Tato diplomová práce se zabývá problematikou obchodování na komoditních trzích. Řešení problematiky spočívá ve využití umělé inteligence, konkrétně neuronových sítí, k technické analýze vývoje ceny vybrané komodity a snaze o co nejpřesnější predikci budoucího vývoje ceny pro podporu investičního rozhodování. Model neuronové sítě je vytvořen a použit pro predikci v programu MATLAB.

Práce byla zaměřena na ověření platnosti koncepce master křivky pro hodnocení heterogenních svarových spojů, resp. teplotně stárnutých svarů. Současně bylo cílem disertace vyvinout kvantitativní model pro predikci referenční teploty lokalizující tranzitní oblast na teplotní ose za použití dat získaných z tahové zkoušky, a to za použití metody umělých neuronových sítí. Studie je současně zaměřena na heterogenní svarový spoj připravený tavným svařováním. Je zacílena na hodnocení lomového chování v tranzitní oblasti nejméně odolné části svaru, tj. tepelně ovlivněné zóny ferritické oceli v blízkosti zóny natavení s vysokolegovaným materiálem. Pro predikci referenční teploty master křivky je použita zmíněná metoda neuronových sítí, a to za použití dat z tahových zkoušek a měření tvrdosti. Predikovaná referenční teplota byla ověřována na základě výsledku experimentálních měření. Vytvoření modelu za použití neuronových sítí vyžaduje dostatečné množství dat a není vždy snadno tuto podmínku splnit. V případě sledovaného problému to znamenalo použití dat z dostatečně věrohodných zdrojů (skupiny Křehký lom ÚFM AVČR) a se známou metalurgickou historií. Smysl práce je tak možno spatřovat ve vývoji modelu neuronové sítě, která bude dostatečně přesně predikovat referenční teplotu. Celkově byla pro tyto účely použita data z 29 nízkolegovaných ocelí. Pro účely vývoje byly použity kromě hladkých zkušebních tyčí, rovněž tahové zkoušky s obvodovým vrubem testované při kritické teplotě křehkosti (mez makroplastických deformací) a při teplotě pokojové. Při tvorbě modelu byla postupně v různých kombinacích využita všechna data z uvedených zkoušek. Studie ukázala, že referenční teplota charakterizující tranzitní chování lomové houževnatosti oceli s převažující feritickou strukturou je jedinečným parametrem predikovatelným na základě vybraných charakteristik tahových zkoušek.

This master’s thesis is about automatic digital modulation recognition using artificial neural networks. The paper briefly describes the issue and existing algorithms for solving the problem of modulation recognition. It was found that the best results are achieved when using the feature-recognition methods and artificial neural networks. The digital modulations that were chosen for recognition are described theoretically and they are ASK, FSK, BPSK, QPSK and 16QAM. These modulations are most commonly used today. Later was briefly described theory of neural networks. In another part was given to the characteristic features of modulation for modulation recognition using artificial neural networks. The penultimate part describes the parameters for signal simulation in Matlab, how to create the key features in Matlab and results after experimental simulation. The last part contains neural network optimization experiments.

Computational Intelligence has found a challenging testbed for various paradigms in the financial sector. Extensive research has resulted in numerous financial applications using neural networks and evolutionary computation, mainly genetic algorithms and genetic programming. More recent advances in the field of computational intelligence have not yet been applied as extensively or have not become available in the public domain, due to the confidentiality requirements of financial institutions. This study investigates how co-evolution together with the combination of par- ticle swarm optimisation and neural networks could be used to discover competitive security trading agents that could enable the timing of buying and selling securities to maximise net profit and minimise risk over time. The investigated model attempts to identify security trend reversals with the help of technical analysis methodologies. Technical market indicators provide the necessary market data to the agents and reflect information such as supply, demand, momentum, volatility, trend, sentiment and retracement. All this is derived from the security price alone, which is one of the strengths of technical analysis and the reason for its use in this study. The model proposed in this thesis evolves trading strategies within a single pop- ulation of competing agents, where each agent is represented by a neural network. The population is governed by a competitive co-evolutionary particle swarm optimi- sation algorithm, with the objective of optimising the weights of the neural networks. A standard feed forward neural network architecture is used, which functions as a market trend reversal confidence. Ultimately, the neural network becomes an amal- gamation of the technical market indicators used as inputs, and hence is capable of detecting trend reversals. Timely trading actions are derived from the confidence output, by buying and short selling securities when the price is expected to rise or fall respectively. No expert trading knowledge is presented to the model, only the technical market indicator data. The co-evolutionary particle swarm optimisation model facilitates the discovery of favourable technical market indicator interpretations, starting with zero knowledge. A competitive fitness function is defined that allows the evaluation of each solution relative to other solutions, based on predefined performance metric objectives. The relative fitness function in this study considers net profit and the Sharpe ratio as a risk measure. For the purposes of this study, the stock prices of eight large market capitalisation companies were chosen. Two benchmarks were used to evaluate the discovered trading agents, consisting of a Bollinger Bands/Relative Strength Index rule-based strategy and the popular buy-and-hold strategy. The agents that were discovered from the proposed hybrid computational intelligence model outperformed both benchmarks by producing higher returns for in-sample and out-sample data at a low risk. This indicates that the introduced model is effective in finding favourable strategies, based on observed historical security price data. Transaction costs were considered in the evaluation of the computational intelligent agents, making this a feasible model for a real-world application. Copyright / Dissertation (MSc)--University of Pretoria, 2010. / Computer Science / unrestricted

The intermittent nature of renewable energy, variations in energy demand, and fluctuations in oil and gas prices have all contributed to variable demand for power generation from coal-burning power plants. The varying demand leads to load-follow and on/off operations referred to as cycling. Cycling causes transients of properties such as pressure and temperature within various components of the steam generation system. The transients can cause increased damage because of fatigue and creep-fatigue interactions shortening the life of components. The data-driven model based on artificial neural networks (ANN) is developed for the first time to estimate properties of the steam generator components during cycling operations of a power plant. This approach utilizes data from the Coal Creek Station power plant located in North Dakota, USA collected over 10 years with a 1-hour resolution. Cycling characteristics of the plant are identified using a time-series of gross power. The ANN model estimates the component properties, for a given gross power profile and initial conditions, as they vary during cycling operations. As a representative example, the ANN estimates are presented for the superheater outlet pressure, reheater inlet temperature, and flue gas temperature at the air heater inlet. The changes in these variables as a function of the gross power over the time duration are compared with measurements to assess the predictive capability of the model. Mean square errors of 4.49E-04 for superheater outlet pressure, 1.62E-03 for reheater inlet temperature, and 4.14E-04 for flue gas temperature at the air heater inlet were observed.

A detailed and comprehensive description of the state of the art in the field of flood forecasting opens this work. Advantages and shortcomings of currently available methods are identified and discussed. Amongst others, one important aspect considers the most exigent weak point of today’s forecasting systems: The representation of all the fundamentally different event specific patterns of flood formation with one single set of model parameters. The study exemplarily proposes an alternative for overcoming this restriction by taking into account the different process characteristics of flood events via a dynamic parameterisation strategy. Other fundamental shortcomings in current approaches especially restrict the potential for real time flash flood forecasting, namely the considerable computational requirements together with the rather cumbersome operation of reliable physically based hydrologic models. The new PAI-OFF methodology (Process Modelling and Artificial Intelligence for Online Flood Forecasting) considers these problems and offers a way out of the general dilemma. It combines the reliability and predictive power of physically based, hydrologic models with the operational advantages of artificial intelligence. These operational advantages feature extremely low computation times, absolute robustness and straightforward operation. Such qualities easily allow for predicting flash floods in small catchments taking into account precipitation forecasts, whilst extremely basic computational requirements open the way for online Monte Carlo analysis of the forecast uncertainty. The study encompasses a detailed analysis of hydrological modeling and a problem specific artificial intelligence approach in the form of artificial neural networks, which build the PAI-OFF methodology. Herein, the synthesis of process modelling and artificial neural networks is achieved by a special training procedure. It optimizes the network according to the patterns of possible catchment reaction to rainstorms. This information is provided by means of a physically based catchment model, thus freeing the artificial neural network from its constriction to the range of observed data – the classical reason for unsatisfactory predictive power of netbased approaches. Instead, the PAI-OFF-net learns to portray the dominant process controls of flood formation in the considered catchment, allowing for a reliable predictive performance. The work ends with an exemplary forecasting of the 2002 flood in a 1700 km² East German watershed.

The optimization of the spinning process and adjustment of the machine settings involve “Trial and Error” method resulting in the wasting of production time and material. This situation becomes worse in the spinning mills where the speed and material changes are frequent. This research includes the use of artificial neural networks to provide the thinking ability to the spinning machines to improve the yarn spinning process. Draw frame, being the central part of the spinning preparation chain and last machine to rectify the variations in the fed slivers is the main focus of the research work. Artificial neural network have been applied to the leveling action point at auto-leveler draw frame and search range of leveling action point has been considerably reduced. Moreover, the sliver and yarn characteristics have been predicted on the basis of draw frame settings using the artificial neural networks. The results of present research work can help the spinning industry in the direction of limiting of “Trial and Error” method, reduction of waste and cutting down the time losses associated with the optimizing of machines. As a vision for the future research work the concept of intelligent spinning machines has also been proposed. / Die Optimierung des Spinnprozesses und die Maschineneinstellung erfolgen häufig mittels „Trial und Error“-Methoden, die mit einem hohen Aufwand an Produktionszeit und Material einhergehen. Diese Situation ist für Spinnereien, in denen häufige Wechsel des eingesetzten Materials oder der Produktionsgeschwindigkeit nötig sind, besonders ungünstig. Die vorliegende Arbeit zeigt das Potenzial Neuronaler Netze, um die Spinnmaschine zum „Denken“ zu befähigen und damit die Garnherstellung effektiver zu machen. Die Strecke ist der zentrale Teil der Spinnereivorbereitungskette und bietet die letzte Möglichkeit, Inhomogenitäten im Faserband zu beseitigen. Der Fokus der Arbeit richtet sich deshalb auf diese Maschine. Künstlich Neuronale Netze werden an der Strecke zur Bestimmung des Regeleinsatzpunktes genutzt, womit eine beträchtliche Reduzierung des Aufwands für die korrekte Festlegung des Regeleinsatzpunkts erreicht wird. Darüber hinaus können mit Hilfe der Neuronalen Netze die Band- und Garneigenschaften auf Basis der Streckeneinstellungen vorausbestimmt werden. Die Resultate der vorliegenden Arbeit machen „Trial und Error“-Methoden überflüssig, reduzieren den Ausschuss und verringern die Zeitverluste bei der Maschinenoptimierung. Als Zukunftsvision wird eine Konzeption für intelligente Spinnmaschinen vorgestellt.

Aus den Mängeln bisher verwendeter Modelle zur Abbildung des Wellenablaufes zu Prognosezwecken im Hochwasserfall wird in dieser Arbeit eine Methodik entwickelt, die die Schnelligkeit und Robustheit künstlicher neuronaler Netze mit der Zuverlässigkeit hydrodynamisch-numerischer Modellierung verbindet. Ein eindimensionales hydrodynamisches Modell beinhaltet die genaue Kenntnis der Geometrie des Flußlaufes und der Vorländer und berücksichtigt die physikalischen Prozesse des Wellenablaufes. Mit diesem deterministischen Modell ist eine Grundlage für umfangreiche Szenarienrechnungen zur Erstellung einer Datenbasis geschaffen, die die weite Spanne theoretisch möglicher Hochwasserereignisse abdeckt. Mit dieser Datenbasis können dann künstliche neuronale Netze trainiert werden, die auch im Bereich extremer Hochwasserereignisse zuverlässige Prognosen liefern. In dieser Arbeit werden mit Multilayer-Feedforward-Netzen und selbstorganisierenden Merkmalskarten zwei Netztypen als Vertreter überwacht und unüberwacht lernender neuronaler Netze auf ihre diesbezügliche Eignung untersucht und beurteilt. Desweiteren wurde die Methodik auf die Einbeziehung von Merkmalen für die Niederschlags-Abfluß-Prozesse im unbeobachteten Zwischengebiet zur Berücksichtigung lateraler Zuflüsse entlang der modellierten Fließstrecken erweitert. Die Datenbasis wurde hierfür mit einem Niederschlags-Abfluß-Modell erstellt. Ein Hauptschwerpunkt liegt in der Überführung der Eingangsdaten in charakteristische Merkmale zur Abbildung der Zielgrößen, in diesem Falle des Durchflusses und Wasserstandes am Zielpegel. So dienen die deterministischen Modelle nicht nur zur Erstellung einer verläßlichen Datenbasis für das Training der Netze, sondern ermöglichen – sowohl für die Niederschlags-Abfluß-Prozesse, als auch für die hydrodynamischen Prozesse – Analysen betreffs der Sensitivität der Modellergebnisse infolge von Änderungen der Inputdaten. Mit Hilfe dieser Analysen werden wichtige Informationen zur Findung der relevanten Merkmale erlangt. Ein Schlüssel für die erfolgreiche Eingliederung der Niederschlags-Abfluß-Prozesse in das Prognosenetz ist die Einführung eines einzigen Zustandsmerkmals, welches die gesamte meteorologische Vorgeschichte des Ereignisses zur Charakterisierung des Gebietszustandes vereinigt. Die entwickelte Methodik wurde anhand des Einzugsgebietes der Freiberger Mulde erfolgreich getestet.

In this paper the brief overview of methods available for individual tree stems quality evaluation and tree species classification has been performed. The use of Near Infrared photometry based on conifer’s canopy reflectance measurement in near infrared range of spectrum has been evaluated for the use in autonomous forest harvesting. Photometric method based on the image processing of the bark pattern has been proposed to perform classification between main construction timber tree species in Scandinavia: Norway spruce (Picea abies) and Scots Pine (Pinus sylvestris). Several feature extraction algorithms have been evaluated, resulting two methods selected: Statistical Analysis using gray level co-occurrence matrix and maximally stable extremal regions feature detector. Feedforward Neural Network with Backpropagation training algorithm and Support Vector Machine classifiers have been implemented and compared. The verification of the proposed algorithm has been performed by real-time testing.

[pt] Na previsão de séries temporais, alguns autores supõem que um método de previsão individual (por exemplo, um modelo ARIMA) produz resíduos (ou erros de previsão) semelhantes a um processo de ruído branco (imprevisível). No entanto, principalmente devido às estruturas de autodependência não mapeadas por um método preditivo individual, tal suposição pode ser facilmente violada na prática. Esta tese propõe um Previsor Híbrido Wavelet (PHW) que integra as seguintes técnicas: decomposição wavelet; modelos ARIMA; redes neurais artificiais (RNAs); combinação de previsões; programação matemática não linear e amostrador Bootstrap. Em termos gerais, o PHW proposto aqui é capaz de capturar, ao mesmo tempo, estruturas com autodependência linear por meio de uma combinação linear wavelet (CLW) de modelos ARIMA, (cujo ajuste numérico ótimo ocorre por programação matemática não linear) e não linear (usando uma RNA wavelet automática) exibidas pela série de tempo a ser predita. Diferentemente de outras abordagens híbridas existentes na literatura, as previsões híbridas produzidas pela PHW proposto levam em conta implicitamente, através da abordagem de decomposição wavelet, as informações oriundas da frequência espectral presentes na série temporal subjacente. Os resultados estatísticos mostram que a metodologia híbrida supracitada alcançou ganhos de precisão relevantes no processo preditivo de quatro séries de tempo diferentes bem conhecidas, quando se compara com outras meteorologistas competitivas. / [en] In time series analysis some authors presume that a single model (an ARIMA for instance) may yield white noise errors. However that assumption can be easily violated, especially in scenarios where unmapped auto dependency structures are present inside the series. With that being said, this thesis proposes a new approach called Hybrid Wavelet Predictor (HWP) which integrates the following techniques: Wavelet Decomposition, ARIMA models, Neural Networks (NN), Combined Prediction, Non-linear mathematical programming and Bootstrap Sampling. In a broad sense, the proposed HWP is able to capture not only the linear auto-dependent structures from ARIMA using linear wavelet combination (where its optimal numerical adjustment is made through non-linear mathematical programming), but also the non-linear structures by using Neural Network. Differently from others hybrid approaches known to date, the hybrid predictions given by HWP model take into account. Statistical tests show that the hybrid approach stated above increased the prediction s effectiveness by a significant amount when compared with four well known processes.