16.3233 15.2674 15.0975 15.8108 16.0035 M03d 16.0686 15.1814 15.0688 15.4859 16.2480 #Steps 100 500 1000 3000 5000 M04a 17.2282 17.1559 17.1559 17.1559 17.1559 M04b 17.7798 11.3929 11.4399 11.7363 11.8890 M04c 11.7956 11.8875 12.0099 12.1321 12.1355 M04d 11.8978 12.3251 12.4722 12.6919 12.7855 M04e 15.9281 10.9266 10.6651 10.8002 10.8726 M04f 11.3494 10.8152 10.7187 10.6816 10.7025 #Steps 100 500 1000 3000 5000 M04g 11.6726 10.9101 10.7480 10.6986 10.7162 M04h 12.4379 11.1285 10.9937 10.8259 10.7277 #Steps 100 500 1000 3000 5000 M05a 11.7696 11.6355 11.5293 11.2261 11.2074 M05b 12.0077 11.5827 11.5422 11.1948 11.1407 M05c 12.5474 11.1635 11.0748 11.1226 11.0501 M05d 12.7423 11.2658 11.1323 10.9521 10.9327 M05e 11.1829 11.6844 11.7748 12.0353 12.3596 M05f 11.5014 11.6142 11.5846 11.3329 11.3798 #Steps 100 500 1000 3000 5000 M05g 10.8132 11.3031 11.2983 10.9417 10.8497 M05h 11.6184 11.5859 11.4888 11.3546 11.3501 M05i 11.6323 11.5733 11.6144 11.3036 11.2098 #Steps 100 500 1000 3000 5000 M06a 12.5614 12.6118 20.2562 18.1967 18.3555 M06b 12.9471 12.5284 12.0667 11.8152 12.0016 M06c 12.1955 11.7043 11.6131 11.7833 11.9388 M06d 12.1056 12.1096 11.5278 11.2674 11.3059 M06e 12.6153 12.2515 11.5542 11.0893 11.9760 M06f 12.4573 12.0379 11.9483 11.3542 10.3186 #Steps 100 500 1000 3000 5000 M07a 12.0761 11.6134 11.0486 10.5150 11.1970 M07b 12.1458 11.5021 10.8558 10.6269 11.4205 M07c 12.3928 11.9894 11.1009 10.9646 11.3466 M07d 16.3523 12.9935 11.6606 11.7003 11.7185 M07e 12.3627 11.6692 11.7382 11.0655 11.2189 M07f 12.6136 11.6748 11.1409 10.8164 11.4517 #Steps 100 500 1000 3000 5000 10000 M08a 13.0530 12.6976 11.9779 11.5771 11.3775 11.6521 M08b 13.2301 12.4605 11.6784 10.2700 10.2378 10.8070 M08c 13.3053 11.6085 11.0446 10.5488 9.9579 10.9865 M08d 12.8895 11.2232 11.0759 10.4890 10.6531 11.5829